xref: /illumos-gate/usr/src/uts/i86pc/io/rootnex.c (revision b1d7ec75953cd517f5b7c3d9cb427ff8ec5d7d07)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * x86 root nexus driver
27  */
28 
29 #include <sys/sysmacros.h>
30 #include <sys/conf.h>
31 #include <sys/autoconf.h>
32 #include <sys/sysmacros.h>
33 #include <sys/debug.h>
34 #include <sys/psw.h>
35 #include <sys/ddidmareq.h>
36 #include <sys/promif.h>
37 #include <sys/devops.h>
38 #include <sys/kmem.h>
39 #include <sys/cmn_err.h>
40 #include <vm/seg.h>
41 #include <vm/seg_kmem.h>
42 #include <vm/seg_dev.h>
43 #include <sys/vmem.h>
44 #include <sys/mman.h>
45 #include <vm/hat.h>
46 #include <vm/as.h>
47 #include <vm/page.h>
48 #include <sys/avintr.h>
49 #include <sys/errno.h>
50 #include <sys/modctl.h>
51 #include <sys/ddi_impldefs.h>
52 #include <sys/sunddi.h>
53 #include <sys/sunndi.h>
54 #include <sys/mach_intr.h>
55 #include <sys/psm.h>
56 #include <sys/ontrap.h>
57 #include <sys/atomic.h>
58 #include <sys/sdt.h>
59 #include <sys/rootnex.h>
60 #include <vm/hat_i86.h>
61 #include <sys/ddifm.h>
62 #include <sys/ddi_isa.h>
63 
64 #ifdef __xpv
65 #include <sys/bootinfo.h>
66 #include <sys/hypervisor.h>
67 #include <sys/bootconf.h>
68 #include <vm/kboot_mmu.h>
69 #endif
70 
71 #if defined(__amd64) && !defined(__xpv)
72 #include <sys/immu.h>
73 #endif
74 
75 
76 /*
77  * enable/disable extra checking of function parameters. Useful for debugging
78  * drivers.
79  */
80 #ifdef	DEBUG
81 int rootnex_alloc_check_parms = 1;
82 int rootnex_bind_check_parms = 1;
83 int rootnex_bind_check_inuse = 1;
84 int rootnex_unbind_verify_buffer = 0;
85 int rootnex_sync_check_parms = 1;
86 #else
87 int rootnex_alloc_check_parms = 0;
88 int rootnex_bind_check_parms = 0;
89 int rootnex_bind_check_inuse = 0;
90 int rootnex_unbind_verify_buffer = 0;
91 int rootnex_sync_check_parms = 0;
92 #endif
93 
94 boolean_t rootnex_dmar_not_setup;
95 
96 /* Master Abort and Target Abort panic flag */
97 int rootnex_fm_ma_ta_panic_flag = 0;
98 
99 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
100 int rootnex_bind_fail = 1;
101 int rootnex_bind_warn = 1;
102 uint8_t *rootnex_warn_list;
103 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
104 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
105 
106 /*
107  * revert back to old broken behavior of always sync'ing entire copy buffer.
108  * This is useful if be have a buggy driver which doesn't correctly pass in
109  * the offset and size into ddi_dma_sync().
110  */
111 int rootnex_sync_ignore_params = 0;
112 
113 /*
114  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
115  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
116  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
117  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
118  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
119  * (< 8K). We will still need to allocate the copy buffer during bind though
120  * (if we need one). These can only be modified in /etc/system before rootnex
121  * attach.
122  */
123 #if defined(__amd64)
124 int rootnex_prealloc_cookies = 65;
125 int rootnex_prealloc_windows = 4;
126 int rootnex_prealloc_copybuf = 2;
127 #else
128 int rootnex_prealloc_cookies = 33;
129 int rootnex_prealloc_windows = 4;
130 int rootnex_prealloc_copybuf = 2;
131 #endif
132 
133 /* driver global state */
134 static rootnex_state_t *rootnex_state;
135 
136 /* shortcut to rootnex counters */
137 static uint64_t *rootnex_cnt;
138 
139 /*
140  * XXX - does x86 even need these or are they left over from the SPARC days?
141  */
142 /* statically defined integer/boolean properties for the root node */
143 static rootnex_intprop_t rootnex_intprp[] = {
144 	{ "PAGESIZE",			PAGESIZE },
145 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
146 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
147 	{ DDI_RELATIVE_ADDRESSING,	1 },
148 };
149 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
150 
151 #ifdef __xpv
152 typedef maddr_t rootnex_addr_t;
153 #define	ROOTNEX_PADDR_TO_RBASE(xinfo, pa)	\
154 	(DOMAIN_IS_INITDOMAIN(xinfo) ? pa_to_ma(pa) : (pa))
155 #else
156 typedef paddr_t rootnex_addr_t;
157 #endif
158 
159 #if !defined(__xpv)
160 char _depends_on[] = "mach/pcplusmp misc/iommulib misc/acpica";
161 #endif
162 
163 static struct cb_ops rootnex_cb_ops = {
164 	nodev,		/* open */
165 	nodev,		/* close */
166 	nodev,		/* strategy */
167 	nodev,		/* print */
168 	nodev,		/* dump */
169 	nodev,		/* read */
170 	nodev,		/* write */
171 	nodev,		/* ioctl */
172 	nodev,		/* devmap */
173 	nodev,		/* mmap */
174 	nodev,		/* segmap */
175 	nochpoll,	/* chpoll */
176 	ddi_prop_op,	/* cb_prop_op */
177 	NULL,		/* struct streamtab */
178 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
179 	CB_REV,		/* Rev */
180 	nodev,		/* cb_aread */
181 	nodev		/* cb_awrite */
182 };
183 
184 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
185     off_t offset, off_t len, caddr_t *vaddrp);
186 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
187     struct hat *hat, struct seg *seg, caddr_t addr,
188     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
189 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
190     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
191 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
192     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
193     ddi_dma_handle_t *handlep);
194 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
195     ddi_dma_handle_t handle);
196 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
197     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
198     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
199 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
200     ddi_dma_handle_t handle);
201 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
202     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
203 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
204     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
205     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
206 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
207     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
208     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
209 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
210     ddi_ctl_enum_t ctlop, void *arg, void *result);
211 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
212     ddi_iblock_cookie_t *ibc);
213 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
214     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
215 
216 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
217     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
218     ddi_dma_handle_t *handlep);
219 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
220     ddi_dma_handle_t handle);
221 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
222     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
223     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
224 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
225     ddi_dma_handle_t handle);
226 #if defined(__amd64) && !defined(__xpv)
227 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
228     ddi_dma_handle_t handle);
229 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
230     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
231 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
232     ddi_dma_cookie_t *cookiep, uint_t ccount);
233 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
234     ddi_dma_handle_t handle);
235 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
236 #endif
237 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
238     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
239 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
240     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
241     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
242 
243 static struct bus_ops rootnex_bus_ops = {
244 	BUSO_REV,
245 	rootnex_map,
246 	NULL,
247 	NULL,
248 	NULL,
249 	rootnex_map_fault,
250 	rootnex_dma_map,
251 	rootnex_dma_allochdl,
252 	rootnex_dma_freehdl,
253 	rootnex_dma_bindhdl,
254 	rootnex_dma_unbindhdl,
255 	rootnex_dma_sync,
256 	rootnex_dma_win,
257 	rootnex_dma_mctl,
258 	rootnex_ctlops,
259 	ddi_bus_prop_op,
260 	i_ddi_rootnex_get_eventcookie,
261 	i_ddi_rootnex_add_eventcall,
262 	i_ddi_rootnex_remove_eventcall,
263 	i_ddi_rootnex_post_event,
264 	0,			/* bus_intr_ctl */
265 	0,			/* bus_config */
266 	0,			/* bus_unconfig */
267 	rootnex_fm_init,	/* bus_fm_init */
268 	NULL,			/* bus_fm_fini */
269 	NULL,			/* bus_fm_access_enter */
270 	NULL,			/* bus_fm_access_exit */
271 	NULL,			/* bus_powr */
272 	rootnex_intr_ops	/* bus_intr_op */
273 };
274 
275 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
276 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
277 static int rootnex_quiesce(dev_info_t *dip);
278 
279 static struct dev_ops rootnex_ops = {
280 	DEVO_REV,
281 	0,
282 	ddi_no_info,
283 	nulldev,
284 	nulldev,
285 	rootnex_attach,
286 	rootnex_detach,
287 	nulldev,
288 	&rootnex_cb_ops,
289 	&rootnex_bus_ops,
290 	NULL,
291 	rootnex_quiesce,		/* quiesce */
292 };
293 
294 static struct modldrv rootnex_modldrv = {
295 	&mod_driverops,
296 	"i86pc root nexus",
297 	&rootnex_ops
298 };
299 
300 static struct modlinkage rootnex_modlinkage = {
301 	MODREV_1,
302 	(void *)&rootnex_modldrv,
303 	NULL
304 };
305 
306 #if defined(__amd64) && !defined(__xpv)
307 static iommulib_nexops_t iommulib_nexops = {
308 	IOMMU_NEXOPS_VERSION,
309 	"Rootnex IOMMU ops Vers 1.1",
310 	NULL,
311 	rootnex_coredma_allochdl,
312 	rootnex_coredma_freehdl,
313 	rootnex_coredma_bindhdl,
314 	rootnex_coredma_unbindhdl,
315 	rootnex_coredma_reset_cookies,
316 	rootnex_coredma_get_cookies,
317 	rootnex_coredma_set_cookies,
318 	rootnex_coredma_clear_cookies,
319 	rootnex_coredma_get_sleep_flags,
320 	rootnex_coredma_sync,
321 	rootnex_coredma_win,
322 	rootnex_dma_map,
323 	rootnex_dma_mctl
324 };
325 #endif
326 
327 /*
328  *  extern hacks
329  */
330 extern struct seg_ops segdev_ops;
331 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
332 #ifdef	DDI_MAP_DEBUG
333 extern int ddi_map_debug_flag;
334 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
335 #endif
336 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
337 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
338 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
339     psm_intr_op_t, int *);
340 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
341 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
342 
343 /*
344  * Use device arena to use for device control register mappings.
345  * Various kernel memory walkers (debugger, dtrace) need to know
346  * to avoid this address range to prevent undesired device activity.
347  */
348 extern void *device_arena_alloc(size_t size, int vm_flag);
349 extern void device_arena_free(void * vaddr, size_t size);
350 
351 
352 /*
353  *  Internal functions
354  */
355 static int rootnex_dma_init();
356 static void rootnex_add_props(dev_info_t *);
357 static int rootnex_ctl_reportdev(dev_info_t *dip);
358 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
359 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
360 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
361 static int rootnex_map_handle(ddi_map_req_t *mp);
362 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
363 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
364 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
365     ddi_dma_attr_t *attr);
366 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
367     rootnex_sglinfo_t *sglinfo);
368 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
369     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag);
370 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
371     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
372 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
373 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
374     ddi_dma_attr_t *attr, int kmflag);
375 static void rootnex_teardown_windows(rootnex_dma_t *dma);
376 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
377     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
378 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
379     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
380     size_t *copybuf_used, page_t **cur_pp);
381 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
382     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
383     ddi_dma_attr_t *attr, off_t cur_offset);
384 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
385     rootnex_dma_t *dma, rootnex_window_t **windowp,
386     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
387 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
388     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
389 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
390     off_t offset, size_t size, uint_t cache_flags);
391 static int rootnex_verify_buffer(rootnex_dma_t *dma);
392 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
393     const void *comp_addr, const void *not_used);
394 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
395     rootnex_sglinfo_t *sglinfo);
396 
397 /*
398  * _init()
399  *
400  */
401 int
402 _init(void)
403 {
404 
405 	rootnex_state = NULL;
406 	return (mod_install(&rootnex_modlinkage));
407 }
408 
409 
410 /*
411  * _info()
412  *
413  */
414 int
415 _info(struct modinfo *modinfop)
416 {
417 	return (mod_info(&rootnex_modlinkage, modinfop));
418 }
419 
420 
421 /*
422  * _fini()
423  *
424  */
425 int
426 _fini(void)
427 {
428 	return (EBUSY);
429 }
430 
431 
432 /*
433  * rootnex_attach()
434  *
435  */
436 static int
437 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
438 {
439 	int fmcap;
440 	int e;
441 
442 	switch (cmd) {
443 	case DDI_ATTACH:
444 		break;
445 	case DDI_RESUME:
446 #if defined(__amd64) && !defined(__xpv)
447 		return (immu_unquiesce());
448 #else
449 		return (DDI_SUCCESS);
450 #endif
451 	default:
452 		return (DDI_FAILURE);
453 	}
454 
455 	/*
456 	 * We should only have one instance of rootnex. Save it away since we
457 	 * don't have an easy way to get it back later.
458 	 */
459 	ASSERT(rootnex_state == NULL);
460 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
461 
462 	rootnex_state->r_dip = dip;
463 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
464 	rootnex_state->r_reserved_msg_printed = B_FALSE;
465 	rootnex_cnt = &rootnex_state->r_counters[0];
466 
467 	/*
468 	 * Set minimum fm capability level for i86pc platforms and then
469 	 * initialize error handling. Since we're the rootnex, we don't
470 	 * care what's returned in the fmcap field.
471 	 */
472 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
473 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
474 	fmcap = ddi_system_fmcap;
475 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
476 
477 	/* initialize DMA related state */
478 	e = rootnex_dma_init();
479 	if (e != DDI_SUCCESS) {
480 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
481 		return (DDI_FAILURE);
482 	}
483 
484 	/* Add static root node properties */
485 	rootnex_add_props(dip);
486 
487 	/* since we can't call ddi_report_dev() */
488 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
489 
490 	/* Initialize rootnex event handle */
491 	i_ddi_rootnex_init_events(dip);
492 
493 #if defined(__amd64) && !defined(__xpv)
494 	e = iommulib_nexus_register(dip, &iommulib_nexops,
495 	    &rootnex_state->r_iommulib_handle);
496 
497 	ASSERT(e == DDI_SUCCESS);
498 #endif
499 
500 	return (DDI_SUCCESS);
501 }
502 
503 
504 /*
505  * rootnex_detach()
506  *
507  */
508 /*ARGSUSED*/
509 static int
510 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
511 {
512 	switch (cmd) {
513 	case DDI_SUSPEND:
514 #if defined(__amd64) && !defined(__xpv)
515 		return (immu_quiesce());
516 #else
517 		return (DDI_SUCCESS);
518 #endif
519 	default:
520 		return (DDI_FAILURE);
521 	}
522 	/*NOTREACHED*/
523 
524 }
525 
526 
527 /*
528  * rootnex_dma_init()
529  *
530  */
531 /*ARGSUSED*/
532 static int
533 rootnex_dma_init()
534 {
535 	size_t bufsize;
536 
537 
538 	/*
539 	 * size of our cookie/window/copybuf state needed in dma bind that we
540 	 * pre-alloc in dma_alloc_handle
541 	 */
542 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
543 	rootnex_state->r_prealloc_size =
544 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
545 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
546 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
547 
548 	/*
549 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
550 	 * allocate 16 extra bytes for struct pointer alignment
551 	 * (p->dmai_private & dma->dp_prealloc_buffer)
552 	 */
553 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
554 	    rootnex_state->r_prealloc_size + 0x10;
555 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
556 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
557 	if (rootnex_state->r_dmahdl_cache == NULL) {
558 		return (DDI_FAILURE);
559 	}
560 
561 	/*
562 	 * allocate array to track which major numbers we have printed warnings
563 	 * for.
564 	 */
565 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
566 	    KM_SLEEP);
567 
568 	return (DDI_SUCCESS);
569 }
570 
571 
572 /*
573  * rootnex_add_props()
574  *
575  */
576 static void
577 rootnex_add_props(dev_info_t *dip)
578 {
579 	rootnex_intprop_t *rpp;
580 	int i;
581 
582 	/* Add static integer/boolean properties to the root node */
583 	rpp = rootnex_intprp;
584 	for (i = 0; i < NROOT_INTPROPS; i++) {
585 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
586 		    rpp[i].prop_name, rpp[i].prop_value);
587 	}
588 }
589 
590 
591 
592 /*
593  * *************************
594  *  ctlops related routines
595  * *************************
596  */
597 
598 /*
599  * rootnex_ctlops()
600  *
601  */
602 /*ARGSUSED*/
603 static int
604 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
605     void *arg, void *result)
606 {
607 	int n, *ptr;
608 	struct ddi_parent_private_data *pdp;
609 
610 	switch (ctlop) {
611 	case DDI_CTLOPS_DMAPMAPC:
612 		/*
613 		 * Return 'partial' to indicate that dma mapping
614 		 * has to be done in the main MMU.
615 		 */
616 		return (DDI_DMA_PARTIAL);
617 
618 	case DDI_CTLOPS_BTOP:
619 		/*
620 		 * Convert byte count input to physical page units.
621 		 * (byte counts that are not a page-size multiple
622 		 * are rounded down)
623 		 */
624 		*(ulong_t *)result = btop(*(ulong_t *)arg);
625 		return (DDI_SUCCESS);
626 
627 	case DDI_CTLOPS_PTOB:
628 		/*
629 		 * Convert size in physical pages to bytes
630 		 */
631 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
632 		return (DDI_SUCCESS);
633 
634 	case DDI_CTLOPS_BTOPR:
635 		/*
636 		 * Convert byte count input to physical page units
637 		 * (byte counts that are not a page-size multiple
638 		 * are rounded up)
639 		 */
640 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
641 		return (DDI_SUCCESS);
642 
643 	case DDI_CTLOPS_INITCHILD:
644 		return (impl_ddi_sunbus_initchild(arg));
645 
646 	case DDI_CTLOPS_UNINITCHILD:
647 		impl_ddi_sunbus_removechild(arg);
648 		return (DDI_SUCCESS);
649 
650 	case DDI_CTLOPS_REPORTDEV:
651 		return (rootnex_ctl_reportdev(rdip));
652 
653 	case DDI_CTLOPS_IOMIN:
654 		/*
655 		 * Nothing to do here but reflect back..
