xref: /titanic_44/usr/src/uts/i86pc/io/rootnex.c (revision 2d6b5ea734bb47d251c82670646fde46af15fd69)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * x86 root nexus driver
28  */
29 
30 #include <sys/sysmacros.h>
31 #include <sys/conf.h>
32 #include <sys/autoconf.h>
33 #include <sys/sysmacros.h>
34 #include <sys/debug.h>
35 #include <sys/psw.h>
36 #include <sys/ddidmareq.h>
37 #include <sys/promif.h>
38 #include <sys/devops.h>
39 #include <sys/kmem.h>
40 #include <sys/cmn_err.h>
41 #include <vm/seg.h>
42 #include <vm/seg_kmem.h>
43 #include <vm/seg_dev.h>
44 #include <sys/vmem.h>
45 #include <sys/mman.h>
46 #include <vm/hat.h>
47 #include <vm/as.h>
48 #include <vm/page.h>
49 #include <sys/avintr.h>
50 #include <sys/errno.h>
51 #include <sys/modctl.h>
52 #include <sys/ddi_impldefs.h>
53 #include <sys/sunddi.h>
54 #include <sys/sunndi.h>
55 #include <sys/mach_intr.h>
56 #include <sys/psm.h>
57 #include <sys/ontrap.h>
58 #include <sys/atomic.h>
59 #include <sys/sdt.h>
60 #include <sys/rootnex.h>
61 #include <vm/hat_i86.h>
62 #include <sys/ddifm.h>
63 #include <sys/ddi_isa.h>
64 
65 #ifdef __xpv
66 #include <sys/bootinfo.h>
67 #include <sys/hypervisor.h>
68 #include <sys/bootconf.h>
69 #include <vm/kboot_mmu.h>
70 #else
71 #include <sys/intel_iommu.h>
72 #endif
73 
74 
75 /*
76  * enable/disable extra checking of function parameters. Useful for debugging
77  * drivers.
78  */
79 #ifdef	DEBUG
80 int rootnex_alloc_check_parms = 1;
81 int rootnex_bind_check_parms = 1;
82 int rootnex_bind_check_inuse = 1;
83 int rootnex_unbind_verify_buffer = 0;
84 int rootnex_sync_check_parms = 1;
85 #else
86 int rootnex_alloc_check_parms = 0;
87 int rootnex_bind_check_parms = 0;
88 int rootnex_bind_check_inuse = 0;
89 int rootnex_unbind_verify_buffer = 0;
90 int rootnex_sync_check_parms = 0;
91 #endif
92 
93 /* Master Abort and Target Abort panic flag */
94 int rootnex_fm_ma_ta_panic_flag = 0;
95 
96 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
97 int rootnex_bind_fail = 1;
98 int rootnex_bind_warn = 1;
99 uint8_t *rootnex_warn_list;
100 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
101 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
102 
103 /*
104  * revert back to old broken behavior of always sync'ing entire copy buffer.
105  * This is useful if be have a buggy driver which doesn't correctly pass in
106  * the offset and size into ddi_dma_sync().
107  */
108 int rootnex_sync_ignore_params = 0;
109 
110 /*
111  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
112  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
113  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
114  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
115  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
116  * (< 8K). We will still need to allocate the copy buffer during bind though
117  * (if we need one). These can only be modified in /etc/system before rootnex
118  * attach.
119  */
120 #if defined(__amd64)
121 int rootnex_prealloc_cookies = 65;
122 int rootnex_prealloc_windows = 4;
123 int rootnex_prealloc_copybuf = 2;
124 #else
125 int rootnex_prealloc_cookies = 33;
126 int rootnex_prealloc_windows = 4;
127 int rootnex_prealloc_copybuf = 2;
128 #endif
129 
130 /* driver global state */
131 static rootnex_state_t *rootnex_state;
132 
133 /* shortcut to rootnex counters */
134 static uint64_t *rootnex_cnt;
135 
136 /*
137  * XXX - does x86 even need these or are they left over from the SPARC days?
138  */
139 /* statically defined integer/boolean properties for the root node */
140 static rootnex_intprop_t rootnex_intprp[] = {
141 	{ "PAGESIZE",			PAGESIZE },
142 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
143 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
144 	{ DDI_RELATIVE_ADDRESSING,	1 },
145 };
146 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
147 
148 #ifdef __xpv
149 typedef maddr_t rootnex_addr_t;
150 #define	ROOTNEX_PADDR_TO_RBASE(xinfo, pa)	\
151 	(DOMAIN_IS_INITDOMAIN(xinfo) ? pa_to_ma(pa) : (pa))
152 #else
153 typedef paddr_t rootnex_addr_t;
154 #endif
155 
156 #if !defined(__xpv)
157 char _depends_on[] = "mach/pcplusmp misc/iommulib";
158 #endif
159 
160 static struct cb_ops rootnex_cb_ops = {
161 	nodev,		/* open */
162 	nodev,		/* close */
163 	nodev,		/* strategy */
164 	nodev,		/* print */
165 	nodev,		/* dump */
166 	nodev,		/* read */
167 	nodev,		/* write */
168 	nodev,		/* ioctl */
169 	nodev,		/* devmap */
170 	nodev,		/* mmap */
171 	nodev,		/* segmap */
172 	nochpoll,	/* chpoll */
173 	ddi_prop_op,	/* cb_prop_op */
174 	NULL,		/* struct streamtab */
175 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
176 	CB_REV,		/* Rev */
177 	nodev,		/* cb_aread */
178 	nodev		/* cb_awrite */
179 };
180 
181 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
182     off_t offset, off_t len, caddr_t *vaddrp);
183 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
184     struct hat *hat, struct seg *seg, caddr_t addr,
185     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
186 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
187     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
188 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
189     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
190     ddi_dma_handle_t *handlep);
191 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
192     ddi_dma_handle_t handle);
193 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
194     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
195     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
196 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
197     ddi_dma_handle_t handle);
198 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
199     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
200 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
201     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
202     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
203 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
204     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
205     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
206 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
207     ddi_ctl_enum_t ctlop, void *arg, void *result);
208 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
209     ddi_iblock_cookie_t *ibc);
210 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
211     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
212 
213 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
214     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
215     ddi_dma_handle_t *handlep);
216 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
217     ddi_dma_handle_t handle);
218 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
219     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
220     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
221 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
222     ddi_dma_handle_t handle);
223 #if !defined(__xpv)
224 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
225     ddi_dma_handle_t handle);
226 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
227     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
228 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
229     ddi_dma_cookie_t *cookiep, uint_t ccount);
230 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
231     ddi_dma_handle_t handle);
232 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
233 #endif
234 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
235     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
236 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
237     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
238     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
239 static int rootnex_coredma_map(dev_info_t *dip, dev_info_t *rdip,
240     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
241 static int rootnex_coredma_mctl(dev_info_t *dip, dev_info_t *rdip,
242     ddi_dma_handle_t handle, enum ddi_dma_ctlops request, off_t *offp,
243     size_t *lenp, caddr_t *objpp, uint_t cache_flags);
244 
245 static struct bus_ops rootnex_bus_ops = {
246 	BUSO_REV,
247 	rootnex_map,
248 	NULL,
249 	NULL,
250 	NULL,
251 	rootnex_map_fault,
252 	rootnex_dma_map,
253 	rootnex_dma_allochdl,
254 	rootnex_dma_freehdl,
255 	rootnex_dma_bindhdl,
256 	rootnex_dma_unbindhdl,
257 	rootnex_dma_sync,
258 	rootnex_dma_win,
259 	rootnex_dma_mctl,
260 	rootnex_ctlops,
261 	ddi_bus_prop_op,
262 	i_ddi_rootnex_get_eventcookie,
263 	i_ddi_rootnex_add_eventcall,
264 	i_ddi_rootnex_remove_eventcall,
265 	i_ddi_rootnex_post_event,
266 	0,			/* bus_intr_ctl */
267 	0,			/* bus_config */
268 	0,			/* bus_unconfig */
269 	rootnex_fm_init,	/* bus_fm_init */
270 	NULL,			/* bus_fm_fini */
271 	NULL,			/* bus_fm_access_enter */
272 	NULL,			/* bus_fm_access_exit */
273 	NULL,			/* bus_powr */
274 	rootnex_intr_ops	/* bus_intr_op */
275 };
276 
277 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
278 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
279 
280 static struct dev_ops rootnex_ops = {
281 	DEVO_REV,
282 	0,
283 	ddi_no_info,
284 	nulldev,
285 	nulldev,
286 	rootnex_attach,
287 	rootnex_detach,
288 	nulldev,
289 	&rootnex_cb_ops,
290 	&rootnex_bus_ops,
291 	NULL,
292 	ddi_quiesce_not_needed,		/* quiesce */
293 };
294 
295 static struct modldrv rootnex_modldrv = {
296 	&mod_driverops,
297 	"i86pc root nexus",
298 	&rootnex_ops
299 };
300 
301 static struct modlinkage rootnex_modlinkage = {
302 	MODREV_1,
303 	(void *)&rootnex_modldrv,
304 	NULL
305 };
306 
307 #if !defined(__xpv)
308 static iommulib_nexops_t iommulib_nexops = {
309 	IOMMU_NEXOPS_VERSION,
310 	"Rootnex IOMMU ops Vers 1.1",
311 	NULL,
312 	rootnex_coredma_allochdl,
313 	rootnex_coredma_freehdl,
314 	rootnex_coredma_bindhdl,
315 	rootnex_coredma_unbindhdl,
316 	rootnex_coredma_reset_cookies,
317 	rootnex_coredma_get_cookies,
318 	rootnex_coredma_set_cookies,
319 	rootnex_coredma_clear_cookies,
320 	rootnex_coredma_get_sleep_flags,
321 	rootnex_coredma_sync,
322 	rootnex_coredma_win,
323 	rootnex_coredma_map,
324 	rootnex_coredma_mctl
325 };
326 #endif
327 
328 /*
329  *  extern hacks
330  */
331 extern struct seg_ops segdev_ops;
332 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
333 #ifdef	DDI_MAP_DEBUG
334 extern int ddi_map_debug_flag;
335 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
336 #endif
337 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
338 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
339 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
340     psm_intr_op_t, int *);
341 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
342 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
343 
344 /*
345  * Use device arena to use for device control register mappings.
346  * Various kernel memory walkers (debugger, dtrace) need to know
347  * to avoid this address range to prevent undesired device activity.
348  */
349 extern void *device_arena_alloc(size_t size, int vm_flag);
350 extern void device_arena_free(void * vaddr, size_t size);
351 
352 
353 /*
354  *  Internal functions
355  */
356 static int rootnex_dma_init();
357 static void rootnex_add_props(dev_info_t *);
358 static int rootnex_ctl_reportdev(dev_info_t *dip);
359 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
360 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
361 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
362 static int rootnex_map_handle(ddi_map_req_t *mp);
363 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
364 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
365 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
366     ddi_dma_attr_t *attr);
367 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
368     rootnex_sglinfo_t *sglinfo);
369 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
370     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag);
371 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
372     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
373 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
374 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
375     ddi_dma_attr_t *attr, int kmflag);
376 static void rootnex_teardown_windows(rootnex_dma_t *dma);
377 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
378     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
379 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
380     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
381     size_t *copybuf_used, page_t **cur_pp);
382 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
383     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
384     ddi_dma_attr_t *attr, off_t cur_offset);
385 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
386     rootnex_dma_t *dma, rootnex_window_t **windowp,
387     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
388 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
389     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
390 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
391     off_t offset, size_t size, uint_t cache_flags);
392 static int rootnex_verify_buffer(rootnex_dma_t *dma);
393 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
394     const void *comp_addr, const void *not_used);
395 
396 /*
397  * _init()
398  *
399  */
400 int
401 _init(void)
402 {
403 
404 	rootnex_state = NULL;
405 	return (mod_install(&rootnex_modlinkage));
406 }
407 
408 
409 /*
410  * _info()
411  *
412  */
413 int
414 _info(struct modinfo *modinfop)
415 {
416 	return (mod_info(&rootnex_modlinkage, modinfop));
417 }
418 
419 
420 /*
421  * _fini()
422  *
423  */
424 int
425 _fini(void)
426 {
427 	return (EBUSY);
428 }
429 
430 
431 /*
432  * rootnex_attach()
433  *
434  */
435 static int
436 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
437 {
438 	int fmcap;
439 	int e;
440 
441 	switch (cmd) {
442 	case DDI_ATTACH:
443 		break;
444 	case DDI_RESUME:
445 		return (DDI_SUCCESS);
446 	default:
447 		return (DDI_FAILURE);
448 	}
449 
450 	/*
451 	 * We should only have one instance of rootnex. Save it away since we
452 	 * don't have an easy way to get it back later.
453 	 */
454 	ASSERT(rootnex_state == NULL);
455 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
456 
457 	rootnex_state->r_dip = dip;
458 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
459 	rootnex_state->r_reserved_msg_printed = B_FALSE;
460 	rootnex_cnt = &rootnex_state->r_counters[0];
461 	rootnex_state->r_intel_iommu_enabled = B_FALSE;
462 
463 	/*
464 	 * Set minimum fm capability level for i86pc platforms and then
465 	 * initialize error handling. Since we're the rootnex, we don't
466 	 * care what's returned in the fmcap field.
467 	 */
468 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
469 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
470 	fmcap = ddi_system_fmcap;
471 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
472 
473 	/* initialize DMA related state */
474 	e = rootnex_dma_init();
475 	if (e != DDI_SUCCESS) {
476 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
477 		return (DDI_FAILURE);
478 	}
479 
480 	/* Add static root node properties */
481 	rootnex_add_props(dip);
482 
483 	/* since we can't call ddi_report_dev() */
484 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
485 
486 	/* Initialize rootnex event handle */
487 	i_ddi_rootnex_init_events(dip);
488 
489 #if !defined(__xpv)
490 #if defined(__amd64)
491 	/* probe intel iommu */
492 	intel_iommu_probe_and_parse();
493 
494 	/* attach the iommu nodes */
495 	if (intel_iommu_support) {
496 		if (intel_iommu_attach_dmar_nodes() == DDI_SUCCESS) {
497 			rootnex_state->r_intel_iommu_enabled = B_TRUE;
498 		} else {
499 			intel_iommu_release_dmar_info();
500 		}
501 	}
502 #endif
503 
504 	e = iommulib_nexus_register(dip, &iommulib_nexops,
505 	    &rootnex_state->r_iommulib_handle);
506 
507 	ASSERT(e == DDI_SUCCESS);
508 #endif
509 
510 	return (DDI_SUCCESS);
511 }
512 
513 
514 /*
515  * rootnex_detach()
516  *
517  */
518 /*ARGSUSED*/
519 static int
520 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
521 {
522 	switch (cmd) {
523 	case DDI_SUSPEND:
524 		break;
525 	default:
526 		return (DDI_FAILURE);
527 	}
528 
529 	return (DDI_SUCCESS);
530 }
531 
532 
533 /*
534  * rootnex_dma_init()
535  *
536  */
537 /*ARGSUSED*/
538 static int
539 rootnex_dma_init()
540 {
541 	size_t bufsize;
542 
543 
544 	/*
545 	 * size of our cookie/window/copybuf state needed in dma bind that we
546 	 * pre-alloc in dma_alloc_handle
547 	 */
548 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
549 	rootnex_state->r_prealloc_size =
550 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
551 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
552 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
553 
554 	/*
555 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
556 	 * allocate 16 extra bytes for struct pointer alignment
557 	 * (p->dmai_private & dma->dp_prealloc_buffer)
558 	 */
559 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
560 	    rootnex_state->r_prealloc_size + 0x10;
561 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
562 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
563 	if (rootnex_state->r_dmahdl_cache == NULL) {
564 		return (DDI_FAILURE);
565 	}
566 
567 	/*
568 	 * allocate array to track which major numbers we have printed warnings
569 	 * for.
570 	 */
571 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
572 	    KM_SLEEP);
573 
574 	return (DDI_SUCCESS);
575 }
576 
577 
578 /*
579  * rootnex_add_props()
580  *
581  */
582 static void
583 rootnex_add_props(dev_info_t *dip)
584 {
585 	rootnex_intprop_t *rpp;
586 	int i;
587 
588 	/* Add static integer/boolean properties to the root node */
589 	rpp = rootnex_intprp;
590 	for (i = 0; i < NROOT_INTPROPS; i++) {
591 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
592 		    rpp[i].prop_name, rpp[i].prop_value);
593 	}
594 }
595 
596 
597 
598 /*
599  * *************************
600  *  ctlops related routines
601  * *************************
602  */
603 
604 /*
605  * rootnex_ctlops()
606  *
607  */
608 /*ARGSUSED*/
609 static int
610 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
611     void *arg, void *result)
612 {
613 	int n, *ptr;
614 	struct ddi_parent_private_data *pdp;
615 
616 	switch (ctlop) {
617 	case DDI_CTLOPS_DMAPMAPC:
618 		/*
619 		 * Return 'partial' to indicate that dma mapping
620 		 * has to be done in the main MMU.
621 		 */
622 		return (DDI_DMA_PARTIAL);
623 
624 	case DDI_CTLOPS_BTOP:
625 		/*
626 		 * Convert byte count input to physical page units.
627 		 * (byte counts that are not a page-size multiple
628 		 * are rounded down)
629 		 */
630 		*(ulong_t *)result = btop(*(ulong_t *)arg);
631 		return (DDI_SUCCESS);
632 
633 	case DDI_CTLOPS_PTOB:
634 		/*
635 		 * Convert size in physical pages to bytes
636 		 */
637 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
638 		return (DDI_SUCCESS);
639 
640 	case DDI_CTLOPS_BTOPR:
641 		/*
642 		 * Convert byte count input to physical page units
643 		 * (byte counts that are not a page-size multiple
644 		 * are rounded up)
645 		 */
646 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
647 		return (DDI_SUCCESS);
648 
649 	case DDI_CTLOPS_INITCHILD:
650 		return (impl_ddi_sunbus_initchild(arg));
651 
652 	case DDI_CTLOPS_UNINITCHILD:
653 		impl_ddi_sunbus_removechild(arg);
654 		return (DDI_SUCCESS);
655 
656 	case DDI_CTLOPS_REPORTDEV:
657 		return (rootnex_ctl_reportdev(rdip));
658 
659 	case DDI_CTLOPS_IOMIN:
660 		/*
661 		 * Nothing to do here but reflect back..
