xref: /illumos-gate/usr/src/uts/i86pc/io/rootnex.c (revision 4f60987d)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * x86 root nexus driver
28  */
29 
30 #include <sys/sysmacros.h>
31 #include <sys/conf.h>
32 #include <sys/autoconf.h>
33 #include <sys/sysmacros.h>
34 #include <sys/debug.h>
35 #include <sys/psw.h>
36 #include <sys/ddidmareq.h>
37 #include <sys/promif.h>
38 #include <sys/devops.h>
39 #include <sys/kmem.h>
40 #include <sys/cmn_err.h>
41 #include <vm/seg.h>
42 #include <vm/seg_kmem.h>
43 #include <vm/seg_dev.h>
44 #include <sys/vmem.h>
45 #include <sys/mman.h>
46 #include <vm/hat.h>
47 #include <vm/as.h>
48 #include <vm/page.h>
49 #include <sys/avintr.h>
50 #include <sys/errno.h>
51 #include <sys/modctl.h>
52 #include <sys/ddi_impldefs.h>
53 #include <sys/sunddi.h>
54 #include <sys/sunndi.h>
55 #include <sys/mach_intr.h>
56 #include <sys/psm.h>
57 #include <sys/ontrap.h>
58 #include <sys/atomic.h>
59 #include <sys/sdt.h>
60 #include <sys/rootnex.h>
61 #include <vm/hat_i86.h>
62 #include <sys/ddifm.h>
63 #include <sys/ddi_isa.h>
64 
65 #ifdef __xpv
66 #include <sys/bootinfo.h>
67 #include <sys/hypervisor.h>
68 #include <sys/bootconf.h>
69 #include <vm/kboot_mmu.h>
70 #endif
71 
72 #if defined(__amd64) && !defined(__xpv)
73 #include <sys/immu.h>
74 #endif
75 
76 
77 /*
78  * enable/disable extra checking of function parameters. Useful for debugging
79  * drivers.
80  */
81 #ifdef	DEBUG
82 int rootnex_alloc_check_parms = 1;
83 int rootnex_bind_check_parms = 1;
84 int rootnex_bind_check_inuse = 1;
85 int rootnex_unbind_verify_buffer = 0;
86 int rootnex_sync_check_parms = 1;
87 #else
88 int rootnex_alloc_check_parms = 0;
89 int rootnex_bind_check_parms = 0;
90 int rootnex_bind_check_inuse = 0;
91 int rootnex_unbind_verify_buffer = 0;
92 int rootnex_sync_check_parms = 0;
93 #endif
94 
95 boolean_t rootnex_dmar_not_setup;
96 
97 /* Master Abort and Target Abort panic flag */
98 int rootnex_fm_ma_ta_panic_flag = 0;
99 
100 /* Semi-temporary patchables to phase in bug fixes, test drivers, etc. */
101 int rootnex_bind_fail = 1;
102 int rootnex_bind_warn = 1;
103 uint8_t *rootnex_warn_list;
104 /* bitmasks for rootnex_warn_list. Up to 8 different warnings with uint8_t */
105 #define	ROOTNEX_BIND_WARNING	(0x1 << 0)
106 
107 /*
108  * revert back to old broken behavior of always sync'ing entire copy buffer.
109  * This is useful if be have a buggy driver which doesn't correctly pass in
110  * the offset and size into ddi_dma_sync().
111  */
112 int rootnex_sync_ignore_params = 0;
113 
114 /*
115  * For the 64-bit kernel, pre-alloc enough cookies for a 256K buffer plus 1
116  * page for alignment. For the 32-bit kernel, pre-alloc enough cookies for a
117  * 64K buffer plus 1 page for alignment (we have less kernel space in a 32-bit
118  * kernel). Allocate enough windows to handle a 256K buffer w/ at least 65
119  * sgllen DMA engine, and enough copybuf buffer state pages to handle 2 pages
120  * (< 8K). We will still need to allocate the copy buffer during bind though
121  * (if we need one). These can only be modified in /etc/system before rootnex
122  * attach.
123  */
124 #if defined(__amd64)
125 int rootnex_prealloc_cookies = 65;
126 int rootnex_prealloc_windows = 4;
127 int rootnex_prealloc_copybuf = 2;
128 #else
129 int rootnex_prealloc_cookies = 33;
130 int rootnex_prealloc_windows = 4;
131 int rootnex_prealloc_copybuf = 2;
132 #endif
133 
134 /* driver global state */
135 static rootnex_state_t *rootnex_state;
136 
137 /* shortcut to rootnex counters */
138 static uint64_t *rootnex_cnt;
139 
140 /*
141  * XXX - does x86 even need these or are they left over from the SPARC days?
142  */
143 /* statically defined integer/boolean properties for the root node */
144 static rootnex_intprop_t rootnex_intprp[] = {
145 	{ "PAGESIZE",			PAGESIZE },
146 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
147 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
148 	{ DDI_RELATIVE_ADDRESSING,	1 },
149 };
150 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
151 
152 #ifdef __xpv
153 typedef maddr_t rootnex_addr_t;
154 #define	ROOTNEX_PADDR_TO_RBASE(xinfo, pa)	\
155 	(DOMAIN_IS_INITDOMAIN(xinfo) ? pa_to_ma(pa) : (pa))
156 #else
157 typedef paddr_t rootnex_addr_t;
158 #endif
159 
160 #if !defined(__xpv)
161 char _depends_on[] = "mach/pcplusmp misc/iommulib misc/acpica";
162 #endif
163 
164 static struct cb_ops rootnex_cb_ops = {
165 	nodev,		/* open */
166 	nodev,		/* close */
167 	nodev,		/* strategy */
168 	nodev,		/* print */
169 	nodev,		/* dump */
170 	nodev,		/* read */
171 	nodev,		/* write */
172 	nodev,		/* ioctl */
173 	nodev,		/* devmap */
174 	nodev,		/* mmap */
175 	nodev,		/* segmap */
176 	nochpoll,	/* chpoll */
177 	ddi_prop_op,	/* cb_prop_op */
178 	NULL,		/* struct streamtab */
179 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
180 	CB_REV,		/* Rev */
181 	nodev,		/* cb_aread */
182 	nodev		/* cb_awrite */
183 };
184 
185 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
186     off_t offset, off_t len, caddr_t *vaddrp);
187 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
188     struct hat *hat, struct seg *seg, caddr_t addr,
189     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
190 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
191     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
192 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
193     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
194     ddi_dma_handle_t *handlep);
195 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
196     ddi_dma_handle_t handle);
197 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
198     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
199     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
200 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
201     ddi_dma_handle_t handle);
202 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
204 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
205     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
206     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
207 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
209     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
210 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
211     ddi_ctl_enum_t ctlop, void *arg, void *result);
212 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
213     ddi_iblock_cookie_t *ibc);
214 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
215     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
216 
217 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
218     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
219     ddi_dma_handle_t *handlep);
220 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
221     ddi_dma_handle_t handle);
222 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
223     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
224     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
225 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
226     ddi_dma_handle_t handle);
227 #if defined(__amd64) && !defined(__xpv)
228 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
229     ddi_dma_handle_t handle);
230 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
231     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
232 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
233     ddi_dma_cookie_t *cookiep, uint_t ccount);
234 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
235     ddi_dma_handle_t handle);
236 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
237 #endif
238 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
239     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
240 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
241     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
242     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
243 
244 static struct bus_ops rootnex_bus_ops = {
245 	BUSO_REV,
246 	rootnex_map,
247 	NULL,
248 	NULL,
249 	NULL,
250 	rootnex_map_fault,
251 	rootnex_dma_map,
252 	rootnex_dma_allochdl,
253 	rootnex_dma_freehdl,
254 	rootnex_dma_bindhdl,
255 	rootnex_dma_unbindhdl,
256 	rootnex_dma_sync,
257 	rootnex_dma_win,
258 	rootnex_dma_mctl,
259 	rootnex_ctlops,
260 	ddi_bus_prop_op,
261 	i_ddi_rootnex_get_eventcookie,
262 	i_ddi_rootnex_add_eventcall,
263 	i_ddi_rootnex_remove_eventcall,
264 	i_ddi_rootnex_post_event,
265 	0,			/* bus_intr_ctl */
266 	0,			/* bus_config */
267 	0,			/* bus_unconfig */
268 	rootnex_fm_init,	/* bus_fm_init */
269 	NULL,			/* bus_fm_fini */
270 	NULL,			/* bus_fm_access_enter */
271 	NULL,			/* bus_fm_access_exit */
272 	NULL,			/* bus_powr */
273 	rootnex_intr_ops	/* bus_intr_op */
274 };
275 
276 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
277 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
278 static int rootnex_quiesce(dev_info_t *dip);
279 
280 static struct dev_ops rootnex_ops = {
281 	DEVO_REV,
282 	0,
283 	ddi_no_info,
284 	nulldev,
285 	nulldev,
286 	rootnex_attach,
287 	rootnex_detach,
288 	nulldev,
289 	&rootnex_cb_ops,
290 	&rootnex_bus_ops,
291 	NULL,
292 	rootnex_quiesce,		/* quiesce */
293 };
294 
295 static struct modldrv rootnex_modldrv = {
296 	&mod_driverops,
297 	"i86pc root nexus",
298 	&rootnex_ops
299 };
300 
301 static struct modlinkage rootnex_modlinkage = {
302 	MODREV_1,
303 	(void *)&rootnex_modldrv,
304 	NULL
305 };
306 
307 #if defined(__amd64) && !defined(__xpv)
308 static iommulib_nexops_t iommulib_nexops = {
309 	IOMMU_NEXOPS_VERSION,
310 	"Rootnex IOMMU ops Vers 1.1",
311 	NULL,
312 	rootnex_coredma_allochdl,
313 	rootnex_coredma_freehdl,
314 	rootnex_coredma_bindhdl,
315 	rootnex_coredma_unbindhdl,
316 	rootnex_coredma_reset_cookies,
317 	rootnex_coredma_get_cookies,
318 	rootnex_coredma_set_cookies,
319 	rootnex_coredma_clear_cookies,
320 	rootnex_coredma_get_sleep_flags,
321 	rootnex_coredma_sync,
322 	rootnex_coredma_win,
323 	rootnex_dma_map,
324 	rootnex_dma_mctl
325 };
326 #endif
327 
328 /*
329  *  extern hacks
330  */
331 extern struct seg_ops segdev_ops;
332 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
333 #ifdef	DDI_MAP_DEBUG
334 extern int ddi_map_debug_flag;
335 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
336 #endif
337 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
338 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
339 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
340     psm_intr_op_t, int *);
341 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
342 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
343 
344 /*
345  * Use device arena to use for device control register mappings.
346  * Various kernel memory walkers (debugger, dtrace) need to know
347  * to avoid this address range to prevent undesired device activity.
348  */
349 extern void *device_arena_alloc(size_t size, int vm_flag);
350 extern void device_arena_free(void * vaddr, size_t size);
351 
352 
353 /*
354  *  Internal functions
355  */
356 static int rootnex_dma_init();
357 static void rootnex_add_props(dev_info_t *);
358 static int rootnex_ctl_reportdev(dev_info_t *dip);
359 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
360 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
361 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
362 static int rootnex_map_handle(ddi_map_req_t *mp);
363 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
364 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
365 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
366     ddi_dma_attr_t *attr);
367 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
368     rootnex_sglinfo_t *sglinfo);
369 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
370     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag);
371 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
372     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
373 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
374 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
375     ddi_dma_attr_t *attr, int kmflag);
376 static void rootnex_teardown_windows(rootnex_dma_t *dma);
377 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
378     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
379 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
380     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
381     size_t *copybuf_used, page_t **cur_pp);
382 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
383     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
384     ddi_dma_attr_t *attr, off_t cur_offset);
385 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
386     rootnex_dma_t *dma, rootnex_window_t **windowp,
387     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
388 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
389     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
390 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
391     off_t offset, size_t size, uint_t cache_flags);
392 static int rootnex_verify_buffer(rootnex_dma_t *dma);
393 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
394     const void *comp_addr, const void *not_used);
395 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
396     rootnex_sglinfo_t *sglinfo);
397 
398 /*
399  * _init()
400  *
401  */
402 int
403 _init(void)
404 {
405 
406 	rootnex_state = NULL;
407 	return (mod_install(&rootnex_modlinkage));
408 }
409 
410 
411 /*
412  * _info()
413  *
414  */
415 int
416 _info(struct modinfo *modinfop)
417 {
418 	return (mod_info(&rootnex_modlinkage, modinfop));
419 }
420 
421 
422 /*
423  * _fini()
424  *
425  */
426 int
427 _fini(void)
428 {
429 	return (EBUSY);
430 }
431 
432 
433 /*
434  * rootnex_attach()
435  *
436  */
437 static int
438 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
439 {
440 	int fmcap;
441 	int e;
442 
443 	switch (cmd) {
444 	case DDI_ATTACH:
445 		break;
446 	case DDI_RESUME:
447 #if defined(__amd64) && !defined(__xpv)
448 		return (immu_unquiesce());
449 #else
450 		return (DDI_SUCCESS);
451 #endif
452 	default:
453 		return (DDI_FAILURE);
454 	}
455 
456 	/*
457 	 * We should only have one instance of rootnex. Save it away since we
458 	 * don't have an easy way to get it back later.
459 	 */
460 	ASSERT(rootnex_state == NULL);
461 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
462 
463 	rootnex_state->r_dip = dip;
464 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
465 	rootnex_state->r_reserved_msg_printed = B_FALSE;
466 	rootnex_cnt = &rootnex_state->r_counters[0];
467 
468 	/*
469 	 * Set minimum fm capability level for i86pc platforms and then
470 	 * initialize error handling. Since we're the rootnex, we don't
471 	 * care what's returned in the fmcap field.
472 	 */
473 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
474 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
475 	fmcap = ddi_system_fmcap;
476 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
477 
478 	/* initialize DMA related state */
479 	e = rootnex_dma_init();
480 	if (e != DDI_SUCCESS) {
481 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
482 		return (DDI_FAILURE);
483 	}
484 
485 	/* Add static root node properties */
486 	rootnex_add_props(dip);
487 
488 	/* since we can't call ddi_report_dev() */
489 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
490 
491 	/* Initialize rootnex event handle */
492 	i_ddi_rootnex_init_events(dip);
493 
494 #if defined(__amd64) && !defined(__xpv)
495 	e = iommulib_nexus_register(dip, &iommulib_nexops,
496 	    &rootnex_state->r_iommulib_handle);
497 
498 	ASSERT(e == DDI_SUCCESS);
499 #endif
500 
501 	return (DDI_SUCCESS);
502 }
503 
504 
505 /*
506  * rootnex_detach()
507  *
508  */
509 /*ARGSUSED*/
510 static int
511 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
512 {
513 	switch (cmd) {
514 	case DDI_SUSPEND:
515 #if defined(__amd64) && !defined(__xpv)
516 		return (immu_quiesce());
517 #else
518 		return (DDI_SUCCESS);
519 #endif
520 	default:
521 		return (DDI_FAILURE);
522 	}
523 	/*NOTREACHED*/
524 
525 }
526 
527 
528 /*
529  * rootnex_dma_init()
530  *
531  */
532 /*ARGSUSED*/
533 static int
534 rootnex_dma_init()
535 {
536 	size_t bufsize;
537 
538 
539 	/*
540 	 * size of our cookie/window/copybuf state needed in dma bind that we
541 	 * pre-alloc in dma_alloc_handle
542 	 */
543 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
544 	rootnex_state->r_prealloc_size =
545 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
546 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
547 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
548 
549 	/*
550 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
551 	 * allocate 16 extra bytes for struct pointer alignment
552 	 * (p->dmai_private & dma->dp_prealloc_buffer)
553 	 */
554 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
555 	    rootnex_state->r_prealloc_size + 0x10;
556 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
557 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
558 	if (rootnex_state->r_dmahdl_cache == NULL) {
559 		return (DDI_FAILURE);
560 	}
561 
562 	/*
563 	 * allocate array to track which major numbers we have printed warnings
564 	 * for.
565 	 */
566 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
567 	    KM_SLEEP);
568 
569 	return (DDI_SUCCESS);
570 }
571 
572 
573 /*
574  * rootnex_add_props()
575  *
576  */
577 static void
578 rootnex_add_props(dev_info_t *dip)
579 {
580 	rootnex_intprop_t *rpp;
581 	int i;
582 
583 	/* Add static integer/boolean properties to the root node */
584 	rpp = rootnex_intprp;
585 	for (i = 0; i < NROOT_INTPROPS; i++) {
586 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
587 		    rpp[i].prop_name, rpp[i].prop_value);
588 	}
589 }
590 
591 
592 
593 /*
594  * *************************
595  *  ctlops related routines
596  * *************************
597  */
598 
599 /*
600  * rootnex_ctlops()
601  *
602  */
603 /*ARGSUSED*/
604 static int
605 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
606     void *arg, void *result)
607 {
608 	int n, *ptr;
609 	struct ddi_parent_private_data *pdp;
610 
611 	switch (ctlop) {
612 	case DDI_CTLOPS_DMAPMAPC:
613 		/*
614 		 * Return 'partial' to indicate that dma mapping
615 		 * has to be done in the main MMU.
616 		 */
617 		return (DDI_DMA_PARTIAL);
618 
619 	case DDI_CTLOPS_BTOP:
620 		/*
621 		 * Convert byte count input to physical page units.