656 		 */
657 		return (DDI_SUCCESS);
658 
659 	case DDI_CTLOPS_REGSIZE:
660 	case DDI_CTLOPS_NREGS:
661 		break;
662 
663 	case DDI_CTLOPS_SIDDEV:
664 		if (ndi_dev_is_prom_node(rdip))
665 			return (DDI_SUCCESS);
666 		if (ndi_dev_is_persistent_node(rdip))
667 			return (DDI_SUCCESS);
668 		return (DDI_FAILURE);
669 
670 	case DDI_CTLOPS_POWER:
671 		return ((*pm_platform_power)((power_req_t *)arg));
672 
673 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
674 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
675 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
676 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
677 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
678 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
679 		if (!rootnex_state->r_reserved_msg_printed) {
680 			rootnex_state->r_reserved_msg_printed = B_TRUE;
681 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
682 			    "1 or more reserved/obsolete operations.");
683 		}
684 		return (DDI_FAILURE);
685 
686 	default:
687 		return (DDI_FAILURE);
688 	}
689 	/*
690 	 * The rest are for "hardware" properties
691 	 */
692 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
693 		return (DDI_FAILURE);
694 
695 	if (ctlop == DDI_CTLOPS_NREGS) {
696 		ptr = (int *)result;
697 		*ptr = pdp->par_nreg;
698 	} else {
699 		off_t *size = (off_t *)result;
700 
701 		ptr = (int *)arg;
702 		n = *ptr;
703 		if (n >= pdp->par_nreg) {
704 			return (DDI_FAILURE);
705 		}
706 		*size = (off_t)pdp->par_reg[n].regspec_size;
707 	}
708 	return (DDI_SUCCESS);
709 }
710 
711 
712 /*
713  * rootnex_ctl_reportdev()
714  *
715  */
716 static int
717 rootnex_ctl_reportdev(dev_info_t *dev)
718 {
719 	int i, n, len, f_len = 0;
720 	char *buf;
721 
722 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
723 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
724 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
725 	len = strlen(buf);
726 
727 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
728 
729 		struct regspec *rp = sparc_pd_getreg(dev, i);
730 
731 		if (i == 0)
732 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
733 			    ": ");
734 		else
735 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
736 			    " and ");
737 		len = strlen(buf);
738 
739 		switch (rp->regspec_bustype) {
740 
741 		case BTEISA:
742 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
743 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
744 			break;
745 
746 		case BTISA:
747 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
748 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
749 			break;
750 
751 		default:
752 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
753 			    "space %x offset %x",
754 			    rp->regspec_bustype, rp->regspec_addr);
755 			break;
756 		}
757 		len = strlen(buf);
758 	}
759 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
760 		int pri;
761 
762 		if (i != 0) {
763 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
764 			    ",");
765 			len = strlen(buf);
766 		}
767 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
768 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
769 		    " sparc ipl %d", pri);
770 		len = strlen(buf);
771 	}
772 #ifdef DEBUG
773 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
774 		cmn_err(CE_NOTE, "next message is truncated: "
775 		    "printed length 1024, real length %d", f_len);
776 	}
777 #endif /* DEBUG */
778 	cmn_err(CE_CONT, "?%s\n", buf);
779 	kmem_free(buf, REPORTDEV_BUFSIZE);
780 	return (DDI_SUCCESS);
781 }
782 
783 
784 /*
785  * ******************
786  *  map related code
787  * ******************
788  */
789 
790 /*
791  * rootnex_map()
792  *
793  */
794 static int
795 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
796     off_t len, caddr_t *vaddrp)
797 {
798 	struct regspec *rp, tmp_reg;
799 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
800 	int error;
801 
802 	mp = &mr;
803 
804 	switch (mp->map_op)  {
805 	case DDI_MO_MAP_LOCKED:
806 	case DDI_MO_UNMAP:
807 	case DDI_MO_MAP_HANDLE:
808 		break;
809 	default:
810 #ifdef	DDI_MAP_DEBUG
811 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
812 		    mp->map_op);
813 #endif	/* DDI_MAP_DEBUG */
814 		return (DDI_ME_UNIMPLEMENTED);
815 	}
816 
817 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
818 #ifdef	DDI_MAP_DEBUG
819 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
820 #endif	/* DDI_MAP_DEBUG */
821 		return (DDI_ME_UNIMPLEMENTED);
822 	}
823 
824 	/*
825 	 * First, if given an rnumber, convert it to a regspec...
826 	 * (Presumably, this is on behalf of a child of the root node?)
827 	 */
828 
829 	if (mp->map_type == DDI_MT_RNUMBER)  {
830 
831 		int rnumber = mp->map_obj.rnumber;
832 #ifdef	DDI_MAP_DEBUG
833 		static char *out_of_range =
834 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
835 #endif	/* DDI_MAP_DEBUG */
836 
837 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
838 		if (rp == NULL)  {
839 #ifdef	DDI_MAP_DEBUG
840 			cmn_err(CE_WARN, out_of_range, rnumber,
841 			    ddi_get_name(rdip));
842 #endif	/* DDI_MAP_DEBUG */
843 			return (DDI_ME_RNUMBER_RANGE);
844 		}
845 
846 		/*
847 		 * Convert the given ddi_map_req_t from rnumber to regspec...
848 		 */
849 
850 		mp->map_type = DDI_MT_REGSPEC;
851 		mp->map_obj.rp = rp;
852 	}
853 
854 	/*
855 	 * Adjust offset and length correspnding to called values...
856 	 * XXX: A non-zero length means override the one in the regspec
857 	 * XXX: (regardless of what's in the parent's range?)
858 	 */
859 
860 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
861 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
862 
863 #ifdef	DDI_MAP_DEBUG
864 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
865 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
866 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
867 	    len, mp->map_handlep);
868 #endif	/* DDI_MAP_DEBUG */
869 
870 	/*
871 	 * I/O or memory mapping:
872 	 *
873 	 *	<bustype=0, addr=x, len=x>: memory
874 	 *	<bustype=1, addr=x, len=x>: i/o
875 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
876 	 */
877 
878 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
879 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
880 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
881 		    ddi_get_name(rdip), rp->regspec_bustype,
882 		    rp->regspec_addr, rp->regspec_size);
883 		return (DDI_ME_INVAL);
884 	}
885 
886 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
887 		/*
888 		 * compatibility i/o mapping
889 		 */
890 		rp->regspec_bustype += (uint_t)offset;
891 	} else {
892 		/*
893 		 * Normal memory or i/o mapping
894 		 */
895 		rp->regspec_addr += (uint_t)offset;
896 	}
897 
898 	if (len != 0)
899 		rp->regspec_size = (uint_t)len;
900 
901 #ifdef	DDI_MAP_DEBUG
902 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
903 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
904 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
905 	    offset, len, mp->map_handlep);
906 #endif	/* DDI_MAP_DEBUG */
907 
908 	/*
909 	 * Apply any parent ranges at this level, if applicable.
910 	 * (This is where nexus specific regspec translation takes place.
911 	 * Use of this function is implicit agreement that translation is
912 	 * provided via ddi_apply_range.)
913 	 */
914 
915 #ifdef	DDI_MAP_DEBUG
916 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
917 	    ddi_get_name(dip), ddi_get_name(rdip));
918 #endif	/* DDI_MAP_DEBUG */
919 
920 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
921 		return (error);
922 
923 	switch (mp->map_op)  {
924 	case DDI_MO_MAP_LOCKED:
925 
926 		/*
927 		 * Set up the locked down kernel mapping to the regspec...
928 		 */
929 
930 		return (rootnex_map_regspec(mp, vaddrp));
931 
932 	case DDI_MO_UNMAP:
933 
934 		/*
935 		 * Release mapping...
936 		 */
937 
938 		return (rootnex_unmap_regspec(mp, vaddrp));
939 
940 	case DDI_MO_MAP_HANDLE:
941 
942 		return (rootnex_map_handle(mp));
943 
944 	default:
945 		return (DDI_ME_UNIMPLEMENTED);
946 	}
947 }
948 
949 
950 /*
951  * rootnex_map_fault()
952  *
953  *	fault in mappings for requestors
954  */
955 /*ARGSUSED*/
956 static int
957 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
958     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
959     uint_t lock)
960 {
961 
962 #ifdef	DDI_MAP_DEBUG
963 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
964 	ddi_map_debug(" Seg <%s>\n",
965 	    seg->s_ops == &segdev_ops ? "segdev" :
966 	    seg == &kvseg ? "segkmem" : "NONE!");
967 #endif	/* DDI_MAP_DEBUG */
968 
969 	/*
970 	 * This is all terribly broken, but it is a start
971 	 *
972 	 * XXX	Note that this test means that segdev_ops
973 	 *	must be exported from seg_dev.c.
974 	 * XXX	What about devices with their own segment drivers?
975 	 */
976 	if (seg->s_ops == &segdev_ops) {
977 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
978 
979 		if (hat == NULL) {
980 			/*
981 			 * This is one plausible interpretation of
982 			 * a null hat i.e. use the first hat on the
983 			 * address space hat list which by convention is
984 			 * the hat of the system MMU.  At alternative
985 			 * would be to panic .. this might well be better ..
986 			 */
987 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
988 			hat = seg->s_as->a_hat;
989 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
990 		}
991 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
992 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
993 	} else if (seg == &kvseg && dp == NULL) {
994 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
995 		    HAT_LOAD_LOCK);
996 	} else
997 		return (DDI_FAILURE);
998 	return (DDI_SUCCESS);
999 }
1000 
1001 
1002 /*
1003  * rootnex_map_regspec()
1004  *     we don't support mapping of I/O cards above 4Gb
1005  */
1006 static int
1007 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1008 {
1009 	rootnex_addr_t rbase;
1010 	void *cvaddr;
1011 	uint_t npages, pgoffset;
1012 	struct regspec *rp;
1013 	ddi_acc_hdl_t *hp;
1014 	ddi_acc_impl_t *ap;
1015 	uint_t	hat_acc_flags;
1016 	paddr_t pbase;
1017 
1018 	rp = mp->map_obj.rp;
1019 	hp = mp->map_handlep;
1020 
1021 #ifdef	DDI_MAP_DEBUG
1022 	ddi_map_debug(
1023 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1024 	    rp->regspec_bustype, rp->regspec_addr,
1025 	    rp->regspec_size, mp->map_handlep);
1026 #endif	/* DDI_MAP_DEBUG */
1027 
1028 	/*
1029 	 * I/O or memory mapping
1030 	 *
1031 	 *	<bustype=0, addr=x, len=x>: memory
1032 	 *	<bustype=1, addr=x, len=x>: i/o
1033 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1034 	 */
1035 
1036 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1037 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1038 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1039 		    rp->regspec_addr, rp->regspec_size);
1040 		return (DDI_FAILURE);
1041 	}
1042 
1043 	if (rp->regspec_bustype != 0) {
1044 		/*
1045 		 * I/O space - needs a handle.
1046 		 */
1047 		if (hp == NULL) {
1048 			return (DDI_FAILURE);
1049 		}
1050 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1051 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1052 		impl_acc_hdl_init(hp);
1053 
1054 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1055 #ifdef  DDI_MAP_DEBUG
1056 			ddi_map_debug("rootnex_map_regspec: mmap() "
1057 			    "to I/O space is not supported.\n");
1058 #endif  /* DDI_MAP_DEBUG */
1059 			return (DDI_ME_INVAL);
1060 		} else {
1061 			/*
1062 			 * 1275-compliant vs. compatibility i/o mapping
1063 			 */
1064 			*vaddrp =
1065 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1066 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1067 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1068 #ifdef __xpv
1069 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1070 				hp->ah_pfn = xen_assign_pfn(
1071 				    mmu_btop((ulong_t)rp->regspec_addr &
1072 				    MMU_PAGEMASK));
1073 			} else {
1074 				hp->ah_pfn = mmu_btop(
1075 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1076 			}
1077 #else
1078 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1079 			    MMU_PAGEMASK);
1080 #endif
1081 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1082 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1083 		}
1084 
1085 #ifdef	DDI_MAP_DEBUG
1086 		ddi_map_debug(
1087 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1088 		    rp->regspec_size, *vaddrp);
1089 #endif	/* DDI_MAP_DEBUG */
1090 		return (DDI_SUCCESS);
1091 	}
1092 
1093 	/*
1094 	 * Memory space
1095 	 */
1096 
1097 	if (hp != NULL) {
1098 		/*
1099 		 * hat layer ignores
1100 		 * hp->ah_acc.devacc_attr_endian_flags.
1101 		 */
1102 		switch (hp->ah_acc.devacc_attr_dataorder) {
1103 		case DDI_STRICTORDER_ACC:
1104 			hat_acc_flags = HAT_STRICTORDER;
1105 			break;
1106 		case DDI_UNORDERED_OK_ACC:
1107 			hat_acc_flags = HAT_UNORDERED_OK;
1108 			break;
1109 		case DDI_MERGING_OK_ACC:
1110 			hat_acc_flags = HAT_MERGING_OK;
1111 			break;
1112 		case DDI_LOADCACHING_OK_ACC:
1113 			hat_acc_flags = HAT_LOADCACHING_OK;
1114 			break;
1115 		case DDI_STORECACHING_OK_ACC:
1116 			hat_acc_flags = HAT_STORECACHING_OK;
1117 			break;
1118 		}
1119 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1120 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1121 		impl_acc_hdl_init(hp);
1122 		hp->ah_hat_flags = hat_acc_flags;
1123 	} else {
1124 		hat_acc_flags = HAT_STRICTORDER;
1125 	}
1126 
1127 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1128 #ifdef __xpv
1129 	/*
1130 	 * If we're dom0, we're using a real device so we need to translate
1131 	 * the MA to a PA.
1132 	 */
1133 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1134 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1135 	} else {
1136 		pbase = rbase;
1137 	}
1138 #else
1139 	pbase = rbase;
1140 #endif
1141 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1142 
1143 	if (rp->regspec_size == 0) {
1144 #ifdef  DDI_MAP_DEBUG
1145 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1146 #endif  /* DDI_MAP_DEBUG */
1147 		return (DDI_ME_INVAL);
1148 	}
1149 
1150 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1151 		/* extra cast to make gcc happy */
1152 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1153 	} else {
1154 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1155 
1156 #ifdef	DDI_MAP_DEBUG
1157 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1158 		    "physical %llx", npages, pbase);
1159 #endif	/* DDI_MAP_DEBUG */
1160 
1161 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1162 		if (cvaddr == NULL)
1163 			return (DDI_ME_NORESOURCES);
1164 
1165 		/*
1166 		 * Now map in the pages we've allocated...
1167 		 */
1168 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1169 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1170 		    HAT_LOAD_LOCK);
1171 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1172 
1173 		/* save away pfn and npages for FMA */
1174 		hp = mp->map_handlep;
1175 		if (hp) {
1176 			hp->ah_pfn = mmu_btop(pbase);
1177 			hp->ah_pnum = npages;
1178 		}
1179 	}
1180 
1181 #ifdef	DDI_MAP_DEBUG
1182 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1183 #endif	/* DDI_MAP_DEBUG */
1184 	return (DDI_SUCCESS);
1185 }
1186 
1187 
1188 /*
1189  * rootnex_unmap_regspec()
1190  *
1191  */
1192 static int
1193 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1194 {
1195 	caddr_t addr = (caddr_t)*vaddrp;
1196 	uint_t npages, pgoffset;
1197 	struct regspec *rp;
1198 
1199 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1200 		return (0);
1201 
1202 	rp = mp->map_obj.rp;
1203 
1204 	if (rp->regspec_size == 0) {
1205 #ifdef  DDI_MAP_DEBUG
1206 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1207 #endif  /* DDI_MAP_DEBUG */
1208 		return (DDI_ME_INVAL);
1209 	}
1210 
1211 	/*
1212 	 * I/O or memory mapping:
1213 	 *
1214 	 *	<bustype=0, addr=x, len=x>: memory
1215 	 *	<bustype=1, addr=x, len=x>: i/o
1216 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1217 	 */
1218 	if (rp->regspec_bustype != 0) {
1219 		/*
1220 		 * This is I/O space, which requires no particular
1221 		 * processing on unmap since it isn't mapped in the
1222 		 * first place.
1223 		 */
1224 		return (DDI_SUCCESS);
1225 	}
1226 
1227 	/*
1228 	 * Memory space
1229 	 */
1230 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1231 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1232 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1233 	device_arena_free(addr - pgoffset, ptob(npages));
1234 
1235 	/*
1236 	 * Destroy the pointer - the mapping has logically gone
1237 	 */
1238 	*vaddrp = NULL;
1239 
1240 	return (DDI_SUCCESS);
1241 }
1242 
1243 
1244 /*
1245  * rootnex_map_handle()
1246  *
1247  */
1248 static int
1249 rootnex_map_handle(ddi_map_req_t *mp)
1250 {
1251 	rootnex_addr_t rbase;
1252 	ddi_acc_hdl_t *hp;
1253 	uint_t pgoffset;
1254 	struct regspec *rp;
1255 	paddr_t pbase;
1256 
1257 	rp = mp->map_obj.rp;
1258 
1259 #ifdef	DDI_MAP_DEBUG
1260 	ddi_map_debug(
1261 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1262 	    rp->regspec_bustype, rp->regspec_addr,
1263 	    rp->regspec_size, mp->map_handlep);
1264 #endif	/* DDI_MAP_DEBUG */
1265 
1266 	/*
1267 	 * I/O or memory mapping:
1268 	 *
1269 	 *	<bustype=0, addr=x, len=x>: memory
1270 	 *	<bustype=1, addr=x, len=x>: i/o
1271 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1272 	 */
1273 	if (rp->regspec_bustype != 0) {
1274 		/*
1275 		 * This refers to I/O space, and we don't support "mapping"
1276 		 * I/O space to a user.
1277 		 */
1278 		return (DDI_FAILURE);
1279 	}
1280 
1281 	/*
1282 	 * Set up the hat_flags for the mapping.
1283 	 */
1284 	hp = mp->map_handlep;
1285 
1286 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1287 	case DDI_NEVERSWAP_ACC:
1288 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1289 		break;
1290 	case DDI_STRUCTURE_LE_ACC:
1291 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1292 		break;
1293 	case DDI_STRUCTURE_BE_ACC:
1294 		return (DDI_FAILURE);
1295 	default:
1296 		return (DDI_REGS_ACC_CONFLICT);
1297 	}
1298 
1299 	switch (hp->ah_acc.devacc_attr_dataorder) {
1300 	case DDI_STRICTORDER_ACC:
1301 		break;
1302 	case DDI_UNORDERED_OK_ACC:
1303 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1304 		break;
1305 	case DDI_MERGING_OK_ACC:
1306 		hp->ah_hat_flags |= HAT_MERGING_OK;
1307 		break;
1308 	case DDI_LOADCACHING_OK_ACC:
1309 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1310 		break;
1311 	case DDI_STORECACHING_OK_ACC:
1312 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1313 		break;
1314 	default:
1315 		return (DDI_FAILURE);
1316 	}
1317 
1318 	rbase = (rootnex_addr_t)rp->regspec_addr &
1319 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1320 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1321 
1322 	if (rp->regspec_size == 0)
1323 		return (DDI_ME_INVAL);
1324 
1325 #ifdef __xpv
1326 	/*
1327 	 * If we're dom0, we're using a real device so we need to translate
1328 	 * the MA to a PA.
1329 	 */
1330 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1331 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1332 		    (rbase & MMU_PAGEOFFSET);
1333 	} else {
1334 		pbase = rbase;
1335 	}
1336 #else
1337 	pbase = rbase;
1338 #endif
1339 
1340 	hp->ah_pfn = mmu_btop(pbase);
1341 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1342 
1343 	return (DDI_SUCCESS);
1344 }
1345 
1346 
1347 
1348 /*
1349  * ************************
1350  *  interrupt related code
1351  * ************************
1352  */
1353 
1354 /*
1355  * rootnex_intr_ops()
1356  *	bus_intr_op() function for interrupt support
1357  */
1358 /* ARGSUSED */
1359 static int
1360 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1361     ddi_intr_handle_impl_t *hdlp, void *result)
1362 {
1363 	struct intrspec			*ispec;
1364 	struct ddi_parent_private_data	*pdp;
1365 
1366 	DDI_INTR_NEXDBG((CE_CONT,
1367 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1368 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1369 
1370 	/* Process the interrupt operation */
1371 	switch (intr_op) {
1372 	case DDI_INTROP_GETCAP:
1373 		/* First check with pcplusmp */
1374 		if (psm_intr_ops == NULL)
1375 			return (DDI_FAILURE);
1376 
1377 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1378 			*(int *)result = 0;
1379 			return (DDI_FAILURE);
1380 		}
1381 		break;
1382 	case DDI_INTROP_SETCAP:
1383 		if (psm_intr_ops == NULL)
1384 			return (DDI_FAILURE);
1385 
1386 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1387 			return (DDI_FAILURE);
1388 		break;
1389 	case DDI_INTROP_ALLOC:
1390 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1391 			return (DDI_FAILURE);
1392 		hdlp->ih_pri = ispec->intrspec_pri;
1393 		*(int *)result = hdlp->ih_scratch1;
1394 		break;
1395 	case DDI_INTROP_FREE:
1396 		pdp = ddi_get_parent_data(rdip);
1397 		/*
1398 		 * Special case for 'pcic' driver' only.