662 		 */
663 		return (DDI_SUCCESS);
664 
665 	case DDI_CTLOPS_REGSIZE:
666 	case DDI_CTLOPS_NREGS:
667 		break;
668 
669 	case DDI_CTLOPS_SIDDEV:
670 		if (ndi_dev_is_prom_node(rdip))
671 			return (DDI_SUCCESS);
672 		if (ndi_dev_is_persistent_node(rdip))
673 			return (DDI_SUCCESS);
674 		return (DDI_FAILURE);
675 
676 	case DDI_CTLOPS_POWER:
677 		return ((*pm_platform_power)((power_req_t *)arg));
678 
679 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
680 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
681 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
682 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
683 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
684 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
685 		if (!rootnex_state->r_reserved_msg_printed) {
686 			rootnex_state->r_reserved_msg_printed = B_TRUE;
687 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
688 			    "1 or more reserved/obsolete operations.");
689 		}
690 		return (DDI_FAILURE);
691 
692 	default:
693 		return (DDI_FAILURE);
694 	}
695 	/*
696 	 * The rest are for "hardware" properties
697 	 */
698 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
699 		return (DDI_FAILURE);
700 
701 	if (ctlop == DDI_CTLOPS_NREGS) {
702 		ptr = (int *)result;
703 		*ptr = pdp->par_nreg;
704 	} else {
705 		off_t *size = (off_t *)result;
706 
707 		ptr = (int *)arg;
708 		n = *ptr;
709 		if (n >= pdp->par_nreg) {
710 			return (DDI_FAILURE);
711 		}
712 		*size = (off_t)pdp->par_reg[n].regspec_size;
713 	}
714 	return (DDI_SUCCESS);
715 }
716 
717 
718 /*
719  * rootnex_ctl_reportdev()
720  *
721  */
722 static int
723 rootnex_ctl_reportdev(dev_info_t *dev)
724 {
725 	int i, n, len, f_len = 0;
726 	char *buf;
727 
728 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
729 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
730 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
731 	len = strlen(buf);
732 
733 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
734 
735 		struct regspec *rp = sparc_pd_getreg(dev, i);
736 
737 		if (i == 0)
738 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
739 			    ": ");
740 		else
741 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
742 			    " and ");
743 		len = strlen(buf);
744 
745 		switch (rp->regspec_bustype) {
746 
747 		case BTEISA:
748 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
749 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
750 			break;
751 
752 		case BTISA:
753 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
754 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
755 			break;
756 
757 		default:
758 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
759 			    "space %x offset %x",
760 			    rp->regspec_bustype, rp->regspec_addr);
761 			break;
762 		}
763 		len = strlen(buf);
764 	}
765 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
766 		int pri;
767 
768 		if (i != 0) {
769 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
770 			    ",");
771 			len = strlen(buf);
772 		}
773 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
774 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
775 		    " sparc ipl %d", pri);
776 		len = strlen(buf);
777 	}
778 #ifdef DEBUG
779 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
780 		cmn_err(CE_NOTE, "next message is truncated: "
781 		    "printed length 1024, real length %d", f_len);
782 	}
783 #endif /* DEBUG */
784 	cmn_err(CE_CONT, "?%s\n", buf);
785 	kmem_free(buf, REPORTDEV_BUFSIZE);
786 	return (DDI_SUCCESS);
787 }
788 
789 
790 /*
791  * ******************
792  *  map related code
793  * ******************
794  */
795 
796 /*
797  * rootnex_map()
798  *
799  */
800 static int
801 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
802     off_t len, caddr_t *vaddrp)
803 {
804 	struct regspec *rp, tmp_reg;
805 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
806 	int error;
807 
808 	mp = &mr;
809 
810 	switch (mp->map_op)  {
811 	case DDI_MO_MAP_LOCKED:
812 	case DDI_MO_UNMAP:
813 	case DDI_MO_MAP_HANDLE:
814 		break;
815 	default:
816 #ifdef	DDI_MAP_DEBUG
817 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
818 		    mp->map_op);
819 #endif	/* DDI_MAP_DEBUG */
820 		return (DDI_ME_UNIMPLEMENTED);
821 	}
822 
823 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
824 #ifdef	DDI_MAP_DEBUG
825 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
826 #endif	/* DDI_MAP_DEBUG */
827 		return (DDI_ME_UNIMPLEMENTED);
828 	}
829 
830 	/*
831 	 * First, if given an rnumber, convert it to a regspec...
832 	 * (Presumably, this is on behalf of a child of the root node?)
833 	 */
834 
835 	if (mp->map_type == DDI_MT_RNUMBER)  {
836 
837 		int rnumber = mp->map_obj.rnumber;
838 #ifdef	DDI_MAP_DEBUG
839 		static char *out_of_range =
840 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
841 #endif	/* DDI_MAP_DEBUG */
842 
843 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
844 		if (rp == NULL)  {
845 #ifdef	DDI_MAP_DEBUG
846 			cmn_err(CE_WARN, out_of_range, rnumber,
847 			    ddi_get_name(rdip));
848 #endif	/* DDI_MAP_DEBUG */
849 			return (DDI_ME_RNUMBER_RANGE);
850 		}
851 
852 		/*
853 		 * Convert the given ddi_map_req_t from rnumber to regspec...
854 		 */
855 
856 		mp->map_type = DDI_MT_REGSPEC;
857 		mp->map_obj.rp = rp;
858 	}
859 
860 	/*
861 	 * Adjust offset and length correspnding to called values...
862 	 * XXX: A non-zero length means override the one in the regspec
863 	 * XXX: (regardless of what's in the parent's range?)
864 	 */
865 
866 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
867 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
868 
869 #ifdef	DDI_MAP_DEBUG
870 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
871 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
872 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
873 	    len, mp->map_handlep);
874 #endif	/* DDI_MAP_DEBUG */
875 
876 	/*
877 	 * I/O or memory mapping:
878 	 *
879 	 *	<bustype=0, addr=x, len=x>: memory
880 	 *	<bustype=1, addr=x, len=x>: i/o
881 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
882 	 */
883 
884 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
885 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
886 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
887 		    ddi_get_name(rdip), rp->regspec_bustype,
888 		    rp->regspec_addr, rp->regspec_size);
889 		return (DDI_ME_INVAL);
890 	}
891 
892 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
893 		/*
894 		 * compatibility i/o mapping
895 		 */
896 		rp->regspec_bustype += (uint_t)offset;
897 	} else {
898 		/*
899 		 * Normal memory or i/o mapping
900 		 */
901 		rp->regspec_addr += (uint_t)offset;
902 	}
903 
904 	if (len != 0)
905 		rp->regspec_size = (uint_t)len;
906 
907 #ifdef	DDI_MAP_DEBUG
908 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
909 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
910 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
911 	    offset, len, mp->map_handlep);
912 #endif	/* DDI_MAP_DEBUG */
913 
914 	/*
915 	 * Apply any parent ranges at this level, if applicable.
916 	 * (This is where nexus specific regspec translation takes place.
917 	 * Use of this function is implicit agreement that translation is
918 	 * provided via ddi_apply_range.)
919 	 */
920 
921 #ifdef	DDI_MAP_DEBUG
922 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
923 	    ddi_get_name(dip), ddi_get_name(rdip));
924 #endif	/* DDI_MAP_DEBUG */
925 
926 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
927 		return (error);
928 
929 	switch (mp->map_op)  {
930 	case DDI_MO_MAP_LOCKED:
931 
932 		/*
933 		 * Set up the locked down kernel mapping to the regspec...
934 		 */
935 
936 		return (rootnex_map_regspec(mp, vaddrp));
937 
938 	case DDI_MO_UNMAP:
939 
940 		/*
941 		 * Release mapping...
942 		 */
943 
944 		return (rootnex_unmap_regspec(mp, vaddrp));
945 
946 	case DDI_MO_MAP_HANDLE:
947 
948 		return (rootnex_map_handle(mp));
949 
950 	default:
951 		return (DDI_ME_UNIMPLEMENTED);
952 	}
953 }
954 
955 
956 /*
957  * rootnex_map_fault()
958  *
959  *	fault in mappings for requestors
960  */
961 /*ARGSUSED*/
962 static int
963 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
964     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
965     uint_t lock)
966 {
967 
968 #ifdef	DDI_MAP_DEBUG
969 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
970 	ddi_map_debug(" Seg <%s>\n",
971 	    seg->s_ops == &segdev_ops ? "segdev" :
972 	    seg == &kvseg ? "segkmem" : "NONE!");
973 #endif	/* DDI_MAP_DEBUG */
974 
975 	/*
976 	 * This is all terribly broken, but it is a start
977 	 *
978 	 * XXX	Note that this test means that segdev_ops
979 	 *	must be exported from seg_dev.c.
980 	 * XXX	What about devices with their own segment drivers?
981 	 */
982 	if (seg->s_ops == &segdev_ops) {
983 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
984 
985 		if (hat == NULL) {
986 			/*
987 			 * This is one plausible interpretation of
988 			 * a null hat i.e. use the first hat on the
989 			 * address space hat list which by convention is
990 			 * the hat of the system MMU.  At alternative
991 			 * would be to panic .. this might well be better ..
992 			 */
993 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
994 			hat = seg->s_as->a_hat;
995 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
996 		}
997 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
998 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
999 	} else if (seg == &kvseg && dp == NULL) {
1000 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
1001 		    HAT_LOAD_LOCK);
1002 	} else
1003 		return (DDI_FAILURE);
1004 	return (DDI_SUCCESS);
1005 }
1006 
1007 
1008 /*
1009  * rootnex_map_regspec()
1010  *     we don't support mapping of I/O cards above 4Gb
1011  */
1012 static int
1013 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1014 {
1015 	rootnex_addr_t rbase;
1016 	void *cvaddr;
1017 	uint_t npages, pgoffset;
1018 	struct regspec *rp;
1019 	ddi_acc_hdl_t *hp;
1020 	ddi_acc_impl_t *ap;
1021 	uint_t	hat_acc_flags;
1022 	paddr_t pbase;
1023 
1024 	rp = mp->map_obj.rp;
1025 	hp = mp->map_handlep;
1026 
1027 #ifdef	DDI_MAP_DEBUG
1028 	ddi_map_debug(
1029 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1030 	    rp->regspec_bustype, rp->regspec_addr,
1031 	    rp->regspec_size, mp->map_handlep);
1032 #endif	/* DDI_MAP_DEBUG */
1033 
1034 	/*
1035 	 * I/O or memory mapping
1036 	 *
1037 	 *	<bustype=0, addr=x, len=x>: memory
1038 	 *	<bustype=1, addr=x, len=x>: i/o
1039 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1040 	 */
1041 
1042 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1043 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1044 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1045 		    rp->regspec_addr, rp->regspec_size);
1046 		return (DDI_FAILURE);
1047 	}
1048 
1049 	if (rp->regspec_bustype != 0) {
1050 		/*
1051 		 * I/O space - needs a handle.
1052 		 */
1053 		if (hp == NULL) {
1054 			return (DDI_FAILURE);
1055 		}
1056 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1057 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1058 		impl_acc_hdl_init(hp);
1059 
1060 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1061 #ifdef  DDI_MAP_DEBUG
1062 			ddi_map_debug("rootnex_map_regspec: mmap() "
1063 			    "to I/O space is not supported.\n");
1064 #endif  /* DDI_MAP_DEBUG */
1065 			return (DDI_ME_INVAL);
1066 		} else {
1067 			/*
1068 			 * 1275-compliant vs. compatibility i/o mapping
1069 			 */
1070 			*vaddrp =
1071 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1072 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1073 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1074 #ifdef __xpv
1075 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1076 				hp->ah_pfn = xen_assign_pfn(
1077 				    mmu_btop((ulong_t)rp->regspec_addr &
1078 				    MMU_PAGEMASK));
1079 			} else {
1080 				hp->ah_pfn = mmu_btop(
1081 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1082 			}
1083 #else
1084 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1085 			    MMU_PAGEMASK);
1086 #endif
1087 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1088 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1089 		}
1090 
1091 #ifdef	DDI_MAP_DEBUG
1092 		ddi_map_debug(
1093 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1094 		    rp->regspec_size, *vaddrp);
1095 #endif	/* DDI_MAP_DEBUG */
1096 		return (DDI_SUCCESS);
1097 	}
1098 
1099 	/*
1100 	 * Memory space
1101 	 */
1102 
1103 	if (hp != NULL) {
1104 		/*
1105 		 * hat layer ignores
1106 		 * hp->ah_acc.devacc_attr_endian_flags.
1107 		 */
1108 		switch (hp->ah_acc.devacc_attr_dataorder) {
1109 		case DDI_STRICTORDER_ACC:
1110 			hat_acc_flags = HAT_STRICTORDER;
1111 			break;
1112 		case DDI_UNORDERED_OK_ACC:
1113 			hat_acc_flags = HAT_UNORDERED_OK;
1114 			break;
1115 		case DDI_MERGING_OK_ACC:
1116 			hat_acc_flags = HAT_MERGING_OK;
1117 			break;
1118 		case DDI_LOADCACHING_OK_ACC:
1119 			hat_acc_flags = HAT_LOADCACHING_OK;
1120 			break;
1121 		case DDI_STORECACHING_OK_ACC:
1122 			hat_acc_flags = HAT_STORECACHING_OK;
1123 			break;
1124 		}
1125 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1126 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1127 		impl_acc_hdl_init(hp);
1128 		hp->ah_hat_flags = hat_acc_flags;
1129 	} else {
1130 		hat_acc_flags = HAT_STRICTORDER;
1131 	}
1132 
1133 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1134 #ifdef __xpv
1135 	/*
1136 	 * If we're dom0, we're using a real device so we need to translate
1137 	 * the MA to a PA.
1138 	 */
1139 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1140 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1141 	} else {
1142 		pbase = rbase;
1143 	}
1144 #else
1145 	pbase = rbase;
1146 #endif
1147 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1148 
1149 	if (rp->regspec_size == 0) {
1150 #ifdef  DDI_MAP_DEBUG
1151 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1152 #endif  /* DDI_MAP_DEBUG */
1153 		return (DDI_ME_INVAL);
1154 	}
1155 
1156 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1157 		/* extra cast to make gcc happy */
1158 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1159 	} else {
1160 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1161 
1162 #ifdef	DDI_MAP_DEBUG
1163 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1164 		    "physical %llx", npages, pbase);
1165 #endif	/* DDI_MAP_DEBUG */
1166 
1167 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1168 		if (cvaddr == NULL)
1169 			return (DDI_ME_NORESOURCES);
1170 
1171 		/*
1172 		 * Now map in the pages we've allocated...
1173 		 */
1174 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1175 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1176 		    HAT_LOAD_LOCK);
1177 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1178 
1179 		/* save away pfn and npages for FMA */
1180 		hp = mp->map_handlep;
1181 		if (hp) {
1182 			hp->ah_pfn = mmu_btop(pbase);
1183 			hp->ah_pnum = npages;
1184 		}
1185 	}
1186 
1187 #ifdef	DDI_MAP_DEBUG
1188 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1189 #endif	/* DDI_MAP_DEBUG */
1190 	return (DDI_SUCCESS);
1191 }
1192 
1193 
1194 /*
1195  * rootnex_unmap_regspec()
1196  *
1197  */
1198 static int
1199 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1200 {
1201 	caddr_t addr = (caddr_t)*vaddrp;
1202 	uint_t npages, pgoffset;
1203 	struct regspec *rp;
1204 
1205 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1206 		return (0);
1207 
1208 	rp = mp->map_obj.rp;
1209 
1210 	if (rp->regspec_size == 0) {
1211 #ifdef  DDI_MAP_DEBUG
1212 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1213 #endif  /* DDI_MAP_DEBUG */
1214 		return (DDI_ME_INVAL);
1215 	}
1216 
1217 	/*
1218 	 * I/O or memory mapping:
1219 	 *
1220 	 *	<bustype=0, addr=x, len=x>: memory
1221 	 *	<bustype=1, addr=x, len=x>: i/o
1222 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1223 	 */
1224 	if (rp->regspec_bustype != 0) {
1225 		/*
1226 		 * This is I/O space, which requires no particular
1227 		 * processing on unmap since it isn't mapped in the
1228 		 * first place.
1229 		 */
1230 		return (DDI_SUCCESS);
1231 	}
1232 
1233 	/*
1234 	 * Memory space
1235 	 */
1236 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1237 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1238 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1239 	device_arena_free(addr - pgoffset, ptob(npages));
1240 
1241 	/*
1242 	 * Destroy the pointer - the mapping has logically gone
1243 	 */
1244 	*vaddrp = NULL;
1245 
1246 	return (DDI_SUCCESS);
1247 }
1248 
1249 
1250 /*
1251  * rootnex_map_handle()
1252  *
1253  */
1254 static int
1255 rootnex_map_handle(ddi_map_req_t *mp)
1256 {
1257 	rootnex_addr_t rbase;
1258 	ddi_acc_hdl_t *hp;
1259 	uint_t pgoffset;
1260 	struct regspec *rp;
1261 	paddr_t pbase;
1262 
1263 	rp = mp->map_obj.rp;
1264 
1265 #ifdef	DDI_MAP_DEBUG
1266 	ddi_map_debug(
1267 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1268 	    rp->regspec_bustype, rp->regspec_addr,
1269 	    rp->regspec_size, mp->map_handlep);
1270 #endif	/* DDI_MAP_DEBUG */
1271 
1272 	/*
1273 	 * I/O or memory mapping:
1274 	 *
1275 	 *	<bustype=0, addr=x, len=x>: memory
1276 	 *	<bustype=1, addr=x, len=x>: i/o
1277 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1278 	 */
1279 	if (rp->regspec_bustype != 0) {
1280 		/*
1281 		 * This refers to I/O space, and we don't support "mapping"
1282 		 * I/O space to a user.
1283 		 */
1284 		return (DDI_FAILURE);
1285 	}
1286 
1287 	/*
1288 	 * Set up the hat_flags for the mapping.
1289 	 */
1290 	hp = mp->map_handlep;
1291 
1292 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1293 	case DDI_NEVERSWAP_ACC:
1294 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1295 		break;
1296 	case DDI_STRUCTURE_LE_ACC:
1297 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1298 		break;
1299 	case DDI_STRUCTURE_BE_ACC:
1300 		return (DDI_FAILURE);
1301 	default:
1302 		return (DDI_REGS_ACC_CONFLICT);
1303 	}
1304 
1305 	switch (hp->ah_acc.devacc_attr_dataorder) {
1306 	case DDI_STRICTORDER_ACC:
1307 		break;
1308 	case DDI_UNORDERED_OK_ACC:
1309 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1310 		break;
1311 	case DDI_MERGING_OK_ACC:
1312 		hp->ah_hat_flags |= HAT_MERGING_OK;
1313 		break;
1314 	case DDI_LOADCACHING_OK_ACC:
1315 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1316 		break;
1317 	case DDI_STORECACHING_OK_ACC:
1318 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1319 		break;
1320 	default:
1321 		return (DDI_FAILURE);
1322 	}
1323 
1324 	rbase = (rootnex_addr_t)rp->regspec_addr &
1325 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1326 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1327 
1328 	if (rp->regspec_size == 0)
1329 		return (DDI_ME_INVAL);
1330 
1331 #ifdef __xpv
1332 	/*
1333 	 * If we're dom0, we're using a real device so we need to translate
1334 	 * the MA to a PA.