622 		 * (byte counts that are not a page-size multiple
623 		 * are rounded down)
624 		 */
625 		*(ulong_t *)result = btop(*(ulong_t *)arg);
626 		return (DDI_SUCCESS);
627 
628 	case DDI_CTLOPS_PTOB:
629 		/*
630 		 * Convert size in physical pages to bytes
631 		 */
632 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
633 		return (DDI_SUCCESS);
634 
635 	case DDI_CTLOPS_BTOPR:
636 		/*
637 		 * Convert byte count input to physical page units
638 		 * (byte counts that are not a page-size multiple
639 		 * are rounded up)
640 		 */
641 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
642 		return (DDI_SUCCESS);
643 
644 	case DDI_CTLOPS_INITCHILD:
645 		return (impl_ddi_sunbus_initchild(arg));
646 
647 	case DDI_CTLOPS_UNINITCHILD:
648 		impl_ddi_sunbus_removechild(arg);
649 		return (DDI_SUCCESS);
650 
651 	case DDI_CTLOPS_REPORTDEV:
652 		return (rootnex_ctl_reportdev(rdip));
653 
654 	case DDI_CTLOPS_IOMIN:
655 		/*
656 		 * Nothing to do here but reflect back..
657 		 */
658 		return (DDI_SUCCESS);
659 
660 	case DDI_CTLOPS_REGSIZE:
661 	case DDI_CTLOPS_NREGS:
662 		break;
663 
664 	case DDI_CTLOPS_SIDDEV:
665 		if (ndi_dev_is_prom_node(rdip))
666 			return (DDI_SUCCESS);
667 		if (ndi_dev_is_persistent_node(rdip))
668 			return (DDI_SUCCESS);
669 		return (DDI_FAILURE);
670 
671 	case DDI_CTLOPS_POWER:
672 		return ((*pm_platform_power)((power_req_t *)arg));
673 
674 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
675 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
676 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
677 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
678 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
679 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
680 		if (!rootnex_state->r_reserved_msg_printed) {
681 			rootnex_state->r_reserved_msg_printed = B_TRUE;
682 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
683 			    "1 or more reserved/obsolete operations.");
684 		}
685 		return (DDI_FAILURE);
686 
687 	default:
688 		return (DDI_FAILURE);
689 	}
690 	/*
691 	 * The rest are for "hardware" properties
692 	 */
693 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
694 		return (DDI_FAILURE);
695 
696 	if (ctlop == DDI_CTLOPS_NREGS) {
697 		ptr = (int *)result;
698 		*ptr = pdp->par_nreg;
699 	} else {
700 		off_t *size = (off_t *)result;
701 
702 		ptr = (int *)arg;
703 		n = *ptr;
704 		if (n >= pdp->par_nreg) {
705 			return (DDI_FAILURE);
706 		}
707 		*size = (off_t)pdp->par_reg[n].regspec_size;
708 	}
709 	return (DDI_SUCCESS);
710 }
711 
712 
713 /*
714  * rootnex_ctl_reportdev()
715  *
716  */
717 static int
718 rootnex_ctl_reportdev(dev_info_t *dev)
719 {
720 	int i, n, len, f_len = 0;
721 	char *buf;
722 
723 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
724 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
725 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
726 	len = strlen(buf);
727 
728 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
729 
730 		struct regspec *rp = sparc_pd_getreg(dev, i);
731 
732 		if (i == 0)
733 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
734 			    ": ");
735 		else
736 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
737 			    " and ");
738 		len = strlen(buf);
739 
740 		switch (rp->regspec_bustype) {
741 
742 		case BTEISA:
743 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
744 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
745 			break;
746 
747 		case BTISA:
748 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
749 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
750 			break;
751 
752 		default:
753 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
754 			    "space %x offset %x",
755 			    rp->regspec_bustype, rp->regspec_addr);
756 			break;
757 		}
758 		len = strlen(buf);
759 	}
760 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
761 		int pri;
762 
763 		if (i != 0) {
764 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
765 			    ",");
766 			len = strlen(buf);
767 		}
768 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
769 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
770 		    " sparc ipl %d", pri);
771 		len = strlen(buf);
772 	}
773 #ifdef DEBUG
774 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
775 		cmn_err(CE_NOTE, "next message is truncated: "
776 		    "printed length 1024, real length %d", f_len);
777 	}
778 #endif /* DEBUG */
779 	cmn_err(CE_CONT, "?%s\n", buf);
780 	kmem_free(buf, REPORTDEV_BUFSIZE);
781 	return (DDI_SUCCESS);
782 }
783 
784 
785 /*
786  * ******************
787  *  map related code
788  * ******************
789  */
790 
791 /*
792  * rootnex_map()
793  *
794  */
795 static int
796 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
797     off_t len, caddr_t *vaddrp)
798 {
799 	struct regspec *rp, tmp_reg;
800 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
801 	int error;
802 
803 	mp = &mr;
804 
805 	switch (mp->map_op)  {
806 	case DDI_MO_MAP_LOCKED:
807 	case DDI_MO_UNMAP:
808 	case DDI_MO_MAP_HANDLE:
809 		break;
810 	default:
811 #ifdef	DDI_MAP_DEBUG
812 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
813 		    mp->map_op);
814 #endif	/* DDI_MAP_DEBUG */
815 		return (DDI_ME_UNIMPLEMENTED);
816 	}
817 
818 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
819 #ifdef	DDI_MAP_DEBUG
820 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
821 #endif	/* DDI_MAP_DEBUG */
822 		return (DDI_ME_UNIMPLEMENTED);
823 	}
824 
825 	/*
826 	 * First, if given an rnumber, convert it to a regspec...
827 	 * (Presumably, this is on behalf of a child of the root node?)
828 	 */
829 
830 	if (mp->map_type == DDI_MT_RNUMBER)  {
831 
832 		int rnumber = mp->map_obj.rnumber;
833 #ifdef	DDI_MAP_DEBUG
834 		static char *out_of_range =
835 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
836 #endif	/* DDI_MAP_DEBUG */
837 
838 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
839 		if (rp == NULL)  {
840 #ifdef	DDI_MAP_DEBUG
841 			cmn_err(CE_WARN, out_of_range, rnumber,
842 			    ddi_get_name(rdip));
843 #endif	/* DDI_MAP_DEBUG */
844 			return (DDI_ME_RNUMBER_RANGE);
845 		}
846 
847 		/*
848 		 * Convert the given ddi_map_req_t from rnumber to regspec...
849 		 */
850 
851 		mp->map_type = DDI_MT_REGSPEC;
852 		mp->map_obj.rp = rp;
853 	}
854 
855 	/*
856 	 * Adjust offset and length correspnding to called values...
857 	 * XXX: A non-zero length means override the one in the regspec
858 	 * XXX: (regardless of what's in the parent's range?)
859 	 */
860 
861 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
862 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
863 
864 #ifdef	DDI_MAP_DEBUG
865 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
866 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
867 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
868 	    len, mp->map_handlep);
869 #endif	/* DDI_MAP_DEBUG */
870 
871 	/*
872 	 * I/O or memory mapping:
873 	 *
874 	 *	<bustype=0, addr=x, len=x>: memory
875 	 *	<bustype=1, addr=x, len=x>: i/o
876 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
877 	 */
878 
879 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
880 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
881 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
882 		    ddi_get_name(rdip), rp->regspec_bustype,
883 		    rp->regspec_addr, rp->regspec_size);
884 		return (DDI_ME_INVAL);
885 	}
886 
887 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
888 		/*
889 		 * compatibility i/o mapping
890 		 */
891 		rp->regspec_bustype += (uint_t)offset;
892 	} else {
893 		/*
894 		 * Normal memory or i/o mapping
895 		 */
896 		rp->regspec_addr += (uint_t)offset;
897 	}
898 
899 	if (len != 0)
900 		rp->regspec_size = (uint_t)len;
901 
902 #ifdef	DDI_MAP_DEBUG
903 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
904 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
905 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
906 	    offset, len, mp->map_handlep);
907 #endif	/* DDI_MAP_DEBUG */
908 
909 	/*
910 	 * Apply any parent ranges at this level, if applicable.
911 	 * (This is where nexus specific regspec translation takes place.
912 	 * Use of this function is implicit agreement that translation is
913 	 * provided via ddi_apply_range.)
914 	 */
915 
916 #ifdef	DDI_MAP_DEBUG
917 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
918 	    ddi_get_name(dip), ddi_get_name(rdip));
919 #endif	/* DDI_MAP_DEBUG */
920 
921 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
922 		return (error);
923 
924 	switch (mp->map_op)  {
925 	case DDI_MO_MAP_LOCKED:
926 
927 		/*
928 		 * Set up the locked down kernel mapping to the regspec...
929 		 */
930 
931 		return (rootnex_map_regspec(mp, vaddrp));
932 
933 	case DDI_MO_UNMAP:
934 
935 		/*
936 		 * Release mapping...
937 		 */
938 
939 		return (rootnex_unmap_regspec(mp, vaddrp));
940 
941 	case DDI_MO_MAP_HANDLE:
942 
943 		return (rootnex_map_handle(mp));
944 
945 	default:
946 		return (DDI_ME_UNIMPLEMENTED);
947 	}
948 }
949 
950 
951 /*
952  * rootnex_map_fault()
953  *
954  *	fault in mappings for requestors
955  */
956 /*ARGSUSED*/
957 static int
958 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
959     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
960     uint_t lock)
961 {
962 
963 #ifdef	DDI_MAP_DEBUG
964 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
965 	ddi_map_debug(" Seg <%s>\n",
966 	    seg->s_ops == &segdev_ops ? "segdev" :
967 	    seg == &kvseg ? "segkmem" : "NONE!");
968 #endif	/* DDI_MAP_DEBUG */
969 
970 	/*
971 	 * This is all terribly broken, but it is a start
972 	 *
973 	 * XXX	Note that this test means that segdev_ops
974 	 *	must be exported from seg_dev.c.
975 	 * XXX	What about devices with their own segment drivers?
976 	 */
977 	if (seg->s_ops == &segdev_ops) {
978 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
979 
980 		if (hat == NULL) {
981 			/*
982 			 * This is one plausible interpretation of
983 			 * a null hat i.e. use the first hat on the
984 			 * address space hat list which by convention is
985 			 * the hat of the system MMU.  At alternative
986 			 * would be to panic .. this might well be better ..
987 			 */
988 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
989 			hat = seg->s_as->a_hat;
990 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
991 		}
992 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
993 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
994 	} else if (seg == &kvseg && dp == NULL) {
995 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
996 		    HAT_LOAD_LOCK);
997 	} else
998 		return (DDI_FAILURE);
999 	return (DDI_SUCCESS);
1000 }
1001 
1002 
1003 /*
1004  * rootnex_map_regspec()
1005  *     we don't support mapping of I/O cards above 4Gb
1006  */
1007 static int
1008 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1009 {
1010 	rootnex_addr_t rbase;
1011 	void *cvaddr;
1012 	uint_t npages, pgoffset;
1013 	struct regspec *rp;
1014 	ddi_acc_hdl_t *hp;
1015 	ddi_acc_impl_t *ap;
1016 	uint_t	hat_acc_flags;
1017 	paddr_t pbase;
1018 
1019 	rp = mp->map_obj.rp;
1020 	hp = mp->map_handlep;
1021 
1022 #ifdef	DDI_MAP_DEBUG
1023 	ddi_map_debug(
1024 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1025 	    rp->regspec_bustype, rp->regspec_addr,
1026 	    rp->regspec_size, mp->map_handlep);
1027 #endif	/* DDI_MAP_DEBUG */
1028 
1029 	/*
1030 	 * I/O or memory mapping
1031 	 *
1032 	 *	<bustype=0, addr=x, len=x>: memory
1033 	 *	<bustype=1, addr=x, len=x>: i/o
1034 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1035 	 */
1036 
1037 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1038 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1039 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1040 		    rp->regspec_addr, rp->regspec_size);
1041 		return (DDI_FAILURE);
1042 	}
1043 
1044 	if (rp->regspec_bustype != 0) {
1045 		/*
1046 		 * I/O space - needs a handle.
1047 		 */
1048 		if (hp == NULL) {
1049 			return (DDI_FAILURE);
1050 		}
1051 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1052 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1053 		impl_acc_hdl_init(hp);
1054 
1055 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1056 #ifdef  DDI_MAP_DEBUG
1057 			ddi_map_debug("rootnex_map_regspec: mmap() "
1058 			    "to I/O space is not supported.\n");
1059 #endif  /* DDI_MAP_DEBUG */
1060 			return (DDI_ME_INVAL);
1061 		} else {
1062 			/*
1063 			 * 1275-compliant vs. compatibility i/o mapping
1064 			 */
1065 			*vaddrp =
1066 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1067 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1068 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1069 #ifdef __xpv
1070 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1071 				hp->ah_pfn = xen_assign_pfn(
1072 				    mmu_btop((ulong_t)rp->regspec_addr &
1073 				    MMU_PAGEMASK));
1074 			} else {
1075 				hp->ah_pfn = mmu_btop(
1076 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1077 			}
1078 #else
1079 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1080 			    MMU_PAGEMASK);
1081 #endif
1082 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1083 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1084 		}
1085 
1086 #ifdef	DDI_MAP_DEBUG
1087 		ddi_map_debug(
1088 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1089 		    rp->regspec_size, *vaddrp);
1090 #endif	/* DDI_MAP_DEBUG */
1091 		return (DDI_SUCCESS);
1092 	}
1093 
1094 	/*
1095 	 * Memory space
1096 	 */
1097 
1098 	if (hp != NULL) {
1099 		/*
1100 		 * hat layer ignores
1101 		 * hp->ah_acc.devacc_attr_endian_flags.
1102 		 */
1103 		switch (hp->ah_acc.devacc_attr_dataorder) {
1104 		case DDI_STRICTORDER_ACC:
1105 			hat_acc_flags = HAT_STRICTORDER;
1106 			break;
1107 		case DDI_UNORDERED_OK_ACC:
1108 			hat_acc_flags = HAT_UNORDERED_OK;
1109 			break;
1110 		case DDI_MERGING_OK_ACC:
1111 			hat_acc_flags = HAT_MERGING_OK;
1112 			break;
1113 		case DDI_LOADCACHING_OK_ACC:
1114 			hat_acc_flags = HAT_LOADCACHING_OK;
1115 			break;
1116 		case DDI_STORECACHING_OK_ACC:
1117 			hat_acc_flags = HAT_STORECACHING_OK;
1118 			break;
1119 		}
1120 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1121 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1122 		impl_acc_hdl_init(hp);
1123 		hp->ah_hat_flags = hat_acc_flags;
1124 	} else {
1125 		hat_acc_flags = HAT_STRICTORDER;
1126 	}
1127 
1128 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1129 #ifdef __xpv
1130 	/*
1131 	 * If we're dom0, we're using a real device so we need to translate
1132 	 * the MA to a PA.
1133 	 */
1134 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1135 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1136 	} else {
1137 		pbase = rbase;
1138 	}
1139 #else
1140 	pbase = rbase;
1141 #endif
1142 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1143 
1144 	if (rp->regspec_size == 0) {
1145 #ifdef  DDI_MAP_DEBUG
1146 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1147 #endif  /* DDI_MAP_DEBUG */
1148 		return (DDI_ME_INVAL);
1149 	}
1150 
1151 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1152 		/* extra cast to make gcc happy */
1153 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1154 	} else {
1155 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1156 
1157 #ifdef	DDI_MAP_DEBUG
1158 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1159 		    "physical %llx", npages, pbase);
1160 #endif	/* DDI_MAP_DEBUG */
1161 
1162 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1163 		if (cvaddr == NULL)
1164 			return (DDI_ME_NORESOURCES);
1165 
1166 		/*
1167 		 * Now map in the pages we've allocated...
1168 		 */
1169 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1170 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1171 		    HAT_LOAD_LOCK);
1172 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1173 
1174 		/* save away pfn and npages for FMA */
1175 		hp = mp->map_handlep;
1176 		if (hp) {
1177 			hp->ah_pfn = mmu_btop(pbase);
1178 			hp->ah_pnum = npages;
1179 		}
1180 	}
1181 
1182 #ifdef	DDI_MAP_DEBUG
1183 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1184 #endif	/* DDI_MAP_DEBUG */
1185 	return (DDI_SUCCESS);
1186 }
1187 
1188 
1189 /*
1190  * rootnex_unmap_regspec()
1191  *
1192  */
1193 static int
1194 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1195 {
1196 	caddr_t addr = (caddr_t)*vaddrp;
1197 	uint_t npages, pgoffset;
1198 	struct regspec *rp;
1199 
1200 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1201 		return (0);
1202 
1203 	rp = mp->map_obj.rp;
1204 
1205 	if (rp->regspec_size == 0) {
1206 #ifdef  DDI_MAP_DEBUG
1207 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1208 #endif  /* DDI_MAP_DEBUG */
1209 		return (DDI_ME_INVAL);
1210 	}
1211 
1212 	/*
1213 	 * I/O or memory mapping:
1214 	 *
1215 	 *	<bustype=0, addr=x, len=x>: memory
1216 	 *	<bustype=1, addr=x, len=x>: i/o
1217 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1218 	 */
1219 	if (rp->regspec_bustype != 0) {
1220 		/*
1221 		 * This is I/O space, which requires no particular
1222 		 * processing on unmap since it isn't mapped in the
1223 		 * first place.
1224 		 */
1225 		return (DDI_SUCCESS);
1226 	}
1227 
1228 	/*
1229 	 * Memory space
1230 	 */
1231 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1232 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1233 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1234 	device_arena_free(addr - pgoffset, ptob(npages));
1235 
1236 	/*
1237 	 * Destroy the pointer - the mapping has logically gone
1238 	 */
1239 	*vaddrp = NULL;
1240 
1241 	return (DDI_SUCCESS);
1242 }
1243 
1244 
1245 /*
1246  * rootnex_map_handle()
1247  *
1248  */
1249 static int
1250 rootnex_map_handle(ddi_map_req_t *mp)
1251 {
1252 	rootnex_addr_t rbase;
1253 	ddi_acc_hdl_t *hp;
1254 	uint_t pgoffset;
1255 	struct regspec *rp;
1256 	paddr_t pbase;
1257 
1258 	rp = mp->map_obj.rp;
1259 
1260 #ifdef	DDI_MAP_DEBUG
1261 	ddi_map_debug(
1262 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1263 	    rp->regspec_bustype, rp->regspec_addr,
1264 	    rp->regspec_size, mp->map_handlep);
1265 #endif	/* DDI_MAP_DEBUG */
1266 
1267 	/*
1268 	 * I/O or memory mapping:
1269 	 *
1270 	 *	<bustype=0, addr=x, len=x>: memory
1271 	 *	<bustype=1, addr=x, len=x>: i/o
1272 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1273 	 */
1274 	if (rp->regspec_bustype != 0) {
1275 		/*
1276 		 * This refers to I/O space, and we don't support "mapping"
1277 		 * I/O space to a user.