1399 		 * If an intrspec was created for it, clean it up here
1400 		 * See detailed comments on this in the function
1401 		 * rootnex_get_ispec().
1402 		 */
1403 		if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1404 			kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1405 			    pdp->par_nintr);
1406 			/*
1407 			 * Set it to zero; so that
1408 			 * DDI framework doesn't free it again
1409 			 */
1410 			pdp->par_intr = NULL;
1411 			pdp->par_nintr = 0;
1412 		}
1413 		break;
1414 	case DDI_INTROP_GETPRI:
1415 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1416 			return (DDI_FAILURE);
1417 		*(int *)result = ispec->intrspec_pri;
1418 		break;
1419 	case DDI_INTROP_SETPRI:
1420 		/* Validate the interrupt priority passed to us */
1421 		if (*(int *)result > LOCK_LEVEL)
1422 			return (DDI_FAILURE);
1423 
1424 		/* Ensure that PSM is all initialized and ispec is ok */
1425 		if ((psm_intr_ops == NULL) ||
1426 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1427 			return (DDI_FAILURE);
1428 
1429 		/* Change the priority */
1430 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1431 		    PSM_FAILURE)
1432 			return (DDI_FAILURE);
1433 
1434 		/* update the ispec with the new priority */
1435 		ispec->intrspec_pri =  *(int *)result;
1436 		break;
1437 	case DDI_INTROP_ADDISR:
1438 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1439 			return (DDI_FAILURE);
1440 		ispec->intrspec_func = hdlp->ih_cb_func;
1441 		break;
1442 	case DDI_INTROP_REMISR:
1443 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1444 			return (DDI_FAILURE);
1445 		ispec->intrspec_func = (uint_t (*)()) 0;
1446 		break;
1447 	case DDI_INTROP_ENABLE:
1448 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1449 			return (DDI_FAILURE);
1450 
1451 		/* Call psmi to translate irq with the dip */
1452 		if (psm_intr_ops == NULL)
1453 			return (DDI_FAILURE);
1454 
1455 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1456 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1457 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1458 			return (DDI_FAILURE);
1459 
1460 		/* Add the interrupt handler */
1461 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1462 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1463 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1464 			return (DDI_FAILURE);
1465 		break;
1466 	case DDI_INTROP_DISABLE:
1467 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1468 			return (DDI_FAILURE);
1469 
1470 		/* Call psm_ops() to translate irq with the dip */
1471 		if (psm_intr_ops == NULL)
1472 			return (DDI_FAILURE);
1473 
1474 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1475 		(void) (*psm_intr_ops)(rdip, hdlp,
1476 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1477 
1478 		/* Remove the interrupt handler */
1479 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1480 		    hdlp->ih_cb_func, hdlp->ih_vector);
1481 		break;
1482 	case DDI_INTROP_SETMASK:
1483 		if (psm_intr_ops == NULL)
1484 			return (DDI_FAILURE);
1485 
1486 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1487 			return (DDI_FAILURE);
1488 		break;
1489 	case DDI_INTROP_CLRMASK:
1490 		if (psm_intr_ops == NULL)
1491 			return (DDI_FAILURE);
1492 
1493 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1494 			return (DDI_FAILURE);
1495 		break;
1496 	case DDI_INTROP_GETPENDING:
1497 		if (psm_intr_ops == NULL)
1498 			return (DDI_FAILURE);
1499 
1500 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1501 		    result)) {
1502 			*(int *)result = 0;
1503 			return (DDI_FAILURE);
1504 		}
1505 		break;
1506 	case DDI_INTROP_NAVAIL:
1507 	case DDI_INTROP_NINTRS:
1508 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1509 		if (*(int *)result == 0) {
1510 			/*
1511 			 * Special case for 'pcic' driver' only. This driver
1512 			 * driver is a child of 'isa' and 'rootnex' drivers.
1513 			 *
1514 			 * See detailed comments on this in the function
1515 			 * rootnex_get_ispec().
1516 			 *
1517 			 * Children of 'pcic' send 'NINITR' request all the
1518 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1519 			 * field may not initialized. So, we fake it here
1520 			 * to return 1 (a la what PCMCIA nexus does).
1521 			 */
1522 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1523 				*(int *)result = 1;
1524 			else
1525 				return (DDI_FAILURE);
1526 		}
1527 		break;
1528 	case DDI_INTROP_SUPPORTED_TYPES:
1529 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1530 		break;
1531 	default:
1532 		return (DDI_FAILURE);
1533 	}
1534 
1535 	return (DDI_SUCCESS);
1536 }
1537 
1538 
1539 /*
1540  * rootnex_get_ispec()
1541  *	convert an interrupt number to an interrupt specification.
1542  *	The interrupt number determines which interrupt spec will be
1543  *	returned if more than one exists.
1544  *
1545  *	Look into the parent private data area of the 'rdip' to find out
1546  *	the interrupt specification.  First check to make sure there is
1547  *	one that matchs "inumber" and then return a pointer to it.
1548  *
1549  *	Return NULL if one could not be found.
1550  *
1551  *	NOTE: This is needed for rootnex_intr_ops()
1552  */
1553 static struct intrspec *
1554 rootnex_get_ispec(dev_info_t *rdip, int inum)
1555 {
1556 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1557 
1558 	/*
1559 	 * Special case handling for drivers that provide their own
1560 	 * intrspec structures instead of relying on the DDI framework.
1561 	 *
1562 	 * A broken hardware driver in ON could potentially provide its
1563 	 * own intrspec structure, instead of relying on the hardware.
1564 	 * If these drivers are children of 'rootnex' then we need to
1565 	 * continue to provide backward compatibility to them here.
1566 	 *
1567 	 * Following check is a special case for 'pcic' driver which
1568 	 * was found to have broken hardwre andby provides its own intrspec.
1569 	 *
1570 	 * Verbatim comments from this driver are shown here:
1571 	 * "Don't use the ddi_add_intr since we don't have a
1572 	 * default intrspec in all cases."
1573 	 *
1574 	 * Since an 'ispec' may not be always created for it,
1575 	 * check for that and create one if so.
1576 	 *
1577 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1578 	 */
1579 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1580 		pdp->par_nintr = 1;
1581 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1582 		    pdp->par_nintr, KM_SLEEP);
1583 	}
1584 
1585 	/* Validate the interrupt number */
1586 	if (inum >= pdp->par_nintr)
1587 		return (NULL);
1588 
1589 	/* Get the interrupt structure pointer and return that */
1590 	return ((struct intrspec *)&pdp->par_intr[inum]);
1591 }
1592 
1593 
1594 /*
1595  * ******************
1596  *  dma related code
1597  * ******************
1598  */
1599 
1600 /*ARGSUSED*/
1601 static int
1602 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1603     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1604     ddi_dma_handle_t *handlep)
1605 {
1606 	uint64_t maxsegmentsize_ll;
1607 	uint_t maxsegmentsize;
1608 	ddi_dma_impl_t *hp;
1609 	rootnex_dma_t *dma;
1610 	uint64_t count_max;
1611 	uint64_t seg;
1612 	int kmflag;
1613 	int e;
1614 
1615 
1616 	/* convert our sleep flags */
1617 	if (waitfp == DDI_DMA_SLEEP) {
1618 		kmflag = KM_SLEEP;
1619 	} else {
1620 		kmflag = KM_NOSLEEP;
1621 	}
1622 
1623 	/*
1624 	 * We try to do only one memory allocation here. We'll do a little
1625 	 * pointer manipulation later. If the bind ends up taking more than
1626 	 * our prealloc's space, we'll have to allocate more memory in the
1627 	 * bind operation. Not great, but much better than before and the
1628 	 * best we can do with the current bind interfaces.
1629 	 */
1630 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1631 	if (hp == NULL) {
1632 		if (waitfp != DDI_DMA_DONTWAIT) {
1633 			ddi_set_callback(waitfp, arg,
1634 			    &rootnex_state->r_dvma_call_list_id);
1635 		}
1636 		return (DDI_DMA_NORESOURCES);
1637 	}
1638 
1639 	/* Do our pointer manipulation now, align the structures */
1640 	hp->dmai_private = (void *)(((uintptr_t)hp +
1641 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1642 	dma = (rootnex_dma_t *)hp->dmai_private;
1643 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1644 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1645 
1646 	/* setup the handle */
1647 	rootnex_clean_dmahdl(hp);
1648 	hp->dmai_error.err_fep = NULL;
1649 	hp->dmai_error.err_cf = NULL;
1650 	dma->dp_dip = rdip;
1651 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1652 	dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1653 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1654 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1655 	hp->dmai_rdip = rdip;
1656 	hp->dmai_attr = *attr;
1657 
1658 	/* we don't need to worry about the SPL since we do a tryenter */
1659 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1660 
1661 	/*
1662 	 * Figure out our maximum segment size. If the segment size is greater
1663 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1664 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1665 	 * dma_attr_count_max are size-1 type values.
1666 	 *
1667 	 * Maximum segment size is the largest physically contiguous chunk of
1668 	 * memory that we can return from a bind (i.e. the maximum size of a
1669 	 * single cookie).
1670 	 */
1671 
1672 	/* handle the rollover cases */
1673 	seg = attr->dma_attr_seg + 1;
1674 	if (seg < attr->dma_attr_seg) {
1675 		seg = attr->dma_attr_seg;
1676 	}
1677 	count_max = attr->dma_attr_count_max + 1;
1678 	if (count_max < attr->dma_attr_count_max) {
1679 		count_max = attr->dma_attr_count_max;
1680 	}
1681 
1682 	/*
1683 	 * granularity may or may not be a power of two. If it isn't, we can't
1684 	 * use a simple mask.
1685 	 */
1686 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1687 		dma->dp_granularity_power_2 = B_FALSE;
1688 	} else {
1689 		dma->dp_granularity_power_2 = B_TRUE;
1690 	}
1691 
1692 	/*
1693 	 * maxxfer should be a whole multiple of granularity. If we're going to
1694 	 * break up a window because we're greater than maxxfer, we might as
1695 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1696 	 * worry about triming the window later on for this case.
1697 	 */
1698 	if (attr->dma_attr_granular > 1) {
1699 		if (dma->dp_granularity_power_2) {
1700 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1701 			    (attr->dma_attr_maxxfer &
1702 			    (attr->dma_attr_granular - 1));
1703 		} else {
1704 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1705 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1706 		}
1707 	} else {
1708 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1709 	}
1710 
1711 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1712 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1713 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1714 		maxsegmentsize = 0xFFFFFFFF;
1715 	} else {
1716 		maxsegmentsize = maxsegmentsize_ll;
1717 	}
1718 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1719 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1720 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1721 
1722 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1723 	if (rootnex_alloc_check_parms) {
1724 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1725 		if (e != DDI_SUCCESS) {
1726 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1727 			(void) rootnex_dma_freehdl(dip, rdip,
1728 			    (ddi_dma_handle_t)hp);
1729 			return (e);
1730 		}
1731 	}
1732 
1733 	*handlep = (ddi_dma_handle_t)hp;
1734 
1735 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1736 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1737 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1738 
1739 	return (DDI_SUCCESS);
1740 }
1741 
1742 
1743 /*
1744  * rootnex_dma_allochdl()
1745  *    called from ddi_dma_alloc_handle().
1746  */
1747 static int
1748 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1749     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1750 {
1751 	int retval;
1752 #if defined(__amd64) && !defined(__xpv)
1753 	uint_t error = ENOTSUP;
1754 
1755 	retval = iommulib_nex_open(rdip, &error);
1756 
1757 	if (retval != DDI_SUCCESS && error == ENOTSUP) {
1758 		/* No IOMMU */
1759 		return (rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1760 		    handlep));
1761 	} else if (retval != DDI_SUCCESS) {
1762 		return (DDI_FAILURE);
1763 	}
1764 
1765 	ASSERT(IOMMU_USED(rdip));
1766 
1767 	/* has an IOMMU */
1768 	retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1769 	    waitfp, arg, handlep);
1770 #else
1771 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1772 	    handlep);
1773 #endif
1774 	if (retval == DDI_SUCCESS)
1775 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1776 	return (retval);
1777 }
1778 
1779 /*ARGSUSED*/
1780 static int
1781 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1782     ddi_dma_handle_t handle)
1783 {
1784 	ddi_dma_impl_t *hp;
1785 	rootnex_dma_t *dma;
1786 
1787 
1788 	hp = (ddi_dma_impl_t *)handle;
1789 	dma = (rootnex_dma_t *)hp->dmai_private;
1790 
1791 	/* unbind should have been called first */
1792 	ASSERT(!dma->dp_inuse);
1793 
1794 	mutex_destroy(&dma->dp_mutex);
1795 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1796 
1797 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1798 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1799 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1800 
1801 	if (rootnex_state->r_dvma_call_list_id)
1802 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1803 
1804 	return (DDI_SUCCESS);
1805 }
1806 
1807 /*
1808  * rootnex_dma_freehdl()
1809  *    called from ddi_dma_free_handle().
1810  */
1811 static int
1812 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1813 {
1814 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1815 #if defined(__amd64) && !defined(__xpv)
1816 	if (IOMMU_USED(rdip)) {
1817 		return (iommulib_nexdma_freehdl(dip, rdip, handle));
1818 	}
1819 #endif
1820 	return (rootnex_coredma_freehdl(dip, rdip, handle));
1821 }
1822 
1823 /*ARGSUSED*/
1824 static int
1825 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1826     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1827     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1828 {
1829 	rootnex_sglinfo_t *sinfo;
1830 	ddi_dma_attr_t *attr;
1831 	ddi_dma_impl_t *hp;
1832 	rootnex_dma_t *dma;
1833 	int kmflag;
1834 	int e;
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(__amd64) && !defined(__xpv)
1887 	e = immu_map_sgl(hp, dmareq, rootnex_prealloc_cookies, rdip);
1888 	switch (e) {
1889 	case DDI_DMA_MAPPED:
1890 		goto out;
1891 	case DDI_DMA_USE_PHYSICAL:
1892 		break;
1893 	case DDI_DMA_PARTIAL:
1894 		ddi_err(DER_PANIC, rdip, "Partial DVMA map");
1895 		e = DDI_DMA_NORESOURCES;
1896 		/*FALLTHROUGH*/
1897 	default:
1898 		ddi_err(DER_MODE, rdip, "DVMA map failed");
1899 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1900 		rootnex_clean_dmahdl(hp);
1901 		return (e);
1902 	}
1903 #endif
1904 
1905 	/*
1906 	 * Figure out a rough estimate of what maximum number of pages this
1907 	 * buffer could use (a high estimate of course).
1908 	 */
1909 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
1910 
1911 	/*
1912 	 * We'll use the pre-allocated cookies for any bind that will *always*
1913 	 * fit (more important to be consistent, we don't want to create
1914 	 * additional degenerate cases).
1915 	 */
1916 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
1917 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
1918 		dma->dp_need_to_free_cookie = B_FALSE;
1919 		DTRACE_PROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
1920 		    uint_t, sinfo->si_max_pages);
1921 
1922 	/*
1923 	 * For anything larger than that, we'll go ahead and allocate the
1924 	 * maximum number of pages we expect to see. Hopefuly, we won't be
1925 	 * seeing this path in the fast path for high performance devices very
1926 	 * frequently.
1927 	 *
1928 	 * a ddi bind interface that allowed the driver to provide storage to
1929 	 * the bind interface would speed this case up.
1930 	 */
1931 	} else {
1932 		/* convert the sleep flags */
1933 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1934 			kmflag =  KM_SLEEP;
1935 		} else {
1936 			kmflag =  KM_NOSLEEP;
1937 		}
1938 
1939 		/*
1940 		 * Save away how much memory we allocated. If we're doing a
1941 		 * nosleep, the alloc could fail...
1942 		 */
1943 		dma->dp_cookie_size = sinfo->si_max_pages *
1944 		    sizeof (ddi_dma_cookie_t);
1945 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
1946 		if (dma->dp_cookies == NULL) {
1947 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1948 			rootnex_clean_dmahdl(hp);
1949 			return (DDI_DMA_NORESOURCES);
1950 		}
1951 		dma->dp_need_to_free_cookie = B_TRUE;
1952 		DTRACE_PROBE2(rootnex__bind__alloc, dev_info_t *, rdip, uint_t,
1953 		    sinfo->si_max_pages);
1954 	}
1955 	hp->dmai_cookie = dma->dp_cookies;
1956 
1957 	/*
1958 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
1959 	 * looking at the constraints in the dma structure. It will then put
1960 	 * some additional state about the sgl in the dma struct (i.e. is
1961 	 * the sgl clean, or do we need to do some munging; how many pages
1962 	 * need to be copied, etc.)
1963 	 */
1964 	rootnex_get_sgl(&dmareq->dmar_object, dma->dp_cookies,
1965 	    &dma->dp_sglinfo);
1966 
1967 out:
1968 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
1969 	/* if we don't need a copy buffer, we don't need to sync */
1970 	if (sinfo->si_copybuf_req == 0) {
1971 		hp->dmai_rflags |= DMP_NOSYNC;
1972 	}
1973 
1974 	/*
1975 	 * if we don't need the copybuf and we don't need to do a partial,  we
1976 	 * hit the fast path. All the high performance devices should be trying
1977 	 * to hit this path. To hit this path, a device should be able to reach
1978 	 * all of memory, shouldn't try to bind more than it can transfer, and
1979 	 * the buffer shouldn't require more cookies than the driver/device can
1980 	 * handle [sgllen]).
1981 	 */
1982 	if ((sinfo->si_copybuf_req == 0) &&
1983 	    (sinfo->si_sgl_size <= attr->dma_attr_sgllen) &&
1984 	    (dma->dp_dma.dmao_size < dma->dp_maxxfer)) {
1985 		/*
1986 		 * If the driver supports FMA, insert the handle in the FMA DMA
1987 		 * handle cache.
1988 		 */
1989 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
1990 			hp->dmai_error.err_cf = rootnex_dma_check;
1991 
1992 		/*
1993 		 * copy out the first cookie and ccountp, set the cookie
1994 		 * pointer to the second cookie. The first cookie is passed
1995 		 * back on the stack. Additional cookies are accessed via
1996 		 * ddi_dma_nextcookie()
1997 		 */
1998 		*cookiep = dma->dp_cookies[0];
1999 		*ccountp = sinfo->si_sgl_size;
2000 		hp->dmai_cookie++;
2001 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2002 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2003 		DTRACE_PROBE3(rootnex__bind__fast, dev_info_t *, rdip,
2004 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2005 		    uint_t, dma->dp_dma.dmao_size);
2006 
2007 
2008 		return (DDI_DMA_MAPPED);
2009 	}
2010 
2011 	/*
2012 	 * go to the slow path, we may need to alloc more memory, create
2013 	 * multiple windows, and munge up a sgl to make the device happy.
2014 	 */
2015 	e = rootnex_bind_slowpath(hp, dmareq, dma, attr, kmflag);
2016 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2017 		if (dma->dp_need_to_free_cookie) {
2018 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2019 		}
2020 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2021 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2022 		return (e);
2023 	}
2024 
2025 	/*
2026 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2027 	 * cache.
2028 	 */
2029 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2030 		hp->dmai_error.err_cf = rootnex_dma_check;
2031 
2032 	/* if the first window uses the copy buffer, sync it for the device */
2033 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2034 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2035 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2036 		    DDI_DMA_SYNC_FORDEV);
2037 	}
2038 
2039 	/*
2040 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2041 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2042 	 * If we have a partial map (i.e. multiple windows), the number of
2043 	 * cookies we return is the number of cookies in the first window.