1335 	 */
1336 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1337 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1338 		    (rbase & MMU_PAGEOFFSET);
1339 	} else {
1340 		pbase = rbase;
1341 	}
1342 #else
1343 	pbase = rbase;
1344 #endif
1345 
1346 	hp->ah_pfn = mmu_btop(pbase);
1347 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1348 
1349 	return (DDI_SUCCESS);
1350 }
1351 
1352 
1353 
1354 /*
1355  * ************************
1356  *  interrupt related code
1357  * ************************
1358  */
1359 
1360 /*
1361  * rootnex_intr_ops()
1362  *	bus_intr_op() function for interrupt support
1363  */
1364 /* ARGSUSED */
1365 static int
1366 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1367     ddi_intr_handle_impl_t *hdlp, void *result)
1368 {
1369 	struct intrspec			*ispec;
1370 	struct ddi_parent_private_data	*pdp;
1371 
1372 	DDI_INTR_NEXDBG((CE_CONT,
1373 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1374 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1375 
1376 	/* Process the interrupt operation */
1377 	switch (intr_op) {
1378 	case DDI_INTROP_GETCAP:
1379 		/* First check with pcplusmp */
1380 		if (psm_intr_ops == NULL)
1381 			return (DDI_FAILURE);
1382 
1383 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1384 			*(int *)result = 0;
1385 			return (DDI_FAILURE);
1386 		}
1387 		break;
1388 	case DDI_INTROP_SETCAP:
1389 		if (psm_intr_ops == NULL)
1390 			return (DDI_FAILURE);
1391 
1392 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1393 			return (DDI_FAILURE);
1394 		break;
1395 	case DDI_INTROP_ALLOC:
1396 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1397 			return (DDI_FAILURE);
1398 		hdlp->ih_pri = ispec->intrspec_pri;
1399 		*(int *)result = hdlp->ih_scratch1;
1400 		break;
1401 	case DDI_INTROP_FREE:
1402 		pdp = ddi_get_parent_data(rdip);
1403 		/*
1404 		 * Special case for 'pcic' driver' only.
1405 		 * If an intrspec was created for it, clean it up here
1406 		 * See detailed comments on this in the function
1407 		 * rootnex_get_ispec().
1408 		 */
1409 		if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1410 			kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1411 			    pdp->par_nintr);
1412 			/*
1413 			 * Set it to zero; so that
1414 			 * DDI framework doesn't free it again
1415 			 */
1416 			pdp->par_intr = NULL;
1417 			pdp->par_nintr = 0;
1418 		}
1419 		break;
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 		(void) (*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1463 		    (int *)&hdlp->ih_vector);
1464 
1465 		/* Add the interrupt handler */
1466 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1467 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1468 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1469 			return (DDI_FAILURE);
1470 		break;
1471 	case DDI_INTROP_DISABLE:
1472 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1473 			return (DDI_FAILURE);
1474 
1475 		/* Call psm_ops() to translate irq with the dip */
1476 		if (psm_intr_ops == NULL)
1477 			return (DDI_FAILURE);
1478 
1479 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1480 		(void) (*psm_intr_ops)(rdip, hdlp,
1481 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1482 
1483 		/* Remove the interrupt handler */
1484 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1485 		    hdlp->ih_cb_func, hdlp->ih_vector);
1486 		break;
1487 	case DDI_INTROP_SETMASK:
1488 		if (psm_intr_ops == NULL)
1489 			return (DDI_FAILURE);
1490 
1491 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1492 			return (DDI_FAILURE);
1493 		break;
1494 	case DDI_INTROP_CLRMASK:
1495 		if (psm_intr_ops == NULL)
1496 			return (DDI_FAILURE);
1497 
1498 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1499 			return (DDI_FAILURE);
1500 		break;
1501 	case DDI_INTROP_GETPENDING:
1502 		if (psm_intr_ops == NULL)
1503 			return (DDI_FAILURE);
1504 
1505 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1506 		    result)) {
1507 			*(int *)result = 0;
1508 			return (DDI_FAILURE);
1509 		}
1510 		break;
1511 	case DDI_INTROP_NAVAIL:
1512 	case DDI_INTROP_NINTRS:
1513 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1514 		if (*(int *)result == 0) {
1515 			/*
1516 			 * Special case for 'pcic' driver' only. This driver
1517 			 * driver is a child of 'isa' and 'rootnex' drivers.
1518 			 *
1519 			 * See detailed comments on this in the function
1520 			 * rootnex_get_ispec().
1521 			 *
1522 			 * Children of 'pcic' send 'NINITR' request all the
1523 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1524 			 * field may not initialized. So, we fake it here
1525 			 * to return 1 (a la what PCMCIA nexus does).
1526 			 */
1527 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1528 				*(int *)result = 1;
1529 			else
1530 				return (DDI_FAILURE);
1531 		}
1532 		break;
1533 	case DDI_INTROP_SUPPORTED_TYPES:
1534 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1535 		break;
1536 	default:
1537 		return (DDI_FAILURE);
1538 	}
1539 
1540 	return (DDI_SUCCESS);
1541 }
1542 
1543 
1544 /*
1545  * rootnex_get_ispec()
1546  *	convert an interrupt number to an interrupt specification.
1547  *	The interrupt number determines which interrupt spec will be
1548  *	returned if more than one exists.
1549  *
1550  *	Look into the parent private data area of the 'rdip' to find out
1551  *	the interrupt specification.  First check to make sure there is
1552  *	one that matchs "inumber" and then return a pointer to it.
1553  *
1554  *	Return NULL if one could not be found.
1555  *
1556  *	NOTE: This is needed for rootnex_intr_ops()
1557  */
1558 static struct intrspec *
1559 rootnex_get_ispec(dev_info_t *rdip, int inum)
1560 {
1561 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1562 
1563 	/*
1564 	 * Special case handling for drivers that provide their own
1565 	 * intrspec structures instead of relying on the DDI framework.
1566 	 *
1567 	 * A broken hardware driver in ON could potentially provide its
1568 	 * own intrspec structure, instead of relying on the hardware.
1569 	 * If these drivers are children of 'rootnex' then we need to
1570 	 * continue to provide backward compatibility to them here.
1571 	 *
1572 	 * Following check is a special case for 'pcic' driver which
1573 	 * was found to have broken hardwre andby provides its own intrspec.
1574 	 *
1575 	 * Verbatim comments from this driver are shown here:
1576 	 * "Don't use the ddi_add_intr since we don't have a
1577 	 * default intrspec in all cases."
1578 	 *
1579 	 * Since an 'ispec' may not be always created for it,
1580 	 * check for that and create one if so.
1581 	 *
1582 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1583 	 */
1584 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1585 		pdp->par_nintr = 1;
1586 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1587 		    pdp->par_nintr, KM_SLEEP);
1588 	}
1589 
1590 	/* Validate the interrupt number */
1591 	if (inum >= pdp->par_nintr)
1592 		return (NULL);
1593 
1594 	/* Get the interrupt structure pointer and return that */
1595 	return ((struct intrspec *)&pdp->par_intr[inum]);
1596 }
1597 
1598 
1599 /*
1600  * ******************
1601  *  dma related code
1602  * ******************
1603  */
1604 
1605 /*ARGSUSED*/
1606 static int
1607 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1608     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1609     ddi_dma_handle_t *handlep)
1610 {
1611 	uint64_t maxsegmentsize_ll;
1612 	uint_t maxsegmentsize;
1613 	ddi_dma_impl_t *hp;
1614 	rootnex_dma_t *dma;
1615 	uint64_t count_max;
1616 	uint64_t seg;
1617 	int kmflag;
1618 	int e;
1619 
1620 
1621 	/* convert our sleep flags */
1622 	if (waitfp == DDI_DMA_SLEEP) {
1623 		kmflag = KM_SLEEP;
1624 	} else {
1625 		kmflag = KM_NOSLEEP;
1626 	}
1627 
1628 	/*
1629 	 * We try to do only one memory allocation here. We'll do a little
1630 	 * pointer manipulation later. If the bind ends up taking more than
1631 	 * our prealloc's space, we'll have to allocate more memory in the
1632 	 * bind operation. Not great, but much better than before and the
1633 	 * best we can do with the current bind interfaces.
1634 	 */
1635 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1636 	if (hp == NULL) {
1637 		if (waitfp != DDI_DMA_DONTWAIT) {
1638 			ddi_set_callback(waitfp, arg,
1639 			    &rootnex_state->r_dvma_call_list_id);
1640 		}
1641 		return (DDI_DMA_NORESOURCES);
1642 	}
1643 
1644 	/* Do our pointer manipulation now, align the structures */
1645 	hp->dmai_private = (void *)(((uintptr_t)hp +
1646 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1647 	dma = (rootnex_dma_t *)hp->dmai_private;
1648 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1649 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1650 
1651 	/* setup the handle */
1652 	rootnex_clean_dmahdl(hp);
1653 	dma->dp_dip = rdip;
1654 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1655 	dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1656 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1657 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1658 	hp->dmai_rdip = rdip;
1659 	hp->dmai_attr = *attr;
1660 
1661 	/* we don't need to worry about the SPL since we do a tryenter */
1662 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1663 
1664 	/*
1665 	 * Figure out our maximum segment size. If the segment size is greater
1666 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1667 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1668 	 * dma_attr_count_max are size-1 type values.
1669 	 *
1670 	 * Maximum segment size is the largest physically contiguous chunk of
1671 	 * memory that we can return from a bind (i.e. the maximum size of a
1672 	 * single cookie).
1673 	 */
1674 
1675 	/* handle the rollover cases */
1676 	seg = attr->dma_attr_seg + 1;
1677 	if (seg < attr->dma_attr_seg) {
1678 		seg = attr->dma_attr_seg;
1679 	}
1680 	count_max = attr->dma_attr_count_max + 1;
1681 	if (count_max < attr->dma_attr_count_max) {
1682 		count_max = attr->dma_attr_count_max;
1683 	}
1684 
1685 	/*
1686 	 * granularity may or may not be a power of two. If it isn't, we can't
1687 	 * use a simple mask.
1688 	 */
1689 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1690 		dma->dp_granularity_power_2 = B_FALSE;
1691 	} else {
1692 		dma->dp_granularity_power_2 = B_TRUE;
1693 	}
1694 
1695 	/*
1696 	 * maxxfer should be a whole multiple of granularity. If we're going to
1697 	 * break up a window because we're greater than maxxfer, we might as
1698 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1699 	 * worry about triming the window later on for this case.
1700 	 */
1701 	if (attr->dma_attr_granular > 1) {
1702 		if (dma->dp_granularity_power_2) {
1703 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1704 			    (attr->dma_attr_maxxfer &
1705 			    (attr->dma_attr_granular - 1));
1706 		} else {
1707 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1708 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1709 		}
1710 	} else {
1711 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1712 	}
1713 
1714 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1715 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1716 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1717 		maxsegmentsize = 0xFFFFFFFF;
1718 	} else {
1719 		maxsegmentsize = maxsegmentsize_ll;
1720 	}
1721 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1722 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1723 
1724 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1725 	if (rootnex_alloc_check_parms) {
1726 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1727 		if (e != DDI_SUCCESS) {
1728 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1729 			(void) rootnex_dma_freehdl(dip, rdip,
1730 			    (ddi_dma_handle_t)hp);
1731 			return (e);
1732 		}
1733 	}
1734 
1735 	*handlep = (ddi_dma_handle_t)hp;
1736 
1737 	ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1738 	DTRACE_PROBE1(rootnex__alloc__handle, uint64_t,
1739 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1740 
1741 	return (DDI_SUCCESS);
1742 }
1743 
1744 
1745 /*
1746  * rootnex_dma_allochdl()
1747  *    called from ddi_dma_alloc_handle().
1748  */
1749 static int
1750 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1751     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1752 {
1753 #if !defined(__xpv)
1754 	uint_t error = ENOTSUP;
1755 	int retval;
1756 
1757 	retval = iommulib_nex_open(rdip, &error);
1758 
1759 	if (retval != DDI_SUCCESS && error == ENOTSUP) {
1760 		/* No IOMMU */
1761 		return (rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1762 		    handlep));
1763 	} else if (retval != DDI_SUCCESS) {
1764 		return (DDI_FAILURE);
1765 	}
1766 
1767 	ASSERT(IOMMU_USED(rdip));
1768 
1769 	/* has an IOMMU */
1770 	return (iommulib_nexdma_allochdl(dip, rdip, attr,
1771 	    waitfp, arg, handlep));
1772 #else
1773 	return (rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1774 	    handlep));
1775 #endif
1776 }
1777 
1778 /*ARGSUSED*/
1779 static int
1780 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1781     ddi_dma_handle_t handle)
1782 {
1783 	ddi_dma_impl_t *hp;
1784 	rootnex_dma_t *dma;
1785 
1786 
1787 	hp = (ddi_dma_impl_t *)handle;
1788 	dma = (rootnex_dma_t *)hp->dmai_private;
1789 
1790 	/* unbind should have been called first */
1791 	ASSERT(!dma->dp_inuse);
1792 
1793 	mutex_destroy(&dma->dp_mutex);
1794 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1795 
1796 	ROOTNEX_PROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1797 	DTRACE_PROBE1(rootnex__free__handle, uint64_t,
1798 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1799 
1800 	if (rootnex_state->r_dvma_call_list_id)
1801 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1802 
1803 	return (DDI_SUCCESS);
1804 }
1805 
1806 /*
1807  * rootnex_dma_freehdl()
1808  *    called from ddi_dma_free_handle().
1809  */
1810 static int
1811 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1812 {
1813 #if !defined(__xpv)
1814 	if (IOMMU_USED(rdip)) {
1815 		return (iommulib_nexdma_freehdl(dip, rdip, handle));
1816 	}
1817 #endif
1818 	return (rootnex_coredma_freehdl(dip, rdip, handle));
1819 }
1820 
1821 
1822 /*ARGSUSED*/
1823 static int
1824 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1825     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1826     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1827 {
1828 	rootnex_sglinfo_t *sinfo;
1829 	ddi_dma_attr_t *attr;
1830 	ddi_dma_impl_t *hp;
1831 	rootnex_dma_t *dma;
1832 	int kmflag;
1833 	int e;
1834 
1835 
1836 	hp = (ddi_dma_impl_t *)handle;
1837 	dma = (rootnex_dma_t *)hp->dmai_private;
1838 	sinfo = &dma->dp_sglinfo;
1839 	attr = &hp->dmai_attr;
1840 
1841 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1842 		dma->dp_sleep_flags = KM_SLEEP;
1843 	} else {
1844 		dma->dp_sleep_flags = KM_NOSLEEP;
1845 	}
1846 
1847 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1848 
1849 	/*
1850 	 * This is useful for debugging a driver. Not as useful in a production
1851 	 * system. The only time this will fail is if you have a driver bug.
1852 	 */
1853 	if (rootnex_bind_check_inuse) {
1854 		/*
1855 		 * No one else should ever have this lock unless someone else
1856 		 * is trying to use this handle. So contention on the lock
1857 		 * is the same as inuse being set.
1858 		 */
1859 		e = mutex_tryenter(&dma->dp_mutex);
1860 		if (e == 0) {
1861 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1862 			return (DDI_DMA_INUSE);
1863 		}
1864 		if (dma->dp_inuse) {
1865 			mutex_exit(&dma->dp_mutex);
1866 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1867 			return (DDI_DMA_INUSE);
1868 		}
1869 		dma->dp_inuse = B_TRUE;
1870 		mutex_exit(&dma->dp_mutex);
1871 	}
1872 
1873 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1874 	if (rootnex_bind_check_parms) {
1875 		e = rootnex_valid_bind_parms(dmareq, attr);
1876 		if (e != DDI_SUCCESS) {
1877 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1878 			rootnex_clean_dmahdl(hp);
1879 			return (e);
1880 		}
1881 	}
1882 
1883 	/* save away the original bind info */
1884 	dma->dp_dma = dmareq->dmar_object;
1885 
1886 #if !defined(__xpv)
1887 	if (rootnex_state->r_intel_iommu_enabled) {
1888 		e = intel_iommu_map_sgl(handle, dmareq,
1889 		    rootnex_state->r_prealloc_cookies);
1890 
1891 		switch (e) {
1892 		case IOMMU_SGL_SUCCESS:
1893 			goto rootnex_sgl_end;
1894 
1895 		case IOMMU_SGL_DISABLE:
1896 			goto rootnex_sgl_start;
1897 
1898 		case IOMMU_SGL_NORESOURCES:
1899 			cmn_err(CE_WARN, "iommu map sgl failed for %s",
1900 			    ddi_node_name(dma->dp_dip));
1901 			rootnex_clean_dmahdl(hp);
1902 			return (DDI_DMA_NORESOURCES);
1903 
1904 		default:
1905 			cmn_err(CE_WARN,
1906 			    "undefined value returned from"
1907 			    " intel_iommu_map_sgl: %d",
1908 			    e);
1909 			rootnex_clean_dmahdl(hp);
1910 			return (DDI_DMA_NORESOURCES);
1911 		}
1912 	}
1913 #endif
1914 
1915 rootnex_sgl_start:
1916 	/*
1917 	 * Figure out a rough estimate of what maximum number of pages this
1918 	 * buffer could use (a high estimate of course).
1919 	 */
1920 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
1921 
1922 	/*
1923 	 * We'll use the pre-allocated cookies for any bind that will *always*
1924 	 * fit (more important to be consistent, we don't want to create
1925 	 * additional degenerate cases).
1926 	 */
1927 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
1928 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
1929 		dma->dp_need_to_free_cookie = B_FALSE;
1930 		DTRACE_PROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
1931 		    uint_t, sinfo->si_max_pages);
1932 
1933 	/*
1934 	 * For anything larger than that, we'll go ahead and allocate the
1935 	 * maximum number of pages we expect to see. Hopefuly, we won't be
1936 	 * seeing this path in the fast path for high performance devices very
1937 	 * frequently.
1938 	 *
1939 	 * a ddi bind interface that allowed the driver to provide storage to
1940 	 * the bind interface would speed this case up.
1941 	 */
1942 	} else {
1943 		/* convert the sleep flags */
1944 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1945 			kmflag =  KM_SLEEP;
1946 		} else {
1947 			kmflag =  KM_NOSLEEP;
1948 		}
1949 
1950 		/*
1951 		 * Save away how much memory we allocated. If we're doing a
1952 		 * nosleep, the alloc could fail...
1953 		 */
1954 		dma->dp_cookie_size = sinfo->si_max_pages *
1955 		    sizeof (ddi_dma_cookie_t);
1956 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
1957 		if (dma->dp_cookies == NULL) {
1958 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1959 			rootnex_clean_dmahdl(hp);
1960 			return (DDI_DMA_NORESOURCES);
1961 		}
1962 		dma->dp_need_to_free_cookie = B_TRUE;
1963 		DTRACE_PROBE2(rootnex__bind__alloc, dev_info_t *, rdip, uint_t,
1964 		    sinfo->si_max_pages);
1965 	}
1966 	hp->dmai_cookie = dma->dp_cookies;
1967 
1968 	/*
1969 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
1970 	 * looking at the contraints in the dma structure. It will then put some
1971 	 * additional state about the sgl in the dma struct (i.e. is the sgl
1972 	 * clean, or do we need to do some munging; how many pages need to be
1973 	 * copied, etc.)
1974 	 */
1975 	rootnex_get_sgl(&dmareq->dmar_object, dma->dp_cookies,
1976 	    &dma->dp_sglinfo);
1977 
1978 rootnex_sgl_end:
1979 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
1980 	/* if we don't need a copy buffer, we don't need to sync */
1981 	if (sinfo->si_copybuf_req == 0) {
1982 		hp->dmai_rflags |= DMP_NOSYNC;
1983 	}
1984 
1985 	/*
1986 	 * if we don't need the copybuf and we don't need to do a partial,  we
1987 	 * hit the fast path. All the high performance devices should be trying
1988 	 * to hit this path. To hit this path, a device should be able to reach
1989 	 * all of memory, shouldn't try to bind more than it can transfer, and
1990 	 * the buffer shouldn't require more cookies than the driver/device can
1991 	 * handle [sgllen]).