1278 		 */
1279 		return (DDI_FAILURE);
1280 	}
1281 
1282 	/*
1283 	 * Set up the hat_flags for the mapping.
1284 	 */
1285 	hp = mp->map_handlep;
1286 
1287 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1288 	case DDI_NEVERSWAP_ACC:
1289 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1290 		break;
1291 	case DDI_STRUCTURE_LE_ACC:
1292 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1293 		break;
1294 	case DDI_STRUCTURE_BE_ACC:
1295 		return (DDI_FAILURE);
1296 	default:
1297 		return (DDI_REGS_ACC_CONFLICT);
1298 	}
1299 
1300 	switch (hp->ah_acc.devacc_attr_dataorder) {
1301 	case DDI_STRICTORDER_ACC:
1302 		break;
1303 	case DDI_UNORDERED_OK_ACC:
1304 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1305 		break;
1306 	case DDI_MERGING_OK_ACC:
1307 		hp->ah_hat_flags |= HAT_MERGING_OK;
1308 		break;
1309 	case DDI_LOADCACHING_OK_ACC:
1310 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1311 		break;
1312 	case DDI_STORECACHING_OK_ACC:
1313 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1314 		break;
1315 	default:
1316 		return (DDI_FAILURE);
1317 	}
1318 
1319 	rbase = (rootnex_addr_t)rp->regspec_addr &
1320 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1321 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1322 
1323 	if (rp->regspec_size == 0)
1324 		return (DDI_ME_INVAL);
1325 
1326 #ifdef __xpv
1327 	/*
1328 	 * If we're dom0, we're using a real device so we need to translate
1329 	 * the MA to a PA.
1330 	 */
1331 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1332 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1333 		    (rbase & MMU_PAGEOFFSET);
1334 	} else {
1335 		pbase = rbase;
1336 	}
1337 #else
1338 	pbase = rbase;
1339 #endif
1340 
1341 	hp->ah_pfn = mmu_btop(pbase);
1342 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1343 
1344 	return (DDI_SUCCESS);
1345 }
1346 
1347 
1348 
1349 /*
1350  * ************************
1351  *  interrupt related code
1352  * ************************
1353  */
1354 
1355 /*
1356  * rootnex_intr_ops()
1357  *	bus_intr_op() function for interrupt support
1358  */
1359 /* ARGSUSED */
1360 static int
1361 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1362     ddi_intr_handle_impl_t *hdlp, void *result)
1363 {
1364 	struct intrspec			*ispec;
1365 	struct ddi_parent_private_data	*pdp;
1366 
1367 	DDI_INTR_NEXDBG((CE_CONT,
1368 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1369 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1370 
1371 	/* Process the interrupt operation */
1372 	switch (intr_op) {
1373 	case DDI_INTROP_GETCAP:
1374 		/* First check with pcplusmp */
1375 		if (psm_intr_ops == NULL)
1376 			return (DDI_FAILURE);
1377 
1378 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1379 			*(int *)result = 0;
1380 			return (DDI_FAILURE);
1381 		}
1382 		break;
1383 	case DDI_INTROP_SETCAP:
1384 		if (psm_intr_ops == NULL)
1385 			return (DDI_FAILURE);
1386 
1387 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1388 			return (DDI_FAILURE);
1389 		break;
1390 	case DDI_INTROP_ALLOC:
1391 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1392 			return (DDI_FAILURE);
1393 		hdlp->ih_pri = ispec->intrspec_pri;
1394 		*(int *)result = hdlp->ih_scratch1;
1395 		break;
1396 	case DDI_INTROP_FREE:
1397 		pdp = ddi_get_parent_data(rdip);
1398 		/*
1399 		 * Special case for 'pcic' driver' only.
1400 		 * If an intrspec was created for it, clean it up here
1401 		 * See detailed comments on this in the function
1402 		 * rootnex_get_ispec().
1403 		 */
1404 		if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1405 			kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1406 			    pdp->par_nintr);
1407 			/*
1408 			 * Set it to zero; so that
1409 			 * DDI framework doesn't free it again
1410 			 */
1411 			pdp->par_intr = NULL;
1412 			pdp->par_nintr = 0;
1413 		}
1414 		break;
1415 	case DDI_INTROP_GETPRI:
1416 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1417 			return (DDI_FAILURE);
1418 		*(int *)result = ispec->intrspec_pri;
1419 		break;
1420 	case DDI_INTROP_SETPRI:
1421 		/* Validate the interrupt priority passed to us */
1422 		if (*(int *)result > LOCK_LEVEL)
1423 			return (DDI_FAILURE);
1424 
1425 		/* Ensure that PSM is all initialized and ispec is ok */
1426 		if ((psm_intr_ops == NULL) ||
1427 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1428 			return (DDI_FAILURE);
1429 
1430 		/* Change the priority */
1431 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1432 		    PSM_FAILURE)
1433 			return (DDI_FAILURE);
1434 
1435 		/* update the ispec with the new priority */
1436 		ispec->intrspec_pri =  *(int *)result;
1437 		break;
1438 	case DDI_INTROP_ADDISR:
1439 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1440 			return (DDI_FAILURE);
1441 		ispec->intrspec_func = hdlp->ih_cb_func;
1442 		break;
1443 	case DDI_INTROP_REMISR:
1444 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1445 			return (DDI_FAILURE);
1446 		ispec->intrspec_func = (uint_t (*)()) 0;
1447 		break;
1448 	case DDI_INTROP_ENABLE:
1449 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1450 			return (DDI_FAILURE);
1451 
1452 		/* Call psmi to translate irq with the dip */
1453 		if (psm_intr_ops == NULL)
1454 			return (DDI_FAILURE);
1455 
1456 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1457 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1458 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1459 			return (DDI_FAILURE);
1460 
1461 		/* Add the interrupt handler */
1462 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1463 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1464 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1465 			return (DDI_FAILURE);
1466 		break;
1467 	case DDI_INTROP_DISABLE:
1468 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1469 			return (DDI_FAILURE);
1470 
1471 		/* Call psm_ops() to translate irq with the dip */
1472 		if (psm_intr_ops == NULL)
1473 			return (DDI_FAILURE);
1474 
1475 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1476 		(void) (*psm_intr_ops)(rdip, hdlp,
1477 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1478 
1479 		/* Remove the interrupt handler */
1480 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1481 		    hdlp->ih_cb_func, hdlp->ih_vector);
1482 		break;
1483 	case DDI_INTROP_SETMASK:
1484 		if (psm_intr_ops == NULL)
1485 			return (DDI_FAILURE);
1486 
1487 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1488 			return (DDI_FAILURE);
1489 		break;
1490 	case DDI_INTROP_CLRMASK:
1491 		if (psm_intr_ops == NULL)
1492 			return (DDI_FAILURE);
1493 
1494 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1495 			return (DDI_FAILURE);
1496 		break;
1497 	case DDI_INTROP_GETPENDING:
1498 		if (psm_intr_ops == NULL)
1499 			return (DDI_FAILURE);
1500 
1501 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1502 		    result)) {
1503 			*(int *)result = 0;
1504 			return (DDI_FAILURE);
1505 		}
1506 		break;
1507 	case DDI_INTROP_NAVAIL:
1508 	case DDI_INTROP_NINTRS:
1509 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1510 		if (*(int *)result == 0) {
1511 			/*
1512 			 * Special case for 'pcic' driver' only. This driver
1513 			 * driver is a child of 'isa' and 'rootnex' drivers.
1514 			 *
1515 			 * See detailed comments on this in the function
1516 			 * rootnex_get_ispec().
1517 			 *
1518 			 * Children of 'pcic' send 'NINITR' request all the
1519 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1520 			 * field may not initialized. So, we fake it here
1521 			 * to return 1 (a la what PCMCIA nexus does).
1522 			 */
1523 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1524 				*(int *)result = 1;
1525 			else
1526 				return (DDI_FAILURE);
1527 		}
1528 		break;
1529 	case DDI_INTROP_SUPPORTED_TYPES:
1530 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1531 		break;
1532 	default:
1533 		return (DDI_FAILURE);
1534 	}
1535 
1536 	return (DDI_SUCCESS);
1537 }
1538 
1539 
1540 /*
1541  * rootnex_get_ispec()
1542  *	convert an interrupt number to an interrupt specification.
1543  *	The interrupt number determines which interrupt spec will be
1544  *	returned if more than one exists.
1545  *
1546  *	Look into the parent private data area of the 'rdip' to find out
1547  *	the interrupt specification.  First check to make sure there is
1548  *	one that matchs "inumber" and then return a pointer to it.
1549  *
1550  *	Return NULL if one could not be found.
1551  *
1552  *	NOTE: This is needed for rootnex_intr_ops()
1553  */
1554 static struct intrspec *
1555 rootnex_get_ispec(dev_info_t *rdip, int inum)
1556 {
1557 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1558 
1559 	/*
1560 	 * Special case handling for drivers that provide their own
1561 	 * intrspec structures instead of relying on the DDI framework.
1562 	 *
1563 	 * A broken hardware driver in ON could potentially provide its
1564 	 * own intrspec structure, instead of relying on the hardware.
1565 	 * If these drivers are children of 'rootnex' then we need to
1566 	 * continue to provide backward compatibility to them here.
1567 	 *
1568 	 * Following check is a special case for 'pcic' driver which
1569 	 * was found to have broken hardwre andby provides its own intrspec.
1570 	 *
1571 	 * Verbatim comments from this driver are shown here:
1572 	 * "Don't use the ddi_add_intr since we don't have a
1573 	 * default intrspec in all cases."
1574 	 *
1575 	 * Since an 'ispec' may not be always created for it,
1576 	 * check for that and create one if so.
1577 	 *
1578 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1579 	 */
1580 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1581 		pdp->par_nintr = 1;
1582 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1583 		    pdp->par_nintr, KM_SLEEP);
1584 	}
1585 
1586 	/* Validate the interrupt number */
1587 	if (inum >= pdp->par_nintr)
1588 		return (NULL);
1589 
1590 	/* Get the interrupt structure pointer and return that */
1591 	return ((struct intrspec *)&pdp->par_intr[inum]);
1592 }
1593 
1594 
1595 /*
1596  * ******************
1597  *  dma related code
1598  * ******************
1599  */
1600 
1601 /*ARGSUSED*/
1602 static int
1603 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1604     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1605     ddi_dma_handle_t *handlep)
1606 {
1607 	uint64_t maxsegmentsize_ll;
1608 	uint_t maxsegmentsize;
1609 	ddi_dma_impl_t *hp;
1610 	rootnex_dma_t *dma;
1611 	uint64_t count_max;
1612 	uint64_t seg;
1613 	int kmflag;
1614 	int e;
1615 
1616 
1617 	/* convert our sleep flags */
1618 	if (waitfp == DDI_DMA_SLEEP) {
1619 		kmflag = KM_SLEEP;
1620 	} else {
1621 		kmflag = KM_NOSLEEP;
1622 	}
1623 
1624 	/*
1625 	 * We try to do only one memory allocation here. We'll do a little
1626 	 * pointer manipulation later. If the bind ends up taking more than
1627 	 * our prealloc's space, we'll have to allocate more memory in the
1628 	 * bind operation. Not great, but much better than before and the
1629 	 * best we can do with the current bind interfaces.
1630 	 */
1631 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1632 	if (hp == NULL) {
1633 		if (waitfp != DDI_DMA_DONTWAIT) {
1634 			ddi_set_callback(waitfp, arg,
1635 			    &rootnex_state->r_dvma_call_list_id);
1636 		}
1637 		return (DDI_DMA_NORESOURCES);
1638 	}
1639 
1640 	/* Do our pointer manipulation now, align the structures */
1641 	hp->dmai_private = (void *)(((uintptr_t)hp +
1642 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1643 	dma = (rootnex_dma_t *)hp->dmai_private;
1644 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1645 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1646 
1647 	/* setup the handle */
1648 	rootnex_clean_dmahdl(hp);
1649 	hp->dmai_error.err_fep = NULL;
1650 	hp->dmai_error.err_cf = NULL;
1651 	dma->dp_dip = rdip;
1652 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1653 	dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1654 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1655 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1656 	hp->dmai_rdip = rdip;
1657 	hp->dmai_attr = *attr;
1658 
1659 	/* we don't need to worry about the SPL since we do a tryenter */
1660 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1661 
1662 	/*
1663 	 * Figure out our maximum segment size. If the segment size is greater
1664 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1665 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1666 	 * dma_attr_count_max are size-1 type values.
1667 	 *
1668 	 * Maximum segment size is the largest physically contiguous chunk of
1669 	 * memory that we can return from a bind (i.e. the maximum size of a
1670 	 * single cookie).
1671 	 */
1672 
1673 	/* handle the rollover cases */
1674 	seg = attr->dma_attr_seg + 1;
1675 	if (seg < attr->dma_attr_seg) {
1676 		seg = attr->dma_attr_seg;
1677 	}
1678 	count_max = attr->dma_attr_count_max + 1;
1679 	if (count_max < attr->dma_attr_count_max) {
1680 		count_max = attr->dma_attr_count_max;
1681 	}
1682 
1683 	/*
1684 	 * granularity may or may not be a power of two. If it isn't, we can't
1685 	 * use a simple mask.
1686 	 */
1687 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1688 		dma->dp_granularity_power_2 = B_FALSE;
1689 	} else {
1690 		dma->dp_granularity_power_2 = B_TRUE;
1691 	}
1692 
1693 	/*
1694 	 * maxxfer should be a whole multiple of granularity. If we're going to
1695 	 * break up a window because we're greater than maxxfer, we might as
1696 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1697 	 * worry about triming the window later on for this case.
1698 	 */
1699 	if (attr->dma_attr_granular > 1) {
1700 		if (dma->dp_granularity_power_2) {
1701 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1702 			    (attr->dma_attr_maxxfer &
1703 			    (attr->dma_attr_granular - 1));
1704 		} else {
1705 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1706 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1707 		}
1708 	} else {
1709 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1710 	}
1711 
1712 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1713 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1714 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1715 		maxsegmentsize = 0xFFFFFFFF;
1716 	} else {
1717 		maxsegmentsize = maxsegmentsize_ll;
1718 	}
1719 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1720 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1721 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1722 
1723 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1724 	if (rootnex_alloc_check_parms) {
1725 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1726 		if (e != DDI_SUCCESS) {
1727 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1728 			(void) rootnex_dma_freehdl(dip, rdip,
1729 			    (ddi_dma_handle_t)hp);
1730 			return (e);
1731 		}
1732 	}
1733 
1734 	*handlep = (ddi_dma_handle_t)hp;
1735 
1736 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1737 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1738 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1739 
1740 	return (DDI_SUCCESS);
1741 }
1742 
1743 
1744 /*
1745  * rootnex_dma_allochdl()
1746  *    called from ddi_dma_alloc_handle().
1747  */
1748 static int
1749 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1750     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1751 {
1752 	int retval;
1753 #if defined(__amd64) && !defined(__xpv)
1754 	uint_t error = ENOTSUP;
1755 
1756 	retval = iommulib_nex_open(rdip, &error);
1757 
1758 	if (retval != DDI_SUCCESS && error == ENOTSUP) {
1759 		/* No IOMMU */
1760 		return (rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1761 		    handlep));
1762 	} else if (retval != DDI_SUCCESS) {
1763 		return (DDI_FAILURE);
1764 	}
1765 
1766 	ASSERT(IOMMU_USED(rdip));
1767 
1768 	/* has an IOMMU */
1769 	retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1770 	    waitfp, arg, handlep);
1771 #else
1772 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1773 	    handlep);
1774 #endif
1775 	if (retval == DDI_SUCCESS)
1776 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1777 	return (retval);
1778 }
1779 
1780 /*ARGSUSED*/
1781 static int
1782 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1783     ddi_dma_handle_t handle)
1784 {
1785 	ddi_dma_impl_t *hp;
1786 	rootnex_dma_t *dma;
1787 
1788 
1789 	hp = (ddi_dma_impl_t *)handle;
1790 	dma = (rootnex_dma_t *)hp->dmai_private;
1791 
1792 	/* unbind should have been called first */
1793 	ASSERT(!dma->dp_inuse);
1794 
1795 	mutex_destroy(&dma->dp_mutex);
1796 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1797 
1798 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1799 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1800 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1801 
1802 	if (rootnex_state->r_dvma_call_list_id)
1803 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1804 
1805 	return (DDI_SUCCESS);
1806 }
1807 
1808 /*
1809  * rootnex_dma_freehdl()
1810  *    called from ddi_dma_free_handle().
1811  */
1812 static int
1813 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1814 {
1815 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1816 #if defined(__amd64) && !defined(__xpv)
1817 	if (IOMMU_USED(rdip)) {
1818 		return (iommulib_nexdma_freehdl(dip, rdip, handle));
1819 	}
1820 #endif
1821 	return (rootnex_coredma_freehdl(dip, rdip, handle));
1822 }
1823 
1824 /*ARGSUSED*/
1825 static int
1826 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1827     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1828     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1829 {
1830 	rootnex_sglinfo_t *sinfo;
1831 	ddi_dma_attr_t *attr;
1832 	ddi_dma_impl_t *hp;
1833 	rootnex_dma_t *dma;
1834 	int kmflag;
1835 	int e;
1836 
1837 	hp = (ddi_dma_impl_t *)handle;
1838 	dma = (rootnex_dma_t *)hp->dmai_private;
1839 	sinfo = &dma->dp_sglinfo;
1840 	attr = &hp->dmai_attr;
1841 
1842 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1843 		dma->dp_sleep_flags = KM_SLEEP;
1844 	} else {
1845 		dma->dp_sleep_flags = KM_NOSLEEP;
1846 	}
1847 
1848 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1849 
1850 	/*
1851 	 * This is useful for debugging a driver. Not as useful in a production
1852 	 * system. The only time this will fail is if you have a driver bug.