2044 	 */
2045 	if (e == DDI_DMA_MAPPED) {
2046 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2047 		*ccountp = sinfo->si_sgl_size;
2048 		hp->dmai_nwin = 1;
2049 	} else {
2050 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2051 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2052 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2053 	}
2054 	*cookiep = dma->dp_cookies[0];
2055 	hp->dmai_cookie++;
2056 
2057 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2058 	ROOTNEX_DPROBE3(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2059 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2060 	    dma->dp_dma.dmao_size);
2061 	return (e);
2062 }
2063 
2064 /*
2065  * rootnex_dma_bindhdl()
2066  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2067  */
2068 static int
2069 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2070     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2071     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2072 {
2073 #if defined(__amd64) && !defined(__xpv)
2074 	if (IOMMU_USED(rdip)) {
2075 		return (iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2076 		    cookiep, ccountp));
2077 	}
2078 #endif
2079 	return (rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2080 	    cookiep, ccountp));
2081 }
2082 
2083 
2084 
2085 /*ARGSUSED*/
2086 static int
2087 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2088     ddi_dma_handle_t handle)
2089 {
2090 	ddi_dma_impl_t *hp;
2091 	rootnex_dma_t *dma;
2092 	int e;
2093 
2094 	hp = (ddi_dma_impl_t *)handle;
2095 	dma = (rootnex_dma_t *)hp->dmai_private;
2096 
2097 	/* make sure the buffer wasn't free'd before calling unbind */
2098 	if (rootnex_unbind_verify_buffer) {
2099 		e = rootnex_verify_buffer(dma);
2100 		if (e != DDI_SUCCESS) {
2101 			ASSERT(0);
2102 			return (DDI_FAILURE);
2103 		}
2104 	}
2105 
2106 	/* sync the current window before unbinding the buffer */
2107 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2108 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2109 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2110 		    DDI_DMA_SYNC_FORCPU);
2111 	}
2112 
2113 	/*
2114 	 * cleanup and copy buffer or window state. if we didn't use the copy
2115 	 * buffer or windows, there won't be much to do :-)
2116 	 */
2117 	rootnex_teardown_copybuf(dma);
2118 	rootnex_teardown_windows(dma);
2119 
2120 #if defined(__amd64) && !defined(__xpv)
2121 	/*
2122 	 * Clean up the page tables and free the dvma
2123 	 */
2124 	e = immu_unmap_sgl(hp, rdip);
2125 	if (e != DDI_DMA_USE_PHYSICAL && e != DDI_SUCCESS) {
2126 		return (e);
2127 	}
2128 #endif
2129 
2130 	/*
2131 	 * If we had to allocate space to for the worse case sgl (it didn't
2132 	 * fit into our pre-allocate buffer), free that up now
2133 	 */
2134 	if (dma->dp_need_to_free_cookie) {
2135 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2136 	}
2137 
2138 	/*
2139 	 * clean up the handle so it's ready for the next bind (i.e. if the
2140 	 * handle is reused).
2141 	 */
2142 	rootnex_clean_dmahdl(hp);
2143 	hp->dmai_error.err_cf = NULL;
2144 
2145 	if (rootnex_state->r_dvma_call_list_id)
2146 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2147 
2148 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2149 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2150 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2151 
2152 	return (DDI_SUCCESS);
2153 }
2154 
2155 /*
2156  * rootnex_dma_unbindhdl()
2157  *    called from ddi_dma_unbind_handle()
2158  */
2159 /*ARGSUSED*/
2160 static int
2161 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2162     ddi_dma_handle_t handle)
2163 {
2164 #if defined(__amd64) && !defined(__xpv)
2165 	if (IOMMU_USED(rdip)) {
2166 		return (iommulib_nexdma_unbindhdl(dip, rdip, handle));
2167 	}
2168 #endif
2169 	return (rootnex_coredma_unbindhdl(dip, rdip, handle));
2170 }
2171 
2172 #if defined(__amd64) && !defined(__xpv)
2173 
2174 static int
2175 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2176 {
2177 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2178 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2179 
2180 	if (dma->dp_sleep_flags != KM_SLEEP &&
2181 	    dma->dp_sleep_flags != KM_NOSLEEP)
2182 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2183 	return (dma->dp_sleep_flags);
2184 }
2185 /*ARGSUSED*/
2186 static void
2187 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2188 {
2189 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2190 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2191 	rootnex_window_t *window;
2192 
2193 	if (dma->dp_window) {
2194 		window = &dma->dp_window[dma->dp_current_win];
2195 		hp->dmai_cookie = window->wd_first_cookie;
2196 	} else {
2197 		hp->dmai_cookie = dma->dp_cookies;
2198 	}
2199 	hp->dmai_cookie++;
2200 }
2201 
2202 /*ARGSUSED*/
2203 static int
2204 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2205     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2206 {
2207 	int i;
2208 	int km_flags;
2209 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2210 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2211 	rootnex_window_t *window;
2212 	ddi_dma_cookie_t *cp;
2213 	ddi_dma_cookie_t *cookie;
2214 
2215 	ASSERT(*cookiepp == NULL);
2216 	ASSERT(*ccountp == 0);
2217 
2218 	if (dma->dp_window) {
2219 		window = &dma->dp_window[dma->dp_current_win];
2220 		cp = window->wd_first_cookie;
2221 		*ccountp = window->wd_cookie_cnt;
2222 	} else {
2223 		cp = dma->dp_cookies;
2224 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2225 	}
2226 
2227 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2228 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2229 	if (cookie == NULL) {
2230 		return (DDI_DMA_NORESOURCES);
2231 	}
2232 
2233 	for (i = 0; i < *ccountp; i++) {
2234 		cookie[i].dmac_notused = cp[i].dmac_notused;
2235 		cookie[i].dmac_type = cp[i].dmac_type;
2236 		cookie[i].dmac_address = cp[i].dmac_address;
2237 		cookie[i].dmac_size = cp[i].dmac_size;
2238 	}
2239 
2240 	*cookiepp = cookie;
2241 
2242 	return (DDI_SUCCESS);
2243 }
2244 
2245 /*ARGSUSED*/
2246 static int
2247 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2248     ddi_dma_cookie_t *cookiep, uint_t ccount)
2249 {
2250 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2251 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2252 	rootnex_window_t *window;
2253 	ddi_dma_cookie_t *cur_cookiep;
2254 
2255 	ASSERT(cookiep);
2256 	ASSERT(ccount != 0);
2257 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2258 
2259 	if (dma->dp_window) {
2260 		window = &dma->dp_window[dma->dp_current_win];
2261 		dma->dp_saved_cookies = window->wd_first_cookie;
2262 		window->wd_first_cookie = cookiep;
2263 		ASSERT(ccount == window->wd_cookie_cnt);
2264 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2265 		    + window->wd_first_cookie;
2266 	} else {
2267 		dma->dp_saved_cookies = dma->dp_cookies;
2268 		dma->dp_cookies = cookiep;
2269 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2270 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2271 		    + dma->dp_cookies;
2272 	}
2273 
2274 	dma->dp_need_to_switch_cookies = B_TRUE;
2275 	hp->dmai_cookie = cur_cookiep;
2276 
2277 	return (DDI_SUCCESS);
2278 }
2279 
2280 /*ARGSUSED*/
2281 static int
2282 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2283 {
2284 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2285 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2286 	rootnex_window_t *window;
2287 	ddi_dma_cookie_t *cur_cookiep;
2288 	ddi_dma_cookie_t *cookie_array;
2289 	uint_t ccount;
2290 
2291 	/* check if cookies have not been switched */
2292 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2293 		return (DDI_SUCCESS);
2294 
2295 	ASSERT(dma->dp_saved_cookies);
2296 
2297 	if (dma->dp_window) {
2298 		window = &dma->dp_window[dma->dp_current_win];
2299 		cookie_array = window->wd_first_cookie;
2300 		window->wd_first_cookie = dma->dp_saved_cookies;
2301 		dma->dp_saved_cookies = NULL;
2302 		ccount = window->wd_cookie_cnt;
2303 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2304 		    + window->wd_first_cookie;
2305 	} else {
2306 		cookie_array = dma->dp_cookies;
2307 		dma->dp_cookies = dma->dp_saved_cookies;
2308 		dma->dp_saved_cookies = NULL;
2309 		ccount = dma->dp_sglinfo.si_sgl_size;
2310 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2311 		    + dma->dp_cookies;
2312 	}
2313 
2314 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2315 
2316 	hp->dmai_cookie = cur_cookiep;
2317 
2318 	dma->dp_need_to_switch_cookies = B_FALSE;
2319 
2320 	return (DDI_SUCCESS);
2321 }
2322 
2323 #endif
2324 
2325 /*
2326  * rootnex_verify_buffer()
2327  *   verify buffer wasn't free'd
2328  */
2329 static int
2330 rootnex_verify_buffer(rootnex_dma_t *dma)
2331 {
2332 	page_t **pplist;
2333 	caddr_t vaddr;
2334 	uint_t pcnt;
2335 	uint_t poff;
2336 	page_t *pp;
2337 	char b;
2338 	int i;
2339 
2340 	/* Figure out how many pages this buffer occupies */
2341 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2342 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2343 	} else {
2344 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2345 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2346 	}
2347 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2348 
2349 	switch (dma->dp_dma.dmao_type) {
2350 	case DMA_OTYP_PAGES:
2351 		/*
2352 		 * for a linked list of pp's walk through them to make sure
2353 		 * they're locked and not free.
2354 		 */
2355 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2356 		for (i = 0; i < pcnt; i++) {
2357 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2358 				return (DDI_FAILURE);
2359 			}
2360 			pp = pp->p_next;
2361 		}
2362 		break;
2363 
2364 	case DMA_OTYP_VADDR:
2365 	case DMA_OTYP_BUFVADDR:
2366 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2367 		/*
2368 		 * for an array of pp's walk through them to make sure they're
2369 		 * not free. It's possible that they may not be locked.
2370 		 */
2371 		if (pplist) {
2372 			for (i = 0; i < pcnt; i++) {
2373 				if (PP_ISFREE(pplist[i])) {
2374 					return (DDI_FAILURE);
2375 				}
2376 			}
2377 
2378 		/* For a virtual address, try to peek at each page */
2379 		} else {
2380 			if (dma->dp_sglinfo.si_asp == &kas) {
2381 				for (i = 0; i < pcnt; i++) {
2382 					if (ddi_peek8(NULL, vaddr, &b) ==
2383 					    DDI_FAILURE)
2384 						return (DDI_FAILURE);
2385 					vaddr += MMU_PAGESIZE;
2386 				}
2387 			}
2388 		}
2389 		break;
2390 
2391 	default:
2392 		ASSERT(0);
2393 		break;
2394 	}
2395 
2396 	return (DDI_SUCCESS);
2397 }
2398 
2399 
2400 /*
2401  * rootnex_clean_dmahdl()
2402  *    Clean the dma handle. This should be called on a handle alloc and an
2403  *    unbind handle. Set the handle state to the default settings.
2404  */
2405 static void
2406 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2407 {
2408 	rootnex_dma_t *dma;
2409 
2410 
2411 	dma = (rootnex_dma_t *)hp->dmai_private;
2412 
2413 	hp->dmai_nwin = 0;
2414 	dma->dp_current_cookie = 0;
2415 	dma->dp_copybuf_size = 0;
2416 	dma->dp_window = NULL;
2417 	dma->dp_cbaddr = NULL;
2418 	dma->dp_inuse = B_FALSE;
2419 	dma->dp_need_to_free_cookie = B_FALSE;
2420 	dma->dp_need_to_switch_cookies = B_FALSE;
2421 	dma->dp_saved_cookies = NULL;
2422 	dma->dp_sleep_flags = KM_PANIC;
2423 	dma->dp_need_to_free_window = B_FALSE;
2424 	dma->dp_partial_required = B_FALSE;
2425 	dma->dp_trim_required = B_FALSE;
2426 	dma->dp_sglinfo.si_copybuf_req = 0;
2427 #if !defined(__amd64)
2428 	dma->dp_cb_remaping = B_FALSE;
2429 	dma->dp_kva = NULL;
2430 #endif
2431 
2432 	/* FMA related initialization */
2433 	hp->dmai_fault = 0;
2434 	hp->dmai_fault_check = NULL;
2435 	hp->dmai_fault_notify = NULL;
2436 	hp->dmai_error.err_ena = 0;
2437 	hp->dmai_error.err_status = DDI_FM_OK;
2438 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2439 	hp->dmai_error.err_ontrap = NULL;
2440 }
2441 
2442 
2443 /*
2444  * rootnex_valid_alloc_parms()
2445  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2446  */
2447 static int
2448 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2449 {
2450 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2451 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2452 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2453 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2454 		return (DDI_DMA_BADATTR);
2455 	}
2456 
2457 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2458 		return (DDI_DMA_BADATTR);
2459 	}
2460 
2461 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2462 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2463 	    attr->dma_attr_sgllen <= 0) {
2464 		return (DDI_DMA_BADATTR);
2465 	}
2466 
2467 	/* We should be able to DMA into every byte offset in a page */
2468 	if (maxsegmentsize < MMU_PAGESIZE) {
2469 		return (DDI_DMA_BADATTR);
2470 	}
2471 
2472 	/* if we're bouncing on seg, seg must be <= addr_hi */
2473 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2474 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2475 		return (DDI_DMA_BADATTR);
2476 	}
2477 	return (DDI_SUCCESS);
2478 }
2479 
2480 /*
2481  * rootnex_valid_bind_parms()
2482  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2483  */
2484 /* ARGSUSED */
2485 static int
2486 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2487 {
2488 #if !defined(__amd64)
2489 	/*
2490 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2491 	 * we can track the offset for the obsoleted interfaces.
2492 	 */
2493 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2494 		return (DDI_DMA_TOOBIG);
2495 	}
2496 #endif
2497 
2498 	return (DDI_SUCCESS);
2499 }
2500 
2501 
2502 /*
2503  * rootnex_need_bounce_seg()
2504  *    check to see if the buffer lives on both side of the seg.
2505  */
2506 static boolean_t
2507 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2508 {
2509 	ddi_dma_atyp_t buftype;
2510 	rootnex_addr_t raddr;
2511 	boolean_t lower_addr;
2512 	boolean_t upper_addr;
2513 	uint64_t offset;
2514 	page_t **pplist;
2515 	uint64_t paddr;
2516 	uint32_t psize;
2517 	uint32_t size;
2518 	caddr_t vaddr;
2519 	uint_t pcnt;
2520 	page_t *pp;
2521 
2522 
2523 	/* shortcuts */
2524 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2525 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2526 	buftype = dmar_object->dmao_type;
2527 	size = dmar_object->dmao_size;
2528 
2529 	lower_addr = B_FALSE;
2530 	upper_addr = B_FALSE;
2531 	pcnt = 0;
2532 
2533 	/*
2534 	 * Process the first page to handle the initial offset of the buffer.
2535 	 * We'll use the base address we get later when we loop through all
2536 	 * the pages.
2537 	 */
2538 	if (buftype == DMA_OTYP_PAGES) {
2539 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2540 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2541 		    MMU_PAGEOFFSET;
2542 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2543 		psize = MIN(size, (MMU_PAGESIZE - offset));
2544 		pp = pp->p_next;
2545 		sglinfo->si_asp = NULL;
2546 	} else if (pplist != NULL) {
2547 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2548 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2549 		if (sglinfo->si_asp == NULL) {
2550 			sglinfo->si_asp = &kas;
2551 		}
2552 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2553 		paddr += offset;
2554 		psize = MIN(size, (MMU_PAGESIZE - offset));
2555 		pcnt++;
2556 	} else {
2557 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2558 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2559 		if (sglinfo->si_asp == NULL) {
2560 			sglinfo->si_asp = &kas;
2561 		}
2562 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2563 		paddr += offset;
2564 		psize = MIN(size, (MMU_PAGESIZE - offset));
2565 		vaddr += psize;
2566 	}
2567 
2568 #ifdef __xpv
2569 	/*
2570 	 * If we're dom0, we're using a real device so we need to load
2571 	 * the cookies with MFNs instead of PFNs.
2572 	 */
2573 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2574 #else
2575 	raddr = paddr;
2576 #endif
2577 
2578 	if ((raddr + psize) > sglinfo->si_segmask) {
2579 		upper_addr = B_TRUE;
2580 	} else {
2581 		lower_addr = B_TRUE;
2582 	}
2583 	size -= psize;
2584 
2585 	/*
2586 	 * Walk through the rest of the pages in the buffer. Track to see
2587 	 * if we have pages on both sides of the segment boundary.
2588 	 */
2589 	while (size > 0) {
2590 		/* partial or full page */
2591 		psize = MIN(size, MMU_PAGESIZE);
2592 
2593 		if (buftype == DMA_OTYP_PAGES) {
2594 			/* get the paddr from the page_t */
2595 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2596 			paddr = pfn_to_pa(pp->p_pagenum);
2597 			pp = pp->p_next;
2598 		} else if (pplist != NULL) {
2599 			/* index into the array of page_t's to get the paddr */
2600 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2601 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2602 			pcnt++;
2603 		} else {
2604 			/* call into the VM to get the paddr */
2605 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2606 			    vaddr));
2607 			vaddr += psize;
2608 		}
2609 
2610 #ifdef __xpv
2611 		/*
2612 		 * If we're dom0, we're using a real device so we need to load
2613 		 * the cookies with MFNs instead of PFNs.
2614 		 */
2615 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2616 #else
2617 		raddr = paddr;
2618 #endif
2619 
2620 		if ((raddr + psize) > sglinfo->si_segmask) {
2621 			upper_addr = B_TRUE;
2622 		} else {
2623 			lower_addr = B_TRUE;
2624 		}
2625 		/*
2626 		 * if the buffer lives both above and below the segment
2627 		 * boundary, or the current page is the page immediately
2628 		 * after the segment, we will use a copy/bounce buffer for
2629 		 * all pages > seg.
2630 		 */
2631 		if ((lower_addr && upper_addr) ||
2632 		    (raddr == (sglinfo->si_segmask + 1))) {
2633 			return (B_TRUE);
2634 		}
2635 
2636 		size -= psize;
2637 	}
2638 
2639 	return (B_FALSE);
2640 }
2641 
2642 
2643 /*
2644  * rootnex_get_sgl()
2645  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2646  *    with a call to the vm layer when vm2.0 comes around...
2647  */
2648 static void
2649 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2650     rootnex_sglinfo_t *sglinfo)
2651 {
2652 	ddi_dma_atyp_t buftype;
2653 	rootnex_addr_t raddr;
2654 	uint64_t last_page;
2655 	uint64_t offset;
2656 	uint64_t addrhi;
2657 	uint64_t addrlo;
2658 	uint64_t maxseg;
2659 	page_t **pplist;
2660 	uint64_t paddr;
2661 	uint32_t psize;
2662 	uint32_t size;
2663 	caddr_t vaddr;
2664 	uint_t pcnt;
2665 	page_t *pp;
2666 	uint_t cnt;
2667 
2668 
2669 	/* shortcuts */
2670 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2671 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2672 	maxseg = sglinfo->si_max_cookie_size;
2673 	buftype = dmar_object->dmao_type;
2674 	addrhi = sglinfo->si_max_addr;
2675 	addrlo = sglinfo->si_min_addr;
2676 	size = dmar_object->dmao_size;
2677 
2678 	pcnt = 0;
2679 	cnt = 0;
2680 
2681 
2682 	/*
2683 	 * check to see if we need to use the copy buffer for pages over
2684 	 * the segment attr.