1992 	 */
1993 	if ((sinfo->si_copybuf_req == 0) &&
1994 	    (sinfo->si_sgl_size <= attr->dma_attr_sgllen) &&
1995 	    (dma->dp_dma.dmao_size < dma->dp_maxxfer)) {
1996 		/*
1997 		 * If the driver supports FMA, insert the handle in the FMA DMA
1998 		 * handle cache.
1999 		 */
2000 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR) {
2001 			hp->dmai_error.err_cf = rootnex_dma_check;
2002 			(void) ndi_fmc_insert(rdip, DMA_HANDLE, hp, NULL);
2003 		}
2004 
2005 		/*
2006 		 * copy out the first cookie and ccountp, set the cookie
2007 		 * pointer to the second cookie. The first cookie is passed
2008 		 * back on the stack. Additional cookies are accessed via
2009 		 * ddi_dma_nextcookie()
2010 		 */
2011 		*cookiep = dma->dp_cookies[0];
2012 		*ccountp = sinfo->si_sgl_size;
2013 		hp->dmai_cookie++;
2014 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2015 		hp->dmai_nwin = 1;
2016 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2017 		DTRACE_PROBE3(rootnex__bind__fast, dev_info_t *, rdip, uint64_t,
2018 		    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2019 		    dma->dp_dma.dmao_size);
2020 		return (DDI_DMA_MAPPED);
2021 	}
2022 
2023 	/*
2024 	 * go to the slow path, we may need to alloc more memory, create
2025 	 * multiple windows, and munge up a sgl to make the device happy.
2026 	 */
2027 	e = rootnex_bind_slowpath(hp, dmareq, dma, attr, kmflag);
2028 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2029 		if (dma->dp_need_to_free_cookie) {
2030 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2031 		}
2032 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2033 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2034 		return (e);
2035 	}
2036 
2037 	/*
2038 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2039 	 * cache.
2040 	 */
2041 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR) {
2042 		hp->dmai_error.err_cf = rootnex_dma_check;
2043 		(void) ndi_fmc_insert(rdip, DMA_HANDLE, hp, NULL);
2044 	}
2045 
2046 	/* if the first window uses the copy buffer, sync it for the device */
2047 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2048 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2049 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2050 		    DDI_DMA_SYNC_FORDEV);
2051 	}
2052 
2053 	/*
2054 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2055 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2056 	 * If we have a partial map (i.e. multiple windows), the number of
2057 	 * cookies we return is the number of cookies in the first window.
2058 	 */
2059 	if (e == DDI_DMA_MAPPED) {
2060 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2061 		*ccountp = sinfo->si_sgl_size;
2062 	} else {
2063 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2064 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2065 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2066 	}
2067 	*cookiep = dma->dp_cookies[0];
2068 	hp->dmai_cookie++;
2069 
2070 	ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2071 	DTRACE_PROBE3(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2072 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2073 	    dma->dp_dma.dmao_size);
2074 	return (e);
2075 }
2076 
2077 
2078 /*
2079  * rootnex_dma_bindhdl()
2080  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2081  */
2082 static int
2083 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2084     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2085     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2086 {
2087 #if !defined(__xpv)
2088 	if (IOMMU_USED(rdip)) {
2089 		return (iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2090 		    cookiep, ccountp));
2091 	}
2092 #endif
2093 	return (rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2094 	    cookiep, ccountp));
2095 }
2096 
2097 /*ARGSUSED*/
2098 static int
2099 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2100     ddi_dma_handle_t handle)
2101 {
2102 	ddi_dma_impl_t *hp;
2103 	rootnex_dma_t *dma;
2104 	int e;
2105 
2106 	hp = (ddi_dma_impl_t *)handle;
2107 	dma = (rootnex_dma_t *)hp->dmai_private;
2108 
2109 	/* make sure the buffer wasn't free'd before calling unbind */
2110 	if (rootnex_unbind_verify_buffer) {
2111 		e = rootnex_verify_buffer(dma);
2112 		if (e != DDI_SUCCESS) {
2113 			ASSERT(0);
2114 			return (DDI_FAILURE);
2115 		}
2116 	}
2117 
2118 	/* sync the current window before unbinding the buffer */
2119 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2120 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2121 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2122 		    DDI_DMA_SYNC_FORCPU);
2123 	}
2124 
2125 	/*
2126 	 * If the driver supports FMA, remove the handle in the FMA DMA handle
2127 	 * cache.
2128 	 */
2129 	if (hp->dmai_attr.dma_attr_flags & DDI_DMA_FLAGERR) {
2130 		if ((DEVI(rdip)->devi_fmhdl != NULL) &&
2131 		    (DDI_FM_DMA_ERR_CAP(DEVI(rdip)->devi_fmhdl->fh_cap))) {
2132 			(void) ndi_fmc_remove(rdip, DMA_HANDLE, hp);
2133 		}
2134 	}
2135 
2136 	/*
2137 	 * cleanup and copy buffer or window state. if we didn't use the copy
2138 	 * buffer or windows, there won't be much to do :-)
2139 	 */
2140 	rootnex_teardown_copybuf(dma);
2141 	rootnex_teardown_windows(dma);
2142 
2143 #if !defined(__xpv)
2144 	/*
2145 	 * If intel iommu enabled, clean up the page tables and free the dvma
2146 	 */
2147 	if (rootnex_state->r_intel_iommu_enabled) {
2148 		intel_iommu_unmap_sgl(handle);
2149 	}
2150 #endif
2151 
2152 	/*
2153 	 * If we had to allocate space to for the worse case sgl (it didn't
2154 	 * fit into our pre-allocate buffer), free that up now
2155 	 */
2156 	if (dma->dp_need_to_free_cookie) {
2157 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2158 	}
2159 
2160 	/*
2161 	 * clean up the handle so it's ready for the next bind (i.e. if the
2162 	 * handle is reused).
2163 	 */
2164 	rootnex_clean_dmahdl(hp);
2165 
2166 	if (rootnex_state->r_dvma_call_list_id)
2167 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2168 
2169 	ROOTNEX_PROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2170 	DTRACE_PROBE1(rootnex__unbind, uint64_t,
2171 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2172 
2173 	return (DDI_SUCCESS);
2174 }
2175 
2176 /*
2177  * rootnex_dma_unbindhdl()
2178  *    called from ddi_dma_unbind_handle()
2179  */
2180 /*ARGSUSED*/
2181 static int
2182 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2183     ddi_dma_handle_t handle)
2184 {
2185 #if !defined(__xpv)
2186 	if (IOMMU_USED(rdip)) {
2187 		return (iommulib_nexdma_unbindhdl(dip, rdip, handle));
2188 	}
2189 #endif
2190 	return (rootnex_coredma_unbindhdl(dip, rdip, handle));
2191 }
2192 
2193 #if !defined(__xpv)
2194 
2195 static int
2196 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2197 {
2198 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2199 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2200 
2201 	if (dma->dp_sleep_flags != KM_SLEEP &&
2202 	    dma->dp_sleep_flags != KM_NOSLEEP)
2203 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2204 	return (dma->dp_sleep_flags);
2205 }
2206 /*ARGSUSED*/
2207 static void
2208 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2209 {
2210 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2211 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2212 	rootnex_window_t *window;
2213 
2214 	if (dma->dp_window) {
2215 		window = &dma->dp_window[dma->dp_current_win];
2216 		hp->dmai_cookie = window->wd_first_cookie;
2217 	} else {
2218 		hp->dmai_cookie = dma->dp_cookies;
2219 	}
2220 	hp->dmai_cookie++;
2221 }
2222 
2223 /*ARGSUSED*/
2224 static int
2225 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2226     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2227 {
2228 	int i;
2229 	int km_flags;
2230 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2231 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2232 	rootnex_window_t *window;
2233 	ddi_dma_cookie_t *cp;
2234 	ddi_dma_cookie_t *cookie;
2235 
2236 	ASSERT(*cookiepp == NULL);
2237 	ASSERT(*ccountp == 0);
2238 
2239 	if (dma->dp_window) {
2240 		window = &dma->dp_window[dma->dp_current_win];
2241 		cp = window->wd_first_cookie;
2242 		*ccountp = window->wd_cookie_cnt;
2243 	} else {
2244 		cp = dma->dp_cookies;
2245 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2246 	}
2247 
2248 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2249 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2250 	if (cookie == NULL) {
2251 		return (DDI_DMA_NORESOURCES);
2252 	}
2253 
2254 	for (i = 0; i < *ccountp; i++) {
2255 		cookie[i].dmac_notused = cp[i].dmac_notused;
2256 		cookie[i].dmac_type = cp[i].dmac_type;
2257 		cookie[i].dmac_address = cp[i].dmac_address;
2258 		cookie[i].dmac_size = cp[i].dmac_size;
2259 	}
2260 
2261 	*cookiepp = cookie;
2262 
2263 	return (DDI_SUCCESS);
2264 }
2265 
2266 /*ARGSUSED*/
2267 static int
2268 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2269     ddi_dma_cookie_t *cookiep, uint_t ccount)
2270 {
2271 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2272 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2273 	rootnex_window_t *window;
2274 	ddi_dma_cookie_t *cur_cookiep;
2275 
2276 	ASSERT(cookiep);
2277 	ASSERT(ccount != 0);
2278 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2279 
2280 	if (dma->dp_window) {
2281 		window = &dma->dp_window[dma->dp_current_win];
2282 		dma->dp_saved_cookies = window->wd_first_cookie;
2283 		window->wd_first_cookie = cookiep;
2284 		ASSERT(ccount == window->wd_cookie_cnt);
2285 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2286 		    + window->wd_first_cookie;
2287 	} else {
2288 		dma->dp_saved_cookies = dma->dp_cookies;
2289 		dma->dp_cookies = cookiep;
2290 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2291 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2292 		    + dma->dp_cookies;
2293 	}
2294 
2295 	dma->dp_need_to_switch_cookies = B_TRUE;
2296 	hp->dmai_cookie = cur_cookiep;
2297 
2298 	return (DDI_SUCCESS);
2299 }
2300 
2301 /*ARGSUSED*/
2302 static int
2303 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2304 {
2305 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2306 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2307 	rootnex_window_t *window;
2308 	ddi_dma_cookie_t *cur_cookiep;
2309 	ddi_dma_cookie_t *cookie_array;
2310 	uint_t ccount;
2311 
2312 	/* check if cookies have not been switched */
2313 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2314 		return (DDI_SUCCESS);
2315 
2316 	ASSERT(dma->dp_saved_cookies);
2317 
2318 	if (dma->dp_window) {
2319 		window = &dma->dp_window[dma->dp_current_win];
2320 		cookie_array = window->wd_first_cookie;
2321 		window->wd_first_cookie = dma->dp_saved_cookies;
2322 		dma->dp_saved_cookies = NULL;
2323 		ccount = window->wd_cookie_cnt;
2324 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2325 		    + window->wd_first_cookie;
2326 	} else {
2327 		cookie_array = dma->dp_cookies;
2328 		dma->dp_cookies = dma->dp_saved_cookies;
2329 		dma->dp_saved_cookies = NULL;
2330 		ccount = dma->dp_sglinfo.si_sgl_size;
2331 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2332 		    + dma->dp_cookies;
2333 	}
2334 
2335 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2336 
2337 	hp->dmai_cookie = cur_cookiep;
2338 
2339 	dma->dp_need_to_switch_cookies = B_FALSE;
2340 
2341 	return (DDI_SUCCESS);
2342 }
2343 
2344 #endif
2345 
2346 /*
2347  * rootnex_verify_buffer()
2348  *   verify buffer wasn't free'd
2349  */
2350 static int
2351 rootnex_verify_buffer(rootnex_dma_t *dma)
2352 {
2353 	page_t **pplist;
2354 	caddr_t vaddr;
2355 	uint_t pcnt;
2356 	uint_t poff;
2357 	page_t *pp;
2358 	char b;
2359 	int i;
2360 
2361 	/* Figure out how many pages this buffer occupies */
2362 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2363 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2364 	} else {
2365 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2366 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2367 	}
2368 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2369 
2370 	switch (dma->dp_dma.dmao_type) {
2371 	case DMA_OTYP_PAGES:
2372 		/*
2373 		 * for a linked list of pp's walk through them to make sure
2374 		 * they're locked and not free.
2375 		 */
2376 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2377 		for (i = 0; i < pcnt; i++) {
2378 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2379 				return (DDI_FAILURE);
2380 			}
2381 			pp = pp->p_next;
2382 		}
2383 		break;
2384 
2385 	case DMA_OTYP_VADDR:
2386 	case DMA_OTYP_BUFVADDR:
2387 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2388 		/*
2389 		 * for an array of pp's walk through them to make sure they're
2390 		 * not free. It's possible that they may not be locked.
2391 		 */
2392 		if (pplist) {
2393 			for (i = 0; i < pcnt; i++) {
2394 				if (PP_ISFREE(pplist[i])) {
2395 					return (DDI_FAILURE);
2396 				}
2397 			}
2398 
2399 		/* For a virtual address, try to peek at each page */
2400 		} else {
2401 			if (dma->dp_sglinfo.si_asp == &kas) {
2402 				for (i = 0; i < pcnt; i++) {
2403 					if (ddi_peek8(NULL, vaddr, &b) ==
2404 					    DDI_FAILURE)
2405 						return (DDI_FAILURE);
2406 					vaddr += MMU_PAGESIZE;
2407 				}
2408 			}
2409 		}
2410 		break;
2411 
2412 	default:
2413 		ASSERT(0);
2414 		break;
2415 	}
2416 
2417 	return (DDI_SUCCESS);
2418 }
2419 
2420 
2421 /*
2422  * rootnex_clean_dmahdl()
2423  *    Clean the dma handle. This should be called on a handle alloc and an
2424  *    unbind handle. Set the handle state to the default settings.
2425  */
2426 static void
2427 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2428 {
2429 	rootnex_dma_t *dma;
2430 
2431 
2432 	dma = (rootnex_dma_t *)hp->dmai_private;
2433 
2434 	hp->dmai_nwin = 0;
2435 	dma->dp_current_cookie = 0;
2436 	dma->dp_copybuf_size = 0;
2437 	dma->dp_window = NULL;
2438 	dma->dp_cbaddr = NULL;
2439 	dma->dp_inuse = B_FALSE;
2440 	dma->dp_need_to_free_cookie = B_FALSE;
2441 	dma->dp_need_to_switch_cookies = B_FALSE;
2442 	dma->dp_saved_cookies = NULL;
2443 	dma->dp_sleep_flags = KM_PANIC;
2444 	dma->dp_need_to_free_window = B_FALSE;
2445 	dma->dp_partial_required = B_FALSE;
2446 	dma->dp_trim_required = B_FALSE;
2447 	dma->dp_sglinfo.si_copybuf_req = 0;
2448 #if !defined(__amd64)
2449 	dma->dp_cb_remaping = B_FALSE;
2450 	dma->dp_kva = NULL;
2451 #endif
2452 
2453 	/* FMA related initialization */
2454 	hp->dmai_fault = 0;
2455 	hp->dmai_fault_check = NULL;
2456 	hp->dmai_fault_notify = NULL;
2457 	hp->dmai_error.err_ena = 0;
2458 	hp->dmai_error.err_status = DDI_FM_OK;
2459 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2460 	hp->dmai_error.err_ontrap = NULL;
2461 	hp->dmai_error.err_fep = NULL;
2462 	hp->dmai_error.err_cf = NULL;
2463 }
2464 
2465 
2466 /*
2467  * rootnex_valid_alloc_parms()
2468  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2469  */
2470 static int
2471 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2472 {
2473 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2474 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2475 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2476 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2477 		return (DDI_DMA_BADATTR);
2478 	}
2479 
2480 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2481 		return (DDI_DMA_BADATTR);
2482 	}
2483 
2484 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2485 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2486 	    attr->dma_attr_sgllen <= 0) {
2487 		return (DDI_DMA_BADATTR);
2488 	}
2489 
2490 	/* We should be able to DMA into every byte offset in a page */
2491 	if (maxsegmentsize < MMU_PAGESIZE) {
2492 		return (DDI_DMA_BADATTR);
2493 	}
2494 
2495 	return (DDI_SUCCESS);
2496 }
2497 
2498 
2499 /*
2500  * rootnex_valid_bind_parms()
2501  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2502  */
2503 /* ARGSUSED */
2504 static int
2505 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2506 {
2507 #if !defined(__amd64)
2508 	/*
2509 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2510 	 * we can track the offset for the obsoleted interfaces.
2511 	 */
2512 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2513 		return (DDI_DMA_TOOBIG);
2514 	}
2515 #endif
2516 
2517 	return (DDI_SUCCESS);
2518 }
2519 
2520 
2521 /*
2522  * rootnex_get_sgl()
2523  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2524  *    with a call to the vm layer when vm2.0 comes around...
2525  */
2526 static void
2527 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2528     rootnex_sglinfo_t *sglinfo)
2529 {
2530 	ddi_dma_atyp_t buftype;
2531 	rootnex_addr_t raddr;
2532 	uint64_t last_page;
2533 	uint64_t offset;
2534 	uint64_t addrhi;
2535 	uint64_t addrlo;
2536 	uint64_t maxseg;
2537 	page_t **pplist;
2538 	uint64_t paddr;
2539 	uint32_t psize;
2540 	uint32_t size;
2541 	caddr_t vaddr;
2542 	uint_t pcnt;
2543 	page_t *pp;
2544 	uint_t cnt;
2545 
2546 
2547 	/* shortcuts */
2548 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2549 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2550 	maxseg = sglinfo->si_max_cookie_size;
2551 	buftype = dmar_object->dmao_type;
2552 	addrhi = sglinfo->si_max_addr;
2553 	addrlo = sglinfo->si_min_addr;
2554 	size = dmar_object->dmao_size;
2555 
2556 	pcnt = 0;
2557 	cnt = 0;
2558 
2559 	/*
2560 	 * if we were passed down a linked list of pages, i.e. pointer to
2561 	 * page_t, use this to get our physical address and buf offset.
2562 	 */
2563 	if (buftype == DMA_OTYP_PAGES) {
2564 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2565 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2566 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2567 		    MMU_PAGEOFFSET;
2568 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2569 		psize = MIN(size, (MMU_PAGESIZE - offset));
2570 		pp = pp->p_next;
2571 		sglinfo->si_asp = NULL;
2572 
2573 	/*
2574 	 * We weren't passed down a linked list of pages, but if we were passed
2575 	 * down an array of pages, use this to get our physical address and buf
2576 	 * offset.
2577 	 */
2578 	} else if (pplist != NULL) {
2579 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2580 		    (buftype == DMA_OTYP_BUFVADDR));
2581 
2582 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2583 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2584 		if (sglinfo->si_asp == NULL) {
2585 			sglinfo->si_asp = &kas;
2586 		}
2587 
2588 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2589 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2590 		paddr += offset;
2591 		psize = MIN(size, (MMU_PAGESIZE - offset));
2592 		pcnt++;
2593 
2594 	/*
2595 	 * All we have is a virtual address, we'll need to call into the VM
2596 	 * to get the physical address.
2597 	 */
2598 	} else {
2599 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2600 		    (buftype == DMA_OTYP_BUFVADDR));
2601 
2602 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2603 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2604 		if (sglinfo->si_asp == NULL) {
2605 			sglinfo->si_asp = &kas;
2606 		}
2607 
2608 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2609 		paddr += offset;
2610 		psize = MIN(size, (MMU_PAGESIZE - offset));
2611 		vaddr += psize;
2612 	}
2613 
2614 #ifdef __xpv
2615 	/*
2616 	 * If we're dom0, we're using a real device so we need to load
2617 	 * the cookies with MFNs instead of PFNs.