1853 	 */
1854 	if (rootnex_bind_check_inuse) {
1855 		/*
1856 		 * No one else should ever have this lock unless someone else
1857 		 * is trying to use this handle. So contention on the lock
1858 		 * is the same as inuse being set.
1859 		 */
1860 		e = mutex_tryenter(&dma->dp_mutex);
1861 		if (e == 0) {
1862 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1863 			return (DDI_DMA_INUSE);
1864 		}
1865 		if (dma->dp_inuse) {
1866 			mutex_exit(&dma->dp_mutex);
1867 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1868 			return (DDI_DMA_INUSE);
1869 		}
1870 		dma->dp_inuse = B_TRUE;
1871 		mutex_exit(&dma->dp_mutex);
1872 	}
1873 
1874 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1875 	if (rootnex_bind_check_parms) {
1876 		e = rootnex_valid_bind_parms(dmareq, attr);
1877 		if (e != DDI_SUCCESS) {
1878 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1879 			rootnex_clean_dmahdl(hp);
1880 			return (e);
1881 		}
1882 	}
1883 
1884 	/* save away the original bind info */
1885 	dma->dp_dma = dmareq->dmar_object;
1886 
1887 #if defined(__amd64) && !defined(__xpv)
1888 	e = immu_map_sgl(hp, dmareq, rootnex_prealloc_cookies, rdip);
1889 	switch (e) {
1890 	case DDI_DMA_MAPPED:
1891 		goto out;
1892 	case DDI_DMA_USE_PHYSICAL:
1893 		break;
1894 	case DDI_DMA_PARTIAL:
1895 		ddi_err(DER_PANIC, rdip, "Partial DVMA map");
1896 		e = DDI_DMA_NORESOURCES;
1897 		/*FALLTHROUGH*/
1898 	default:
1899 		ddi_err(DER_MODE, rdip, "DVMA map failed");
1900 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1901 		rootnex_clean_dmahdl(hp);
1902 		return (e);
1903 	}
1904 #endif
1905 
1906 	/*
1907 	 * Figure out a rough estimate of what maximum number of pages this
1908 	 * buffer could use (a high estimate of course).
1909 	 */
1910 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
1911 
1912 	/*
1913 	 * We'll use the pre-allocated cookies for any bind that will *always*
1914 	 * fit (more important to be consistent, we don't want to create
1915 	 * additional degenerate cases).
1916 	 */
1917 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
1918 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
1919 		dma->dp_need_to_free_cookie = B_FALSE;
1920 		DTRACE_PROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
1921 		    uint_t, sinfo->si_max_pages);
1922 
1923 	/*
1924 	 * For anything larger than that, we'll go ahead and allocate the
1925 	 * maximum number of pages we expect to see. Hopefuly, we won't be
1926 	 * seeing this path in the fast path for high performance devices very
1927 	 * frequently.
1928 	 *
1929 	 * a ddi bind interface that allowed the driver to provide storage to
1930 	 * the bind interface would speed this case up.
1931 	 */
1932 	} else {
1933 		/* convert the sleep flags */
1934 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1935 			kmflag =  KM_SLEEP;
1936 		} else {
1937 			kmflag =  KM_NOSLEEP;
1938 		}
1939 
1940 		/*
1941 		 * Save away how much memory we allocated. If we're doing a
1942 		 * nosleep, the alloc could fail...
1943 		 */
1944 		dma->dp_cookie_size = sinfo->si_max_pages *
1945 		    sizeof (ddi_dma_cookie_t);
1946 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
1947 		if (dma->dp_cookies == NULL) {
1948 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
1949 			rootnex_clean_dmahdl(hp);
1950 			return (DDI_DMA_NORESOURCES);
1951 		}
1952 		dma->dp_need_to_free_cookie = B_TRUE;
1953 		DTRACE_PROBE2(rootnex__bind__alloc, dev_info_t *, rdip, uint_t,
1954 		    sinfo->si_max_pages);
1955 	}
1956 	hp->dmai_cookie = dma->dp_cookies;
1957 
1958 	/*
1959 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
1960 	 * looking at the constraints in the dma structure. It will then put
1961 	 * some additional state about the sgl in the dma struct (i.e. is
1962 	 * the sgl clean, or do we need to do some munging; how many pages
1963 	 * need to be copied, etc.)
1964 	 */
1965 	rootnex_get_sgl(&dmareq->dmar_object, dma->dp_cookies,
1966 	    &dma->dp_sglinfo);
1967 
1968 out:
1969 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
1970 	/* if we don't need a copy buffer, we don't need to sync */
1971 	if (sinfo->si_copybuf_req == 0) {
1972 		hp->dmai_rflags |= DMP_NOSYNC;
1973 	}
1974 
1975 	/*
1976 	 * if we don't need the copybuf and we don't need to do a partial,  we
1977 	 * hit the fast path. All the high performance devices should be trying
1978 	 * to hit this path. To hit this path, a device should be able to reach
1979 	 * all of memory, shouldn't try to bind more than it can transfer, and
1980 	 * the buffer shouldn't require more cookies than the driver/device can
1981 	 * handle [sgllen]).
1982 	 */
1983 	if ((sinfo->si_copybuf_req == 0) &&
1984 	    (sinfo->si_sgl_size <= attr->dma_attr_sgllen) &&
1985 	    (dma->dp_dma.dmao_size < dma->dp_maxxfer)) {
1986 		/*
1987 		 * If the driver supports FMA, insert the handle in the FMA DMA
1988 		 * handle cache.
1989 		 */
1990 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
1991 			hp->dmai_error.err_cf = rootnex_dma_check;
1992 
1993 		/*
1994 		 * copy out the first cookie and ccountp, set the cookie
1995 		 * pointer to the second cookie. The first cookie is passed
1996 		 * back on the stack. Additional cookies are accessed via
1997 		 * ddi_dma_nextcookie()
1998 		 */
1999 		*cookiep = dma->dp_cookies[0];
2000 		*ccountp = sinfo->si_sgl_size;
2001 		hp->dmai_cookie++;
2002 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2003 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2004 		DTRACE_PROBE3(rootnex__bind__fast, dev_info_t *, rdip,
2005 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2006 		    uint_t, dma->dp_dma.dmao_size);
2007 
2008 
2009 		return (DDI_DMA_MAPPED);
2010 	}
2011 
2012 	/*
2013 	 * go to the slow path, we may need to alloc more memory, create
2014 	 * multiple windows, and munge up a sgl to make the device happy.
2015 	 */
2016 	e = rootnex_bind_slowpath(hp, dmareq, dma, attr, kmflag);
2017 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2018 		if (dma->dp_need_to_free_cookie) {
2019 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2020 		}
2021 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2022 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2023 		return (e);
2024 	}
2025 
2026 	/*
2027 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2028 	 * cache.
2029 	 */
2030 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2031 		hp->dmai_error.err_cf = rootnex_dma_check;
2032 
2033 	/* if the first window uses the copy buffer, sync it for the device */
2034 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2035 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2036 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2037 		    DDI_DMA_SYNC_FORDEV);
2038 	}
2039 
2040 	/*
2041 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2042 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2043 	 * If we have a partial map (i.e. multiple windows), the number of
2044 	 * cookies we return is the number of cookies in the first window.
2045 	 */
2046 	if (e == DDI_DMA_MAPPED) {
2047 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2048 		*ccountp = sinfo->si_sgl_size;
2049 		hp->dmai_nwin = 1;
2050 	} else {
2051 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2052 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2053 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2054 	}
2055 	*cookiep = dma->dp_cookies[0];
2056 	hp->dmai_cookie++;
2057 
2058 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2059 	ROOTNEX_DPROBE3(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2060 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2061 	    dma->dp_dma.dmao_size);
2062 	return (e);
2063 }
2064 
2065 /*
2066  * rootnex_dma_bindhdl()
2067  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2068  */
2069 static int
2070 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2071     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2072     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2073 {
2074 #if defined(__amd64) && !defined(__xpv)
2075 	if (IOMMU_USED(rdip)) {
2076 		return (iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2077 		    cookiep, ccountp));
2078 	}
2079 #endif
2080 	return (rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2081 	    cookiep, ccountp));
2082 }
2083 
2084 
2085 
2086 /*ARGSUSED*/
2087 static int
2088 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2089     ddi_dma_handle_t handle)
2090 {
2091 	ddi_dma_impl_t *hp;
2092 	rootnex_dma_t *dma;
2093 	int e;
2094 
2095 	hp = (ddi_dma_impl_t *)handle;
2096 	dma = (rootnex_dma_t *)hp->dmai_private;
2097 
2098 	/* make sure the buffer wasn't free'd before calling unbind */
2099 	if (rootnex_unbind_verify_buffer) {
2100 		e = rootnex_verify_buffer(dma);
2101 		if (e != DDI_SUCCESS) {
2102 			ASSERT(0);
2103 			return (DDI_FAILURE);
2104 		}
2105 	}
2106 
2107 	/* sync the current window before unbinding the buffer */
2108 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2109 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2110 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2111 		    DDI_DMA_SYNC_FORCPU);
2112 	}
2113 
2114 	/*
2115 	 * cleanup and copy buffer or window state. if we didn't use the copy
2116 	 * buffer or windows, there won't be much to do :-)
2117 	 */
2118 	rootnex_teardown_copybuf(dma);
2119 	rootnex_teardown_windows(dma);
2120 
2121 #if defined(__amd64) && !defined(__xpv)
2122 	/*
2123 	 * Clean up the page tables and free the dvma
2124 	 */
2125 	e = immu_unmap_sgl(hp, rdip);
2126 	if (e != DDI_DMA_USE_PHYSICAL && e != DDI_SUCCESS) {
2127 		return (e);
2128 	}
2129 #endif
2130 
2131 	/*
2132 	 * If we had to allocate space to for the worse case sgl (it didn't
2133 	 * fit into our pre-allocate buffer), free that up now
2134 	 */
2135 	if (dma->dp_need_to_free_cookie) {
2136 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2137 	}
2138 
2139 	/*
2140 	 * clean up the handle so it's ready for the next bind (i.e. if the
2141 	 * handle is reused).
2142 	 */
2143 	rootnex_clean_dmahdl(hp);
2144 	hp->dmai_error.err_cf = NULL;
2145 
2146 	if (rootnex_state->r_dvma_call_list_id)
2147 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2148 
2149 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2150 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2151 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2152 
2153 	return (DDI_SUCCESS);
2154 }
2155 
2156 /*
2157  * rootnex_dma_unbindhdl()
2158  *    called from ddi_dma_unbind_handle()
2159  */
2160 /*ARGSUSED*/
2161 static int
2162 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2163     ddi_dma_handle_t handle)
2164 {
2165 #if defined(__amd64) && !defined(__xpv)
2166 	if (IOMMU_USED(rdip)) {
2167 		return (iommulib_nexdma_unbindhdl(dip, rdip, handle));
2168 	}
2169 #endif
2170 	return (rootnex_coredma_unbindhdl(dip, rdip, handle));
2171 }
2172 
2173 #if defined(__amd64) && !defined(__xpv)
2174 
2175 static int
2176 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2177 {
2178 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2179 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2180 
2181 	if (dma->dp_sleep_flags != KM_SLEEP &&
2182 	    dma->dp_sleep_flags != KM_NOSLEEP)
2183 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2184 	return (dma->dp_sleep_flags);
2185 }
2186 /*ARGSUSED*/
2187 static void
2188 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2189 {
2190 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2191 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2192 	rootnex_window_t *window;
2193 
2194 	if (dma->dp_window) {
2195 		window = &dma->dp_window[dma->dp_current_win];
2196 		hp->dmai_cookie = window->wd_first_cookie;
2197 	} else {
2198 		hp->dmai_cookie = dma->dp_cookies;
2199 	}
2200 	hp->dmai_cookie++;
2201 }
2202 
2203 /*ARGSUSED*/
2204 static int
2205 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2206     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2207 {
2208 	int i;
2209 	int km_flags;
2210 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2211 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2212 	rootnex_window_t *window;
2213 	ddi_dma_cookie_t *cp;
2214 	ddi_dma_cookie_t *cookie;
2215 
2216 	ASSERT(*cookiepp == NULL);
2217 	ASSERT(*ccountp == 0);
2218 
2219 	if (dma->dp_window) {
2220 		window = &dma->dp_window[dma->dp_current_win];
2221 		cp = window->wd_first_cookie;
2222 		*ccountp = window->wd_cookie_cnt;
2223 	} else {
2224 		cp = dma->dp_cookies;
2225 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2226 	}
2227 
2228 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2229 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2230 	if (cookie == NULL) {
2231 		return (DDI_DMA_NORESOURCES);
2232 	}
2233 
2234 	for (i = 0; i < *ccountp; i++) {
2235 		cookie[i].dmac_notused = cp[i].dmac_notused;
2236 		cookie[i].dmac_type = cp[i].dmac_type;
2237 		cookie[i].dmac_address = cp[i].dmac_address;
2238 		cookie[i].dmac_size = cp[i].dmac_size;
2239 	}
2240 
2241 	*cookiepp = cookie;
2242 
2243 	return (DDI_SUCCESS);
2244 }
2245 
2246 /*ARGSUSED*/
2247 static int
2248 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2249     ddi_dma_cookie_t *cookiep, uint_t ccount)
2250 {
2251 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2252 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2253 	rootnex_window_t *window;
2254 	ddi_dma_cookie_t *cur_cookiep;
2255 
2256 	ASSERT(cookiep);
2257 	ASSERT(ccount != 0);
2258 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2259 
2260 	if (dma->dp_window) {
2261 		window = &dma->dp_window[dma->dp_current_win];
2262 		dma->dp_saved_cookies = window->wd_first_cookie;
2263 		window->wd_first_cookie = cookiep;
2264 		ASSERT(ccount == window->wd_cookie_cnt);
2265 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2266 		    + window->wd_first_cookie;
2267 	} else {
2268 		dma->dp_saved_cookies = dma->dp_cookies;
2269 		dma->dp_cookies = cookiep;
2270 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2271 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2272 		    + dma->dp_cookies;
2273 	}
2274 
2275 	dma->dp_need_to_switch_cookies = B_TRUE;
2276 	hp->dmai_cookie = cur_cookiep;
2277 
2278 	return (DDI_SUCCESS);
2279 }
2280 
2281 /*ARGSUSED*/
2282 static int
2283 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2284 {
2285 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2286 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2287 	rootnex_window_t *window;
2288 	ddi_dma_cookie_t *cur_cookiep;
2289 	ddi_dma_cookie_t *cookie_array;
2290 	uint_t ccount;
2291 
2292 	/* check if cookies have not been switched */
2293 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2294 		return (DDI_SUCCESS);
2295 
2296 	ASSERT(dma->dp_saved_cookies);
2297 
2298 	if (dma->dp_window) {
2299 		window = &dma->dp_window[dma->dp_current_win];
2300 		cookie_array = window->wd_first_cookie;
2301 		window->wd_first_cookie = dma->dp_saved_cookies;
2302 		dma->dp_saved_cookies = NULL;
2303 		ccount = window->wd_cookie_cnt;
2304 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2305 		    + window->wd_first_cookie;
2306 	} else {
2307 		cookie_array = dma->dp_cookies;
2308 		dma->dp_cookies = dma->dp_saved_cookies;
2309 		dma->dp_saved_cookies = NULL;
2310 		ccount = dma->dp_sglinfo.si_sgl_size;
2311 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2312 		    + dma->dp_cookies;
2313 	}
2314 
2315 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2316 
2317 	hp->dmai_cookie = cur_cookiep;
2318 
2319 	dma->dp_need_to_switch_cookies = B_FALSE;
2320 
2321 	return (DDI_SUCCESS);
2322 }
2323 
2324 #endif
2325 
2326 /*
2327  * rootnex_verify_buffer()
2328  *   verify buffer wasn't free'd
2329  */
2330 static int
2331 rootnex_verify_buffer(rootnex_dma_t *dma)
2332 {
2333 	page_t **pplist;
2334 	caddr_t vaddr;
2335 	uint_t pcnt;
2336 	uint_t poff;
2337 	page_t *pp;
2338 	char b;
2339 	int i;
2340 
2341 	/* Figure out how many pages this buffer occupies */
2342 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2343 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2344 	} else {
2345 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2346 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2347 	}
2348 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2349 
2350 	switch (dma->dp_dma.dmao_type) {
2351 	case DMA_OTYP_PAGES:
2352 		/*
2353 		 * for a linked list of pp's walk through them to make sure
2354 		 * they're locked and not free.
2355 		 */
2356 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2357 		for (i = 0; i < pcnt; i++) {
2358 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2359 				return (DDI_FAILURE);
2360 			}
2361 			pp = pp->p_next;
2362 		}
2363 		break;
2364 
2365 	case DMA_OTYP_VADDR:
2366 	case DMA_OTYP_BUFVADDR:
2367 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2368 		/*
2369 		 * for an array of pp's walk through them to make sure they're
2370 		 * not free. It's possible that they may not be locked.