2685 	 */
2686 	sglinfo->si_bounce_on_seg = B_FALSE;
2687 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2688 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2689 		    dmar_object, sglinfo);
2690 	}
2691 
2692 	/*
2693 	 * if we were passed down a linked list of pages, i.e. pointer to
2694 	 * page_t, use this to get our physical address and buf offset.
2695 	 */
2696 	if (buftype == DMA_OTYP_PAGES) {
2697 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2698 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2699 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2700 		    MMU_PAGEOFFSET;
2701 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2702 		psize = MIN(size, (MMU_PAGESIZE - offset));
2703 		pp = pp->p_next;
2704 		sglinfo->si_asp = NULL;
2705 
2706 	/*
2707 	 * We weren't passed down a linked list of pages, but if we were passed
2708 	 * down an array of pages, use this to get our physical address and buf
2709 	 * offset.
2710 	 */
2711 	} else if (pplist != NULL) {
2712 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2713 		    (buftype == DMA_OTYP_BUFVADDR));
2714 
2715 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2716 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2717 		if (sglinfo->si_asp == NULL) {
2718 			sglinfo->si_asp = &kas;
2719 		}
2720 
2721 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2722 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2723 		paddr += offset;
2724 		psize = MIN(size, (MMU_PAGESIZE - offset));
2725 		pcnt++;
2726 
2727 	/*
2728 	 * All we have is a virtual address, we'll need to call into the VM
2729 	 * to get the physical address.
2730 	 */
2731 	} else {
2732 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2733 		    (buftype == DMA_OTYP_BUFVADDR));
2734 
2735 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2736 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2737 		if (sglinfo->si_asp == NULL) {
2738 			sglinfo->si_asp = &kas;
2739 		}
2740 
2741 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2742 		paddr += offset;
2743 		psize = MIN(size, (MMU_PAGESIZE - offset));
2744 		vaddr += psize;
2745 	}
2746 
2747 #ifdef __xpv
2748 	/*
2749 	 * If we're dom0, we're using a real device so we need to load
2750 	 * the cookies with MFNs instead of PFNs.
2751 	 */
2752 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2753 #else
2754 	raddr = paddr;
2755 #endif
2756 
2757 	/*
2758 	 * Setup the first cookie with the physical address of the page and the
2759 	 * size of the page (which takes into account the initial offset into
2760 	 * the page.
2761 	 */
2762 	sgl[cnt].dmac_laddress = raddr;
2763 	sgl[cnt].dmac_size = psize;
2764 	sgl[cnt].dmac_type = 0;
2765 
2766 	/*
2767 	 * Save away the buffer offset into the page. We'll need this later in
2768 	 * the copy buffer code to help figure out the page index within the
2769 	 * buffer and the offset into the current page.
2770 	 */
2771 	sglinfo->si_buf_offset = offset;
2772 
2773 	/*
2774 	 * If we are using the copy buffer for anything over the segment
2775 	 * boundary, and this page is over the segment boundary.
2776 	 *   OR
2777 	 * if the DMA engine can't reach the physical address.
2778 	 */
2779 	if (((sglinfo->si_bounce_on_seg) &&
2780 	    ((raddr + psize) > sglinfo->si_segmask)) ||
2781 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2782 		/*
2783 		 * Increase how much copy buffer we use. We always increase by
2784 		 * pagesize so we don't have to worry about converting offsets.
2785 		 * Set a flag in the cookies dmac_type to indicate that it uses
2786 		 * the copy buffer. If this isn't the last cookie, go to the
2787 		 * next cookie (since we separate each page which uses the copy
2788 		 * buffer in case the copy buffer is not physically contiguous.
2789 		 */
2790 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
2791 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2792 		if ((cnt + 1) < sglinfo->si_max_pages) {
2793 			cnt++;
2794 			sgl[cnt].dmac_laddress = 0;
2795 			sgl[cnt].dmac_size = 0;
2796 			sgl[cnt].dmac_type = 0;
2797 		}
2798 	}
2799 
2800 	/*
2801 	 * save this page's physical address so we can figure out if the next
2802 	 * page is physically contiguous. Keep decrementing size until we are
2803 	 * done with the buffer.
2804 	 */
2805 	last_page = raddr & MMU_PAGEMASK;
2806 	size -= psize;
2807 
2808 	while (size > 0) {
2809 		/* Get the size for this page (i.e. partial or full page) */
2810 		psize = MIN(size, MMU_PAGESIZE);
2811 
2812 		if (buftype == DMA_OTYP_PAGES) {
2813 			/* get the paddr from the page_t */
2814 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2815 			paddr = pfn_to_pa(pp->p_pagenum);
2816 			pp = pp->p_next;
2817 		} else if (pplist != NULL) {
2818 			/* index into the array of page_t's to get the paddr */
2819 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2820 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2821 			pcnt++;
2822 		} else {
2823 			/* call into the VM to get the paddr */
2824 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2825 			    vaddr));
2826 			vaddr += psize;
2827 		}
2828 
2829 #ifdef __xpv
2830 		/*
2831 		 * If we're dom0, we're using a real device so we need to load
2832 		 * the cookies with MFNs instead of PFNs.
2833 		 */
2834 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2835 #else
2836 		raddr = paddr;
2837 #endif
2838 
2839 		/*
2840 		 * If we are using the copy buffer for anything over the
2841 		 * segment boundary, and this page is over the segment
2842 		 * boundary.
2843 		 *   OR
2844 		 * if the DMA engine can't reach the physical address.
2845 		 */
2846 		if (((sglinfo->si_bounce_on_seg) &&
2847 		    ((raddr + psize) > sglinfo->si_segmask)) ||
2848 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2849 
2850 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
2851 
2852 			/*
2853 			 * if there is something in the current cookie, go to
2854 			 * the next one. We only want one page in a cookie which
2855 			 * uses the copybuf since the copybuf doesn't have to
2856 			 * be physically contiguous.
2857 			 */
2858 			if (sgl[cnt].dmac_size != 0) {
2859 				cnt++;
2860 			}
2861 			sgl[cnt].dmac_laddress = raddr;
2862 			sgl[cnt].dmac_size = psize;
2863 #if defined(__amd64)
2864 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2865 #else
2866 			/*
2867 			 * save the buf offset for 32-bit kernel. used in the
2868 			 * obsoleted interfaces.
2869 			 */
2870 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
2871 			    (dmar_object->dmao_size - size);
2872 #endif
2873 			/* if this isn't the last cookie, go to the next one */
2874 			if ((cnt + 1) < sglinfo->si_max_pages) {
2875 				cnt++;
2876 				sgl[cnt].dmac_laddress = 0;
2877 				sgl[cnt].dmac_size = 0;
2878 				sgl[cnt].dmac_type = 0;
2879 			}
2880 
2881 		/*
2882 		 * this page didn't need the copy buffer, if it's not physically
2883 		 * contiguous, or it would put us over a segment boundary, or it
2884 		 * puts us over the max cookie size, or the current sgl doesn't
2885 		 * have anything in it.
2886 		 */
2887 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
2888 		    !(raddr & sglinfo->si_segmask) ||
2889 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
2890 		    (sgl[cnt].dmac_size == 0)) {
2891 			/*
2892 			 * if we're not already in a new cookie, go to the next
2893 			 * cookie.
2894 			 */
2895 			if (sgl[cnt].dmac_size != 0) {
2896 				cnt++;
2897 			}
2898 
2899 			/* save the cookie information */
2900 			sgl[cnt].dmac_laddress = raddr;
2901 			sgl[cnt].dmac_size = psize;
2902 #if defined(__amd64)
2903 			sgl[cnt].dmac_type = 0;
2904 #else
2905 			/*
2906 			 * save the buf offset for 32-bit kernel. used in the
2907 			 * obsoleted interfaces.
2908 			 */
2909 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
2910 #endif
2911 
2912 		/*
2913 		 * this page didn't need the copy buffer, it is physically
2914 		 * contiguous with the last page, and it's <= the max cookie
2915 		 * size.
2916 		 */
2917 		} else {
2918 			sgl[cnt].dmac_size += psize;
2919 
2920 			/*
2921 			 * if this exactly ==  the maximum cookie size, and
2922 			 * it isn't the last cookie, go to the next cookie.
2923 			 */
2924 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
2925 			    ((cnt + 1) < sglinfo->si_max_pages)) {
2926 				cnt++;
2927 				sgl[cnt].dmac_laddress = 0;
2928 				sgl[cnt].dmac_size = 0;
2929 				sgl[cnt].dmac_type = 0;
2930 			}
2931 		}
2932 
2933 		/*
2934 		 * save this page's physical address so we can figure out if the
2935 		 * next page is physically contiguous. Keep decrementing size
2936 		 * until we are done with the buffer.
2937 		 */
2938 		last_page = raddr;
2939 		size -= psize;
2940 	}
2941 
2942 	/* we're done, save away how many cookies the sgl has */
2943 	if (sgl[cnt].dmac_size == 0) {
2944 		ASSERT(cnt < sglinfo->si_max_pages);
2945 		sglinfo->si_sgl_size = cnt;
2946 	} else {
2947 		sglinfo->si_sgl_size = cnt + 1;
2948 	}
2949 }
2950 
2951 /*
2952  * rootnex_bind_slowpath()
2953  *    Call in the bind path if the calling driver can't use the sgl without
2954  *    modifying it. We either need to use the copy buffer and/or we will end up
2955  *    with a partial bind.
2956  */
2957 static int
2958 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
2959     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag)
2960 {
2961 	rootnex_sglinfo_t *sinfo;
2962 	rootnex_window_t *window;
2963 	ddi_dma_cookie_t *cookie;
2964 	size_t copybuf_used;
2965 	size_t dmac_size;
2966 	boolean_t partial;
2967 	off_t cur_offset;
2968 	page_t *cur_pp;
2969 	major_t mnum;
2970 	int e;
2971 	int i;
2972 
2973 
2974 	sinfo = &dma->dp_sglinfo;
2975 	copybuf_used = 0;
2976 	partial = B_FALSE;
2977 
2978 	/*
2979 	 * If we're using the copybuf, set the copybuf state in dma struct.
2980 	 * Needs to be first since it sets the copy buffer size.
2981 	 */
2982 	if (sinfo->si_copybuf_req != 0) {
2983 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
2984 		if (e != DDI_SUCCESS) {
2985 			return (e);
2986 		}
2987 	} else {
2988 		dma->dp_copybuf_size = 0;
2989 	}
2990 
2991 	/*
2992 	 * Figure out if we need to do a partial mapping. If so, figure out
2993 	 * if we need to trim the buffers when we munge the sgl.
2994 	 */
2995 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
2996 	    (dma->dp_dma.dmao_size > dma->dp_maxxfer) ||
2997 	    (attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
2998 		dma->dp_partial_required = B_TRUE;
2999 		if (attr->dma_attr_granular != 1) {
3000 			dma->dp_trim_required = B_TRUE;
3001 		}
3002 	} else {
3003 		dma->dp_partial_required = B_FALSE;
3004 		dma->dp_trim_required = B_FALSE;
3005 	}
3006 
3007 	/* If we need to do a partial bind, make sure the driver supports it */
3008 	if (dma->dp_partial_required &&
3009 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3010 
3011 		mnum = ddi_driver_major(dma->dp_dip);
3012 		/*
3013 		 * patchable which allows us to print one warning per major
3014 		 * number.
3015 		 */
3016 		if ((rootnex_bind_warn) &&
3017 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3018 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3019 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3020 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3021 			    "There is a small risk of data corruption in "
3022 			    "particular with large I/Os. The driver should be "
3023 			    "replaced with a corrected version for proper "
3024 			    "system operation. To disable this warning, add "
3025 			    "'set rootnex:rootnex_bind_warn=0' to "
3026 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3027 		}
3028 		return (DDI_DMA_TOOBIG);
3029 	}
3030 
3031 	/*
3032 	 * we might need multiple windows, setup state to handle them. In this
3033 	 * code path, we will have at least one window.
3034 	 */
3035 	e = rootnex_setup_windows(hp, dma, attr, kmflag);
3036 	if (e != DDI_SUCCESS) {
3037 		rootnex_teardown_copybuf(dma);
3038 		return (e);
3039 	}
3040 
3041 	window = &dma->dp_window[0];
3042 	cookie = &dma->dp_cookies[0];
3043 	cur_offset = 0;
3044 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3045 	if (dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) {
3046 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3047 	}
3048 
3049 	/* loop though all the cookies we got back from get_sgl() */
3050 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3051 		/*
3052 		 * If we're using the copy buffer, check this cookie and setup
3053 		 * its associated copy buffer state. If this cookie uses the
3054 		 * copy buffer, make sure we sync this window during dma_sync.
3055 		 */
3056 		if (dma->dp_copybuf_size > 0) {
3057 			rootnex_setup_cookie(&dmareq->dmar_object, dma, cookie,
3058 			    cur_offset, &copybuf_used, &cur_pp);
3059 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3060 				window->wd_dosync = B_TRUE;
3061 			}
3062 		}
3063 
3064 		/*
3065 		 * save away the cookie size, since it could be modified in
3066 		 * the windowing code.
3067 		 */
3068 		dmac_size = cookie->dmac_size;
3069 
3070 		/* if we went over max copybuf size */
3071 		if (dma->dp_copybuf_size &&
3072 		    (copybuf_used > dma->dp_copybuf_size)) {
3073 			partial = B_TRUE;
3074 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3075 			    cookie, cur_offset, &copybuf_used);
3076 			if (e != DDI_SUCCESS) {
3077 				rootnex_teardown_copybuf(dma);
3078 				rootnex_teardown_windows(dma);
3079 				return (e);
3080 			}
3081 
3082 			/*
3083 			 * if the coookie uses the copy buffer, make sure the
3084 			 * new window we just moved to is set to sync.
3085 			 */
3086 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3087 				window->wd_dosync = B_TRUE;
3088 			}
3089 			DTRACE_PROBE1(rootnex__copybuf__window, dev_info_t *,
3090 			    dma->dp_dip);
3091 
3092 		/* if the cookie cnt == max sgllen, move to the next window */
3093 		} else if (window->wd_cookie_cnt >= attr->dma_attr_sgllen) {
3094 			partial = B_TRUE;
3095 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3096 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3097 			    cookie, attr, cur_offset);
3098 			if (e != DDI_SUCCESS) {
3099 				rootnex_teardown_copybuf(dma);
3100 				rootnex_teardown_windows(dma);
3101 				return (e);
3102 			}
3103 
3104 			/*
3105 			 * if the coookie uses the copy buffer, make sure the
3106 			 * new window we just moved to is set to sync.
3107 			 */
3108 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3109 				window->wd_dosync = B_TRUE;
3110 			}
3111 			DTRACE_PROBE1(rootnex__sgllen__window, dev_info_t *,
3112 			    dma->dp_dip);
3113 
3114 		/* else if we will be over maxxfer */
3115 		} else if ((window->wd_size + dmac_size) >
3116 		    dma->dp_maxxfer) {
3117 			partial = B_TRUE;
3118 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3119 			    cookie);
3120 			if (e != DDI_SUCCESS) {
3121 				rootnex_teardown_copybuf(dma);
3122 				rootnex_teardown_windows(dma);
3123 				return (e);
3124 			}
3125 
3126 			/*
3127 			 * if the coookie uses the copy buffer, make sure the
3128 			 * new window we just moved to is set to sync.
3129 			 */
3130 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3131 				window->wd_dosync = B_TRUE;
3132 			}
3133 			DTRACE_PROBE1(rootnex__maxxfer__window, dev_info_t *,
3134 			    dma->dp_dip);
3135 
3136 		/* else this cookie fits in the current window */
3137 		} else {
3138 			window->wd_cookie_cnt++;
3139 			window->wd_size += dmac_size;
3140 		}
3141 
3142 		/* track our offset into the buffer, go to the next cookie */
3143 		ASSERT(dmac_size <= dma->dp_dma.dmao_size);
3144 		ASSERT(cookie->dmac_size <= dmac_size);
3145 		cur_offset += dmac_size;
3146 		cookie++;
3147 	}
3148 
3149 	/* if we ended up with a zero sized window in the end, clean it up */
3150 	if (window->wd_size == 0) {
3151 		hp->dmai_nwin--;
3152 		window--;
3153 	}
3154 
3155 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3156 
3157 	if (!partial) {
3158 		return (DDI_DMA_MAPPED);
3159 	}
3160 
3161 	ASSERT(dma->dp_partial_required);
3162 	return (DDI_DMA_PARTIAL_MAP);
3163 }
3164 
3165 
3166 /*
3167  * rootnex_setup_copybuf()
3168  *    Called in bind slowpath. Figures out if we're going to use the copy
3169  *    buffer, and if we do, sets up the basic state to handle it.
3170  */
3171 static int
3172 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3173     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3174 {
3175 	rootnex_sglinfo_t *sinfo;
3176 	ddi_dma_attr_t lattr;
3177 	size_t max_copybuf;
3178 	int cansleep;
3179 	int e;
3180 #if !defined(__amd64)
3181 	int vmflag;
3182 #endif
3183 
3184 
3185 	sinfo = &dma->dp_sglinfo;
3186 
3187 	/* read this first so it's consistent through the routine  */
3188 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3189 
3190 	/* We need to call into the rootnex on ddi_dma_sync() */
3191 	hp->dmai_rflags &= ~DMP_NOSYNC;
3192 
3193 	/* make sure the copybuf size <= the max size */
3194 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3195 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3196 
3197 #if !defined(__amd64)
3198 	/*
3199 	 * if we don't have kva space to copy to/from, allocate the KVA space
3200 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3201 	 * the 64-bit kernel.
3202 	 */
3203 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3204 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3205 
3206 		/* convert the sleep flags */
3207 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3208 			vmflag = VM_SLEEP;
3209 		} else {
3210 			vmflag = VM_NOSLEEP;
3211 		}
3212 
3213 		/* allocate Kernel VA space that we can bcopy to/from */
3214 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3215 		    vmflag);
3216 		if (dma->dp_kva == NULL) {
3217 			return (DDI_DMA_NORESOURCES);
3218 		}
3219 	}
3220 #endif
3221 
3222 	/* convert the sleep flags */
3223 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3224 		cansleep = 1;
3225 	} else {
3226 		cansleep = 0;
3227 	}
3228 
3229 	/*
3230 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3231 	 * engine limits, so we can't use kmem_alloc... We don't need
3232 	 * contiguous memory (sgllen) since we will be forcing windows on
3233 	 * sgllen anyway.
3234 	 */
3235 	lattr = *attr;
3236 	lattr.dma_attr_align = MMU_PAGESIZE;
3237 	/*
3238 	 * this should be < 0 to indicate no limit, but due to a bug in
3239 	 * the rootnex, we'll set it to the maximum positive int.
3240 	 */
3241 	lattr.dma_attr_sgllen = 0x7fffffff;
3242 	/*
3243 	 * if we're using the copy buffer because of seg, use that for our
3244 	 * upper address limit.
3245 	 */
3246 	if (sinfo->si_bounce_on_seg) {
3247 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3248 	}
3249 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3250 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3251 	if (e != DDI_SUCCESS) {
3252 #if !defined(__amd64)
3253 		if (dma->dp_kva != NULL) {
3254 			vmem_free(heap_arena, dma->dp_kva,
3255 			    dma->dp_copybuf_size);
3256 		}
3257 #endif
3258 		return (DDI_DMA_NORESOURCES);
3259 	}
3260 
3261 	DTRACE_PROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3262 	    size_t, dma->dp_copybuf_size);
3263 
3264 	return (DDI_SUCCESS);
3265 }
3266 
3267 
3268 /*
3269  * rootnex_setup_windows()
3270  *    Called in bind slowpath to setup the window state. We always have windows
3271  *    in the slowpath. Even if the window count = 1.