2618 	 */
2619 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2620 #else
2621 	raddr = paddr;
2622 #endif
2623 
2624 	/*
2625 	 * Setup the first cookie with the physical address of the page and the
2626 	 * size of the page (which takes into account the initial offset into
2627 	 * the page.
2628 	 */
2629 	sgl[cnt].dmac_laddress = raddr;
2630 	sgl[cnt].dmac_size = psize;
2631 	sgl[cnt].dmac_type = 0;
2632 
2633 	/*
2634 	 * Save away the buffer offset into the page. We'll need this later in
2635 	 * the copy buffer code to help figure out the page index within the
2636 	 * buffer and the offset into the current page.
2637 	 */
2638 	sglinfo->si_buf_offset = offset;
2639 
2640 	/*
2641 	 * If the DMA engine can't reach the physical address, increase how
2642 	 * much copy buffer we need. We always increase by pagesize so we don't
2643 	 * have to worry about converting offsets. Set a flag in the cookies
2644 	 * dmac_type to indicate that it uses the copy buffer. If this isn't the
2645 	 * last cookie, go to the next cookie (since we separate each page which
2646 	 * uses the copy buffer in case the copy buffer is not physically
2647 	 * contiguous.
2648 	 */
2649 	if ((raddr < addrlo) || ((raddr + psize) > addrhi)) {
2650 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
2651 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2652 		if ((cnt + 1) < sglinfo->si_max_pages) {
2653 			cnt++;
2654 			sgl[cnt].dmac_laddress = 0;
2655 			sgl[cnt].dmac_size = 0;
2656 			sgl[cnt].dmac_type = 0;
2657 		}
2658 	}
2659 
2660 	/*
2661 	 * save this page's physical address so we can figure out if the next
2662 	 * page is physically contiguous. Keep decrementing size until we are
2663 	 * done with the buffer.
2664 	 */
2665 	last_page = raddr & MMU_PAGEMASK;
2666 	size -= psize;
2667 
2668 	while (size > 0) {
2669 		/* Get the size for this page (i.e. partial or full page) */
2670 		psize = MIN(size, MMU_PAGESIZE);
2671 
2672 		if (buftype == DMA_OTYP_PAGES) {
2673 			/* get the paddr from the page_t */
2674 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2675 			paddr = pfn_to_pa(pp->p_pagenum);
2676 			pp = pp->p_next;
2677 		} else if (pplist != NULL) {
2678 			/* index into the array of page_t's to get the paddr */
2679 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2680 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2681 			pcnt++;
2682 		} else {
2683 			/* call into the VM to get the paddr */
2684 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2685 			    vaddr));
2686 			vaddr += psize;
2687 		}
2688 
2689 #ifdef __xpv
2690 		/*
2691 		 * If we're dom0, we're using a real device so we need to load
2692 		 * the cookies with MFNs instead of PFNs.
2693 		 */
2694 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2695 #else
2696 		raddr = paddr;
2697 #endif
2698 		/* check to see if this page needs the copy buffer */
2699 		if ((raddr < addrlo) || ((raddr + psize) > addrhi)) {
2700 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
2701 
2702 			/*
2703 			 * if there is something in the current cookie, go to
2704 			 * the next one. We only want one page in a cookie which
2705 			 * uses the copybuf since the copybuf doesn't have to
2706 			 * be physically contiguous.
2707 			 */
2708 			if (sgl[cnt].dmac_size != 0) {
2709 				cnt++;
2710 			}
2711 			sgl[cnt].dmac_laddress = raddr;
2712 			sgl[cnt].dmac_size = psize;
2713 #if defined(__amd64)
2714 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2715 #else
2716 			/*
2717 			 * save the buf offset for 32-bit kernel. used in the
2718 			 * obsoleted interfaces.
2719 			 */
2720 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
2721 			    (dmar_object->dmao_size - size);
2722 #endif
2723 			/* if this isn't the last cookie, go to the next one */
2724 			if ((cnt + 1) < sglinfo->si_max_pages) {
2725 				cnt++;
2726 				sgl[cnt].dmac_laddress = 0;
2727 				sgl[cnt].dmac_size = 0;
2728 				sgl[cnt].dmac_type = 0;
2729 			}
2730 
2731 		/*
2732 		 * this page didn't need the copy buffer, if it's not physically
2733 		 * contiguous, or it would put us over a segment boundary, or it
2734 		 * puts us over the max cookie size, or the current sgl doesn't
2735 		 * have anything in it.
2736 		 */
2737 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
2738 		    !(raddr & sglinfo->si_segmask) ||
2739 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
2740 		    (sgl[cnt].dmac_size == 0)) {
2741 			/*
2742 			 * if we're not already in a new cookie, go to the next
2743 			 * cookie.
2744 			 */
2745 			if (sgl[cnt].dmac_size != 0) {
2746 				cnt++;
2747 			}
2748 
2749 			/* save the cookie information */
2750 			sgl[cnt].dmac_laddress = raddr;
2751 			sgl[cnt].dmac_size = psize;
2752 #if defined(__amd64)
2753 			sgl[cnt].dmac_type = 0;
2754 #else
2755 			/*
2756 			 * save the buf offset for 32-bit kernel. used in the
2757 			 * obsoleted interfaces.
2758 			 */
2759 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
2760 #endif
2761 
2762 		/*
2763 		 * this page didn't need the copy buffer, it is physically
2764 		 * contiguous with the last page, and it's <= the max cookie
2765 		 * size.
2766 		 */
2767 		} else {
2768 			sgl[cnt].dmac_size += psize;
2769 
2770 			/*
2771 			 * if this exactly ==  the maximum cookie size, and
2772 			 * it isn't the last cookie, go to the next cookie.
2773 			 */
2774 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
2775 			    ((cnt + 1) < sglinfo->si_max_pages)) {
2776 				cnt++;
2777 				sgl[cnt].dmac_laddress = 0;
2778 				sgl[cnt].dmac_size = 0;
2779 				sgl[cnt].dmac_type = 0;
2780 			}
2781 		}
2782 
2783 		/*
2784 		 * save this page's physical address so we can figure out if the
2785 		 * next page is physically contiguous. Keep decrementing size
2786 		 * until we are done with the buffer.
2787 		 */
2788 		last_page = raddr;
2789 		size -= psize;
2790 	}
2791 
2792 	/* we're done, save away how many cookies the sgl has */
2793 	if (sgl[cnt].dmac_size == 0) {
2794 		ASSERT(cnt < sglinfo->si_max_pages);
2795 		sglinfo->si_sgl_size = cnt;
2796 	} else {
2797 		sglinfo->si_sgl_size = cnt + 1;
2798 	}
2799 }
2800 
2801 
2802 /*
2803  * rootnex_bind_slowpath()
2804  *    Call in the bind path if the calling driver can't use the sgl without
2805  *    modifying it. We either need to use the copy buffer and/or we will end up
2806  *    with a partial bind.
2807  */
2808 static int
2809 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
2810     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag)
2811 {
2812 	rootnex_sglinfo_t *sinfo;
2813 	rootnex_window_t *window;
2814 	ddi_dma_cookie_t *cookie;
2815 	size_t copybuf_used;
2816 	size_t dmac_size;
2817 	boolean_t partial;
2818 	off_t cur_offset;
2819 	page_t *cur_pp;
2820 	major_t mnum;
2821 	int e;
2822 	int i;
2823 
2824 
2825 	sinfo = &dma->dp_sglinfo;
2826 	copybuf_used = 0;
2827 	partial = B_FALSE;
2828 
2829 	/*
2830 	 * If we're using the copybuf, set the copybuf state in dma struct.
2831 	 * Needs to be first since it sets the copy buffer size.
2832 	 */
2833 	if (sinfo->si_copybuf_req != 0) {
2834 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
2835 		if (e != DDI_SUCCESS) {
2836 			return (e);
2837 		}
2838 	} else {
2839 		dma->dp_copybuf_size = 0;
2840 	}
2841 
2842 	/*
2843 	 * Figure out if we need to do a partial mapping. If so, figure out
2844 	 * if we need to trim the buffers when we munge the sgl.
2845 	 */
2846 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
2847 	    (dma->dp_dma.dmao_size > dma->dp_maxxfer) ||
2848 	    (attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
2849 		dma->dp_partial_required = B_TRUE;
2850 		if (attr->dma_attr_granular != 1) {
2851 			dma->dp_trim_required = B_TRUE;
2852 		}
2853 	} else {
2854 		dma->dp_partial_required = B_FALSE;
2855 		dma->dp_trim_required = B_FALSE;
2856 	}
2857 
2858 	/* If we need to do a partial bind, make sure the driver supports it */
2859 	if (dma->dp_partial_required &&
2860 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
2861 
2862 		mnum = ddi_driver_major(dma->dp_dip);
2863 		/*
2864 		 * patchable which allows us to print one warning per major
2865 		 * number.
2866 		 */
2867 		if ((rootnex_bind_warn) &&
2868 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
2869 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
2870 			cmn_err(CE_WARN, "!%s: coding error detected, the "
2871 			    "driver is using ddi_dma_attr(9S) incorrectly. "
2872 			    "There is a small risk of data corruption in "
2873 			    "particular with large I/Os. The driver should be "
2874 			    "replaced with a corrected version for proper "
2875 			    "system operation. To disable this warning, add "
2876 			    "'set rootnex:rootnex_bind_warn=0' to "
2877 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
2878 		}
2879 		return (DDI_DMA_TOOBIG);
2880 	}
2881 
2882 	/*
2883 	 * we might need multiple windows, setup state to handle them. In this
2884 	 * code path, we will have at least one window.
2885 	 */
2886 	e = rootnex_setup_windows(hp, dma, attr, kmflag);
2887 	if (e != DDI_SUCCESS) {
2888 		rootnex_teardown_copybuf(dma);
2889 		return (e);
2890 	}
2891 
2892 	window = &dma->dp_window[0];
2893 	cookie = &dma->dp_cookies[0];
2894 	cur_offset = 0;
2895 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
2896 	if (dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) {
2897 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
2898 	}
2899 
2900 	/* loop though all the cookies we got back from get_sgl() */
2901 	for (i = 0; i < sinfo->si_sgl_size; i++) {
2902 		/*
2903 		 * If we're using the copy buffer, check this cookie and setup
2904 		 * its associated copy buffer state. If this cookie uses the
2905 		 * copy buffer, make sure we sync this window during dma_sync.
2906 		 */
2907 		if (dma->dp_copybuf_size > 0) {
2908 			rootnex_setup_cookie(&dmareq->dmar_object, dma, cookie,
2909 			    cur_offset, &copybuf_used, &cur_pp);
2910 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
2911 				window->wd_dosync = B_TRUE;
2912 			}
2913 		}
2914 
2915 		/*
2916 		 * save away the cookie size, since it could be modified in
2917 		 * the windowing code.
2918 		 */
2919 		dmac_size = cookie->dmac_size;
2920 
2921 		/* if we went over max copybuf size */
2922 		if (dma->dp_copybuf_size &&
2923 		    (copybuf_used > dma->dp_copybuf_size)) {
2924 			partial = B_TRUE;
2925 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
2926 			    cookie, cur_offset, &copybuf_used);
2927 			if (e != DDI_SUCCESS) {
2928 				rootnex_teardown_copybuf(dma);
2929 				rootnex_teardown_windows(dma);
2930 				return (e);
2931 			}
2932 
2933 			/*
2934 			 * if the coookie uses the copy buffer, make sure the
2935 			 * new window we just moved to is set to sync.
2936 			 */
2937 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
2938 				window->wd_dosync = B_TRUE;
2939 			}
2940 			DTRACE_PROBE1(rootnex__copybuf__window, dev_info_t *,
2941 			    dma->dp_dip);
2942 
2943 		/* if the cookie cnt == max sgllen, move to the next window */
2944 		} else if (window->wd_cookie_cnt >= attr->dma_attr_sgllen) {
2945 			partial = B_TRUE;
2946 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
2947 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
2948 			    cookie, attr, cur_offset);
2949 			if (e != DDI_SUCCESS) {
2950 				rootnex_teardown_copybuf(dma);
2951 				rootnex_teardown_windows(dma);
2952 				return (e);
2953 			}
2954 
2955 			/*
2956 			 * if the coookie uses the copy buffer, make sure the
2957 			 * new window we just moved to is set to sync.
2958 			 */
2959 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
2960 				window->wd_dosync = B_TRUE;
2961 			}
2962 			DTRACE_PROBE1(rootnex__sgllen__window, dev_info_t *,
2963 			    dma->dp_dip);
2964 
2965 		/* else if we will be over maxxfer */
2966 		} else if ((window->wd_size + dmac_size) >
2967 		    dma->dp_maxxfer) {
2968 			partial = B_TRUE;
2969 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
2970 			    cookie);
2971 			if (e != DDI_SUCCESS) {
2972 				rootnex_teardown_copybuf(dma);
2973 				rootnex_teardown_windows(dma);
2974 				return (e);
2975 			}
2976 
2977 			/*
2978 			 * if the coookie uses the copy buffer, make sure the
2979 			 * new window we just moved to is set to sync.
2980 			 */
2981 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
2982 				window->wd_dosync = B_TRUE;
2983 			}
2984 			DTRACE_PROBE1(rootnex__maxxfer__window, dev_info_t *,
2985 			    dma->dp_dip);
2986 
2987 		/* else this cookie fits in the current window */
2988 		} else {
2989 			window->wd_cookie_cnt++;
2990 			window->wd_size += dmac_size;
2991 		}
2992 
2993 		/* track our offset into the buffer, go to the next cookie */
2994 		ASSERT(dmac_size <= dma->dp_dma.dmao_size);
2995 		ASSERT(cookie->dmac_size <= dmac_size);
2996 		cur_offset += dmac_size;
2997 		cookie++;
2998 	}
2999 
3000 	/* if we ended up with a zero sized window in the end, clean it up */
3001 	if (window->wd_size == 0) {
3002 		hp->dmai_nwin--;
3003 		window--;
3004 	}
3005 
3006 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3007 
3008 	if (!partial) {
3009 		return (DDI_DMA_MAPPED);
3010 	}
3011 
3012 	ASSERT(dma->dp_partial_required);
3013 	return (DDI_DMA_PARTIAL_MAP);
3014 }
3015 
3016 
3017 /*
3018  * rootnex_setup_copybuf()
3019  *    Called in bind slowpath. Figures out if we're going to use the copy
3020  *    buffer, and if we do, sets up the basic state to handle it.
3021  */
3022 static int
3023 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3024     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3025 {
3026 	rootnex_sglinfo_t *sinfo;
3027 	ddi_dma_attr_t lattr;
3028 	size_t max_copybuf;
3029 	int cansleep;
3030 	int e;
3031 #if !defined(__amd64)
3032 	int vmflag;
3033 #endif
3034 
3035 
3036 	sinfo = &dma->dp_sglinfo;
3037 
3038 	/* read this first so it's consistent through the routine  */
3039 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3040 
3041 	/* We need to call into the rootnex on ddi_dma_sync() */
3042 	hp->dmai_rflags &= ~DMP_NOSYNC;
3043 
3044 	/* make sure the copybuf size <= the max size */
3045 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3046 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3047 
3048 #if !defined(__amd64)
3049 	/*
3050 	 * if we don't have kva space to copy to/from, allocate the KVA space
3051 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3052 	 * the 64-bit kernel.
3053 	 */
3054 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3055 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3056 
3057 		/* convert the sleep flags */
3058 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3059 			vmflag = VM_SLEEP;
3060 		} else {
3061 			vmflag = VM_NOSLEEP;
3062 		}
3063 
3064 		/* allocate Kernel VA space that we can bcopy to/from */
3065 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3066 		    vmflag);
3067 		if (dma->dp_kva == NULL) {
3068 			return (DDI_DMA_NORESOURCES);
3069 		}
3070 	}
3071 #endif
3072 
3073 	/* convert the sleep flags */
3074 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3075 		cansleep = 1;
3076 	} else {
3077 		cansleep = 0;
3078 	}
3079 
3080 	/*
3081 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3082 	 * engine limits, so we can't use kmem_alloc... We don't need
3083 	 * contiguous memory (sgllen) since we will be forcing windows on
3084 	 * sgllen anyway.
3085 	 */
3086 	lattr = *attr;
3087 	lattr.dma_attr_align = MMU_PAGESIZE;
3088 	/*
3089 	 * this should be < 0 to indicate no limit, but due to a bug in
3090 	 * the rootnex, we'll set it to the maximum positive int.
3091 	 */
3092 	lattr.dma_attr_sgllen = 0x7fffffff;
3093 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3094 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3095 	if (e != DDI_SUCCESS) {
3096 #if !defined(__amd64)
3097 		if (dma->dp_kva != NULL) {
3098 			vmem_free(heap_arena, dma->dp_kva,
3099 			    dma->dp_copybuf_size);
3100 		}
3101 #endif
3102 		return (DDI_DMA_NORESOURCES);
3103 	}
3104 
3105 	DTRACE_PROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3106 	    size_t, dma->dp_copybuf_size);
3107 
3108 	return (DDI_SUCCESS);
3109 }
3110 
3111 
3112 /*
3113  * rootnex_setup_windows()
3114  *    Called in bind slowpath to setup the window state. We always have windows
3115  *    in the slowpath. Even if the window count = 1.
3116  */
3117 static int
3118 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3119     ddi_dma_attr_t *attr, int kmflag)
3120 {
3121 	rootnex_window_t *windowp;
3122 	rootnex_sglinfo_t *sinfo;
3123 	size_t copy_state_size;
3124 	size_t win_state_size;
3125 	size_t state_available;
3126 	size_t space_needed;
3127 	uint_t copybuf_win;
3128 	uint_t maxxfer_win;
3129 	size_t space_used;
3130 	uint_t sglwin;
3131 
3132 
3133 	sinfo = &dma->dp_sglinfo;
3134 
3135 	dma->dp_current_win = 0;
3136 	hp->dmai_nwin = 0;
3137 
3138 	/* If we don't need to do a partial, we only have one window */
3139 	if (!dma->dp_partial_required) {
3140 		dma->dp_max_win = 1;
3141 
3142 	/*
3143 	 * we need multiple windows, need to figure out the worse case number
3144 	 * of windows.
3145 	 */
3146 	} else {
3147 		/*
3148 		 * if we need windows because we need more copy buffer that
3149 		 * we allow, the worse case number of windows we could need
3150 		 * here would be (copybuf space required / copybuf space that
3151 		 * we have) plus one for remainder, and plus 2 to handle the
3152 		 * extra pages on the trim for the first and last pages of the
3153 		 * buffer (a page is the minimum window size so under the right
3154 		 * attr settings, you could have a window for each page).
3155 		 * The last page will only be hit here if the size is not a
3156 		 * multiple of the granularity (which theoretically shouldn't
3157 		 * be the case but never has been enforced, so we could have
3158 		 * broken things without it).