2371 		 */
2372 		if (pplist) {
2373 			for (i = 0; i < pcnt; i++) {
2374 				if (PP_ISFREE(pplist[i])) {
2375 					return (DDI_FAILURE);
2376 				}
2377 			}
2378 
2379 		/* For a virtual address, try to peek at each page */
2380 		} else {
2381 			if (dma->dp_sglinfo.si_asp == &kas) {
2382 				for (i = 0; i < pcnt; i++) {
2383 					if (ddi_peek8(NULL, vaddr, &b) ==
2384 					    DDI_FAILURE)
2385 						return (DDI_FAILURE);
2386 					vaddr += MMU_PAGESIZE;
2387 				}
2388 			}
2389 		}
2390 		break;
2391 
2392 	default:
2393 		ASSERT(0);
2394 		break;
2395 	}
2396 
2397 	return (DDI_SUCCESS);
2398 }
2399 
2400 
2401 /*
2402  * rootnex_clean_dmahdl()
2403  *    Clean the dma handle. This should be called on a handle alloc and an
2404  *    unbind handle. Set the handle state to the default settings.
2405  */
2406 static void
2407 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2408 {
2409 	rootnex_dma_t *dma;
2410 
2411 
2412 	dma = (rootnex_dma_t *)hp->dmai_private;
2413 
2414 	hp->dmai_nwin = 0;
2415 	dma->dp_current_cookie = 0;
2416 	dma->dp_copybuf_size = 0;
2417 	dma->dp_window = NULL;
2418 	dma->dp_cbaddr = NULL;
2419 	dma->dp_inuse = B_FALSE;
2420 	dma->dp_need_to_free_cookie = B_FALSE;
2421 	dma->dp_need_to_switch_cookies = B_FALSE;
2422 	dma->dp_saved_cookies = NULL;
2423 	dma->dp_sleep_flags = KM_PANIC;
2424 	dma->dp_need_to_free_window = B_FALSE;
2425 	dma->dp_partial_required = B_FALSE;
2426 	dma->dp_trim_required = B_FALSE;
2427 	dma->dp_sglinfo.si_copybuf_req = 0;
2428 #if !defined(__amd64)
2429 	dma->dp_cb_remaping = B_FALSE;
2430 	dma->dp_kva = NULL;
2431 #endif
2432 
2433 	/* FMA related initialization */
2434 	hp->dmai_fault = 0;
2435 	hp->dmai_fault_check = NULL;
2436 	hp->dmai_fault_notify = NULL;
2437 	hp->dmai_error.err_ena = 0;
2438 	hp->dmai_error.err_status = DDI_FM_OK;
2439 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2440 	hp->dmai_error.err_ontrap = NULL;
2441 }
2442 
2443 
2444 /*
2445  * rootnex_valid_alloc_parms()
2446  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2447  */
2448 static int
2449 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2450 {
2451 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2452 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2453 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2454 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2455 		return (DDI_DMA_BADATTR);
2456 	}
2457 
2458 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2459 		return (DDI_DMA_BADATTR);
2460 	}
2461 
2462 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2463 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2464 	    attr->dma_attr_sgllen <= 0) {
2465 		return (DDI_DMA_BADATTR);
2466 	}
2467 
2468 	/* We should be able to DMA into every byte offset in a page */
2469 	if (maxsegmentsize < MMU_PAGESIZE) {
2470 		return (DDI_DMA_BADATTR);
2471 	}
2472 
2473 	/* if we're bouncing on seg, seg must be <= addr_hi */
2474 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2475 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2476 		return (DDI_DMA_BADATTR);
2477 	}
2478 	return (DDI_SUCCESS);
2479 }
2480 
2481 /*
2482  * rootnex_valid_bind_parms()
2483  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2484  */
2485 /* ARGSUSED */
2486 static int
2487 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2488 {
2489 #if !defined(__amd64)
2490 	/*
2491 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2492 	 * we can track the offset for the obsoleted interfaces.
2493 	 */
2494 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2495 		return (DDI_DMA_TOOBIG);
2496 	}
2497 #endif
2498 
2499 	return (DDI_SUCCESS);
2500 }
2501 
2502 
2503 /*
2504  * rootnex_need_bounce_seg()
2505  *    check to see if the buffer lives on both side of the seg.
2506  */
2507 static boolean_t
2508 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2509 {
2510 	ddi_dma_atyp_t buftype;
2511 	rootnex_addr_t raddr;
2512 	boolean_t lower_addr;
2513 	boolean_t upper_addr;
2514 	uint64_t offset;
2515 	page_t **pplist;
2516 	uint64_t paddr;
2517 	uint32_t psize;
2518 	uint32_t size;
2519 	caddr_t vaddr;
2520 	uint_t pcnt;
2521 	page_t *pp;
2522 
2523 
2524 	/* shortcuts */
2525 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2526 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2527 	buftype = dmar_object->dmao_type;
2528 	size = dmar_object->dmao_size;
2529 
2530 	lower_addr = B_FALSE;
2531 	upper_addr = B_FALSE;
2532 	pcnt = 0;
2533 
2534 	/*
2535 	 * Process the first page to handle the initial offset of the buffer.
2536 	 * We'll use the base address we get later when we loop through all
2537 	 * the pages.
2538 	 */
2539 	if (buftype == DMA_OTYP_PAGES) {
2540 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2541 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2542 		    MMU_PAGEOFFSET;
2543 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2544 		psize = MIN(size, (MMU_PAGESIZE - offset));
2545 		pp = pp->p_next;
2546 		sglinfo->si_asp = NULL;
2547 	} else if (pplist != NULL) {
2548 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2549 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2550 		if (sglinfo->si_asp == NULL) {
2551 			sglinfo->si_asp = &kas;
2552 		}
2553 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2554 		paddr += offset;
2555 		psize = MIN(size, (MMU_PAGESIZE - offset));
2556 		pcnt++;
2557 	} else {
2558 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2559 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2560 		if (sglinfo->si_asp == NULL) {
2561 			sglinfo->si_asp = &kas;
2562 		}
2563 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2564 		paddr += offset;
2565 		psize = MIN(size, (MMU_PAGESIZE - offset));
2566 		vaddr += psize;
2567 	}
2568 
2569 #ifdef __xpv
2570 	/*
2571 	 * If we're dom0, we're using a real device so we need to load
2572 	 * the cookies with MFNs instead of PFNs.
2573 	 */
2574 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2575 #else
2576 	raddr = paddr;
2577 #endif
2578 
2579 	if ((raddr + psize) > sglinfo->si_segmask) {
2580 		upper_addr = B_TRUE;
2581 	} else {
2582 		lower_addr = B_TRUE;
2583 	}
2584 	size -= psize;
2585 
2586 	/*
2587 	 * Walk through the rest of the pages in the buffer. Track to see
2588 	 * if we have pages on both sides of the segment boundary.
2589 	 */
2590 	while (size > 0) {
2591 		/* partial or full page */
2592 		psize = MIN(size, MMU_PAGESIZE);
2593 
2594 		if (buftype == DMA_OTYP_PAGES) {
2595 			/* get the paddr from the page_t */
2596 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2597 			paddr = pfn_to_pa(pp->p_pagenum);
2598 			pp = pp->p_next;
2599 		} else if (pplist != NULL) {
2600 			/* index into the array of page_t's to get the paddr */
2601 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2602 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2603 			pcnt++;
2604 		} else {
2605 			/* call into the VM to get the paddr */
2606 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2607 			    vaddr));
2608 			vaddr += psize;
2609 		}
2610 
2611 #ifdef __xpv
2612 		/*
2613 		 * If we're dom0, we're using a real device so we need to load
2614 		 * the cookies with MFNs instead of PFNs.
2615 		 */
2616 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2617 #else
2618 		raddr = paddr;
2619 #endif
2620 
2621 		if ((raddr + psize) > sglinfo->si_segmask) {
2622 			upper_addr = B_TRUE;
2623 		} else {
2624 			lower_addr = B_TRUE;
2625 		}
2626 		/*
2627 		 * if the buffer lives both above and below the segment
2628 		 * boundary, or the current page is the page immediately
2629 		 * after the segment, we will use a copy/bounce buffer for
2630 		 * all pages > seg.
2631 		 */
2632 		if ((lower_addr && upper_addr) ||
2633 		    (raddr == (sglinfo->si_segmask + 1))) {
2634 			return (B_TRUE);
2635 		}
2636 
2637 		size -= psize;
2638 	}
2639 
2640 	return (B_FALSE);
2641 }
2642 
2643 
2644 /*
2645  * rootnex_get_sgl()
2646  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2647  *    with a call to the vm layer when vm2.0 comes around...
2648  */
2649 static void
2650 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2651     rootnex_sglinfo_t *sglinfo)
2652 {
2653 	ddi_dma_atyp_t buftype;
2654 	rootnex_addr_t raddr;
2655 	uint64_t last_page;
2656 	uint64_t offset;
2657 	uint64_t addrhi;
2658 	uint64_t addrlo;
2659 	uint64_t maxseg;
2660 	page_t **pplist;
2661 	uint64_t paddr;
2662 	uint32_t psize;
2663 	uint32_t size;
2664 	caddr_t vaddr;
2665 	uint_t pcnt;
2666 	page_t *pp;
2667 	uint_t cnt;
2668 
2669 
2670 	/* shortcuts */
2671 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2672 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2673 	maxseg = sglinfo->si_max_cookie_size;
2674 	buftype = dmar_object->dmao_type;
2675 	addrhi = sglinfo->si_max_addr;
2676 	addrlo = sglinfo->si_min_addr;
2677 	size = dmar_object->dmao_size;
2678 
2679 	pcnt = 0;
2680 	cnt = 0;
2681 
2682 
2683 	/*
2684 	 * check to see if we need to use the copy buffer for pages over
2685 	 * the segment attr.
2686 	 */
2687 	sglinfo->si_bounce_on_seg = B_FALSE;
2688 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2689 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2690 		    dmar_object, sglinfo);
2691 	}
2692 
2693 	/*
2694 	 * if we were passed down a linked list of pages, i.e. pointer to
2695 	 * page_t, use this to get our physical address and buf offset.
2696 	 */
2697 	if (buftype == DMA_OTYP_PAGES) {
2698 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2699 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2700 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2701 		    MMU_PAGEOFFSET;
2702 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2703 		psize = MIN(size, (MMU_PAGESIZE - offset));
2704 		pp = pp->p_next;
2705 		sglinfo->si_asp = NULL;
2706 
2707 	/*
2708 	 * We weren't passed down a linked list of pages, but if we were passed
2709 	 * down an array of pages, use this to get our physical address and buf
2710 	 * offset.
2711 	 */
2712 	} else if (pplist != NULL) {
2713 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2714 		    (buftype == DMA_OTYP_BUFVADDR));
2715 
2716 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2717 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2718 		if (sglinfo->si_asp == NULL) {
2719 			sglinfo->si_asp = &kas;
2720 		}
2721 
2722 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2723 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2724 		paddr += offset;
2725 		psize = MIN(size, (MMU_PAGESIZE - offset));
2726 		pcnt++;
2727 
2728 	/*
2729 	 * All we have is a virtual address, we'll need to call into the VM
2730 	 * to get the physical address.
2731 	 */
2732 	} else {
2733 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2734 		    (buftype == DMA_OTYP_BUFVADDR));
2735 
2736 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2737 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2738 		if (sglinfo->si_asp == NULL) {
2739 			sglinfo->si_asp = &kas;
2740 		}
2741 
2742 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2743 		paddr += offset;
2744 		psize = MIN(size, (MMU_PAGESIZE - offset));
2745 		vaddr += psize;
2746 	}
2747 
2748 #ifdef __xpv
2749 	/*
2750 	 * If we're dom0, we're using a real device so we need to load
2751 	 * the cookies with MFNs instead of PFNs.
2752 	 */
2753 	raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2754 #else
2755 	raddr = paddr;
2756 #endif
2757 
2758 	/*
2759 	 * Setup the first cookie with the physical address of the page and the
2760 	 * size of the page (which takes into account the initial offset into
2761 	 * the page.
2762 	 */
2763 	sgl[cnt].dmac_laddress = raddr;
2764 	sgl[cnt].dmac_size = psize;
2765 	sgl[cnt].dmac_type = 0;
2766 
2767 	/*
2768 	 * Save away the buffer offset into the page. We'll need this later in
2769 	 * the copy buffer code to help figure out the page index within the
2770 	 * buffer and the offset into the current page.
2771 	 */
2772 	sglinfo->si_buf_offset = offset;
2773 
2774 	/*
2775 	 * If we are using the copy buffer for anything over the segment
2776 	 * boundary, and this page is over the segment boundary.
2777 	 *   OR
2778 	 * if the DMA engine can't reach the physical address.
2779 	 */
2780 	if (((sglinfo->si_bounce_on_seg) &&
2781 	    ((raddr + psize) > sglinfo->si_segmask)) ||
2782 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2783 		/*
2784 		 * Increase how much copy buffer we use. We always increase by
2785 		 * pagesize so we don't have to worry about converting offsets.
2786 		 * Set a flag in the cookies dmac_type to indicate that it uses
2787 		 * the copy buffer. If this isn't the last cookie, go to the
2788 		 * next cookie (since we separate each page which uses the copy
2789 		 * buffer in case the copy buffer is not physically contiguous.
2790 		 */
2791 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
2792 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2793 		if ((cnt + 1) < sglinfo->si_max_pages) {
2794 			cnt++;
2795 			sgl[cnt].dmac_laddress = 0;
2796 			sgl[cnt].dmac_size = 0;
2797 			sgl[cnt].dmac_type = 0;
2798 		}
2799 	}
2800 
2801 	/*
2802 	 * save this page's physical address so we can figure out if the next
2803 	 * page is physically contiguous. Keep decrementing size until we are
2804 	 * done with the buffer.
2805 	 */
2806 	last_page = raddr & MMU_PAGEMASK;
2807 	size -= psize;
2808 
2809 	while (size > 0) {
2810 		/* Get the size for this page (i.e. partial or full page) */
2811 		psize = MIN(size, MMU_PAGESIZE);
2812 
2813 		if (buftype == DMA_OTYP_PAGES) {
2814 			/* get the paddr from the page_t */
2815 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2816 			paddr = pfn_to_pa(pp->p_pagenum);
2817 			pp = pp->p_next;
2818 		} else if (pplist != NULL) {
2819 			/* index into the array of page_t's to get the paddr */
2820 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2821 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2822 			pcnt++;
2823 		} else {
2824 			/* call into the VM to get the paddr */
2825 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2826 			    vaddr));
2827 			vaddr += psize;
2828 		}
2829 
2830 #ifdef __xpv
2831 		/*
2832 		 * If we're dom0, we're using a real device so we need to load
2833 		 * the cookies with MFNs instead of PFNs.
2834 		 */
2835 		raddr = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
2836 #else
2837 		raddr = paddr;
2838 #endif
2839 
2840 		/*
2841 		 * If we are using the copy buffer for anything over the
2842 		 * segment boundary, and this page is over the segment
2843 		 * boundary.
2844 		 *   OR
2845 		 * if the DMA engine can't reach the physical address.
2846 		 */
2847 		if (((sglinfo->si_bounce_on_seg) &&
2848 		    ((raddr + psize) > sglinfo->si_segmask)) ||
2849 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2850 
2851 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
2852 
2853 			/*
2854 			 * if there is something in the current cookie, go to
2855 			 * the next one. We only want one page in a cookie which
2856 			 * uses the copybuf since the copybuf doesn't have to
2857 			 * be physically contiguous.
2858 			 */
2859 			if (sgl[cnt].dmac_size != 0) {
2860 				cnt++;
2861 			}
2862 			sgl[cnt].dmac_laddress = raddr;
2863 			sgl[cnt].dmac_size = psize;
2864 #if defined(__amd64)
2865 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
2866 #else
2867 			/*
2868 			 * save the buf offset for 32-bit kernel. used in the
2869 			 * obsoleted interfaces.
2870 			 */
2871 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
2872 			    (dmar_object->dmao_size - size);
2873 #endif
2874 			/* if this isn't the last cookie, go to the next one */
2875 			if ((cnt + 1) < sglinfo->si_max_pages) {
2876 				cnt++;
2877 				sgl[cnt].dmac_laddress = 0;
2878 				sgl[cnt].dmac_size = 0;
2879 				sgl[cnt].dmac_type = 0;
2880 			}
2881 
2882 		/*
2883 		 * this page didn't need the copy buffer, if it's not physically
2884 		 * contiguous, or it would put us over a segment boundary, or it
2885 		 * puts us over the max cookie size, or the current sgl doesn't
2886 		 * have anything in it.
2887 		 */
2888 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
2889 		    !(raddr & sglinfo->si_segmask) ||
2890 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
2891 		    (sgl[cnt].dmac_size == 0)) {
2892 			/*
2893 			 * if we're not already in a new cookie, go to the next
2894 			 * cookie.
2895 			 */
2896 			if (sgl[cnt].dmac_size != 0) {
2897 				cnt++;
2898 			}
2899 
2900 			/* save the cookie information */
2901 			sgl[cnt].dmac_laddress = raddr;
2902 			sgl[cnt].dmac_size = psize;
2903 #if defined(__amd64)
2904 			sgl[cnt].dmac_type = 0;
2905 #else
2906 			/*
2907 			 * save the buf offset for 32-bit kernel. used in the
2908 			 * obsoleted interfaces.
2909 			 */
2910 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
2911 #endif
2912 
2913 		/*
2914 		 * this page didn't need the copy buffer, it is physically
2915 		 * contiguous with the last page, and it's <= the max cookie
2916 		 * size.
2917 		 */
2918 		} else {
2919 			sgl[cnt].dmac_size += psize;
2920 
2921 			/*
2922 			 * if this exactly ==  the maximum cookie size, and
2923 			 * it isn't the last cookie, go to the next cookie.
2924 			 */
2925 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
2926 			    ((cnt + 1) < sglinfo->si_max_pages)) {
2927 				cnt++;
2928 				sgl[cnt].dmac_laddress = 0;
2929 				sgl[cnt].dmac_size = 0;
2930 				sgl[cnt].dmac_type = 0;
2931 			}
2932 		}
2933 
2934 		/*
2935 		 * save this page's physical address so we can figure out if the
2936 		 * next page is physically contiguous. Keep decrementing size
2937 		 * until we are done with the buffer.