3272  */
3273 static int
3274 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3275     ddi_dma_attr_t *attr, int kmflag)
3276 {
3277 	rootnex_window_t *windowp;
3278 	rootnex_sglinfo_t *sinfo;
3279 	size_t copy_state_size;
3280 	size_t win_state_size;
3281 	size_t state_available;
3282 	size_t space_needed;
3283 	uint_t copybuf_win;
3284 	uint_t maxxfer_win;
3285 	size_t space_used;
3286 	uint_t sglwin;
3287 
3288 
3289 	sinfo = &dma->dp_sglinfo;
3290 
3291 	dma->dp_current_win = 0;
3292 	hp->dmai_nwin = 0;
3293 
3294 	/* If we don't need to do a partial, we only have one window */
3295 	if (!dma->dp_partial_required) {
3296 		dma->dp_max_win = 1;
3297 
3298 	/*
3299 	 * we need multiple windows, need to figure out the worse case number
3300 	 * of windows.
3301 	 */
3302 	} else {
3303 		/*
3304 		 * if we need windows because we need more copy buffer that
3305 		 * we allow, the worse case number of windows we could need
3306 		 * here would be (copybuf space required / copybuf space that
3307 		 * we have) plus one for remainder, and plus 2 to handle the
3308 		 * extra pages on the trim for the first and last pages of the
3309 		 * buffer (a page is the minimum window size so under the right
3310 		 * attr settings, you could have a window for each page).
3311 		 * The last page will only be hit here if the size is not a
3312 		 * multiple of the granularity (which theoretically shouldn't
3313 		 * be the case but never has been enforced, so we could have
3314 		 * broken things without it).
3315 		 */
3316 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3317 			ASSERT(dma->dp_copybuf_size > 0);
3318 			copybuf_win = (sinfo->si_copybuf_req /
3319 			    dma->dp_copybuf_size) + 1 + 2;
3320 		} else {
3321 			copybuf_win = 0;
3322 		}
3323 
3324 		/*
3325 		 * if we need windows because we have more cookies than the H/W
3326 		 * can handle, the number of windows we would need here would
3327 		 * be (cookie count / cookies count H/W supports minus 1[for
3328 		 * trim]) plus one for remainder.
3329 		 */
3330 		if (attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3331 			sglwin = (sinfo->si_sgl_size /
3332 			    (attr->dma_attr_sgllen - 1)) + 1;
3333 		} else {
3334 			sglwin = 0;
3335 		}
3336 
3337 		/*
3338 		 * if we need windows because we're binding more memory than the
3339 		 * H/W can transfer at once, the number of windows we would need
3340 		 * here would be (xfer count / max xfer H/W supports) plus one
3341 		 * for remainder, and plus 2 to handle the extra pages on the
3342 		 * trim (see above comment about trim)
3343 		 */
3344 		if (dma->dp_dma.dmao_size > dma->dp_maxxfer) {
3345 			maxxfer_win = (dma->dp_dma.dmao_size /
3346 			    dma->dp_maxxfer) + 1 + 2;
3347 		} else {
3348 			maxxfer_win = 0;
3349 		}
3350 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3351 		ASSERT(dma->dp_max_win > 0);
3352 	}
3353 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3354 
3355 	/*
3356 	 * Get space for window and potential copy buffer state. Before we
3357 	 * go and allocate memory, see if we can get away with using what's
3358 	 * left in the pre-allocted state or the dynamically allocated sgl.
3359 	 */
3360 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3361 	    sizeof (ddi_dma_cookie_t));
3362 
3363 	/* if we dynamically allocated space for the cookies */
3364 	if (dma->dp_need_to_free_cookie) {
3365 		/* if we have more space in the pre-allocted buffer, use it */
3366 		ASSERT(space_used <= dma->dp_cookie_size);
3367 		if ((dma->dp_cookie_size - space_used) <=
3368 		    rootnex_state->r_prealloc_size) {
3369 			state_available = rootnex_state->r_prealloc_size;
3370 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3371 
3372 		/*
3373 		 * else, we have more free space in the dynamically allocated
3374 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3375 		 * didn't need a lot of cookies.
3376 		 */
3377 		} else {
3378 			state_available = dma->dp_cookie_size - space_used;
3379 			windowp = (rootnex_window_t *)
3380 			    &dma->dp_cookies[sinfo->si_sgl_size];
3381 		}
3382 
3383 	/* we used the pre-alloced buffer */
3384 	} else {
3385 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3386 		state_available = rootnex_state->r_prealloc_size - space_used;
3387 		windowp = (rootnex_window_t *)
3388 		    &dma->dp_cookies[sinfo->si_sgl_size];
3389 	}
3390 
3391 	/*
3392 	 * figure out how much state we need to track the copy buffer. Add an
3393 	 * addition 8 bytes for pointer alignemnt later.
3394 	 */
3395 	if (dma->dp_copybuf_size > 0) {
3396 		copy_state_size = sinfo->si_max_pages *
3397 		    sizeof (rootnex_pgmap_t);
3398 	} else {
3399 		copy_state_size = 0;
3400 	}
3401 	/* add an additional 8 bytes for pointer alignment */
3402 	space_needed = win_state_size + copy_state_size + 0x8;
3403 
3404 	/* if we have enough space already, use it */
3405 	if (state_available >= space_needed) {
3406 		dma->dp_window = windowp;
3407 		dma->dp_need_to_free_window = B_FALSE;
3408 
3409 	/* not enough space, need to allocate more. */
3410 	} else {
3411 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3412 		if (dma->dp_window == NULL) {
3413 			return (DDI_DMA_NORESOURCES);
3414 		}
3415 		dma->dp_need_to_free_window = B_TRUE;
3416 		dma->dp_window_size = space_needed;
3417 		DTRACE_PROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3418 		    dma->dp_dip, size_t, space_needed);
3419 	}
3420 
3421 	/*
3422 	 * we allocate copy buffer state and window state at the same time.
3423 	 * setup our copy buffer state pointers. Make sure it's aligned.
3424 	 */
3425 	if (dma->dp_copybuf_size > 0) {
3426 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3427 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3428 
3429 #if !defined(__amd64)
3430 		/*
3431 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3432 		 * false/NULL. Should be quicker to bzero vs loop and set.
3433 		 */
3434 		bzero(dma->dp_pgmap, copy_state_size);
3435 #endif
3436 	} else {
3437 		dma->dp_pgmap = NULL;
3438 	}
3439 
3440 	return (DDI_SUCCESS);
3441 }
3442 
3443 
3444 /*
3445  * rootnex_teardown_copybuf()
3446  *    cleans up after rootnex_setup_copybuf()
3447  */
3448 static void
3449 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3450 {
3451 #if !defined(__amd64)
3452 	int i;
3453 
3454 	/*
3455 	 * if we allocated kernel heap VMEM space, go through all the pages and
3456 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3457 	 * arena. Then free the VMEM space.
3458 	 */
3459 	if (dma->dp_kva != NULL) {
3460 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3461 			if (dma->dp_pgmap[i].pm_mapped) {
3462 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3463 				    MMU_PAGESIZE, HAT_UNLOAD);
3464 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3465 			}
3466 		}
3467 
3468 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3469 	}
3470 
3471 #endif
3472 
3473 	/* if we allocated a copy buffer, free it */
3474 	if (dma->dp_cbaddr != NULL) {
3475 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3476 	}
3477 }
3478 
3479 
3480 /*
3481  * rootnex_teardown_windows()
3482  *    cleans up after rootnex_setup_windows()
3483  */
3484 static void
3485 rootnex_teardown_windows(rootnex_dma_t *dma)
3486 {
3487 	/*
3488 	 * if we had to allocate window state on the last bind (because we
3489 	 * didn't have enough pre-allocated space in the handle), free it.
3490 	 */
3491 	if (dma->dp_need_to_free_window) {
3492 		kmem_free(dma->dp_window, dma->dp_window_size);
3493 	}
3494 }
3495 
3496 
3497 /*
3498  * rootnex_init_win()
3499  *    Called in bind slow path during creation of a new window. Initializes
3500  *    window state to default values.
3501  */
3502 /*ARGSUSED*/
3503 static void
3504 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3505     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3506 {
3507 	hp->dmai_nwin++;
3508 	window->wd_dosync = B_FALSE;
3509 	window->wd_offset = cur_offset;
3510 	window->wd_size = 0;
3511 	window->wd_first_cookie = cookie;
3512 	window->wd_cookie_cnt = 0;
3513 	window->wd_trim.tr_trim_first = B_FALSE;
3514 	window->wd_trim.tr_trim_last = B_FALSE;
3515 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3516 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3517 #if !defined(__amd64)
3518 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3519 #endif
3520 }
3521 
3522 
3523 /*
3524  * rootnex_setup_cookie()
3525  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3526  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3527  *    out if it uses the copy buffer, and if it does, save away everything we'll
3528  *    need during sync.
3529  */
3530 static void
3531 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3532     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3533     page_t **cur_pp)
3534 {
3535 	boolean_t copybuf_sz_power_2;
3536 	rootnex_sglinfo_t *sinfo;
3537 	paddr_t paddr;
3538 	uint_t pidx;
3539 	uint_t pcnt;
3540 	off_t poff;
3541 #if defined(__amd64)
3542 	pfn_t pfn;
3543 #else
3544 	page_t **pplist;
3545 #endif
3546 
3547 	sinfo = &dma->dp_sglinfo;
3548 
3549 	/*
3550 	 * Calculate the page index relative to the start of the buffer. The
3551 	 * index to the current page for our buffer is the offset into the
3552 	 * first page of the buffer plus our current offset into the buffer
3553 	 * itself, shifted of course...
3554 	 */
3555 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3556 	ASSERT(pidx < sinfo->si_max_pages);
3557 
3558 	/* if this cookie uses the copy buffer */
3559 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3560 		/*
3561 		 * NOTE: we know that since this cookie uses the copy buffer, it
3562 		 * is <= MMU_PAGESIZE.
3563 		 */
3564 
3565 		/*
3566 		 * get the offset into the page. For the 64-bit kernel, get the
3567 		 * pfn which we'll use with seg kpm.
3568 		 */
3569 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3570 #if defined(__amd64)
3571 		/* mfn_to_pfn() is a NOP on i86pc */
3572 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3573 #endif /* __amd64 */
3574 
3575 		/* figure out if the copybuf size is a power of 2 */
3576 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3577 			copybuf_sz_power_2 = B_FALSE;
3578 		} else {
3579 			copybuf_sz_power_2 = B_TRUE;
3580 		}
3581 
3582 		/* This page uses the copy buffer */
3583 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3584 
3585 		/*
3586 		 * save the copy buffer KVA that we'll use with this page.
3587 		 * if we still fit within the copybuf, it's a simple add.
3588 		 * otherwise, we need to wrap over using & or % accordingly.
3589 		 */
3590 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3591 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3592 			    *copybuf_used;
3593 		} else {
3594 			if (copybuf_sz_power_2) {
3595 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3596 				    (uintptr_t)dma->dp_cbaddr +
3597 				    (*copybuf_used &
3598 				    (dma->dp_copybuf_size - 1)));
3599 			} else {
3600 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3601 				    (uintptr_t)dma->dp_cbaddr +
3602 				    (*copybuf_used % dma->dp_copybuf_size));
3603 			}
3604 		}
3605 
3606 		/*
3607 		 * over write the cookie physical address with the address of
3608 		 * the physical address of the copy buffer page that we will
3609 		 * use.
3610 		 */
3611 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3612 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3613 
3614 #ifdef __xpv
3615 		/*
3616 		 * If we're dom0, we're using a real device so we need to load
3617 		 * the cookies with MAs instead of PAs.
3618 		 */
3619 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3620 #else
3621 		cookie->dmac_laddress = paddr;
3622 #endif
3623 
3624 		/* if we have a kernel VA, it's easy, just save that address */
3625 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3626 		    (sinfo->si_asp == &kas)) {
3627 			/*
3628 			 * save away the page aligned virtual address of the
3629 			 * driver buffer. Offsets are handled in the sync code.
3630 			 */
3631 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3632 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3633 			    & MMU_PAGEMASK);
3634 #if !defined(__amd64)
3635 			/*
3636 			 * we didn't need to, and will never need to map this
3637 			 * page.
3638 			 */
3639 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3640 #endif
3641 
3642 		/* we don't have a kernel VA. We need one for the bcopy. */
3643 		} else {
3644 #if defined(__amd64)
3645 			/*
3646 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3647 			 * get a Kernel VA for the corresponding pfn.
3648 			 */
3649 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3650 #else
3651 			/*
3652 			 * for the 32-bit kernel, this is a pain. First we'll
3653 			 * save away the page_t or user VA for this page. This
3654 			 * is needed in rootnex_dma_win() when we switch to a
3655 			 * new window which requires us to re-map the copy
3656 			 * buffer.
3657 			 */
3658 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3659 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3660 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3661 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3662 			} else if (pplist != NULL) {
3663 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3664 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3665 			} else {
3666 				dma->dp_pgmap[pidx].pm_pp = NULL;
3667 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3668 				    (((uintptr_t)
3669 				    dmar_object->dmao_obj.virt_obj.v_addr +
3670 				    cur_offset) & MMU_PAGEMASK);
3671 			}
3672 
3673 			/*
3674 			 * save away the page aligned virtual address which was
3675 			 * allocated from the kernel heap arena (taking into
3676 			 * account if we need more copy buffer than we alloced
3677 			 * and use multiple windows to handle this, i.e. &,%).
3678 			 * NOTE: there isn't and physical memory backing up this
3679 			 * virtual address space currently.
3680 			 */
3681 			if ((*copybuf_used + MMU_PAGESIZE) <=
3682 			    dma->dp_copybuf_size) {
3683 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3684 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3685 				    MMU_PAGEMASK);
3686 			} else {
3687 				if (copybuf_sz_power_2) {
3688 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3689 					    (((uintptr_t)dma->dp_kva +
3690 					    (*copybuf_used &
3691 					    (dma->dp_copybuf_size - 1))) &
3692 					    MMU_PAGEMASK);
3693 				} else {
3694 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3695 					    (((uintptr_t)dma->dp_kva +
3696 					    (*copybuf_used %
3697 					    dma->dp_copybuf_size)) &
3698 					    MMU_PAGEMASK);
3699 				}
3700 			}
3701 
3702 			/*
3703 			 * if we haven't used up the available copy buffer yet,
3704 			 * map the kva to the physical page.
3705 			 */
3706 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3707 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3708 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
3709 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
3710 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
3711 					    dma->dp_pgmap[pidx].pm_kaddr);
3712 				} else {
3713 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
3714 					    sinfo->si_asp,
3715 					    dma->dp_pgmap[pidx].pm_kaddr);
3716 				}
3717 
3718 			/*
3719 			 * we've used up the available copy buffer, this page
3720 			 * will have to be mapped during rootnex_dma_win() when
3721 			 * we switch to a new window which requires a re-map
3722 			 * the copy buffer. (32-bit kernel only)
3723 			 */
3724 			} else {
3725 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3726 			}
3727 #endif
3728 			/* go to the next page_t */
3729 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3730 				*cur_pp = (*cur_pp)->p_next;
3731 			}
3732 		}
3733 
3734 		/* add to the copy buffer count */
3735 		*copybuf_used += MMU_PAGESIZE;
3736 
3737 	/*
3738 	 * This cookie doesn't use the copy buffer. Walk through the pages this
3739 	 * cookie occupies to reflect this.
3740 	 */
3741 	} else {
3742 		/*
3743 		 * figure out how many pages the cookie occupies. We need to
3744 		 * use the original page offset of the buffer and the cookies
3745 		 * offset in the buffer to do this.
3746 		 */
3747 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
3748 		pcnt = mmu_btopr(cookie->dmac_size + poff);
3749 
3750 		while (pcnt > 0) {
3751 #if !defined(__amd64)
3752 			/*
3753 			 * the 32-bit kernel doesn't have seg kpm, so we need
3754 			 * to map in the driver buffer (if it didn't come down
3755 			 * with a kernel VA) on the fly. Since this page doesn't
3756 			 * use the copy buffer, it's not, or will it ever, have
3757 			 * to be mapped in.
3758 			 */
3759 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3760 #endif
3761 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
3762 
3763 			/*
3764 			 * we need to update pidx and cur_pp or we'll loose
3765 			 * track of where we are.
3766 			 */
3767 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3768 				*cur_pp = (*cur_pp)->p_next;
3769 			}
3770 			pidx++;
3771 			pcnt--;
3772 		}
3773 	}
3774 }
3775 
3776 
3777 /*
3778  * rootnex_sgllen_window_boundary()
3779  *    Called in the bind slow path when the next cookie causes us to exceed (in
3780  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
3781  *    length supported by the DMA H/W.
3782  */
3783 static int
3784 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3785     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
3786     off_t cur_offset)
3787 {
3788 	off_t new_offset;
3789 	size_t trim_sz;
3790 	off_t coffset;
3791 
3792 
3793 	/*
3794 	 * if we know we'll never have to trim, it's pretty easy. Just move to
3795 	 * the next window and init it. We're done.
3796 	 */
3797 	if (!dma->dp_trim_required) {
3798 		(*windowp)++;
3799 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3800 		(*windowp)->wd_cookie_cnt++;
3801 		(*windowp)->wd_size = cookie->dmac_size;
3802 		return (DDI_SUCCESS);
3803 	}
3804 
3805 	/* figure out how much we need to trim from the window */
3806 	ASSERT(attr->dma_attr_granular != 0);
3807 	if (dma->dp_granularity_power_2) {
3808 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
3809 	} else {
3810 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
3811 	}
3812 
3813 	/* The window's a whole multiple of granularity. We're done */
3814 	if (trim_sz == 0) {
3815 		(*windowp)++;
3816 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3817 		(*windowp)->wd_cookie_cnt++;
3818 		(*windowp)->wd_size = cookie->dmac_size;
3819 		return (DDI_SUCCESS);
3820 	}
3821 
3822 	/*
3823 	 * The window's not a whole multiple of granularity, since we know this
3824 	 * is due to the sgllen, we need to go back to the last cookie and trim
3825 	 * that one, add the left over part of the old cookie into the new
3826 	 * window, and then add in the new cookie into the new window.
3827 	 */
3828 
3829 	/*
3830 	 * make sure the driver isn't making us do something bad... Trimming and
3831 	 * sgllen == 1 don't go together.
3832 	 */
3833 	if (attr->dma_attr_sgllen == 1) {
3834 		return (DDI_DMA_NOMAPPING);
3835 	}
3836 
3837 	/*
3838 	 * first, setup the current window to account for the trim. Need to go
3839 	 * back to the last cookie for this.
3840 	 */
3841 	cookie--;
3842 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3843 	(*windowp)->wd_trim.tr_last_cookie = cookie;
3844 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3845 	ASSERT(cookie->dmac_size > trim_sz);
3846 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3847 	(*windowp)->wd_size -= trim_sz;
3848 
3849 	/* save the buffer offsets for the next window */
3850 	coffset = cookie->dmac_size - trim_sz;
3851 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3852 
3853 	/*
3854 	 * set this now in case this is the first window. all other cases are
3855 	 * set in dma_win()
3856 	 */
3857 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
3858 
3859 	/*
3860 	 * initialize the next window using what's left over in the previous
3861 	 * cookie.