3159 		 */
3160 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3161 			ASSERT(dma->dp_copybuf_size > 0);
3162 			copybuf_win = (sinfo->si_copybuf_req /
3163 			    dma->dp_copybuf_size) + 1 + 2;
3164 		} else {
3165 			copybuf_win = 0;
3166 		}
3167 
3168 		/*
3169 		 * if we need windows because we have more cookies than the H/W
3170 		 * can handle, the number of windows we would need here would
3171 		 * be (cookie count / cookies count H/W supports) plus one for
3172 		 * remainder, and plus 2 to handle the extra pages on the trim
3173 		 * (see above comment about trim)
3174 		 */
3175 		if (attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3176 			sglwin = ((sinfo->si_sgl_size / attr->dma_attr_sgllen)
3177 			    + 1) + 2;
3178 		} else {
3179 			sglwin = 0;
3180 		}
3181 
3182 		/*
3183 		 * if we need windows because we're binding more memory than the
3184 		 * H/W can transfer at once, the number of windows we would need
3185 		 * here would be (xfer count / max xfer H/W supports) plus one
3186 		 * for remainder, and plus 2 to handle the extra pages on the
3187 		 * trim (see above comment about trim)
3188 		 */
3189 		if (dma->dp_dma.dmao_size > dma->dp_maxxfer) {
3190 			maxxfer_win = (dma->dp_dma.dmao_size /
3191 			    dma->dp_maxxfer) + 1 + 2;
3192 		} else {
3193 			maxxfer_win = 0;
3194 		}
3195 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3196 		ASSERT(dma->dp_max_win > 0);
3197 	}
3198 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3199 
3200 	/*
3201 	 * Get space for window and potential copy buffer state. Before we
3202 	 * go and allocate memory, see if we can get away with using what's
3203 	 * left in the pre-allocted state or the dynamically allocated sgl.
3204 	 */
3205 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3206 	    sizeof (ddi_dma_cookie_t));
3207 
3208 	/* if we dynamically allocated space for the cookies */
3209 	if (dma->dp_need_to_free_cookie) {
3210 		/* if we have more space in the pre-allocted buffer, use it */
3211 		ASSERT(space_used <= dma->dp_cookie_size);
3212 		if ((dma->dp_cookie_size - space_used) <=
3213 		    rootnex_state->r_prealloc_size) {
3214 			state_available = rootnex_state->r_prealloc_size;
3215 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3216 
3217 		/*
3218 		 * else, we have more free space in the dynamically allocated
3219 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3220 		 * didn't need a lot of cookies.
3221 		 */
3222 		} else {
3223 			state_available = dma->dp_cookie_size - space_used;
3224 			windowp = (rootnex_window_t *)
3225 			    &dma->dp_cookies[sinfo->si_sgl_size];
3226 		}
3227 
3228 	/* we used the pre-alloced buffer */
3229 	} else {
3230 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3231 		state_available = rootnex_state->r_prealloc_size - space_used;
3232 		windowp = (rootnex_window_t *)
3233 		    &dma->dp_cookies[sinfo->si_sgl_size];
3234 	}
3235 
3236 	/*
3237 	 * figure out how much state we need to track the copy buffer. Add an
3238 	 * addition 8 bytes for pointer alignemnt later.
3239 	 */
3240 	if (dma->dp_copybuf_size > 0) {
3241 		copy_state_size = sinfo->si_max_pages *
3242 		    sizeof (rootnex_pgmap_t);
3243 	} else {
3244 		copy_state_size = 0;
3245 	}
3246 	/* add an additional 8 bytes for pointer alignment */
3247 	space_needed = win_state_size + copy_state_size + 0x8;
3248 
3249 	/* if we have enough space already, use it */
3250 	if (state_available >= space_needed) {
3251 		dma->dp_window = windowp;
3252 		dma->dp_need_to_free_window = B_FALSE;
3253 
3254 	/* not enough space, need to allocate more. */
3255 	} else {
3256 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3257 		if (dma->dp_window == NULL) {
3258 			return (DDI_DMA_NORESOURCES);
3259 		}
3260 		dma->dp_need_to_free_window = B_TRUE;
3261 		dma->dp_window_size = space_needed;
3262 		DTRACE_PROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3263 		    dma->dp_dip, size_t, space_needed);
3264 	}
3265 
3266 	/*
3267 	 * we allocate copy buffer state and window state at the same time.
3268 	 * setup our copy buffer state pointers. Make sure it's aligned.
3269 	 */
3270 	if (dma->dp_copybuf_size > 0) {
3271 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3272 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3273 
3274 #if !defined(__amd64)
3275 		/*
3276 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3277 		 * false/NULL. Should be quicker to bzero vs loop and set.
3278 		 */
3279 		bzero(dma->dp_pgmap, copy_state_size);
3280 #endif
3281 	} else {
3282 		dma->dp_pgmap = NULL;
3283 	}
3284 
3285 	return (DDI_SUCCESS);
3286 }
3287 
3288 
3289 /*
3290  * rootnex_teardown_copybuf()
3291  *    cleans up after rootnex_setup_copybuf()
3292  */
3293 static void
3294 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3295 {
3296 #if !defined(__amd64)
3297 	int i;
3298 
3299 	/*
3300 	 * if we allocated kernel heap VMEM space, go through all the pages and
3301 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3302 	 * arena. Then free the VMEM space.
3303 	 */
3304 	if (dma->dp_kva != NULL) {
3305 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3306 			if (dma->dp_pgmap[i].pm_mapped) {
3307 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3308 				    MMU_PAGESIZE, HAT_UNLOAD);
3309 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3310 			}
3311 		}
3312 
3313 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3314 	}
3315 
3316 #endif
3317 
3318 	/* if we allocated a copy buffer, free it */
3319 	if (dma->dp_cbaddr != NULL) {
3320 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3321 	}
3322 }
3323 
3324 
3325 /*
3326  * rootnex_teardown_windows()
3327  *    cleans up after rootnex_setup_windows()
3328  */
3329 static void
3330 rootnex_teardown_windows(rootnex_dma_t *dma)
3331 {
3332 	/*
3333 	 * if we had to allocate window state on the last bind (because we
3334 	 * didn't have enough pre-allocated space in the handle), free it.
3335 	 */
3336 	if (dma->dp_need_to_free_window) {
3337 		kmem_free(dma->dp_window, dma->dp_window_size);
3338 	}
3339 }
3340 
3341 
3342 /*
3343  * rootnex_init_win()
3344  *    Called in bind slow path during creation of a new window. Initializes
3345  *    window state to default values.
3346  */
3347 /*ARGSUSED*/
3348 static void
3349 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3350     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3351 {
3352 	hp->dmai_nwin++;
3353 	window->wd_dosync = B_FALSE;
3354 	window->wd_offset = cur_offset;
3355 	window->wd_size = 0;
3356 	window->wd_first_cookie = cookie;
3357 	window->wd_cookie_cnt = 0;
3358 	window->wd_trim.tr_trim_first = B_FALSE;
3359 	window->wd_trim.tr_trim_last = B_FALSE;
3360 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3361 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3362 #if !defined(__amd64)
3363 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3364 #endif
3365 }
3366 
3367 
3368 /*
3369  * rootnex_setup_cookie()
3370  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3371  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3372  *    out if it uses the copy buffer, and if it does, save away everything we'll
3373  *    need during sync.
3374  */
3375 static void
3376 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3377     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3378     page_t **cur_pp)
3379 {
3380 	boolean_t copybuf_sz_power_2;
3381 	rootnex_sglinfo_t *sinfo;
3382 	paddr_t paddr;
3383 	uint_t pidx;
3384 	uint_t pcnt;
3385 	off_t poff;
3386 #if defined(__amd64)
3387 	pfn_t pfn;
3388 #else
3389 	page_t **pplist;
3390 #endif
3391 
3392 	sinfo = &dma->dp_sglinfo;
3393 
3394 	/*
3395 	 * Calculate the page index relative to the start of the buffer. The
3396 	 * index to the current page for our buffer is the offset into the
3397 	 * first page of the buffer plus our current offset into the buffer
3398 	 * itself, shifted of course...
3399 	 */
3400 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3401 	ASSERT(pidx < sinfo->si_max_pages);
3402 
3403 	/* if this cookie uses the copy buffer */
3404 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3405 		/*
3406 		 * NOTE: we know that since this cookie uses the copy buffer, it
3407 		 * is <= MMU_PAGESIZE.
3408 		 */
3409 
3410 		/*
3411 		 * get the offset into the page. For the 64-bit kernel, get the
3412 		 * pfn which we'll use with seg kpm.
3413 		 */
3414 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3415 #if defined(__amd64)
3416 		/* mfn_to_pfn() is a NOP on i86pc */
3417 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3418 #endif /* __amd64 */
3419 
3420 		/* figure out if the copybuf size is a power of 2 */
3421 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3422 			copybuf_sz_power_2 = B_FALSE;
3423 		} else {
3424 			copybuf_sz_power_2 = B_TRUE;
3425 		}
3426 
3427 		/* This page uses the copy buffer */
3428 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3429 
3430 		/*
3431 		 * save the copy buffer KVA that we'll use with this page.
3432 		 * if we still fit within the copybuf, it's a simple add.
3433 		 * otherwise, we need to wrap over using & or % accordingly.
3434 		 */
3435 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3436 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3437 			    *copybuf_used;
3438 		} else {
3439 			if (copybuf_sz_power_2) {
3440 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3441 				    (uintptr_t)dma->dp_cbaddr +
3442 				    (*copybuf_used &
3443 				    (dma->dp_copybuf_size - 1)));
3444 			} else {
3445 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3446 				    (uintptr_t)dma->dp_cbaddr +
3447 				    (*copybuf_used % dma->dp_copybuf_size));
3448 			}
3449 		}
3450 
3451 		/*
3452 		 * over write the cookie physical address with the address of
3453 		 * the physical address of the copy buffer page that we will
3454 		 * use.
3455 		 */
3456 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3457 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3458 
3459 #ifdef __xpv
3460 		/*
3461 		 * If we're dom0, we're using a real device so we need to load
3462 		 * the cookies with MAs instead of PAs.
3463 		 */
3464 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3465 #else
3466 		cookie->dmac_laddress = paddr;
3467 #endif
3468 
3469 		/* if we have a kernel VA, it's easy, just save that address */
3470 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3471 		    (sinfo->si_asp == &kas)) {
3472 			/*
3473 			 * save away the page aligned virtual address of the
3474 			 * driver buffer. Offsets are handled in the sync code.
3475 			 */
3476 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3477 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3478 			    & MMU_PAGEMASK);
3479 #if !defined(__amd64)
3480 			/*
3481 			 * we didn't need to, and will never need to map this
3482 			 * page.
3483 			 */
3484 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3485 #endif
3486 
3487 		/* we don't have a kernel VA. We need one for the bcopy. */
3488 		} else {
3489 #if defined(__amd64)
3490 			/*
3491 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3492 			 * get a Kernel VA for the corresponding pfn.
3493 			 */
3494 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3495 #else
3496 			/*
3497 			 * for the 32-bit kernel, this is a pain. First we'll
3498 			 * save away the page_t or user VA for this page. This
3499 			 * is needed in rootnex_dma_win() when we switch to a
3500 			 * new window which requires us to re-map the copy
3501 			 * buffer.
3502 			 */
3503 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3504 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3505 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3506 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3507 			} else if (pplist != NULL) {
3508 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3509 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3510 			} else {
3511 				dma->dp_pgmap[pidx].pm_pp = NULL;
3512 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3513 				    (((uintptr_t)
3514 				    dmar_object->dmao_obj.virt_obj.v_addr +
3515 				    cur_offset) & MMU_PAGEMASK);
3516 			}
3517 
3518 			/*
3519 			 * save away the page aligned virtual address which was
3520 			 * allocated from the kernel heap arena (taking into
3521 			 * account if we need more copy buffer than we alloced
3522 			 * and use multiple windows to handle this, i.e. &,%).
3523 			 * NOTE: there isn't and physical memory backing up this
3524 			 * virtual address space currently.
3525 			 */
3526 			if ((*copybuf_used + MMU_PAGESIZE) <=
3527 			    dma->dp_copybuf_size) {
3528 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3529 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3530 				    MMU_PAGEMASK);
3531 			} else {
3532 				if (copybuf_sz_power_2) {
3533 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3534 					    (((uintptr_t)dma->dp_kva +
3535 					    (*copybuf_used &
3536 					    (dma->dp_copybuf_size - 1))) &
3537 					    MMU_PAGEMASK);
3538 				} else {
3539 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3540 					    (((uintptr_t)dma->dp_kva +
3541 					    (*copybuf_used %
3542 					    dma->dp_copybuf_size)) &
3543 					    MMU_PAGEMASK);
3544 				}
3545 			}
3546 
3547 			/*
3548 			 * if we haven't used up the available copy buffer yet,
3549 			 * map the kva to the physical page.
3550 			 */
3551 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3552 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3553 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
3554 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
3555 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
3556 					    dma->dp_pgmap[pidx].pm_kaddr);
3557 				} else {
3558 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
3559 					    sinfo->si_asp,
3560 					    dma->dp_pgmap[pidx].pm_kaddr);
3561 				}
3562 
3563 			/*
3564 			 * we've used up the available copy buffer, this page
3565 			 * will have to be mapped during rootnex_dma_win() when
3566 			 * we switch to a new window which requires a re-map
3567 			 * the copy buffer. (32-bit kernel only)
3568 			 */
3569 			} else {
3570 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3571 			}
3572 #endif
3573 			/* go to the next page_t */
3574 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3575 				*cur_pp = (*cur_pp)->p_next;
3576 			}
3577 		}
3578 
3579 		/* add to the copy buffer count */
3580 		*copybuf_used += MMU_PAGESIZE;
3581 
3582 	/*
3583 	 * This cookie doesn't use the copy buffer. Walk through the pages this
3584 	 * cookie occupies to reflect this.
3585 	 */
3586 	} else {
3587 		/*
3588 		 * figure out how many pages the cookie occupies. We need to
3589 		 * use the original page offset of the buffer and the cookies
3590 		 * offset in the buffer to do this.
3591 		 */
3592 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
3593 		pcnt = mmu_btopr(cookie->dmac_size + poff);
3594 
3595 		while (pcnt > 0) {
3596 #if !defined(__amd64)
3597 			/*
3598 			 * the 32-bit kernel doesn't have seg kpm, so we need
3599 			 * to map in the driver buffer (if it didn't come down
3600 			 * with a kernel VA) on the fly. Since this page doesn't
3601 			 * use the copy buffer, it's not, or will it ever, have
3602 			 * to be mapped in.
3603 			 */
3604 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3605 #endif
3606 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
3607 
3608 			/*
3609 			 * we need to update pidx and cur_pp or we'll loose
3610 			 * track of where we are.
3611 			 */
3612 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3613 				*cur_pp = (*cur_pp)->p_next;
3614 			}
3615 			pidx++;
3616 			pcnt--;
3617 		}
3618 	}
3619 }
3620 
3621 
3622 /*
3623  * rootnex_sgllen_window_boundary()
3624  *    Called in the bind slow path when the next cookie causes us to exceed (in
3625  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
3626  *    length supported by the DMA H/W.
3627  */
3628 static int
3629 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3630     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
3631     off_t cur_offset)
3632 {
3633 	off_t new_offset;
3634 	size_t trim_sz;
3635 	off_t coffset;
3636 
3637 
3638 	/*
3639 	 * if we know we'll never have to trim, it's pretty easy. Just move to
3640 	 * the next window and init it. We're done.
3641 	 */
3642 	if (!dma->dp_trim_required) {
3643 		(*windowp)++;
3644 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3645 		(*windowp)->wd_cookie_cnt++;
3646 		(*windowp)->wd_size = cookie->dmac_size;
3647 		return (DDI_SUCCESS);
3648 	}
3649 
3650 	/* figure out how much we need to trim from the window */
3651 	ASSERT(attr->dma_attr_granular != 0);
3652 	if (dma->dp_granularity_power_2) {
3653 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
3654 	} else {
3655 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
3656 	}
3657 
3658 	/* The window's a whole multiple of granularity. We're done */
3659 	if (trim_sz == 0) {
3660 		(*windowp)++;
3661 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3662 		(*windowp)->wd_cookie_cnt++;
3663 		(*windowp)->wd_size = cookie->dmac_size;
3664 		return (DDI_SUCCESS);
3665 	}
3666 
3667 	/*
3668 	 * The window's not a whole multiple of granularity, since we know this
3669 	 * is due to the sgllen, we need to go back to the last cookie and trim
3670 	 * that one, add the left over part of the old cookie into the new
3671 	 * window, and then add in the new cookie into the new window.
3672 	 */
3673 
3674 	/*
3675 	 * make sure the driver isn't making us do something bad... Trimming and
3676 	 * sgllen == 1 don't go together.
3677 	 */
3678 	if (attr->dma_attr_sgllen == 1) {
3679 		return (DDI_DMA_NOMAPPING);
3680 	}
3681 
3682 	/*
3683 	 * first, setup the current window to account for the trim. Need to go
3684 	 * back to the last cookie for this.
3685 	 */
3686 	cookie--;
3687 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3688 	(*windowp)->wd_trim.tr_last_cookie = cookie;
3689 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3690 	ASSERT(cookie->dmac_size > trim_sz);
3691 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3692 	(*windowp)->wd_size -= trim_sz;
3693 
3694 	/* save the buffer offsets for the next window */
3695 	coffset = cookie->dmac_size - trim_sz;
3696 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3697 
3698 	/*
3699 	 * set this now in case this is the first window. all other cases are
3700 	 * set in dma_win()
3701 	 */
3702 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
3703 
3704 	/*
3705 	 * initialize the next window using what's left over in the previous
3706 	 * cookie.
3707 	 */
3708 	(*windowp)++;
3709 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3710 	(*windowp)->wd_cookie_cnt++;
3711 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3712 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
3713 	(*windowp)->wd_trim.tr_first_size = trim_sz;
3714 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3715 		(*windowp)->wd_dosync = B_TRUE;
3716 	}
3717 
3718 	/*
3719 	 * now go back to the current cookie and add it to the new window. set
3720 	 * the new window size to the what was left over from the previous
3721 	 * cookie and what's in the current cookie.
3722 	 */
3723 	cookie++;
3724 	(*windowp)->wd_cookie_cnt++;
3725 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
3726 
3727 	/*
3728 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
3729 	 * a max size of maxxfer). Handle that case.
3730 	 */
3731 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
3732 		/*
3733 		 * maxxfer is already a whole multiple of granularity, and this
3734 		 * trim will be <= the previous trim (since a cookie can't be
3735 		 * larger than maxxfer). Make things simple here.
3736 		 */
3737 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
3738 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3739 		(*windowp)->wd_trim.tr_last_cookie = cookie;
3740 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3741 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3742 		(*windowp)->wd_size -= trim_sz;
3743 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
3744 
3745 		/* save the buffer offsets for the next window */
3746 		coffset = cookie->dmac_size - trim_sz;
3747 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3748 
3749 		/* setup the next window */
3750 		(*windowp)++;
3751 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3752 		(*windowp)->wd_cookie_cnt++;
3753 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3754 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
3755 		    coffset;
3756 		(*windowp)->wd_trim.tr_first_size = trim_sz;
3757 	}
3758 
3759 	return (DDI_SUCCESS);
3760 }
3761 
3762 
3763 /*
3764  * rootnex_copybuf_window_boundary()
3765  *    Called in bind slowpath when we get to a window boundary because we used
3766  *    up all the copy buffer that we have.