2938 		 */
2939 		last_page = raddr;
2940 		size -= psize;
2941 	}
2942 
2943 	/* we're done, save away how many cookies the sgl has */
2944 	if (sgl[cnt].dmac_size == 0) {
2945 		ASSERT(cnt < sglinfo->si_max_pages);
2946 		sglinfo->si_sgl_size = cnt;
2947 	} else {
2948 		sglinfo->si_sgl_size = cnt + 1;
2949 	}
2950 }
2951 
2952 /*
2953  * rootnex_bind_slowpath()
2954  *    Call in the bind path if the calling driver can't use the sgl without
2955  *    modifying it. We either need to use the copy buffer and/or we will end up
2956  *    with a partial bind.
2957  */
2958 static int
2959 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
2960     rootnex_dma_t *dma, ddi_dma_attr_t *attr, int kmflag)
2961 {
2962 	rootnex_sglinfo_t *sinfo;
2963 	rootnex_window_t *window;
2964 	ddi_dma_cookie_t *cookie;
2965 	size_t copybuf_used;
2966 	size_t dmac_size;
2967 	boolean_t partial;
2968 	off_t cur_offset;
2969 	page_t *cur_pp;
2970 	major_t mnum;
2971 	int e;
2972 	int i;
2973 
2974 
2975 	sinfo = &dma->dp_sglinfo;
2976 	copybuf_used = 0;
2977 	partial = B_FALSE;
2978 
2979 	/*
2980 	 * If we're using the copybuf, set the copybuf state in dma struct.
2981 	 * Needs to be first since it sets the copy buffer size.
2982 	 */
2983 	if (sinfo->si_copybuf_req != 0) {
2984 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
2985 		if (e != DDI_SUCCESS) {
2986 			return (e);
2987 		}
2988 	} else {
2989 		dma->dp_copybuf_size = 0;
2990 	}
2991 
2992 	/*
2993 	 * Figure out if we need to do a partial mapping. If so, figure out
2994 	 * if we need to trim the buffers when we munge the sgl.
2995 	 */
2996 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
2997 	    (dma->dp_dma.dmao_size > dma->dp_maxxfer) ||
2998 	    (attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
2999 		dma->dp_partial_required = B_TRUE;
3000 		if (attr->dma_attr_granular != 1) {
3001 			dma->dp_trim_required = B_TRUE;
3002 		}
3003 	} else {
3004 		dma->dp_partial_required = B_FALSE;
3005 		dma->dp_trim_required = B_FALSE;
3006 	}
3007 
3008 	/* If we need to do a partial bind, make sure the driver supports it */
3009 	if (dma->dp_partial_required &&
3010 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3011 
3012 		mnum = ddi_driver_major(dma->dp_dip);
3013 		/*
3014 		 * patchable which allows us to print one warning per major
3015 		 * number.
3016 		 */
3017 		if ((rootnex_bind_warn) &&
3018 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3019 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3020 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3021 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3022 			    "There is a small risk of data corruption in "
3023 			    "particular with large I/Os. The driver should be "
3024 			    "replaced with a corrected version for proper "
3025 			    "system operation. To disable this warning, add "
3026 			    "'set rootnex:rootnex_bind_warn=0' to "
3027 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3028 		}
3029 		return (DDI_DMA_TOOBIG);
3030 	}
3031 
3032 	/*
3033 	 * we might need multiple windows, setup state to handle them. In this
3034 	 * code path, we will have at least one window.
3035 	 */
3036 	e = rootnex_setup_windows(hp, dma, attr, kmflag);
3037 	if (e != DDI_SUCCESS) {
3038 		rootnex_teardown_copybuf(dma);
3039 		return (e);
3040 	}
3041 
3042 	window = &dma->dp_window[0];
3043 	cookie = &dma->dp_cookies[0];
3044 	cur_offset = 0;
3045 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3046 	if (dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) {
3047 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3048 	}
3049 
3050 	/* loop though all the cookies we got back from get_sgl() */
3051 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3052 		/*
3053 		 * If we're using the copy buffer, check this cookie and setup
3054 		 * its associated copy buffer state. If this cookie uses the
3055 		 * copy buffer, make sure we sync this window during dma_sync.
3056 		 */
3057 		if (dma->dp_copybuf_size > 0) {
3058 			rootnex_setup_cookie(&dmareq->dmar_object, dma, cookie,
3059 			    cur_offset, &copybuf_used, &cur_pp);
3060 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3061 				window->wd_dosync = B_TRUE;
3062 			}
3063 		}
3064 
3065 		/*
3066 		 * save away the cookie size, since it could be modified in
3067 		 * the windowing code.
3068 		 */
3069 		dmac_size = cookie->dmac_size;
3070 
3071 		/* if we went over max copybuf size */
3072 		if (dma->dp_copybuf_size &&
3073 		    (copybuf_used > dma->dp_copybuf_size)) {
3074 			partial = B_TRUE;
3075 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3076 			    cookie, cur_offset, &copybuf_used);
3077 			if (e != DDI_SUCCESS) {
3078 				rootnex_teardown_copybuf(dma);
3079 				rootnex_teardown_windows(dma);
3080 				return (e);
3081 			}
3082 
3083 			/*
3084 			 * if the coookie uses the copy buffer, make sure the
3085 			 * new window we just moved to is set to sync.
3086 			 */
3087 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3088 				window->wd_dosync = B_TRUE;
3089 			}
3090 			DTRACE_PROBE1(rootnex__copybuf__window, dev_info_t *,
3091 			    dma->dp_dip);
3092 
3093 		/* if the cookie cnt == max sgllen, move to the next window */
3094 		} else if (window->wd_cookie_cnt >= attr->dma_attr_sgllen) {
3095 			partial = B_TRUE;
3096 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3097 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3098 			    cookie, attr, cur_offset);
3099 			if (e != DDI_SUCCESS) {
3100 				rootnex_teardown_copybuf(dma);
3101 				rootnex_teardown_windows(dma);
3102 				return (e);
3103 			}
3104 
3105 			/*
3106 			 * if the coookie uses the copy buffer, make sure the
3107 			 * new window we just moved to is set to sync.
3108 			 */
3109 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3110 				window->wd_dosync = B_TRUE;
3111 			}
3112 			DTRACE_PROBE1(rootnex__sgllen__window, dev_info_t *,
3113 			    dma->dp_dip);
3114 
3115 		/* else if we will be over maxxfer */
3116 		} else if ((window->wd_size + dmac_size) >
3117 		    dma->dp_maxxfer) {
3118 			partial = B_TRUE;
3119 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3120 			    cookie);
3121 			if (e != DDI_SUCCESS) {
3122 				rootnex_teardown_copybuf(dma);
3123 				rootnex_teardown_windows(dma);
3124 				return (e);
3125 			}
3126 
3127 			/*
3128 			 * if the coookie uses the copy buffer, make sure the
3129 			 * new window we just moved to is set to sync.
3130 			 */
3131 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3132 				window->wd_dosync = B_TRUE;
3133 			}
3134 			DTRACE_PROBE1(rootnex__maxxfer__window, dev_info_t *,
3135 			    dma->dp_dip);
3136 
3137 		/* else this cookie fits in the current window */
3138 		} else {
3139 			window->wd_cookie_cnt++;
3140 			window->wd_size += dmac_size;
3141 		}
3142 
3143 		/* track our offset into the buffer, go to the next cookie */
3144 		ASSERT(dmac_size <= dma->dp_dma.dmao_size);
3145 		ASSERT(cookie->dmac_size <= dmac_size);
3146 		cur_offset += dmac_size;
3147 		cookie++;
3148 	}
3149 
3150 	/* if we ended up with a zero sized window in the end, clean it up */
3151 	if (window->wd_size == 0) {
3152 		hp->dmai_nwin--;
3153 		window--;
3154 	}
3155 
3156 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3157 
3158 	if (!partial) {
3159 		return (DDI_DMA_MAPPED);
3160 	}
3161 
3162 	ASSERT(dma->dp_partial_required);
3163 	return (DDI_DMA_PARTIAL_MAP);
3164 }
3165 
3166 
3167 /*
3168  * rootnex_setup_copybuf()
3169  *    Called in bind slowpath. Figures out if we're going to use the copy
3170  *    buffer, and if we do, sets up the basic state to handle it.
3171  */
3172 static int
3173 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3174     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3175 {
3176 	rootnex_sglinfo_t *sinfo;
3177 	ddi_dma_attr_t lattr;
3178 	size_t max_copybuf;
3179 	int cansleep;
3180 	int e;
3181 #if !defined(__amd64)
3182 	int vmflag;
3183 #endif
3184 
3185 
3186 	sinfo = &dma->dp_sglinfo;
3187 
3188 	/* read this first so it's consistent through the routine  */
3189 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3190 
3191 	/* We need to call into the rootnex on ddi_dma_sync() */
3192 	hp->dmai_rflags &= ~DMP_NOSYNC;
3193 
3194 	/* make sure the copybuf size <= the max size */
3195 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3196 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3197 
3198 #if !defined(__amd64)
3199 	/*
3200 	 * if we don't have kva space to copy to/from, allocate the KVA space
3201 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3202 	 * the 64-bit kernel.
3203 	 */
3204 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3205 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3206 
3207 		/* convert the sleep flags */
3208 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3209 			vmflag = VM_SLEEP;
3210 		} else {
3211 			vmflag = VM_NOSLEEP;
3212 		}
3213 
3214 		/* allocate Kernel VA space that we can bcopy to/from */
3215 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3216 		    vmflag);
3217 		if (dma->dp_kva == NULL) {
3218 			return (DDI_DMA_NORESOURCES);
3219 		}
3220 	}
3221 #endif
3222 
3223 	/* convert the sleep flags */
3224 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3225 		cansleep = 1;
3226 	} else {
3227 		cansleep = 0;
3228 	}
3229 
3230 	/*
3231 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3232 	 * engine limits, so we can't use kmem_alloc... We don't need
3233 	 * contiguous memory (sgllen) since we will be forcing windows on
3234 	 * sgllen anyway.
3235 	 */
3236 	lattr = *attr;
3237 	lattr.dma_attr_align = MMU_PAGESIZE;
3238 	/*
3239 	 * this should be < 0 to indicate no limit, but due to a bug in
3240 	 * the rootnex, we'll set it to the maximum positive int.
3241 	 */
3242 	lattr.dma_attr_sgllen = 0x7fffffff;
3243 	/*
3244 	 * if we're using the copy buffer because of seg, use that for our
3245 	 * upper address limit.
3246 	 */
3247 	if (sinfo->si_bounce_on_seg) {
3248 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3249 	}
3250 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3251 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3252 	if (e != DDI_SUCCESS) {
3253 #if !defined(__amd64)
3254 		if (dma->dp_kva != NULL) {
3255 			vmem_free(heap_arena, dma->dp_kva,
3256 			    dma->dp_copybuf_size);
3257 		}
3258 #endif
3259 		return (DDI_DMA_NORESOURCES);
3260 	}
3261 
3262 	DTRACE_PROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3263 	    size_t, dma->dp_copybuf_size);
3264 
3265 	return (DDI_SUCCESS);
3266 }
3267 
3268 
3269 /*
3270  * rootnex_setup_windows()
3271  *    Called in bind slowpath to setup the window state. We always have windows
3272  *    in the slowpath. Even if the window count = 1.
3273  */
3274 static int
3275 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3276     ddi_dma_attr_t *attr, int kmflag)
3277 {
3278 	rootnex_window_t *windowp;
3279 	rootnex_sglinfo_t *sinfo;
3280 	size_t copy_state_size;
3281 	size_t win_state_size;
3282 	size_t state_available;
3283 	size_t space_needed;
3284 	uint_t copybuf_win;
3285 	uint_t maxxfer_win;
3286 	size_t space_used;
3287 	uint_t sglwin;
3288 
3289 
3290 	sinfo = &dma->dp_sglinfo;
3291 
3292 	dma->dp_current_win = 0;
3293 	hp->dmai_nwin = 0;
3294 
3295 	/* If we don't need to do a partial, we only have one window */
3296 	if (!dma->dp_partial_required) {
3297 		dma->dp_max_win = 1;
3298 
3299 	/*
3300 	 * we need multiple windows, need to figure out the worse case number
3301 	 * of windows.
3302 	 */
3303 	} else {
3304 		/*
3305 		 * if we need windows because we need more copy buffer that
3306 		 * we allow, the worse case number of windows we could need
3307 		 * here would be (copybuf space required / copybuf space that
3308 		 * we have) plus one for remainder, and plus 2 to handle the
3309 		 * extra pages on the trim for the first and last pages of the
3310 		 * buffer (a page is the minimum window size so under the right
3311 		 * attr settings, you could have a window for each page).
3312 		 * The last page will only be hit here if the size is not a
3313 		 * multiple of the granularity (which theoretically shouldn't
3314 		 * be the case but never has been enforced, so we could have
3315 		 * broken things without it).
3316 		 */
3317 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3318 			ASSERT(dma->dp_copybuf_size > 0);
3319 			copybuf_win = (sinfo->si_copybuf_req /
3320 			    dma->dp_copybuf_size) + 1 + 2;
3321 		} else {
3322 			copybuf_win = 0;
3323 		}
3324 
3325 		/*
3326 		 * if we need windows because we have more cookies than the H/W
3327 		 * can handle, the number of windows we would need here would
3328 		 * be (cookie count / cookies count H/W supports) plus one for
3329 		 * remainder, and plus 2 to handle the extra pages on the trim
3330 		 * (see above comment about trim)
3331 		 */
3332 		if (attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3333 			sglwin = ((sinfo->si_sgl_size / attr->dma_attr_sgllen)
3334 			    + 1) + 2;
3335 		} else {
3336 			sglwin = 0;
3337 		}
3338 
3339 		/*
3340 		 * if we need windows because we're binding more memory than the
3341 		 * H/W can transfer at once, the number of windows we would need
3342 		 * here would be (xfer count / max xfer H/W supports) plus one
3343 		 * for remainder, and plus 2 to handle the extra pages on the
3344 		 * trim (see above comment about trim)
3345 		 */
3346 		if (dma->dp_dma.dmao_size > dma->dp_maxxfer) {
3347 			maxxfer_win = (dma->dp_dma.dmao_size /
3348 			    dma->dp_maxxfer) + 1 + 2;
3349 		} else {
3350 			maxxfer_win = 0;
3351 		}
3352 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3353 		ASSERT(dma->dp_max_win > 0);
3354 	}
3355 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3356 
3357 	/*
3358 	 * Get space for window and potential copy buffer state. Before we
3359 	 * go and allocate memory, see if we can get away with using what's
3360 	 * left in the pre-allocted state or the dynamically allocated sgl.
3361 	 */
3362 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3363 	    sizeof (ddi_dma_cookie_t));
3364 
3365 	/* if we dynamically allocated space for the cookies */
3366 	if (dma->dp_need_to_free_cookie) {
3367 		/* if we have more space in the pre-allocted buffer, use it */
3368 		ASSERT(space_used <= dma->dp_cookie_size);
3369 		if ((dma->dp_cookie_size - space_used) <=
3370 		    rootnex_state->r_prealloc_size) {
3371 			state_available = rootnex_state->r_prealloc_size;
3372 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3373 
3374 		/*
3375 		 * else, we have more free space in the dynamically allocated
3376 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3377 		 * didn't need a lot of cookies.
3378 		 */
3379 		} else {
3380 			state_available = dma->dp_cookie_size - space_used;
3381 			windowp = (rootnex_window_t *)
3382 			    &dma->dp_cookies[sinfo->si_sgl_size];
3383 		}
3384 
3385 	/* we used the pre-alloced buffer */
3386 	} else {
3387 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3388 		state_available = rootnex_state->r_prealloc_size - space_used;
3389 		windowp = (rootnex_window_t *)
3390 		    &dma->dp_cookies[sinfo->si_sgl_size];
3391 	}
3392 
3393 	/*
3394 	 * figure out how much state we need to track the copy buffer. Add an
3395 	 * addition 8 bytes for pointer alignemnt later.
3396 	 */
3397 	if (dma->dp_copybuf_size > 0) {
3398 		copy_state_size = sinfo->si_max_pages *
3399 		    sizeof (rootnex_pgmap_t);
3400 	} else {
3401 		copy_state_size = 0;
3402 	}
3403 	/* add an additional 8 bytes for pointer alignment */
3404 	space_needed = win_state_size + copy_state_size + 0x8;
3405 
3406 	/* if we have enough space already, use it */
3407 	if (state_available >= space_needed) {
3408 		dma->dp_window = windowp;
3409 		dma->dp_need_to_free_window = B_FALSE;
3410 
3411 	/* not enough space, need to allocate more. */
3412 	} else {
3413 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3414 		if (dma->dp_window == NULL) {
3415 			return (DDI_DMA_NORESOURCES);
3416 		}
3417 		dma->dp_need_to_free_window = B_TRUE;
3418 		dma->dp_window_size = space_needed;
3419 		DTRACE_PROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3420 		    dma->dp_dip, size_t, space_needed);
3421 	}
3422 
3423 	/*
3424 	 * we allocate copy buffer state and window state at the same time.
3425 	 * setup our copy buffer state pointers. Make sure it's aligned.
3426 	 */
3427 	if (dma->dp_copybuf_size > 0) {
3428 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3429 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3430 
3431 #if !defined(__amd64)
3432 		/*
3433 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3434 		 * false/NULL. Should be quicker to bzero vs loop and set.
3435 		 */
3436 		bzero(dma->dp_pgmap, copy_state_size);
3437 #endif
3438 	} else {
3439 		dma->dp_pgmap = NULL;
3440 	}
3441 
3442 	return (DDI_SUCCESS);
3443 }
3444 
3445 
3446 /*
3447  * rootnex_teardown_copybuf()
3448  *    cleans up after rootnex_setup_copybuf()
3449  */
3450 static void
3451 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3452 {
3453 #if !defined(__amd64)
3454 	int i;
3455 
3456 	/*
3457 	 * if we allocated kernel heap VMEM space, go through all the pages and
3458 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3459 	 * arena. Then free the VMEM space.