3862 	 */
3863 	(*windowp)++;
3864 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3865 	(*windowp)->wd_cookie_cnt++;
3866 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3867 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
3868 	(*windowp)->wd_trim.tr_first_size = trim_sz;
3869 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3870 		(*windowp)->wd_dosync = B_TRUE;
3871 	}
3872 
3873 	/*
3874 	 * now go back to the current cookie and add it to the new window. set
3875 	 * the new window size to the what was left over from the previous
3876 	 * cookie and what's in the current cookie.
3877 	 */
3878 	cookie++;
3879 	(*windowp)->wd_cookie_cnt++;
3880 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
3881 
3882 	/*
3883 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
3884 	 * a max size of maxxfer). Handle that case.
3885 	 */
3886 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
3887 		/*
3888 		 * maxxfer is already a whole multiple of granularity, and this
3889 		 * trim will be <= the previous trim (since a cookie can't be
3890 		 * larger than maxxfer). Make things simple here.
3891 		 */
3892 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
3893 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3894 		(*windowp)->wd_trim.tr_last_cookie = cookie;
3895 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3896 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3897 		(*windowp)->wd_size -= trim_sz;
3898 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
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 		/* setup the next window */
3905 		(*windowp)++;
3906 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3907 		(*windowp)->wd_cookie_cnt++;
3908 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3909 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
3910 		    coffset;
3911 		(*windowp)->wd_trim.tr_first_size = trim_sz;
3912 	}
3913 
3914 	return (DDI_SUCCESS);
3915 }
3916 
3917 
3918 /*
3919  * rootnex_copybuf_window_boundary()
3920  *    Called in bind slowpath when we get to a window boundary because we used
3921  *    up all the copy buffer that we have.
3922  */
3923 static int
3924 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3925     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
3926     size_t *copybuf_used)
3927 {
3928 	rootnex_sglinfo_t *sinfo;
3929 	off_t new_offset;
3930 	size_t trim_sz;
3931 	paddr_t paddr;
3932 	off_t coffset;
3933 	uint_t pidx;
3934 	off_t poff;
3935 
3936 
3937 	sinfo = &dma->dp_sglinfo;
3938 
3939 	/*
3940 	 * the copy buffer should be a whole multiple of page size. We know that
3941 	 * this cookie is <= MMU_PAGESIZE.
3942 	 */
3943 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
3944 
3945 	/*
3946 	 * from now on, all new windows in this bind need to be re-mapped during
3947 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
3948 	 * space...
3949 	 */
3950 #if !defined(__amd64)
3951 	dma->dp_cb_remaping = B_TRUE;
3952 #endif
3953 
3954 	/* reset copybuf used */
3955 	*copybuf_used = 0;
3956 
3957 	/*
3958 	 * if we don't have to trim (since granularity is set to 1), go to the
3959 	 * next window and add the current cookie to it. We know the current
3960 	 * cookie uses the copy buffer since we're in this code path.
3961 	 */
3962 	if (!dma->dp_trim_required) {
3963 		(*windowp)++;
3964 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3965 
3966 		/* Add this cookie to the new window */
3967 		(*windowp)->wd_cookie_cnt++;
3968 		(*windowp)->wd_size += cookie->dmac_size;
3969 		*copybuf_used += MMU_PAGESIZE;
3970 		return (DDI_SUCCESS);
3971 	}
3972 
3973 	/*
3974 	 * *** may need to trim, figure it out.
3975 	 */
3976 
3977 	/* figure out how much we need to trim from the window */
3978 	if (dma->dp_granularity_power_2) {
3979 		trim_sz = (*windowp)->wd_size &
3980 		    (hp->dmai_attr.dma_attr_granular - 1);
3981 	} else {
3982 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
3983 	}
3984 
3985 	/*
3986 	 * if the window's a whole multiple of granularity, go to the next
3987 	 * window, init it, then add in the current cookie. We know the current
3988 	 * cookie uses the copy buffer since we're in this code path.
3989 	 */
3990 	if (trim_sz == 0) {
3991 		(*windowp)++;
3992 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3993 
3994 		/* Add this cookie to the new window */
3995 		(*windowp)->wd_cookie_cnt++;
3996 		(*windowp)->wd_size += cookie->dmac_size;
3997 		*copybuf_used += MMU_PAGESIZE;
3998 		return (DDI_SUCCESS);
3999 	}
4000 
4001 	/*
4002 	 * *** We figured it out, we definitly need to trim
4003 	 */
4004 
4005 	/*
4006 	 * make sure the driver isn't making us do something bad...
4007 	 * Trimming and sgllen == 1 don't go together.
4008 	 */
4009 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4010 		return (DDI_DMA_NOMAPPING);
4011 	}
4012 
4013 	/*
4014 	 * first, setup the current window to account for the trim. Need to go
4015 	 * back to the last cookie for this. Some of the last cookie will be in
4016 	 * the current window, and some of the last cookie will be in the new
4017 	 * window. All of the current cookie will be in the new window.
4018 	 */
4019 	cookie--;
4020 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4021 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4022 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4023 	ASSERT(cookie->dmac_size > trim_sz);
4024 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4025 	(*windowp)->wd_size -= trim_sz;
4026 
4027 	/*
4028 	 * we're trimming the last cookie (not the current cookie). So that
4029 	 * last cookie may have or may not have been using the copy buffer (
4030 	 * we know the cookie passed in uses the copy buffer since we're in
4031 	 * this code path).
4032 	 *
4033 	 * If the last cookie doesn't use the copy buffer, nothing special to
4034 	 * do. However, if it does uses the copy buffer, it will be both the
4035 	 * last page in the current window and the first page in the next
4036 	 * window. Since we are reusing the copy buffer (and KVA space on the
4037 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4038 	 * current window, and the start of the copy buffer in the next window.
4039 	 * Track that info... The cookie physical address was already set to
4040 	 * the copy buffer physical address in setup_cookie..
4041 	 */
4042 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4043 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4044 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4045 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4046 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4047 		(*windowp)->wd_trim.tr_last_cbaddr =
4048 		    dma->dp_pgmap[pidx].pm_cbaddr;
4049 #if !defined(__amd64)
4050 		(*windowp)->wd_trim.tr_last_kaddr =
4051 		    dma->dp_pgmap[pidx].pm_kaddr;
4052 #endif
4053 	}
4054 
4055 	/* save the buffer offsets for the next window */
4056 	coffset = cookie->dmac_size - trim_sz;
4057 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4058 
4059 	/*
4060 	 * set this now in case this is the first window. all other cases are
4061 	 * set in dma_win()
4062 	 */
4063 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4064 
4065 	/*
4066 	 * initialize the next window using what's left over in the previous
4067 	 * cookie.
4068 	 */
4069 	(*windowp)++;
4070 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4071 	(*windowp)->wd_cookie_cnt++;
4072 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4073 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4074 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4075 
4076 	/*
4077 	 * again, we're tracking if the last cookie uses the copy buffer.
4078 	 * read the comment above for more info on why we need to track
4079 	 * additional state.
4080 	 *
4081 	 * For the first cookie in the new window, we need reset the physical
4082 	 * address to DMA into to the start of the copy buffer plus any
4083 	 * initial page offset which may be present.
4084 	 */
4085 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4086 		(*windowp)->wd_dosync = B_TRUE;
4087 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4088 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4089 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4090 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4091 
4092 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4093 		    poff;
4094 #ifdef __xpv
4095 		/*
4096 		 * If we're dom0, we're using a real device so we need to load
4097 		 * the cookies with MAs instead of PAs.
4098 		 */
4099 		(*windowp)->wd_trim.tr_first_paddr =
4100 		    ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4101 #else
4102 		(*windowp)->wd_trim.tr_first_paddr = paddr;
4103 #endif
4104 
4105 #if !defined(__amd64)
4106 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4107 #endif
4108 		/* account for the cookie copybuf usage in the new window */
4109 		*copybuf_used += MMU_PAGESIZE;
4110 
4111 		/*
4112 		 * every piece of code has to have a hack, and here is this
4113 		 * ones :-)
4114 		 *
4115 		 * There is a complex interaction between setup_cookie and the
4116 		 * copybuf window boundary. The complexity had to be in either
4117 		 * the maxxfer window, or the copybuf window, and I chose the
4118 		 * copybuf code.
4119 		 *
4120 		 * So in this code path, we have taken the last cookie,
4121 		 * virtually broken it in half due to the trim, and it happens
4122 		 * to use the copybuf which further complicates life. At the
4123 		 * same time, we have already setup the current cookie, which
4124 		 * is now wrong. More background info: the current cookie uses
4125 		 * the copybuf, so it is only a page long max. So we need to
4126 		 * fix the current cookies copy buffer address, physical
4127 		 * address, and kva for the 32-bit kernel. We due this by
4128 		 * bumping them by page size (of course, we can't due this on
4129 		 * the physical address since the copy buffer may not be
4130 		 * physically contiguous).
4131 		 */
4132 		cookie++;
4133 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4134 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4135 
4136 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4137 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4138 #ifdef __xpv
4139 		/*
4140 		 * If we're dom0, we're using a real device so we need to load
4141 		 * the cookies with MAs instead of PAs.
4142 		 */
4143 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4144 #else
4145 		cookie->dmac_laddress = paddr;
4146 #endif
4147 
4148 #if !defined(__amd64)
4149 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4150 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4151 #endif
4152 	} else {
4153 		/* go back to the current cookie */
4154 		cookie++;
4155 	}
4156 
4157 	/*
4158 	 * add the current cookie to the new window. set the new window size to
4159 	 * the what was left over from the previous cookie and what's in the
4160 	 * current cookie.
4161 	 */
4162 	(*windowp)->wd_cookie_cnt++;
4163 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4164 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4165 
4166 	/*
4167 	 * we know that the cookie passed in always uses the copy buffer. We
4168 	 * wouldn't be here if it didn't.
4169 	 */
4170 	*copybuf_used += MMU_PAGESIZE;
4171 
4172 	return (DDI_SUCCESS);
4173 }
4174 
4175 
4176 /*
4177  * rootnex_maxxfer_window_boundary()
4178  *    Called in bind slowpath when we get to a window boundary because we will
4179  *    go over maxxfer.
4180  */
4181 static int
4182 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4183     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4184 {
4185 	size_t dmac_size;
4186 	off_t new_offset;
4187 	size_t trim_sz;
4188 	off_t coffset;
4189 
4190 
4191 	/*
4192 	 * calculate how much we have to trim off of the current cookie to equal
4193 	 * maxxfer. We don't have to account for granularity here since our
4194 	 * maxxfer already takes that into account.
4195 	 */
4196 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4197 	ASSERT(trim_sz <= cookie->dmac_size);
4198 	ASSERT(trim_sz <= dma->dp_maxxfer);
4199 
4200 	/* save cookie size since we need it later and we might change it */
4201 	dmac_size = cookie->dmac_size;
4202 
4203 	/*
4204 	 * if we're not trimming the entire cookie, setup the current window to
4205 	 * account for the trim.
4206 	 */
4207 	if (trim_sz < cookie->dmac_size) {
4208 		(*windowp)->wd_cookie_cnt++;
4209 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4210 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4211 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4212 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4213 		(*windowp)->wd_size = dma->dp_maxxfer;
4214 
4215 		/*
4216 		 * set the adjusted cookie size now in case this is the first
4217 		 * window. All other windows are taken care of in get win
4218 		 */
4219 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4220 	}
4221 
4222 	/*
4223 	 * coffset is the current offset within the cookie, new_offset is the
4224 	 * current offset with the entire buffer.
4225 	 */
4226 	coffset = dmac_size - trim_sz;
4227 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4228 
4229 	/* initialize the next window */
4230 	(*windowp)++;
4231 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4232 	(*windowp)->wd_cookie_cnt++;
4233 	(*windowp)->wd_size = trim_sz;
4234 	if (trim_sz < dmac_size) {
4235 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4236 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4237 		    coffset;
4238 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4239 	}
4240 
4241 	return (DDI_SUCCESS);
4242 }
4243 
4244 
4245 /*ARGSUSED*/
4246 static int
4247 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4248     off_t off, size_t len, uint_t cache_flags)
4249 {
4250 	rootnex_sglinfo_t *sinfo;
4251 	rootnex_pgmap_t *cbpage;
4252 	rootnex_window_t *win;
4253 	ddi_dma_impl_t *hp;
4254 	rootnex_dma_t *dma;
4255 	caddr_t fromaddr;
4256 	caddr_t toaddr;
4257 	uint_t psize;
4258 	off_t offset;
4259 	uint_t pidx;
4260 	size_t size;
4261 	off_t poff;
4262 	int e;
4263 
4264 
4265 	hp = (ddi_dma_impl_t *)handle;
4266 	dma = (rootnex_dma_t *)hp->dmai_private;
4267 	sinfo = &dma->dp_sglinfo;
4268 
4269 	/*
4270 	 * if we don't have any windows, we don't need to sync. A copybuf
4271 	 * will cause us to have at least one window.
4272 	 */
4273 	if (dma->dp_window == NULL) {
4274 		return (DDI_SUCCESS);
4275 	}
4276 
4277 	/* This window may not need to be sync'd */
4278 	win = &dma->dp_window[dma->dp_current_win];
4279 	if (!win->wd_dosync) {
4280 		return (DDI_SUCCESS);
4281 	}
4282 
4283 	/* handle off and len special cases */
4284 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4285 		offset = win->wd_offset;
4286 	} else {
4287 		offset = off;
4288 	}
4289 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4290 		size = win->wd_size;
4291 	} else {
4292 		size = len;
4293 	}
4294 
4295 	/* check the sync args to make sure they make a little sense */
4296 	if (rootnex_sync_check_parms) {
4297 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4298 		    cache_flags);
4299 		if (e != DDI_SUCCESS) {
4300 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4301 			return (DDI_FAILURE);
4302 		}
4303 	}
4304 
4305 	/*
4306 	 * special case the first page to handle the offset into the page. The
4307 	 * offset to the current page for our buffer is the offset into the
4308 	 * first page of the buffer plus our current offset into the buffer
4309 	 * itself, masked of course.
4310 	 */
4311 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4312 	psize = MIN((MMU_PAGESIZE - poff), size);
4313 
4314 	/* go through all the pages that we want to sync */
4315 	while (size > 0) {
4316 		/*
4317 		 * Calculate the page index relative to the start of the buffer.
4318 		 * The index to the current page for our buffer is the offset
4319 		 * into the first page of the buffer plus our current offset
4320 		 * into the buffer itself, shifted of course...
4321 		 */
4322 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4323 		ASSERT(pidx < sinfo->si_max_pages);
4324 
4325 		/*
4326 		 * if this page uses the copy buffer, we need to sync it,
4327 		 * otherwise, go on to the next page.
4328 		 */
4329 		cbpage = &dma->dp_pgmap[pidx];
4330 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4331 		    (cbpage->pm_uses_copybuf == B_FALSE));
4332 		if (cbpage->pm_uses_copybuf) {
4333 			/* cbaddr and kaddr should be page aligned */
4334 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4335 			    MMU_PAGEOFFSET) == 0);
4336 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4337 			    MMU_PAGEOFFSET) == 0);
4338 
4339 			/*
4340 			 * if we're copying for the device, we are going to
4341 			 * copy from the drivers buffer and to the rootnex
4342 			 * allocated copy buffer.
4343 			 */
4344 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4345 				fromaddr = cbpage->pm_kaddr + poff;
4346 				toaddr = cbpage->pm_cbaddr + poff;
4347 				DTRACE_PROBE2(rootnex__sync__dev,
4348 				    dev_info_t *, dma->dp_dip, size_t, psize);
4349 
4350 			/*
4351 			 * if we're copying for the cpu/kernel, we are going to
4352 			 * copy from the rootnex allocated copy buffer to the
4353 			 * drivers buffer.
4354 			 */
4355 			} else {
4356 				fromaddr = cbpage->pm_cbaddr + poff;
4357 				toaddr = cbpage->pm_kaddr + poff;
4358 				DTRACE_PROBE2(rootnex__sync__cpu,
4359 				    dev_info_t *, dma->dp_dip, size_t, psize);
4360 			}
4361 
4362 			bcopy(fromaddr, toaddr, psize);
4363 		}
4364 
4365 		/*
4366 		 * decrement size until we're done, update our offset into the
4367 		 * buffer, and get the next page size.
4368 		 */
4369 		size -= psize;
4370 		offset += psize;
4371 		psize = MIN(MMU_PAGESIZE, size);
4372 
4373 		/* page offset is zero for the rest of this loop */
4374 		poff = 0;
4375 	}
4376 
4377 	return (DDI_SUCCESS);
4378 }
4379 
4380 /*
4381  * rootnex_dma_sync()
4382  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4383  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4384  *    is set, ddi_dma_sync() returns immediately passing back success.
4385  */
4386 /*ARGSUSED*/
4387 static int
4388 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4389     off_t off, size_t len, uint_t cache_flags)
4390 {
4391 #if defined(__amd64) && !defined(__xpv)
4392 	if (IOMMU_USED(rdip)) {
4393 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4394 		    cache_flags));
4395 	}
4396 #endif
4397 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4398 	    cache_flags));
4399 }
4400 
4401 /*
4402  * rootnex_valid_sync_parms()
4403  *    checks the parameters passed to sync to verify they are correct.
4404  */
4405 static int
4406 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4407     off_t offset, size_t size, uint_t cache_flags)
4408 {
4409 	off_t woffset;
4410 
4411 
4412 	/*
4413 	 * the first part of the test to make sure the offset passed in is
4414 	 * within the window.
4415 	 */
4416 	if (offset < win->wd_offset) {
4417 		return (DDI_FAILURE);
4418 	}
4419 
4420 	/*
4421 	 * second and last part of the test to make sure the offset and length
4422 	 * passed in is within the window.
4423 	 */
4424 	woffset = offset - win->wd_offset;
4425 	if ((woffset + size) > win->wd_size) {
4426 		return (DDI_FAILURE);
4427 	}
4428 
4429 	/*
4430 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4431 	 * be set too.
4432 	 */
4433 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4434 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4435 		return (DDI_SUCCESS);
4436 	}
4437 
4438 	/*
4439 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4440 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4441 	 */
4442 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4443 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4444 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4445 		return (DDI_SUCCESS);
4446 	}
4447 
4448 	return (DDI_FAILURE);
4449 }
4450 
4451 
4452 /*ARGSUSED*/
4453 static int
4454 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4455     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4456     uint_t *ccountp)
4457 {
4458 	rootnex_window_t *window;
4459 	rootnex_trim_t *trim;
4460 	ddi_dma_impl_t *hp;
4461 	rootnex_dma_t *dma;
4462 #if !defined(__amd64)
4463 	rootnex_sglinfo_t *sinfo;
4464 	rootnex_pgmap_t *pmap;
4465 	uint_t pidx;
4466 	uint_t pcnt;
4467 	off_t poff;
4468 	int i;
4469 #endif
4470 
4471 
4472 	hp = (ddi_dma_impl_t *)handle;
4473 	dma = (rootnex_dma_t *)hp->dmai_private;
4474 #if !defined(__amd64)
4475 	sinfo = &dma->dp_sglinfo;
4476 #endif
4477 
4478 	/* If we try and get a window which doesn't exist, return failure */
4479 	if (win >= hp->dmai_nwin) {
4480 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4481 		return (DDI_FAILURE);
4482 	}
4483 
4484 	/*
4485 	 * if we don't have any windows, and they're asking for the first
4486 	 * window, setup the cookie pointer to the first cookie in the bind.
4487 	 * setup our return values, then increment the cookie since we return
4488 	 * the first cookie on the stack.