3767  */
3768 static int
3769 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3770     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
3771     size_t *copybuf_used)
3772 {
3773 	rootnex_sglinfo_t *sinfo;
3774 	off_t new_offset;
3775 	size_t trim_sz;
3776 	paddr_t paddr;
3777 	off_t coffset;
3778 	uint_t pidx;
3779 	off_t poff;
3780 
3781 
3782 	sinfo = &dma->dp_sglinfo;
3783 
3784 	/*
3785 	 * the copy buffer should be a whole multiple of page size. We know that
3786 	 * this cookie is <= MMU_PAGESIZE.
3787 	 */
3788 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
3789 
3790 	/*
3791 	 * from now on, all new windows in this bind need to be re-mapped during
3792 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
3793 	 * space...
3794 	 */
3795 #if !defined(__amd64)
3796 	dma->dp_cb_remaping = B_TRUE;
3797 #endif
3798 
3799 	/* reset copybuf used */
3800 	*copybuf_used = 0;
3801 
3802 	/*
3803 	 * if we don't have to trim (since granularity is set to 1), go to the
3804 	 * next window and add the current cookie to it. We know the current
3805 	 * cookie uses the copy buffer since we're in this code path.
3806 	 */
3807 	if (!dma->dp_trim_required) {
3808 		(*windowp)++;
3809 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3810 
3811 		/* Add this cookie to the new window */
3812 		(*windowp)->wd_cookie_cnt++;
3813 		(*windowp)->wd_size += cookie->dmac_size;
3814 		*copybuf_used += MMU_PAGESIZE;
3815 		return (DDI_SUCCESS);
3816 	}
3817 
3818 	/*
3819 	 * *** may need to trim, figure it out.
3820 	 */
3821 
3822 	/* figure out how much we need to trim from the window */
3823 	if (dma->dp_granularity_power_2) {
3824 		trim_sz = (*windowp)->wd_size &
3825 		    (hp->dmai_attr.dma_attr_granular - 1);
3826 	} else {
3827 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
3828 	}
3829 
3830 	/*
3831 	 * if the window's a whole multiple of granularity, go to the next
3832 	 * window, init it, then add in the current cookie. We know the current
3833 	 * cookie uses the copy buffer since we're in this code path.
3834 	 */
3835 	if (trim_sz == 0) {
3836 		(*windowp)++;
3837 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3838 
3839 		/* Add this cookie to the new window */
3840 		(*windowp)->wd_cookie_cnt++;
3841 		(*windowp)->wd_size += cookie->dmac_size;
3842 		*copybuf_used += MMU_PAGESIZE;
3843 		return (DDI_SUCCESS);
3844 	}
3845 
3846 	/*
3847 	 * *** We figured it out, we definitly need to trim
3848 	 */
3849 
3850 	/*
3851 	 * make sure the driver isn't making us do something bad...
3852 	 * Trimming and sgllen == 1 don't go together.
3853 	 */
3854 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
3855 		return (DDI_DMA_NOMAPPING);
3856 	}
3857 
3858 	/*
3859 	 * first, setup the current window to account for the trim. Need to go
3860 	 * back to the last cookie for this. Some of the last cookie will be in
3861 	 * the current window, and some of the last cookie will be in the new
3862 	 * window. All of the current cookie will be in the new window.
3863 	 */
3864 	cookie--;
3865 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3866 	(*windowp)->wd_trim.tr_last_cookie = cookie;
3867 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3868 	ASSERT(cookie->dmac_size > trim_sz);
3869 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3870 	(*windowp)->wd_size -= trim_sz;
3871 
3872 	/*
3873 	 * we're trimming the last cookie (not the current cookie). So that
3874 	 * last cookie may have or may not have been using the copy buffer (
3875 	 * we know the cookie passed in uses the copy buffer since we're in
3876 	 * this code path).
3877 	 *
3878 	 * If the last cookie doesn't use the copy buffer, nothing special to
3879 	 * do. However, if it does uses the copy buffer, it will be both the
3880 	 * last page in the current window and the first page in the next
3881 	 * window. Since we are reusing the copy buffer (and KVA space on the
3882 	 * 32-bit kernel), this page will use the end of the copy buffer in the
3883 	 * current window, and the start of the copy buffer in the next window.
3884 	 * Track that info... The cookie physical address was already set to
3885 	 * the copy buffer physical address in setup_cookie..
3886 	 */
3887 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3888 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
3889 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
3890 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
3891 		(*windowp)->wd_trim.tr_last_pidx = pidx;
3892 		(*windowp)->wd_trim.tr_last_cbaddr =
3893 		    dma->dp_pgmap[pidx].pm_cbaddr;
3894 #if !defined(__amd64)
3895 		(*windowp)->wd_trim.tr_last_kaddr =
3896 		    dma->dp_pgmap[pidx].pm_kaddr;
3897 #endif
3898 	}
3899 
3900 	/* save the buffer offsets for the next window */
3901 	coffset = cookie->dmac_size - trim_sz;
3902 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3903 
3904 	/*
3905 	 * set this now in case this is the first window. all other cases are
3906 	 * set in dma_win()
3907 	 */
3908 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
3909 
3910 	/*
3911 	 * initialize the next window using what's left over in the previous
3912 	 * cookie.
3913 	 */
3914 	(*windowp)++;
3915 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3916 	(*windowp)->wd_cookie_cnt++;
3917 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3918 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
3919 	(*windowp)->wd_trim.tr_first_size = trim_sz;
3920 
3921 	/*
3922 	 * again, we're tracking if the last cookie uses the copy buffer.
3923 	 * read the comment above for more info on why we need to track
3924 	 * additional state.
3925 	 *
3926 	 * For the first cookie in the new window, we need reset the physical
3927 	 * address to DMA into to the start of the copy buffer plus any
3928 	 * initial page offset which may be present.
3929 	 */
3930 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3931 		(*windowp)->wd_dosync = B_TRUE;
3932 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
3933 		(*windowp)->wd_trim.tr_first_pidx = pidx;
3934 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
3935 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
3936 
3937 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
3938 		    poff;
3939 #ifdef __xpv
3940 		/*
3941 		 * If we're dom0, we're using a real device so we need to load
3942 		 * the cookies with MAs instead of PAs.
3943 		 */
3944 		(*windowp)->wd_trim.tr_first_paddr =
3945 		    ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3946 #else
3947 		(*windowp)->wd_trim.tr_first_paddr = paddr;
3948 #endif
3949 
3950 #if !defined(__amd64)
3951 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
3952 #endif
3953 		/* account for the cookie copybuf usage in the new window */
3954 		*copybuf_used += MMU_PAGESIZE;
3955 
3956 		/*
3957 		 * every piece of code has to have a hack, and here is this
3958 		 * ones :-)
3959 		 *
3960 		 * There is a complex interaction between setup_cookie and the
3961 		 * copybuf window boundary. The complexity had to be in either
3962 		 * the maxxfer window, or the copybuf window, and I chose the
3963 		 * copybuf code.
3964 		 *
3965 		 * So in this code path, we have taken the last cookie,
3966 		 * virtually broken it in half due to the trim, and it happens
3967 		 * to use the copybuf which further complicates life. At the
3968 		 * same time, we have already setup the current cookie, which
3969 		 * is now wrong. More background info: the current cookie uses
3970 		 * the copybuf, so it is only a page long max. So we need to
3971 		 * fix the current cookies copy buffer address, physical
3972 		 * address, and kva for the 32-bit kernel. We due this by
3973 		 * bumping them by page size (of course, we can't due this on
3974 		 * the physical address since the copy buffer may not be
3975 		 * physically contiguous).
3976 		 */
3977 		cookie++;
3978 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
3979 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3980 
3981 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3982 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
3983 #ifdef __xpv
3984 		/*
3985 		 * If we're dom0, we're using a real device so we need to load
3986 		 * the cookies with MAs instead of PAs.
3987 		 */
3988 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3989 #else
3990 		cookie->dmac_laddress = paddr;
3991 #endif
3992 
3993 #if !defined(__amd64)
3994 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
3995 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
3996 #endif
3997 	} else {
3998 		/* go back to the current cookie */
3999 		cookie++;
4000 	}
4001 
4002 	/*
4003 	 * add the current cookie to the new window. set the new window size to
4004 	 * the what was left over from the previous cookie and what's in the
4005 	 * current cookie.
4006 	 */
4007 	(*windowp)->wd_cookie_cnt++;
4008 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4009 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4010 
4011 	/*
4012 	 * we know that the cookie passed in always uses the copy buffer. We
4013 	 * wouldn't be here if it didn't.
4014 	 */
4015 	*copybuf_used += MMU_PAGESIZE;
4016 
4017 	return (DDI_SUCCESS);
4018 }
4019 
4020 
4021 /*
4022  * rootnex_maxxfer_window_boundary()
4023  *    Called in bind slowpath when we get to a window boundary because we will
4024  *    go over maxxfer.
4025  */
4026 static int
4027 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4028     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4029 {
4030 	size_t dmac_size;
4031 	off_t new_offset;
4032 	size_t trim_sz;
4033 	off_t coffset;
4034 
4035 
4036 	/*
4037 	 * calculate how much we have to trim off of the current cookie to equal
4038 	 * maxxfer. We don't have to account for granularity here since our
4039 	 * maxxfer already takes that into account.
4040 	 */
4041 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4042 	ASSERT(trim_sz <= cookie->dmac_size);
4043 	ASSERT(trim_sz <= dma->dp_maxxfer);
4044 
4045 	/* save cookie size since we need it later and we might change it */
4046 	dmac_size = cookie->dmac_size;
4047 
4048 	/*
4049 	 * if we're not trimming the entire cookie, setup the current window to
4050 	 * account for the trim.
4051 	 */
4052 	if (trim_sz < cookie->dmac_size) {
4053 		(*windowp)->wd_cookie_cnt++;
4054 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4055 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4056 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4057 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4058 		(*windowp)->wd_size = dma->dp_maxxfer;
4059 
4060 		/*
4061 		 * set the adjusted cookie size now in case this is the first
4062 		 * window. All other windows are taken care of in get win
4063 		 */
4064 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4065 	}
4066 
4067 	/*
4068 	 * coffset is the current offset within the cookie, new_offset is the
4069 	 * current offset with the entire buffer.
4070 	 */
4071 	coffset = dmac_size - trim_sz;
4072 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4073 
4074 	/* initialize the next window */
4075 	(*windowp)++;
4076 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4077 	(*windowp)->wd_cookie_cnt++;
4078 	(*windowp)->wd_size = trim_sz;
4079 	if (trim_sz < dmac_size) {
4080 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4081 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4082 		    coffset;
4083 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4084 	}
4085 
4086 	return (DDI_SUCCESS);
4087 }
4088 
4089 
4090 /*ARGSUSED*/
4091 static int
4092 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4093     off_t off, size_t len, uint_t cache_flags)
4094 {
4095 	rootnex_sglinfo_t *sinfo;
4096 	rootnex_pgmap_t *cbpage;
4097 	rootnex_window_t *win;
4098 	ddi_dma_impl_t *hp;
4099 	rootnex_dma_t *dma;
4100 	caddr_t fromaddr;
4101 	caddr_t toaddr;
4102 	uint_t psize;
4103 	off_t offset;
4104 	uint_t pidx;
4105 	size_t size;
4106 	off_t poff;
4107 	int e;
4108 
4109 
4110 	hp = (ddi_dma_impl_t *)handle;
4111 	dma = (rootnex_dma_t *)hp->dmai_private;
4112 	sinfo = &dma->dp_sglinfo;
4113 
4114 	/*
4115 	 * if we don't have any windows, we don't need to sync. A copybuf
4116 	 * will cause us to have at least one window.
4117 	 */
4118 	if (dma->dp_window == NULL) {
4119 		return (DDI_SUCCESS);
4120 	}
4121 
4122 	/* This window may not need to be sync'd */
4123 	win = &dma->dp_window[dma->dp_current_win];
4124 	if (!win->wd_dosync) {
4125 		return (DDI_SUCCESS);
4126 	}
4127 
4128 	if (strcmp(ddi_driver_name(rdip), "bnx") == 0 ||
4129 	    strcmp(ddi_driver_name(rdip), "ohci") == 0)
4130 		cmn_err(CE_WARN, "%s: syncing DMA ...",
4131 		    ddi_driver_name(rdip));
4132 
4133 	/* handle off and len special cases */
4134 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4135 		offset = win->wd_offset;
4136 	} else {
4137 		offset = off;
4138 	}
4139 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4140 		size = win->wd_size;
4141 	} else {
4142 		size = len;
4143 	}
4144 
4145 	/* check the sync args to make sure they make a little sense */
4146 	if (rootnex_sync_check_parms) {
4147 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4148 		    cache_flags);
4149 		if (e != DDI_SUCCESS) {
4150 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4151 			return (DDI_FAILURE);
4152 		}
4153 	}
4154 
4155 	/*
4156 	 * special case the first page to handle the offset into the page. The
4157 	 * offset to the current page for our buffer is the offset into the
4158 	 * first page of the buffer plus our current offset into the buffer
4159 	 * itself, masked of course.
4160 	 */
4161 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4162 	psize = MIN((MMU_PAGESIZE - poff), size);
4163 
4164 	/* go through all the pages that we want to sync */
4165 	while (size > 0) {
4166 		/*
4167 		 * Calculate the page index relative to the start of the buffer.
4168 		 * The index to the current page for our buffer is the offset
4169 		 * into the first page of the buffer plus our current offset
4170 		 * into the buffer itself, shifted of course...
4171 		 */
4172 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4173 		ASSERT(pidx < sinfo->si_max_pages);
4174 
4175 		/*
4176 		 * if this page uses the copy buffer, we need to sync it,
4177 		 * otherwise, go on to the next page.
4178 		 */
4179 		cbpage = &dma->dp_pgmap[pidx];
4180 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4181 		    (cbpage->pm_uses_copybuf == B_FALSE));
4182 		if (cbpage->pm_uses_copybuf) {
4183 			/* cbaddr and kaddr should be page aligned */
4184 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4185 			    MMU_PAGEOFFSET) == 0);
4186 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4187 			    MMU_PAGEOFFSET) == 0);
4188 
4189 			/*
4190 			 * if we're copying for the device, we are going to
4191 			 * copy from the drivers buffer and to the rootnex
4192 			 * allocated copy buffer.
4193 			 */
4194 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4195 				fromaddr = cbpage->pm_kaddr + poff;
4196 				toaddr = cbpage->pm_cbaddr + poff;
4197 				DTRACE_PROBE2(rootnex__sync__dev,
4198 				    dev_info_t *, dma->dp_dip, size_t, psize);
4199 
4200 			/*
4201 			 * if we're copying for the cpu/kernel, we are going to
4202 			 * copy from the rootnex allocated copy buffer to the
4203 			 * drivers buffer.
4204 			 */
4205 			} else {
4206 				fromaddr = cbpage->pm_cbaddr + poff;
4207 				toaddr = cbpage->pm_kaddr + poff;
4208 				DTRACE_PROBE2(rootnex__sync__cpu,
4209 				    dev_info_t *, dma->dp_dip, size_t, psize);
4210 			}
4211 
4212 			bcopy(fromaddr, toaddr, psize);
4213 		}
4214 
4215 		/*
4216 		 * decrement size until we're done, update our offset into the
4217 		 * buffer, and get the next page size.
4218 		 */
4219 		size -= psize;
4220 		offset += psize;
4221 		psize = MIN(MMU_PAGESIZE, size);
4222 
4223 		/* page offset is zero for the rest of this loop */
4224 		poff = 0;
4225 	}
4226 
4227 	return (DDI_SUCCESS);
4228 }
4229 
4230 /*
4231  * rootnex_dma_sync()
4232  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4233  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4234  *    is set, ddi_dma_sync() returns immediately passing back success.
4235  */
4236 /*ARGSUSED*/
4237 static int
4238 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4239     off_t off, size_t len, uint_t cache_flags)
4240 {
4241 #if !defined(__xpv)
4242 	if (IOMMU_USED(rdip)) {
4243 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4244 		    cache_flags));
4245 	}
4246 #endif
4247 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4248 	    cache_flags));
4249 }
4250 
4251 /*
4252  * rootnex_valid_sync_parms()
4253  *    checks the parameters passed to sync to verify they are correct.
4254  */
4255 static int
4256 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4257     off_t offset, size_t size, uint_t cache_flags)
4258 {
4259 	off_t woffset;
4260 
4261 
4262 	/*
4263 	 * the first part of the test to make sure the offset passed in is
4264 	 * within the window.
4265 	 */
4266 	if (offset < win->wd_offset) {
4267 		return (DDI_FAILURE);
4268 	}
4269 
4270 	/*
4271 	 * second and last part of the test to make sure the offset and length
4272 	 * passed in is within the window.
4273 	 */
4274 	woffset = offset - win->wd_offset;
4275 	if ((woffset + size) > win->wd_size) {
4276 		return (DDI_FAILURE);
4277 	}
4278 
4279 	/*
4280 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4281 	 * be set too.
4282 	 */
4283 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4284 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4285 		return (DDI_SUCCESS);
4286 	}
4287 
4288 	/*
4289 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4290 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4291 	 */
4292 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4293 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4294 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4295 		return (DDI_SUCCESS);
4296 	}
4297 
4298 	return (DDI_FAILURE);
4299 }
4300 
4301 
4302 /*ARGSUSED*/
4303 static int
4304 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4305     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4306     uint_t *ccountp)
4307 {
4308 	rootnex_window_t *window;
4309 	rootnex_trim_t *trim;
4310 	ddi_dma_impl_t *hp;
4311 	rootnex_dma_t *dma;
4312 #if !defined(__amd64)
4313 	rootnex_sglinfo_t *sinfo;
4314 	rootnex_pgmap_t *pmap;
4315 	uint_t pidx;
4316 	uint_t pcnt;
4317 	off_t poff;
4318 	int i;
4319 #endif
4320 
4321 
4322 	hp = (ddi_dma_impl_t *)handle;
4323 	dma = (rootnex_dma_t *)hp->dmai_private;
4324 #if !defined(__amd64)
4325 	sinfo = &dma->dp_sglinfo;
4326 #endif
4327 
4328 	/* If we try and get a window which doesn't exist, return failure */
4329 	if (win >= hp->dmai_nwin) {
4330 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4331 		return (DDI_FAILURE);
4332 	}
4333 
4334 	/*
4335 	 * if we don't have any windows, and they're asking for the first
4336 	 * window, setup the cookie pointer to the first cookie in the bind.
4337 	 * setup our return values, then increment the cookie since we return
4338 	 * the first cookie on the stack.
4339 	 */
4340 	if (dma->dp_window == NULL) {
4341 		if (win != 0) {
4342 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4343 			return (DDI_FAILURE);
4344 		}
4345 		hp->dmai_cookie = dma->dp_cookies;
4346 		*offp = 0;
4347 		*lenp = dma->dp_dma.dmao_size;
4348 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4349 		*cookiep = hp->dmai_cookie[0];
4350 		hp->dmai_cookie++;
4351 		return (DDI_SUCCESS);
4352 	}
4353 
4354 	/* sync the old window before moving on to the new one */
4355 	window = &dma->dp_window[dma->dp_current_win];
4356 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4357 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4358 		    DDI_DMA_SYNC_FORCPU);
4359 	}
4360 
4361 #if !defined(__amd64)
4362 	/*
4363 	 * before we move to the next window, if we need to re-map, unmap all
4364 	 * the pages in this window.