3460 	 */
3461 	if (dma->dp_kva != NULL) {
3462 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3463 			if (dma->dp_pgmap[i].pm_mapped) {
3464 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3465 				    MMU_PAGESIZE, HAT_UNLOAD);
3466 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3467 			}
3468 		}
3469 
3470 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3471 	}
3472 
3473 #endif
3474 
3475 	/* if we allocated a copy buffer, free it */
3476 	if (dma->dp_cbaddr != NULL) {
3477 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3478 	}
3479 }
3480 
3481 
3482 /*
3483  * rootnex_teardown_windows()
3484  *    cleans up after rootnex_setup_windows()
3485  */
3486 static void
3487 rootnex_teardown_windows(rootnex_dma_t *dma)
3488 {
3489 	/*
3490 	 * if we had to allocate window state on the last bind (because we
3491 	 * didn't have enough pre-allocated space in the handle), free it.
3492 	 */
3493 	if (dma->dp_need_to_free_window) {
3494 		kmem_free(dma->dp_window, dma->dp_window_size);
3495 	}
3496 }
3497 
3498 
3499 /*
3500  * rootnex_init_win()
3501  *    Called in bind slow path during creation of a new window. Initializes
3502  *    window state to default values.
3503  */
3504 /*ARGSUSED*/
3505 static void
3506 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3507     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3508 {
3509 	hp->dmai_nwin++;
3510 	window->wd_dosync = B_FALSE;
3511 	window->wd_offset = cur_offset;
3512 	window->wd_size = 0;
3513 	window->wd_first_cookie = cookie;
3514 	window->wd_cookie_cnt = 0;
3515 	window->wd_trim.tr_trim_first = B_FALSE;
3516 	window->wd_trim.tr_trim_last = B_FALSE;
3517 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3518 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3519 #if !defined(__amd64)
3520 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3521 #endif
3522 }
3523 
3524 
3525 /*
3526  * rootnex_setup_cookie()
3527  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3528  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3529  *    out if it uses the copy buffer, and if it does, save away everything we'll
3530  *    need during sync.
3531  */
3532 static void
3533 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3534     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3535     page_t **cur_pp)
3536 {
3537 	boolean_t copybuf_sz_power_2;
3538 	rootnex_sglinfo_t *sinfo;
3539 	paddr_t paddr;
3540 	uint_t pidx;
3541 	uint_t pcnt;
3542 	off_t poff;
3543 #if defined(__amd64)
3544 	pfn_t pfn;
3545 #else
3546 	page_t **pplist;
3547 #endif
3548 
3549 	sinfo = &dma->dp_sglinfo;
3550 
3551 	/*
3552 	 * Calculate the page index relative to the start of the buffer. The
3553 	 * index to the current page for our buffer is the offset into the
3554 	 * first page of the buffer plus our current offset into the buffer
3555 	 * itself, shifted of course...
3556 	 */
3557 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3558 	ASSERT(pidx < sinfo->si_max_pages);
3559 
3560 	/* if this cookie uses the copy buffer */
3561 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3562 		/*
3563 		 * NOTE: we know that since this cookie uses the copy buffer, it
3564 		 * is <= MMU_PAGESIZE.
3565 		 */
3566 
3567 		/*
3568 		 * get the offset into the page. For the 64-bit kernel, get the
3569 		 * pfn which we'll use with seg kpm.
3570 		 */
3571 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3572 #if defined(__amd64)
3573 		/* mfn_to_pfn() is a NOP on i86pc */
3574 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3575 #endif /* __amd64 */
3576 
3577 		/* figure out if the copybuf size is a power of 2 */
3578 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3579 			copybuf_sz_power_2 = B_FALSE;
3580 		} else {
3581 			copybuf_sz_power_2 = B_TRUE;
3582 		}
3583 
3584 		/* This page uses the copy buffer */
3585 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3586 
3587 		/*
3588 		 * save the copy buffer KVA that we'll use with this page.
3589 		 * if we still fit within the copybuf, it's a simple add.
3590 		 * otherwise, we need to wrap over using & or % accordingly.
3591 		 */
3592 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3593 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3594 			    *copybuf_used;
3595 		} else {
3596 			if (copybuf_sz_power_2) {
3597 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3598 				    (uintptr_t)dma->dp_cbaddr +
3599 				    (*copybuf_used &
3600 				    (dma->dp_copybuf_size - 1)));
3601 			} else {
3602 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3603 				    (uintptr_t)dma->dp_cbaddr +
3604 				    (*copybuf_used % dma->dp_copybuf_size));
3605 			}
3606 		}
3607 
3608 		/*
3609 		 * over write the cookie physical address with the address of
3610 		 * the physical address of the copy buffer page that we will
3611 		 * use.
3612 		 */
3613 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3614 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3615 
3616 #ifdef __xpv
3617 		/*
3618 		 * If we're dom0, we're using a real device so we need to load
3619 		 * the cookies with MAs instead of PAs.
3620 		 */
3621 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
3622 #else
3623 		cookie->dmac_laddress = paddr;
3624 #endif
3625 
3626 		/* if we have a kernel VA, it's easy, just save that address */
3627 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3628 		    (sinfo->si_asp == &kas)) {
3629 			/*
3630 			 * save away the page aligned virtual address of the
3631 			 * driver buffer. Offsets are handled in the sync code.
3632 			 */
3633 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3634 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3635 			    & MMU_PAGEMASK);
3636 #if !defined(__amd64)
3637 			/*
3638 			 * we didn't need to, and will never need to map this
3639 			 * page.
3640 			 */
3641 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3642 #endif
3643 
3644 		/* we don't have a kernel VA. We need one for the bcopy. */
3645 		} else {
3646 #if defined(__amd64)
3647 			/*
3648 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3649 			 * get a Kernel VA for the corresponding pfn.
3650 			 */
3651 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3652 #else
3653 			/*
3654 			 * for the 32-bit kernel, this is a pain. First we'll
3655 			 * save away the page_t or user VA for this page. This
3656 			 * is needed in rootnex_dma_win() when we switch to a
3657 			 * new window which requires us to re-map the copy
3658 			 * buffer.
3659 			 */
3660 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3661 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3662 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3663 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3664 			} else if (pplist != NULL) {
3665 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3666 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3667 			} else {
3668 				dma->dp_pgmap[pidx].pm_pp = NULL;
3669 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3670 				    (((uintptr_t)
3671 				    dmar_object->dmao_obj.virt_obj.v_addr +
3672 				    cur_offset) & MMU_PAGEMASK);
3673 			}
3674 
3675 			/*
3676 			 * save away the page aligned virtual address which was
3677 			 * allocated from the kernel heap arena (taking into
3678 			 * account if we need more copy buffer than we alloced
3679 			 * and use multiple windows to handle this, i.e. &,%).
3680 			 * NOTE: there isn't and physical memory backing up this
3681 			 * virtual address space currently.
3682 			 */
3683 			if ((*copybuf_used + MMU_PAGESIZE) <=
3684 			    dma->dp_copybuf_size) {
3685 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3686 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3687 				    MMU_PAGEMASK);
3688 			} else {
3689 				if (copybuf_sz_power_2) {
3690 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3691 					    (((uintptr_t)dma->dp_kva +
3692 					    (*copybuf_used &
3693 					    (dma->dp_copybuf_size - 1))) &
3694 					    MMU_PAGEMASK);
3695 				} else {
3696 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3697 					    (((uintptr_t)dma->dp_kva +
3698 					    (*copybuf_used %
3699 					    dma->dp_copybuf_size)) &
3700 					    MMU_PAGEMASK);
3701 				}
3702 			}
3703 
3704 			/*
3705 			 * if we haven't used up the available copy buffer yet,
3706 			 * map the kva to the physical page.
3707 			 */
3708 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3709 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3710 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
3711 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
3712 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
3713 					    dma->dp_pgmap[pidx].pm_kaddr);
3714 				} else {
3715 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
3716 					    sinfo->si_asp,
3717 					    dma->dp_pgmap[pidx].pm_kaddr);
3718 				}
3719 
3720 			/*
3721 			 * we've used up the available copy buffer, this page
3722 			 * will have to be mapped during rootnex_dma_win() when
3723 			 * we switch to a new window which requires a re-map
3724 			 * the copy buffer. (32-bit kernel only)
3725 			 */
3726 			} else {
3727 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3728 			}
3729 #endif
3730 			/* go to the next page_t */
3731 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3732 				*cur_pp = (*cur_pp)->p_next;
3733 			}
3734 		}
3735 
3736 		/* add to the copy buffer count */
3737 		*copybuf_used += MMU_PAGESIZE;
3738 
3739 	/*
3740 	 * This cookie doesn't use the copy buffer. Walk through the pages this
3741 	 * cookie occupies to reflect this.
3742 	 */
3743 	} else {
3744 		/*
3745 		 * figure out how many pages the cookie occupies. We need to
3746 		 * use the original page offset of the buffer and the cookies
3747 		 * offset in the buffer to do this.
3748 		 */
3749 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
3750 		pcnt = mmu_btopr(cookie->dmac_size + poff);
3751 
3752 		while (pcnt > 0) {
3753 #if !defined(__amd64)
3754 			/*
3755 			 * the 32-bit kernel doesn't have seg kpm, so we need
3756 			 * to map in the driver buffer (if it didn't come down
3757 			 * with a kernel VA) on the fly. Since this page doesn't
3758 			 * use the copy buffer, it's not, or will it ever, have
3759 			 * to be mapped in.
3760 			 */
3761 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3762 #endif
3763 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
3764 
3765 			/*
3766 			 * we need to update pidx and cur_pp or we'll loose
3767 			 * track of where we are.
3768 			 */
3769 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3770 				*cur_pp = (*cur_pp)->p_next;
3771 			}
3772 			pidx++;
3773 			pcnt--;
3774 		}
3775 	}
3776 }
3777 
3778 
3779 /*
3780  * rootnex_sgllen_window_boundary()
3781  *    Called in the bind slow path when the next cookie causes us to exceed (in
3782  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
3783  *    length supported by the DMA H/W.
3784  */
3785 static int
3786 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3787     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
3788     off_t cur_offset)
3789 {
3790 	off_t new_offset;
3791 	size_t trim_sz;
3792 	off_t coffset;
3793 
3794 
3795 	/*
3796 	 * if we know we'll never have to trim, it's pretty easy. Just move to
3797 	 * the next window and init it. We're done.
3798 	 */
3799 	if (!dma->dp_trim_required) {
3800 		(*windowp)++;
3801 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3802 		(*windowp)->wd_cookie_cnt++;
3803 		(*windowp)->wd_size = cookie->dmac_size;
3804 		return (DDI_SUCCESS);
3805 	}
3806 
3807 	/* figure out how much we need to trim from the window */
3808 	ASSERT(attr->dma_attr_granular != 0);
3809 	if (dma->dp_granularity_power_2) {
3810 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
3811 	} else {
3812 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
3813 	}
3814 
3815 	/* The window's a whole multiple of granularity. We're done */
3816 	if (trim_sz == 0) {
3817 		(*windowp)++;
3818 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3819 		(*windowp)->wd_cookie_cnt++;
3820 		(*windowp)->wd_size = cookie->dmac_size;
3821 		return (DDI_SUCCESS);
3822 	}
3823 
3824 	/*
3825 	 * The window's not a whole multiple of granularity, since we know this
3826 	 * is due to the sgllen, we need to go back to the last cookie and trim
3827 	 * that one, add the left over part of the old cookie into the new
3828 	 * window, and then add in the new cookie into the new window.
3829 	 */
3830 
3831 	/*
3832 	 * make sure the driver isn't making us do something bad... Trimming and
3833 	 * sgllen == 1 don't go together.
3834 	 */
3835 	if (attr->dma_attr_sgllen == 1) {
3836 		return (DDI_DMA_NOMAPPING);
3837 	}
3838 
3839 	/*
3840 	 * first, setup the current window to account for the trim. Need to go
3841 	 * back to the last cookie for this.
3842 	 */
3843 	cookie--;
3844 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3845 	(*windowp)->wd_trim.tr_last_cookie = cookie;
3846 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3847 	ASSERT(cookie->dmac_size > trim_sz);
3848 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3849 	(*windowp)->wd_size -= trim_sz;
3850 
3851 	/* save the buffer offsets for the next window */
3852 	coffset = cookie->dmac_size - trim_sz;
3853 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3854 
3855 	/*
3856 	 * set this now in case this is the first window. all other cases are
3857 	 * set in dma_win()
3858 	 */
3859 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
3860 
3861 	/*
3862 	 * initialize the next window using what's left over in the previous
3863 	 * cookie.
3864 	 */
3865 	(*windowp)++;
3866 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3867 	(*windowp)->wd_cookie_cnt++;
3868 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3869 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
3870 	(*windowp)->wd_trim.tr_first_size = trim_sz;
3871 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3872 		(*windowp)->wd_dosync = B_TRUE;
3873 	}
3874 
3875 	/*
3876 	 * now go back to the current cookie and add it to the new window. set
3877 	 * the new window size to the what was left over from the previous
3878 	 * cookie and what's in the current cookie.
3879 	 */
3880 	cookie++;
3881 	(*windowp)->wd_cookie_cnt++;
3882 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
3883 
3884 	/*
3885 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
3886 	 * a max size of maxxfer). Handle that case.
3887 	 */
3888 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
3889 		/*
3890 		 * maxxfer is already a whole multiple of granularity, and this
3891 		 * trim will be <= the previous trim (since a cookie can't be
3892 		 * larger than maxxfer). Make things simple here.
3893 		 */
3894 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
3895 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
3896 		(*windowp)->wd_trim.tr_last_cookie = cookie;
3897 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
3898 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
3899 		(*windowp)->wd_size -= trim_sz;
3900 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
3901 
3902 		/* save the buffer offsets for the next window */
3903 		coffset = cookie->dmac_size - trim_sz;
3904 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
3905 
3906 		/* setup the next window */
3907 		(*windowp)++;
3908 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
3909 		(*windowp)->wd_cookie_cnt++;
3910 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
3911 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
3912 		    coffset;
3913 		(*windowp)->wd_trim.tr_first_size = trim_sz;
3914 	}
3915 
3916 	return (DDI_SUCCESS);
3917 }
3918 
3919 
3920 /*
3921  * rootnex_copybuf_window_boundary()
3922  *    Called in bind slowpath when we get to a window boundary because we used
3923  *    up all the copy buffer that we have.
3924  */
3925 static int
3926 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3927     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
3928     size_t *copybuf_used)
3929 {
3930 	rootnex_sglinfo_t *sinfo;
3931 	off_t new_offset;
3932 	size_t trim_sz;
3933 	paddr_t paddr;
3934 	off_t coffset;
3935 	uint_t pidx;
3936 	off_t poff;
3937 
3938 
3939 	sinfo = &dma->dp_sglinfo;
3940 
3941 	/*
3942 	 * the copy buffer should be a whole multiple of page size. We know that
3943 	 * this cookie is <= MMU_PAGESIZE.
3944 	 */
3945 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
3946 
3947 	/*
3948 	 * from now on, all new windows in this bind need to be re-mapped during
3949 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
3950 	 * space...
3951 	 */
3952 #if !defined(__amd64)
3953 	dma->dp_cb_remaping = B_TRUE;
3954 #endif
3955 
3956 	/* reset copybuf used */
3957 	*copybuf_used = 0;
3958 
3959 	/*
3960 	 * if we don't have to trim (since granularity is set to 1), go to the
3961 	 * next window and add the current cookie to it. We know the current
3962 	 * cookie uses the copy buffer since we're in this code path.
3963 	 */
3964 	if (!dma->dp_trim_required) {
3965 		(*windowp)++;
3966 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3967 
3968 		/* Add this cookie to the new window */
3969 		(*windowp)->wd_cookie_cnt++;
3970 		(*windowp)->wd_size += cookie->dmac_size;
3971 		*copybuf_used += MMU_PAGESIZE;
3972 		return (DDI_SUCCESS);
3973 	}
3974 
3975 	/*
3976 	 * *** may need to trim, figure it out.
3977 	 */
3978 
3979 	/* figure out how much we need to trim from the window */
3980 	if (dma->dp_granularity_power_2) {
3981 		trim_sz = (*windowp)->wd_size &
3982 		    (hp->dmai_attr.dma_attr_granular - 1);
3983 	} else {
3984 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
3985 	}
3986 
3987 	/*
3988 	 * if the window's a whole multiple of granularity, go to the next
3989 	 * window, init it, then add in the current cookie. We know the current
3990 	 * cookie uses the copy buffer since we're in this code path.
3991 	 */
3992 	if (trim_sz == 0) {
3993 		(*windowp)++;
3994 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
3995 
3996 		/* Add this cookie to the new window */
3997 		(*windowp)->wd_cookie_cnt++;
3998 		(*windowp)->wd_size += cookie->dmac_size;
3999 		*copybuf_used += MMU_PAGESIZE;
4000 		return (DDI_SUCCESS);
4001 	}
4002 
4003 	/*
4004 	 * *** We figured it out, we definitly need to trim
4005 	 */
4006 
4007 	/*
4008 	 * make sure the driver isn't making us do something bad...
4009 	 * Trimming and sgllen == 1 don't go together.
4010 	 */
4011 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4012 		return (DDI_DMA_NOMAPPING);
4013 	}
4014 
4015 	/*
4016 	 * first, setup the current window to account for the trim. Need to go
4017 	 * back to the last cookie for this. Some of the last cookie will be in
4018 	 * the current window, and some of the last cookie will be in the new
4019 	 * window. All of the current cookie will be in the new window.
4020 	 */
4021 	cookie--;
4022 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4023 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4024 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4025 	ASSERT(cookie->dmac_size > trim_sz);
4026 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4027 	(*windowp)->wd_size -= trim_sz;
4028 
4029 	/*
4030 	 * we're trimming the last cookie (not the current cookie). So that
4031 	 * last cookie may have or may not have been using the copy buffer (
4032 	 * we know the cookie passed in uses the copy buffer since we're in
4033 	 * this code path).