4489 	 */
4490 	if (dma->dp_window == NULL) {
4491 		if (win != 0) {
4492 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4493 			return (DDI_FAILURE);
4494 		}
4495 		hp->dmai_cookie = dma->dp_cookies;
4496 		*offp = 0;
4497 		*lenp = dma->dp_dma.dmao_size;
4498 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4499 		*cookiep = hp->dmai_cookie[0];
4500 		hp->dmai_cookie++;
4501 		return (DDI_SUCCESS);
4502 	}
4503 
4504 	/* sync the old window before moving on to the new one */
4505 	window = &dma->dp_window[dma->dp_current_win];
4506 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4507 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4508 		    DDI_DMA_SYNC_FORCPU);
4509 	}
4510 
4511 #if !defined(__amd64)
4512 	/*
4513 	 * before we move to the next window, if we need to re-map, unmap all
4514 	 * the pages in this window.
4515 	 */
4516 	if (dma->dp_cb_remaping) {
4517 		/*
4518 		 * If we switch to this window again, we'll need to map in
4519 		 * on the fly next time.
4520 		 */
4521 		window->wd_remap_copybuf = B_TRUE;
4522 
4523 		/*
4524 		 * calculate the page index into the buffer where this window
4525 		 * starts, and the number of pages this window takes up.
4526 		 */
4527 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4528 		    MMU_PAGESHIFT;
4529 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4530 		    MMU_PAGEOFFSET;
4531 		pcnt = mmu_btopr(window->wd_size + poff);
4532 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4533 
4534 		/* unmap pages which are currently mapped in this window */
4535 		for (i = 0; i < pcnt; i++) {
4536 			if (dma->dp_pgmap[pidx].pm_mapped) {
4537 				hat_unload(kas.a_hat,
4538 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4539 				    HAT_UNLOAD);
4540 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4541 			}
4542 			pidx++;
4543 		}
4544 	}
4545 #endif
4546 
4547 	/*
4548 	 * Move to the new window.
4549 	 * NOTE: current_win must be set for sync to work right
4550 	 */
4551 	dma->dp_current_win = win;
4552 	window = &dma->dp_window[win];
4553 
4554 	/* if needed, adjust the first and/or last cookies for trim */
4555 	trim = &window->wd_trim;
4556 	if (trim->tr_trim_first) {
4557 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4558 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4559 #if !defined(__amd64)
4560 		window->wd_first_cookie->dmac_type =
4561 		    (window->wd_first_cookie->dmac_type &
4562 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4563 #endif
4564 		if (trim->tr_first_copybuf_win) {
4565 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4566 			    trim->tr_first_cbaddr;
4567 #if !defined(__amd64)
4568 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4569 			    trim->tr_first_kaddr;
4570 #endif
4571 		}
4572 	}
4573 	if (trim->tr_trim_last) {
4574 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4575 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4576 		if (trim->tr_last_copybuf_win) {
4577 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4578 			    trim->tr_last_cbaddr;
4579 #if !defined(__amd64)
4580 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4581 			    trim->tr_last_kaddr;
4582 #endif
4583 		}
4584 	}
4585 
4586 	/*
4587 	 * setup the cookie pointer to the first cookie in the window. setup
4588 	 * our return values, then increment the cookie since we return the
4589 	 * first cookie on the stack.
4590 	 */
4591 	hp->dmai_cookie = window->wd_first_cookie;
4592 	*offp = window->wd_offset;
4593 	*lenp = window->wd_size;
4594 	*ccountp = window->wd_cookie_cnt;
4595 	*cookiep = hp->dmai_cookie[0];
4596 	hp->dmai_cookie++;
4597 
4598 #if !defined(__amd64)
4599 	/* re-map copybuf if required for this window */
4600 	if (dma->dp_cb_remaping) {
4601 		/*
4602 		 * calculate the page index into the buffer where this
4603 		 * window starts.
4604 		 */
4605 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4606 		    MMU_PAGESHIFT;
4607 		ASSERT(pidx < sinfo->si_max_pages);
4608 
4609 		/*
4610 		 * the first page can get unmapped if it's shared with the
4611 		 * previous window. Even if the rest of this window is already
4612 		 * mapped in, we need to still check this one.
4613 		 */
4614 		pmap = &dma->dp_pgmap[pidx];
4615 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4616 			if (pmap->pm_pp != NULL) {
4617 				pmap->pm_mapped = B_TRUE;
4618 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4619 			} else if (pmap->pm_vaddr != NULL) {
4620 				pmap->pm_mapped = B_TRUE;
4621 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4622 				    pmap->pm_kaddr);
4623 			}
4624 		}
4625 		pidx++;
4626 
4627 		/* map in the rest of the pages if required */
4628 		if (window->wd_remap_copybuf) {
4629 			window->wd_remap_copybuf = B_FALSE;
4630 
4631 			/* figure out many pages this window takes up */
4632 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4633 			    MMU_PAGEOFFSET;
4634 			pcnt = mmu_btopr(window->wd_size + poff);
4635 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4636 
4637 			/* map pages which require it */
4638 			for (i = 1; i < pcnt; i++) {
4639 				pmap = &dma->dp_pgmap[pidx];
4640 				if (pmap->pm_uses_copybuf) {
4641 					ASSERT(pmap->pm_mapped == B_FALSE);
4642 					if (pmap->pm_pp != NULL) {
4643 						pmap->pm_mapped = B_TRUE;
4644 						i86_pp_map(pmap->pm_pp,
4645 						    pmap->pm_kaddr);
4646 					} else if (pmap->pm_vaddr != NULL) {
4647 						pmap->pm_mapped = B_TRUE;
4648 						i86_va_map(pmap->pm_vaddr,
4649 						    sinfo->si_asp,
4650 						    pmap->pm_kaddr);
4651 					}
4652 				}
4653 				pidx++;
4654 			}
4655 		}
4656 	}
4657 #endif
4658 
4659 	/* if the new window uses the copy buffer, sync it for the device */
4660 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4661 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4662 		    DDI_DMA_SYNC_FORDEV);
4663 	}
4664 
4665 	return (DDI_SUCCESS);
4666 }
4667 
4668 /*
4669  * rootnex_dma_win()
4670  *    called from ddi_dma_getwin()
4671  */
4672 /*ARGSUSED*/
4673 static int
4674 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4675     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4676     uint_t *ccountp)
4677 {
4678 #if defined(__amd64) && !defined(__xpv)
4679 	if (IOMMU_USED(rdip)) {
4680 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4681 		    cookiep, ccountp));
4682 	}
4683 #endif
4684 
4685 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4686 	    cookiep, ccountp));
4687 }
4688 
4689 /*
4690  * ************************
4691  *  obsoleted dma routines
4692  * ************************
4693  */
4694 
4695 /*
4696  * rootnex_dma_map()
4697  *    called from ddi_dma_setup()
4698  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
4699  */
4700 /* ARGSUSED */
4701 static int
4702 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
4703     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
4704 {
4705 #if defined(__amd64)
4706 	/*
4707 	 * this interface is not supported in 64-bit x86 kernel. See comment in
4708 	 * rootnex_dma_mctl()
4709 	 */
4710 	return (DDI_DMA_NORESOURCES);
4711 
4712 #else /* 32-bit x86 kernel */
4713 	ddi_dma_handle_t *lhandlep;
4714 	ddi_dma_handle_t lhandle;
4715 	ddi_dma_cookie_t cookie;
4716 	ddi_dma_attr_t dma_attr;
4717 	ddi_dma_lim_t *dma_lim;
4718 	uint_t ccnt;
4719 	int e;
4720 
4721 
4722 	/*
4723 	 * if the driver is just testing to see if it's possible to do the bind,
4724 	 * we'll use local state. Otherwise, use the handle pointer passed in.
4725 	 */
4726 	if (handlep == NULL) {
4727 		lhandlep = &lhandle;
4728 	} else {
4729 		lhandlep = handlep;
4730 	}
4731 
4732 	/* convert the limit structure to a dma_attr one */
4733 	dma_lim = dmareq->dmar_limits;
4734 	dma_attr.dma_attr_version = DMA_ATTR_V0;
4735 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
4736 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
4737 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
4738 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
4739 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
4740 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
4741 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
4742 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
4743 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
4744 	dma_attr.dma_attr_align = MMU_PAGESIZE;
4745 	dma_attr.dma_attr_flags = 0;
4746 
4747 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
4748 	    dmareq->dmar_arg, lhandlep);
4749 	if (e != DDI_SUCCESS) {
4750 		return (e);
4751 	}
4752 
4753 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
4754 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
4755 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4756 		return (e);
4757 	}
4758 
4759 	/*
4760 	 * if the driver is just testing to see if it's possible to do the bind,
4761 	 * free up the local state and return the result.
4762 	 */
4763 	if (handlep == NULL) {
4764 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
4765 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4766 		if (e == DDI_DMA_MAPPED) {
4767 			return (DDI_DMA_MAPOK);
4768 		} else {
4769 			return (DDI_DMA_NOMAPPING);
4770 		}
4771 	}
4772 
4773 	return (e);
4774 #endif /* defined(__amd64) */
4775 }
4776 
4777 /*
4778  * rootnex_dma_mctl()
4779  *
4780  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
4781  */
4782 /* ARGSUSED */
4783 static int
4784 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4785     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
4786     uint_t cache_flags)
4787 {
4788 #if defined(__amd64)
4789 	/*
4790 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
4791 	 * common implementation in genunix, so they no longer have x86
4792 	 * specific functionality which called into dma_ctl.
4793 	 *
4794 	 * The rest of the obsoleted interfaces were never supported in the
4795 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
4796 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
4797 	 * implementation issues.
4798 	 *
4799 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
4800 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
4801 	 * reflect that now too...
4802 	 *
4803 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
4804 	 * not going to put this functionality into the 64-bit x86 kernel now.
4805 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
4806 	 * that in a future release.
4807 	 */
4808 	return (DDI_FAILURE);
4809 
4810 #else /* 32-bit x86 kernel */
4811 	ddi_dma_cookie_t lcookie;
4812 	ddi_dma_cookie_t *cookie;
4813 	rootnex_window_t *window;
4814 	ddi_dma_impl_t *hp;
4815 	rootnex_dma_t *dma;
4816 	uint_t nwin;
4817 	uint_t ccnt;
4818 	size_t len;
4819 	off_t off;
4820 	int e;
4821 
4822 
4823 	/*
4824 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
4825 	 * hacky since were optimizing for the current interfaces and so we can
4826 	 * cleanup the mess in genunix. Hopefully we will remove the this
4827 	 * obsoleted routines someday soon.
4828 	 */
4829 
4830 	switch (request) {
4831 
4832 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
4833 		hp = (ddi_dma_impl_t *)handle;
4834 		cookie = (ddi_dma_cookie_t *)objpp;
4835 
4836 		/*
4837 		 * convert segment to cookie. We don't distinguish between the
4838 		 * two :-)
4839 		 */
4840 		*cookie = *hp->dmai_cookie;
4841 		*lenp = cookie->dmac_size;
4842 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
4843 		return (DDI_SUCCESS);
4844 
4845 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
4846 		hp = (ddi_dma_impl_t *)handle;
4847 		dma = (rootnex_dma_t *)hp->dmai_private;
4848 
4849 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
4850 			return (DDI_DMA_STALE);
4851 		}
4852 
4853 		/* handle the case where we don't have any windows */
4854 		if (dma->dp_window == NULL) {
4855 			/*
4856 			 * if seg == NULL, and we don't have any windows,
4857 			 * return the first cookie in the sgl.
4858 			 */
4859 			if (*lenp == NULL) {
4860 				dma->dp_current_cookie = 0;
4861 				hp->dmai_cookie = dma->dp_cookies;
4862 				*objpp = (caddr_t)handle;
4863 				return (DDI_SUCCESS);
4864 
4865 			/* if we have more cookies, go to the next cookie */
4866 			} else {
4867 				if ((dma->dp_current_cookie + 1) >=
4868 				    dma->dp_sglinfo.si_sgl_size) {
4869 					return (DDI_DMA_DONE);
4870 				}
4871 				dma->dp_current_cookie++;
4872 				hp->dmai_cookie++;
4873 				return (DDI_SUCCESS);
4874 			}
4875 		}
4876 
4877 		/* We have one or more windows */
4878 		window = &dma->dp_window[dma->dp_current_win];
4879 
4880 		/*
4881 		 * if seg == NULL, return the first cookie in the current
4882 		 * window
4883 		 */
4884 		if (*lenp == NULL) {
4885 			dma->dp_current_cookie = 0;
4886 			hp->dmai_cookie = window->wd_first_cookie;
4887 
4888 		/*
4889 		 * go to the next cookie in the window then see if we done with
4890 		 * this window.
4891 		 */
4892 		} else {
4893 			if ((dma->dp_current_cookie + 1) >=
4894 			    window->wd_cookie_cnt) {
4895 				return (DDI_DMA_DONE);
4896 			}
4897 			dma->dp_current_cookie++;
4898 			hp->dmai_cookie++;
4899 		}
4900 		*objpp = (caddr_t)handle;
4901 		return (DDI_SUCCESS);
4902 
4903 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
4904 		hp = (ddi_dma_impl_t *)handle;
4905 		dma = (rootnex_dma_t *)hp->dmai_private;
4906 
4907 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
4908 			return (DDI_DMA_STALE);
4909 		}
4910 
4911 		/* if win == NULL, return the first window in the bind */
4912 		if (*offp == NULL) {
4913 			nwin = 0;
4914 
4915 		/*
4916 		 * else, go to the next window then see if we're done with all
4917 		 * the windows.
4918 		 */
4919 		} else {
4920 			nwin = dma->dp_current_win + 1;
4921 			if (nwin >= hp->dmai_nwin) {
4922 				return (DDI_DMA_DONE);
4923 			}
4924 		}
4925 
4926 		/* switch to the next window */
4927 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
4928 		    &lcookie, &ccnt);
4929 		ASSERT(e == DDI_SUCCESS);
4930 		if (e != DDI_SUCCESS) {
4931 			return (DDI_DMA_STALE);
4932 		}
4933 
4934 		/* reset the cookie back to the first cookie in the window */
4935 		if (dma->dp_window != NULL) {
4936 			window = &dma->dp_window[dma->dp_current_win];
4937 			hp->dmai_cookie = window->wd_first_cookie;
4938 		} else {
4939 			hp->dmai_cookie = dma->dp_cookies;
4940 		}
4941 
4942 		*objpp = (caddr_t)handle;
4943 		return (DDI_SUCCESS);
4944 
4945 	case DDI_DMA_FREE: /* ddi_dma_free() */
4946 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
4947 		(void) rootnex_dma_freehdl(dip, rdip, handle);
4948 		if (rootnex_state->r_dvma_call_list_id) {
4949 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
4950 		}
4951 		return (DDI_SUCCESS);
4952 
4953 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
4954 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
4955 		/* should never get here, handled in genunix */
4956 		ASSERT(0);
4957 		return (DDI_FAILURE);
4958 
4959 	case DDI_DMA_KVADDR:
4960 	case DDI_DMA_GETERR:
4961 	case DDI_DMA_COFF:
4962 		return (DDI_FAILURE);
4963 	}
4964 
4965 	return (DDI_FAILURE);
4966 #endif /* defined(__amd64) */
4967 }
4968 
4969 /*
4970  * *********
4971  *  FMA Code
4972  * *********
4973  */
4974 
4975 /*
4976  * rootnex_fm_init()
4977  *    FMA init busop
4978  */
4979 /* ARGSUSED */
4980 static int
4981 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
4982     ddi_iblock_cookie_t *ibc)
4983 {
4984 	*ibc = rootnex_state->r_err_ibc;
4985 
4986 	return (ddi_system_fmcap);
4987 }
4988 
4989 /*
4990  * rootnex_dma_check()
4991  *    Function called after a dma fault occurred to find out whether the
4992  *    fault address is associated with a driver that is able to handle faults
4993  *    and recover from faults.
4994  */
4995 /* ARGSUSED */
4996 static int
4997 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
4998     const void *not_used)
4999 {
5000 	rootnex_window_t *window;
5001 	uint64_t start_addr;
5002 	uint64_t fault_addr;
5003 	ddi_dma_impl_t *hp;
5004 	rootnex_dma_t *dma;
5005 	uint64_t end_addr;
5006 	size_t csize;
5007 	int i;
5008 	int j;
5009 
5010 
5011 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5012 	hp = (ddi_dma_impl_t *)handle;
5013 	ASSERT(hp);
5014 
5015 	dma = (rootnex_dma_t *)hp->dmai_private;
5016 
5017 	/* Get the address that we need to search for */
5018 	fault_addr = *(uint64_t *)addr;
5019 
5020 	/*
5021 	 * if we don't have any windows, we can just walk through all the
5022 	 * cookies.
5023 	 */
5024 	if (dma->dp_window == NULL) {
5025 		/* for each cookie */
5026 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5027 			/*
5028 			 * if the faulted address is within the physical address
5029 			 * range of the cookie, return DDI_FM_NONFATAL.
5030 			 */
5031 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5032 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5033 			    dma->dp_cookies[i].dmac_size))) {
5034 				return (DDI_FM_NONFATAL);
5035 			}
5036 		}
5037 
5038 		/* fault_addr not within this DMA handle */
5039 		return (DDI_FM_UNKNOWN);
5040 	}
5041 
5042 	/* we have mutiple windows, walk through each window */
5043 	for (i = 0; i < hp->dmai_nwin; i++) {
5044 		window = &dma->dp_window[i];
5045 
5046 		/* Go through all the cookies in the window */
5047 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5048 
5049 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5050 			csize = window->wd_first_cookie[j].dmac_size;
5051 
5052 			/*
5053 			 * if we are trimming the first cookie in the window,
5054 			 * and this is the first cookie, adjust the start
5055 			 * address and size of the cookie to account for the
5056 			 * trim.
5057 			 */
5058 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5059 				start_addr = window->wd_trim.tr_first_paddr;
5060 				csize = window->wd_trim.tr_first_size;
5061 			}
5062 
5063 			/*
5064 			 * if we are trimming the last cookie in the window,
5065 			 * and this is the last cookie, adjust the start
5066 			 * address and size of the cookie to account for the
5067 			 * trim.
5068 			 */
5069 			if (window->wd_trim.tr_trim_last &&
5070 			    (j == (window->wd_cookie_cnt - 1))) {
5071 				start_addr = window->wd_trim.tr_last_paddr;
5072 				csize = window->wd_trim.tr_last_size;
5073 			}
5074 
5075 			end_addr = start_addr + csize;
5076 
5077 			/*
5078 			 * if the faulted address is within the physical
5079 			 * address of the cookie, return DDI_FM_NONFATAL.
5080 			 */
5081 			if ((fault_addr >= start_addr) &&
5082 			    (fault_addr <= end_addr)) {
5083 				return (DDI_FM_NONFATAL);
5084 			}
5085 		}
5086 	}
5087 
5088 	/* fault_addr not within this DMA handle */
5089 	return (DDI_FM_UNKNOWN);
5090 }
5091 
5092 /*ARGSUSED*/
5093 static int
5094 rootnex_quiesce(dev_info_t *dip)
5095 {
5096 #if defined(__amd64) && !defined(__xpv)
5097 	return (immu_quiesce());
5098 #else
5099 	return (DDI_SUCCESS);
5100 #endif
5101 }
5102 
5103 #if defined(__xpv)
5104 void
5105 immu_init(void)
5106 {
5107 	;
5108 }
5109 
5110 void
5111 immu_startup(void)
5112 {
5113 	;
5114 }
5115 /*ARGSUSED*/
5116 void
5117 immu_physmem_update(uint64_t addr, uint64_t size)
5118 {
5119 	;
5120 }
5121 #endif
5122