4365 	 */
4366 	if (dma->dp_cb_remaping) {
4367 		/*
4368 		 * If we switch to this window again, we'll need to map in
4369 		 * on the fly next time.
4370 		 */
4371 		window->wd_remap_copybuf = B_TRUE;
4372 
4373 		/*
4374 		 * calculate the page index into the buffer where this window
4375 		 * starts, and the number of pages this window takes up.
4376 		 */
4377 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4378 		    MMU_PAGESHIFT;
4379 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4380 		    MMU_PAGEOFFSET;
4381 		pcnt = mmu_btopr(window->wd_size + poff);
4382 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4383 
4384 		/* unmap pages which are currently mapped in this window */
4385 		for (i = 0; i < pcnt; i++) {
4386 			if (dma->dp_pgmap[pidx].pm_mapped) {
4387 				hat_unload(kas.a_hat,
4388 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4389 				    HAT_UNLOAD);
4390 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4391 			}
4392 			pidx++;
4393 		}
4394 	}
4395 #endif
4396 
4397 	/*
4398 	 * Move to the new window.
4399 	 * NOTE: current_win must be set for sync to work right
4400 	 */
4401 	dma->dp_current_win = win;
4402 	window = &dma->dp_window[win];
4403 
4404 	/* if needed, adjust the first and/or last cookies for trim */
4405 	trim = &window->wd_trim;
4406 	if (trim->tr_trim_first) {
4407 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4408 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4409 #if !defined(__amd64)
4410 		window->wd_first_cookie->dmac_type =
4411 		    (window->wd_first_cookie->dmac_type &
4412 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4413 #endif
4414 		if (trim->tr_first_copybuf_win) {
4415 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4416 			    trim->tr_first_cbaddr;
4417 #if !defined(__amd64)
4418 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4419 			    trim->tr_first_kaddr;
4420 #endif
4421 		}
4422 	}
4423 	if (trim->tr_trim_last) {
4424 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4425 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4426 		if (trim->tr_last_copybuf_win) {
4427 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4428 			    trim->tr_last_cbaddr;
4429 #if !defined(__amd64)
4430 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4431 			    trim->tr_last_kaddr;
4432 #endif
4433 		}
4434 	}
4435 
4436 	/*
4437 	 * setup the cookie pointer to the first cookie in the window. setup
4438 	 * our return values, then increment the cookie since we return the
4439 	 * first cookie on the stack.
4440 	 */
4441 	hp->dmai_cookie = window->wd_first_cookie;
4442 	*offp = window->wd_offset;
4443 	*lenp = window->wd_size;
4444 	*ccountp = window->wd_cookie_cnt;
4445 	*cookiep = hp->dmai_cookie[0];
4446 	hp->dmai_cookie++;
4447 
4448 #if !defined(__amd64)
4449 	/* re-map copybuf if required for this window */
4450 	if (dma->dp_cb_remaping) {
4451 		/*
4452 		 * calculate the page index into the buffer where this
4453 		 * window starts.
4454 		 */
4455 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4456 		    MMU_PAGESHIFT;
4457 		ASSERT(pidx < sinfo->si_max_pages);
4458 
4459 		/*
4460 		 * the first page can get unmapped if it's shared with the
4461 		 * previous window. Even if the rest of this window is already
4462 		 * mapped in, we need to still check this one.
4463 		 */
4464 		pmap = &dma->dp_pgmap[pidx];
4465 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4466 			if (pmap->pm_pp != NULL) {
4467 				pmap->pm_mapped = B_TRUE;
4468 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4469 			} else if (pmap->pm_vaddr != NULL) {
4470 				pmap->pm_mapped = B_TRUE;
4471 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4472 				    pmap->pm_kaddr);
4473 			}
4474 		}
4475 		pidx++;
4476 
4477 		/* map in the rest of the pages if required */
4478 		if (window->wd_remap_copybuf) {
4479 			window->wd_remap_copybuf = B_FALSE;
4480 
4481 			/* figure out many pages this window takes up */
4482 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4483 			    MMU_PAGEOFFSET;
4484 			pcnt = mmu_btopr(window->wd_size + poff);
4485 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4486 
4487 			/* map pages which require it */
4488 			for (i = 1; i < pcnt; i++) {
4489 				pmap = &dma->dp_pgmap[pidx];
4490 				if (pmap->pm_uses_copybuf) {
4491 					ASSERT(pmap->pm_mapped == B_FALSE);
4492 					if (pmap->pm_pp != NULL) {
4493 						pmap->pm_mapped = B_TRUE;
4494 						i86_pp_map(pmap->pm_pp,
4495 						    pmap->pm_kaddr);
4496 					} else if (pmap->pm_vaddr != NULL) {
4497 						pmap->pm_mapped = B_TRUE;
4498 						i86_va_map(pmap->pm_vaddr,
4499 						    sinfo->si_asp,
4500 						    pmap->pm_kaddr);
4501 					}
4502 				}
4503 				pidx++;
4504 			}
4505 		}
4506 	}
4507 #endif
4508 
4509 	/* if the new window uses the copy buffer, sync it for the device */
4510 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4511 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4512 		    DDI_DMA_SYNC_FORDEV);
4513 	}
4514 
4515 	return (DDI_SUCCESS);
4516 }
4517 
4518 /*
4519  * rootnex_dma_win()
4520  *    called from ddi_dma_getwin()
4521  */
4522 /*ARGSUSED*/
4523 static int
4524 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4525     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4526     uint_t *ccountp)
4527 {
4528 #if !defined(__xpv)
4529 	if (IOMMU_USED(rdip)) {
4530 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4531 		    cookiep, ccountp));
4532 	}
4533 #endif
4534 
4535 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4536 	    cookiep, ccountp));
4537 }
4538 
4539 /*
4540  * ************************
4541  *  obsoleted dma routines
4542  * ************************
4543  */
4544 
4545 /* ARGSUSED */
4546 static int
4547 rootnex_coredma_map(dev_info_t *dip, dev_info_t *rdip,
4548     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
4549 {
4550 #if defined(__amd64)
4551 	/*
4552 	 * this interface is not supported in 64-bit x86 kernel. See comment in
4553 	 * rootnex_dma_mctl()
4554 	 */
4555 	return (DDI_DMA_NORESOURCES);
4556 
4557 #else /* 32-bit x86 kernel */
4558 	ddi_dma_handle_t *lhandlep;
4559 	ddi_dma_handle_t lhandle;
4560 	ddi_dma_cookie_t cookie;
4561 	ddi_dma_attr_t dma_attr;
4562 	ddi_dma_lim_t *dma_lim;
4563 	uint_t ccnt;
4564 	int e;
4565 
4566 
4567 	/*
4568 	 * if the driver is just testing to see if it's possible to do the bind,
4569 	 * we'll use local state. Otherwise, use the handle pointer passed in.
4570 	 */
4571 	if (handlep == NULL) {
4572 		lhandlep = &lhandle;
4573 	} else {
4574 		lhandlep = handlep;
4575 	}
4576 
4577 	/* convert the limit structure to a dma_attr one */
4578 	dma_lim = dmareq->dmar_limits;
4579 	dma_attr.dma_attr_version = DMA_ATTR_V0;
4580 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
4581 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
4582 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
4583 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
4584 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
4585 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
4586 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
4587 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
4588 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
4589 	dma_attr.dma_attr_align = MMU_PAGESIZE;
4590 	dma_attr.dma_attr_flags = 0;
4591 
4592 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
4593 	    dmareq->dmar_arg, lhandlep);
4594 	if (e != DDI_SUCCESS) {
4595 		return (e);
4596 	}
4597 
4598 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
4599 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
4600 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4601 		return (e);
4602 	}
4603 
4604 	/*
4605 	 * if the driver is just testing to see if it's possible to do the bind,
4606 	 * free up the local state and return the result.
4607 	 */
4608 	if (handlep == NULL) {
4609 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
4610 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4611 		if (e == DDI_DMA_MAPPED) {
4612 			return (DDI_DMA_MAPOK);
4613 		} else {
4614 			return (DDI_DMA_NOMAPPING);
4615 		}
4616 	}
4617 
4618 	return (e);
4619 #endif /* defined(__amd64) */
4620 }
4621 
4622 /*
4623  * rootnex_dma_map()
4624  *    called from ddi_dma_setup()
4625  */
4626 /* ARGSUSED */
4627 static int
4628 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
4629     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
4630 {
4631 #if !defined(__xpv)
4632 	if (IOMMU_USED(rdip)) {
4633 		return (iommulib_nexdma_map(dip, rdip, dmareq, handlep));
4634 	}
4635 #endif
4636 	return (rootnex_coredma_map(dip, rdip, dmareq, handlep));
4637 }
4638 
4639 /*
4640  * rootnex_dma_mctl()
4641  *
4642  */
4643 /* ARGSUSED */
4644 static int
4645 rootnex_coredma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4646     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
4647     uint_t cache_flags)
4648 {
4649 #if defined(__amd64)
4650 	/*
4651 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
4652 	 * common implementation in genunix, so they no longer have x86
4653 	 * specific functionality which called into dma_ctl.
4654 	 *
4655 	 * The rest of the obsoleted interfaces were never supported in the
4656 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
4657 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
4658 	 * implementation issues.
4659 	 *
4660 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
4661 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
4662 	 * reflect that now too...
4663 	 *
4664 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
4665 	 * not going to put this functionality into the 64-bit x86 kernel now.
4666 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
4667 	 * that in a future release.
4668 	 */
4669 	return (DDI_FAILURE);
4670 
4671 #else /* 32-bit x86 kernel */
4672 	ddi_dma_cookie_t lcookie;
4673 	ddi_dma_cookie_t *cookie;
4674 	rootnex_window_t *window;
4675 	ddi_dma_impl_t *hp;
4676 	rootnex_dma_t *dma;
4677 	uint_t nwin;
4678 	uint_t ccnt;
4679 	size_t len;
4680 	off_t off;
4681 	int e;
4682 
4683 
4684 	/*
4685 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
4686 	 * hacky since were optimizing for the current interfaces and so we can
4687 	 * cleanup the mess in genunix. Hopefully we will remove the this
4688 	 * obsoleted routines someday soon.
4689 	 */
4690 
4691 	switch (request) {
4692 
4693 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
4694 		hp = (ddi_dma_impl_t *)handle;
4695 		cookie = (ddi_dma_cookie_t *)objpp;
4696 
4697 		/*
4698 		 * convert segment to cookie. We don't distinguish between the
4699 		 * two :-)
4700 		 */
4701 		*cookie = *hp->dmai_cookie;
4702 		*lenp = cookie->dmac_size;
4703 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
4704 		return (DDI_SUCCESS);
4705 
4706 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
4707 		hp = (ddi_dma_impl_t *)handle;
4708 		dma = (rootnex_dma_t *)hp->dmai_private;
4709 
4710 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
4711 			return (DDI_DMA_STALE);
4712 		}
4713 
4714 		/* handle the case where we don't have any windows */
4715 		if (dma->dp_window == NULL) {
4716 			/*
4717 			 * if seg == NULL, and we don't have any windows,
4718 			 * return the first cookie in the sgl.
4719 			 */
4720 			if (*lenp == NULL) {
4721 				dma->dp_current_cookie = 0;
4722 				hp->dmai_cookie = dma->dp_cookies;
4723 				*objpp = (caddr_t)handle;
4724 				return (DDI_SUCCESS);
4725 
4726 			/* if we have more cookies, go to the next cookie */
4727 			} else {
4728 				if ((dma->dp_current_cookie + 1) >=
4729 				    dma->dp_sglinfo.si_sgl_size) {
4730 					return (DDI_DMA_DONE);
4731 				}
4732 				dma->dp_current_cookie++;
4733 				hp->dmai_cookie++;
4734 				return (DDI_SUCCESS);
4735 			}
4736 		}
4737 
4738 		/* We have one or more windows */
4739 		window = &dma->dp_window[dma->dp_current_win];
4740 
4741 		/*
4742 		 * if seg == NULL, return the first cookie in the current
4743 		 * window
4744 		 */
4745 		if (*lenp == NULL) {
4746 			dma->dp_current_cookie = 0;
4747 			hp->dmai_cookie = window->wd_first_cookie;
4748 
4749 		/*
4750 		 * go to the next cookie in the window then see if we done with
4751 		 * this window.
4752 		 */
4753 		} else {
4754 			if ((dma->dp_current_cookie + 1) >=
4755 			    window->wd_cookie_cnt) {
4756 				return (DDI_DMA_DONE);
4757 			}
4758 			dma->dp_current_cookie++;
4759 			hp->dmai_cookie++;
4760 		}
4761 		*objpp = (caddr_t)handle;
4762 		return (DDI_SUCCESS);
4763 
4764 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
4765 		hp = (ddi_dma_impl_t *)handle;
4766 		dma = (rootnex_dma_t *)hp->dmai_private;
4767 
4768 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
4769 			return (DDI_DMA_STALE);
4770 		}
4771 
4772 		/* if win == NULL, return the first window in the bind */
4773 		if (*offp == NULL) {
4774 			nwin = 0;
4775 
4776 		/*
4777 		 * else, go to the next window then see if we're done with all
4778 		 * the windows.
4779 		 */
4780 		} else {
4781 			nwin = dma->dp_current_win + 1;
4782 			if (nwin >= hp->dmai_nwin) {
4783 				return (DDI_DMA_DONE);
4784 			}
4785 		}
4786 
4787 		/* switch to the next window */
4788 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
4789 		    &lcookie, &ccnt);
4790 		ASSERT(e == DDI_SUCCESS);
4791 		if (e != DDI_SUCCESS) {
4792 			return (DDI_DMA_STALE);
4793 		}
4794 
4795 		/* reset the cookie back to the first cookie in the window */
4796 		if (dma->dp_window != NULL) {
4797 			window = &dma->dp_window[dma->dp_current_win];
4798 			hp->dmai_cookie = window->wd_first_cookie;
4799 		} else {
4800 			hp->dmai_cookie = dma->dp_cookies;
4801 		}
4802 
4803 		*objpp = (caddr_t)handle;
4804 		return (DDI_SUCCESS);
4805 
4806 	case DDI_DMA_FREE: /* ddi_dma_free() */
4807 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
4808 		(void) rootnex_dma_freehdl(dip, rdip, handle);
4809 		if (rootnex_state->r_dvma_call_list_id) {
4810 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
4811 		}
4812 		return (DDI_SUCCESS);
4813 
4814 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
4815 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
4816 		/* should never get here, handled in genunix */
4817 		ASSERT(0);
4818 		return (DDI_FAILURE);
4819 
4820 	case DDI_DMA_KVADDR:
4821 	case DDI_DMA_GETERR:
4822 	case DDI_DMA_COFF:
4823 		return (DDI_FAILURE);
4824 	}
4825 
4826 	return (DDI_FAILURE);
4827 #endif /* defined(__amd64) */
4828 }
4829 
4830 /*
4831  * rootnex_dma_mctl()
4832  *
4833  */
4834 /* ARGSUSED */
4835 static int
4836 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4837     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
4838     uint_t cache_flags)
4839 {
4840 #if !defined(__xpv)
4841 	if (IOMMU_USED(rdip)) {
4842 		return (iommulib_nexdma_mctl(dip, rdip, handle, request, offp,
4843 		    lenp, objpp, cache_flags));
4844 	}
4845 #endif
4846 
4847 	return (rootnex_coredma_mctl(dip, rdip, handle, request, offp,
4848 	    lenp, objpp, cache_flags));
4849 }
4850 
4851 /*
4852  * *********
4853  *  FMA Code
4854  * *********
4855  */
4856 
4857 /*
4858  * rootnex_fm_init()
4859  *    FMA init busop
4860  */
4861 /* ARGSUSED */
4862 static int
4863 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
4864     ddi_iblock_cookie_t *ibc)
4865 {
4866 	*ibc = rootnex_state->r_err_ibc;
4867 
4868 	return (ddi_system_fmcap);
4869 }
4870 
4871 /*
4872  * rootnex_dma_check()
4873  *    Function called after a dma fault occurred to find out whether the
4874  *    fault address is associated with a driver that is able to handle faults
4875  *    and recover from faults.
4876  */
4877 /* ARGSUSED */
4878 static int
4879 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
4880     const void *not_used)
4881 {
4882 	rootnex_window_t *window;
4883 	uint64_t start_addr;
4884 	uint64_t fault_addr;
4885 	ddi_dma_impl_t *hp;
4886 	rootnex_dma_t *dma;
4887 	uint64_t end_addr;
4888 	size_t csize;
4889 	int i;
4890 	int j;
4891 
4892 
4893 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
4894 	hp = (ddi_dma_impl_t *)handle;
4895 	ASSERT(hp);
4896 
4897 	dma = (rootnex_dma_t *)hp->dmai_private;
4898 
4899 	/* Get the address that we need to search for */
4900 	fault_addr = *(uint64_t *)addr;
4901 
4902 	/*
4903 	 * if we don't have any windows, we can just walk through all the
4904 	 * cookies.
4905 	 */
4906 	if (dma->dp_window == NULL) {
4907 		/* for each cookie */
4908 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
4909 			/*
4910 			 * if the faulted address is within the physical address
4911 			 * range of the cookie, return DDI_FM_NONFATAL.
4912 			 */
4913 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
4914 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
4915 			    dma->dp_cookies[i].dmac_size))) {
4916 				return (DDI_FM_NONFATAL);
4917 			}
4918 		}
4919 
4920 		/* fault_addr not within this DMA handle */
4921 		return (DDI_FM_UNKNOWN);
4922 	}
4923 
4924 	/* we have mutiple windows, walk through each window */
4925 	for (i = 0; i < hp->dmai_nwin; i++) {
4926 		window = &dma->dp_window[i];
4927 
4928 		/* Go through all the cookies in the window */
4929 		for (j = 0; j < window->wd_cookie_cnt; j++) {
4930 
4931 			start_addr = window->wd_first_cookie[j].dmac_laddress;
4932 			csize = window->wd_first_cookie[j].dmac_size;
4933 
4934 			/*
4935 			 * if we are trimming the first cookie in the window,
4936 			 * and this is the first cookie, adjust the start
4937 			 * address and size of the cookie to account for the
4938 			 * trim.
4939 			 */
4940 			if (window->wd_trim.tr_trim_first && (j == 0)) {
4941 				start_addr = window->wd_trim.tr_first_paddr;
4942 				csize = window->wd_trim.tr_first_size;
4943 			}
4944 
4945 			/*
4946 			 * if we are trimming the last cookie in the window,
4947 			 * and this is the last cookie, adjust the start
4948 			 * address and size of the cookie to account for the
4949 			 * trim.
4950 			 */
4951 			if (window->wd_trim.tr_trim_last &&
4952 			    (j == (window->wd_cookie_cnt - 1))) {
4953 				start_addr = window->wd_trim.tr_last_paddr;
4954 				csize = window->wd_trim.tr_last_size;
4955 			}
4956 
4957 			end_addr = start_addr + csize;
4958 
4959 			/*
4960 			 * if the faulted address is within the physical address
4961 			 * range of the cookie, return DDI_FM_NONFATAL.
4962 			 */
4963 			if ((fault_addr >= start_addr) &&
4964 			    (fault_addr <= end_addr)) {
4965 				return (DDI_FM_NONFATAL);
4966 			}
4967 		}
4968 	}
4969 
4970 	/* fault_addr not within this DMA handle */
4971 	return (DDI_FM_UNKNOWN);
4972 }
4973