4034 	 *
4035 	 * If the last cookie doesn't use the copy buffer, nothing special to
4036 	 * do. However, if it does uses the copy buffer, it will be both the
4037 	 * last page in the current window and the first page in the next
4038 	 * window. Since we are reusing the copy buffer (and KVA space on the
4039 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4040 	 * current window, and the start of the copy buffer in the next window.
4041 	 * Track that info... The cookie physical address was already set to
4042 	 * the copy buffer physical address in setup_cookie..
4043 	 */
4044 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4045 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4046 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4047 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4048 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4049 		(*windowp)->wd_trim.tr_last_cbaddr =
4050 		    dma->dp_pgmap[pidx].pm_cbaddr;
4051 #if !defined(__amd64)
4052 		(*windowp)->wd_trim.tr_last_kaddr =
4053 		    dma->dp_pgmap[pidx].pm_kaddr;
4054 #endif
4055 	}
4056 
4057 	/* save the buffer offsets for the next window */
4058 	coffset = cookie->dmac_size - trim_sz;
4059 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4060 
4061 	/*
4062 	 * set this now in case this is the first window. all other cases are
4063 	 * set in dma_win()
4064 	 */
4065 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4066 
4067 	/*
4068 	 * initialize the next window using what's left over in the previous
4069 	 * cookie.
4070 	 */
4071 	(*windowp)++;
4072 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4073 	(*windowp)->wd_cookie_cnt++;
4074 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4075 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4076 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4077 
4078 	/*
4079 	 * again, we're tracking if the last cookie uses the copy buffer.
4080 	 * read the comment above for more info on why we need to track
4081 	 * additional state.
4082 	 *
4083 	 * For the first cookie in the new window, we need reset the physical
4084 	 * address to DMA into to the start of the copy buffer plus any
4085 	 * initial page offset which may be present.
4086 	 */
4087 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4088 		(*windowp)->wd_dosync = B_TRUE;
4089 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4090 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4091 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4092 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4093 
4094 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4095 		    poff;
4096 #ifdef __xpv
4097 		/*
4098 		 * If we're dom0, we're using a real device so we need to load
4099 		 * the cookies with MAs instead of PAs.
4100 		 */
4101 		(*windowp)->wd_trim.tr_first_paddr =
4102 		    ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4103 #else
4104 		(*windowp)->wd_trim.tr_first_paddr = paddr;
4105 #endif
4106 
4107 #if !defined(__amd64)
4108 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4109 #endif
4110 		/* account for the cookie copybuf usage in the new window */
4111 		*copybuf_used += MMU_PAGESIZE;
4112 
4113 		/*
4114 		 * every piece of code has to have a hack, and here is this
4115 		 * ones :-)
4116 		 *
4117 		 * There is a complex interaction between setup_cookie and the
4118 		 * copybuf window boundary. The complexity had to be in either
4119 		 * the maxxfer window, or the copybuf window, and I chose the
4120 		 * copybuf code.
4121 		 *
4122 		 * So in this code path, we have taken the last cookie,
4123 		 * virtually broken it in half due to the trim, and it happens
4124 		 * to use the copybuf which further complicates life. At the
4125 		 * same time, we have already setup the current cookie, which
4126 		 * is now wrong. More background info: the current cookie uses
4127 		 * the copybuf, so it is only a page long max. So we need to
4128 		 * fix the current cookies copy buffer address, physical
4129 		 * address, and kva for the 32-bit kernel. We due this by
4130 		 * bumping them by page size (of course, we can't due this on
4131 		 * the physical address since the copy buffer may not be
4132 		 * physically contiguous).
4133 		 */
4134 		cookie++;
4135 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4136 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4137 
4138 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4139 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4140 #ifdef __xpv
4141 		/*
4142 		 * If we're dom0, we're using a real device so we need to load
4143 		 * the cookies with MAs instead of PAs.
4144 		 */
4145 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(xen_info, paddr);
4146 #else
4147 		cookie->dmac_laddress = paddr;
4148 #endif
4149 
4150 #if !defined(__amd64)
4151 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4152 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4153 #endif
4154 	} else {
4155 		/* go back to the current cookie */
4156 		cookie++;
4157 	}
4158 
4159 	/*
4160 	 * add the current cookie to the new window. set the new window size to
4161 	 * the what was left over from the previous cookie and what's in the
4162 	 * current cookie.
4163 	 */
4164 	(*windowp)->wd_cookie_cnt++;
4165 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4166 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4167 
4168 	/*
4169 	 * we know that the cookie passed in always uses the copy buffer. We
4170 	 * wouldn't be here if it didn't.
4171 	 */
4172 	*copybuf_used += MMU_PAGESIZE;
4173 
4174 	return (DDI_SUCCESS);
4175 }
4176 
4177 
4178 /*
4179  * rootnex_maxxfer_window_boundary()
4180  *    Called in bind slowpath when we get to a window boundary because we will
4181  *    go over maxxfer.
4182  */
4183 static int
4184 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4185     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4186 {
4187 	size_t dmac_size;
4188 	off_t new_offset;
4189 	size_t trim_sz;
4190 	off_t coffset;
4191 
4192 
4193 	/*
4194 	 * calculate how much we have to trim off of the current cookie to equal
4195 	 * maxxfer. We don't have to account for granularity here since our
4196 	 * maxxfer already takes that into account.
4197 	 */
4198 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4199 	ASSERT(trim_sz <= cookie->dmac_size);
4200 	ASSERT(trim_sz <= dma->dp_maxxfer);
4201 
4202 	/* save cookie size since we need it later and we might change it */
4203 	dmac_size = cookie->dmac_size;
4204 
4205 	/*
4206 	 * if we're not trimming the entire cookie, setup the current window to
4207 	 * account for the trim.
4208 	 */
4209 	if (trim_sz < cookie->dmac_size) {
4210 		(*windowp)->wd_cookie_cnt++;
4211 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4212 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4213 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4214 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4215 		(*windowp)->wd_size = dma->dp_maxxfer;
4216 
4217 		/*
4218 		 * set the adjusted cookie size now in case this is the first
4219 		 * window. All other windows are taken care of in get win
4220 		 */
4221 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4222 	}
4223 
4224 	/*
4225 	 * coffset is the current offset within the cookie, new_offset is the
4226 	 * current offset with the entire buffer.
4227 	 */
4228 	coffset = dmac_size - trim_sz;
4229 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4230 
4231 	/* initialize the next window */
4232 	(*windowp)++;
4233 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4234 	(*windowp)->wd_cookie_cnt++;
4235 	(*windowp)->wd_size = trim_sz;
4236 	if (trim_sz < dmac_size) {
4237 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4238 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4239 		    coffset;
4240 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4241 	}
4242 
4243 	return (DDI_SUCCESS);
4244 }
4245 
4246 
4247 /*ARGSUSED*/
4248 static int
4249 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4250     off_t off, size_t len, uint_t cache_flags)
4251 {
4252 	rootnex_sglinfo_t *sinfo;
4253 	rootnex_pgmap_t *cbpage;
4254 	rootnex_window_t *win;
4255 	ddi_dma_impl_t *hp;
4256 	rootnex_dma_t *dma;
4257 	caddr_t fromaddr;
4258 	caddr_t toaddr;
4259 	uint_t psize;
4260 	off_t offset;
4261 	uint_t pidx;
4262 	size_t size;
4263 	off_t poff;
4264 	int e;
4265 
4266 
4267 	hp = (ddi_dma_impl_t *)handle;
4268 	dma = (rootnex_dma_t *)hp->dmai_private;
4269 	sinfo = &dma->dp_sglinfo;
4270 
4271 	/*
4272 	 * if we don't have any windows, we don't need to sync. A copybuf
4273 	 * will cause us to have at least one window.
4274 	 */
4275 	if (dma->dp_window == NULL) {
4276 		return (DDI_SUCCESS);
4277 	}
4278 
4279 	/* This window may not need to be sync'd */
4280 	win = &dma->dp_window[dma->dp_current_win];
4281 	if (!win->wd_dosync) {
4282 		return (DDI_SUCCESS);
4283 	}
4284 
4285 	/* handle off and len special cases */
4286 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4287 		offset = win->wd_offset;
4288 	} else {
4289 		offset = off;
4290 	}
4291 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4292 		size = win->wd_size;
4293 	} else {
4294 		size = len;
4295 	}
4296 
4297 	/* check the sync args to make sure they make a little sense */
4298 	if (rootnex_sync_check_parms) {
4299 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4300 		    cache_flags);
4301 		if (e != DDI_SUCCESS) {
4302 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4303 			return (DDI_FAILURE);
4304 		}
4305 	}
4306 
4307 	/*
4308 	 * special case the first page to handle the offset into the page. The
4309 	 * offset to the current page for our buffer is the offset into the
4310 	 * first page of the buffer plus our current offset into the buffer
4311 	 * itself, masked of course.
4312 	 */
4313 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4314 	psize = MIN((MMU_PAGESIZE - poff), size);
4315 
4316 	/* go through all the pages that we want to sync */
4317 	while (size > 0) {
4318 		/*
4319 		 * Calculate the page index relative to the start of the buffer.
4320 		 * The index to the current page for our buffer is the offset
4321 		 * into the first page of the buffer plus our current offset
4322 		 * into the buffer itself, shifted of course...
4323 		 */
4324 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4325 		ASSERT(pidx < sinfo->si_max_pages);
4326 
4327 		/*
4328 		 * if this page uses the copy buffer, we need to sync it,
4329 		 * otherwise, go on to the next page.
4330 		 */
4331 		cbpage = &dma->dp_pgmap[pidx];
4332 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4333 		    (cbpage->pm_uses_copybuf == B_FALSE));
4334 		if (cbpage->pm_uses_copybuf) {
4335 			/* cbaddr and kaddr should be page aligned */
4336 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4337 			    MMU_PAGEOFFSET) == 0);
4338 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4339 			    MMU_PAGEOFFSET) == 0);
4340 
4341 			/*
4342 			 * if we're copying for the device, we are going to
4343 			 * copy from the drivers buffer and to the rootnex
4344 			 * allocated copy buffer.
4345 			 */
4346 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4347 				fromaddr = cbpage->pm_kaddr + poff;
4348 				toaddr = cbpage->pm_cbaddr + poff;
4349 				DTRACE_PROBE2(rootnex__sync__dev,
4350 				    dev_info_t *, dma->dp_dip, size_t, psize);
4351 
4352 			/*
4353 			 * if we're copying for the cpu/kernel, we are going to
4354 			 * copy from the rootnex allocated copy buffer to the
4355 			 * drivers buffer.
4356 			 */
4357 			} else {
4358 				fromaddr = cbpage->pm_cbaddr + poff;
4359 				toaddr = cbpage->pm_kaddr + poff;
4360 				DTRACE_PROBE2(rootnex__sync__cpu,
4361 				    dev_info_t *, dma->dp_dip, size_t, psize);
4362 			}
4363 
4364 			bcopy(fromaddr, toaddr, psize);
4365 		}
4366 
4367 		/*
4368 		 * decrement size until we're done, update our offset into the
4369 		 * buffer, and get the next page size.
4370 		 */
4371 		size -= psize;
4372 		offset += psize;
4373 		psize = MIN(MMU_PAGESIZE, size);
4374 
4375 		/* page offset is zero for the rest of this loop */
4376 		poff = 0;
4377 	}
4378 
4379 	return (DDI_SUCCESS);
4380 }
4381 
4382 /*
4383  * rootnex_dma_sync()
4384  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4385  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4386  *    is set, ddi_dma_sync() returns immediately passing back success.
4387  */
4388 /*ARGSUSED*/
4389 static int
4390 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4391     off_t off, size_t len, uint_t cache_flags)
4392 {
4393 #if defined(__amd64) && !defined(__xpv)
4394 	if (IOMMU_USED(rdip)) {
4395 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4396 		    cache_flags));
4397 	}
4398 #endif
4399 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4400 	    cache_flags));
4401 }
4402 
4403 /*
4404  * rootnex_valid_sync_parms()
4405  *    checks the parameters passed to sync to verify they are correct.
4406  */
4407 static int
4408 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4409     off_t offset, size_t size, uint_t cache_flags)
4410 {
4411 	off_t woffset;
4412 
4413 
4414 	/*
4415 	 * the first part of the test to make sure the offset passed in is
4416 	 * within the window.
4417 	 */
4418 	if (offset < win->wd_offset) {
4419 		return (DDI_FAILURE);
4420 	}
4421 
4422 	/*
4423 	 * second and last part of the test to make sure the offset and length
4424 	 * passed in is within the window.
4425 	 */
4426 	woffset = offset - win->wd_offset;
4427 	if ((woffset + size) > win->wd_size) {
4428 		return (DDI_FAILURE);
4429 	}
4430 
4431 	/*
4432 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4433 	 * be set too.
4434 	 */
4435 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4436 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4437 		return (DDI_SUCCESS);
4438 	}
4439 
4440 	/*
4441 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4442 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4443 	 */
4444 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4445 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4446 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4447 		return (DDI_SUCCESS);
4448 	}
4449 
4450 	return (DDI_FAILURE);
4451 }
4452 
4453 
4454 /*ARGSUSED*/
4455 static int
4456 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4457     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4458     uint_t *ccountp)
4459 {
4460 	rootnex_window_t *window;
4461 	rootnex_trim_t *trim;
4462 	ddi_dma_impl_t *hp;
4463 	rootnex_dma_t *dma;
4464 #if !defined(__amd64)
4465 	rootnex_sglinfo_t *sinfo;
4466 	rootnex_pgmap_t *pmap;
4467 	uint_t pidx;
4468 	uint_t pcnt;
4469 	off_t poff;
4470 	int i;
4471 #endif
4472 
4473 
4474 	hp = (ddi_dma_impl_t *)handle;
4475 	dma = (rootnex_dma_t *)hp->dmai_private;
4476 #if !defined(__amd64)
4477 	sinfo = &dma->dp_sglinfo;
4478 #endif
4479 
4480 	/* If we try and get a window which doesn't exist, return failure */
4481 	if (win >= hp->dmai_nwin) {
4482 		ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4483 		return (DDI_FAILURE);
4484 	}
4485 
4486 	/*
4487 	 * if we don't have any windows, and they're asking for the first
4488 	 * window, setup the cookie pointer to the first cookie in the bind.
4489 	 * setup our return values, then increment the cookie since we return
4490 	 * the first cookie on the stack.
4491 	 */
4492 	if (dma->dp_window == NULL) {
4493 		if (win != 0) {
4494 			ROOTNEX_PROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4495 			return (DDI_FAILURE);
4496 		}
4497 		hp->dmai_cookie = dma->dp_cookies;
4498 		*offp = 0;
4499 		*lenp = dma->dp_dma.dmao_size;
4500 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4501 		*cookiep = hp->dmai_cookie[0];
4502 		hp->dmai_cookie++;
4503 		return (DDI_SUCCESS);
4504 	}
4505 
4506 	/* sync the old window before moving on to the new one */
4507 	window = &dma->dp_window[dma->dp_current_win];
4508 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4509 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4510 		    DDI_DMA_SYNC_FORCPU);
4511 	}
4512 
4513 #if !defined(__amd64)
4514 	/*
4515 	 * before we move to the next window, if we need to re-map, unmap all
4516 	 * the pages in this window.
4517 	 */
4518 	if (dma->dp_cb_remaping) {
4519 		/*
4520 		 * If we switch to this window again, we'll need to map in
4521 		 * on the fly next time.
4522 		 */
4523 		window->wd_remap_copybuf = B_TRUE;
4524 
4525 		/*
4526 		 * calculate the page index into the buffer where this window
4527 		 * starts, and the number of pages this window takes up.
4528 		 */
4529 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4530 		    MMU_PAGESHIFT;
4531 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4532 		    MMU_PAGEOFFSET;
4533 		pcnt = mmu_btopr(window->wd_size + poff);
4534 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4535 
4536 		/* unmap pages which are currently mapped in this window */
4537 		for (i = 0; i < pcnt; i++) {
4538 			if (dma->dp_pgmap[pidx].pm_mapped) {
4539 				hat_unload(kas.a_hat,
4540 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4541 				    HAT_UNLOAD);
4542 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4543 			}
4544 			pidx++;
4545 		}
4546 	}
4547 #endif
4548 
4549 	/*
4550 	 * Move to the new window.
4551 	 * NOTE: current_win must be set for sync to work right
4552 	 */
4553 	dma->dp_current_win = win;
4554 	window = &dma->dp_window[win];
4555 
4556 	/* if needed, adjust the first and/or last cookies for trim */
4557 	trim = &window->wd_trim;
4558 	if (trim->tr_trim_first) {
4559 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4560 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4561 #if !defined(__amd64)
4562 		window->wd_first_cookie->dmac_type =
4563 		    (window->wd_first_cookie->dmac_type &
4564 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4565 #endif
4566 		if (trim->tr_first_copybuf_win) {
4567 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4568 			    trim->tr_first_cbaddr;
4569 #if !defined(__amd64)
4570 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4571 			    trim->tr_first_kaddr;
4572 #endif
4573 		}
4574 	}
4575 	if (trim->tr_trim_last) {
4576 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4577 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4578 		if (trim->tr_last_copybuf_win) {
4579 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4580 			    trim->tr_last_cbaddr;
4581 #if !defined(__amd64)
4582 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4583 			    trim->tr_last_kaddr;
4584 #endif
4585 		}
4586 	}
4587 
4588 	/*
4589 	 * setup the cookie pointer to the first cookie in the window. setup
4590 	 * our return values, then increment the cookie since we return the
4591 	 * first cookie on the stack.
4592 	 */
4593 	hp->dmai_cookie = window->wd_first_cookie;
4594 	*offp = window->wd_offset;
4595 	*lenp = window->wd_size;
4596 	*ccountp = window->wd_cookie_cnt;
4597 	*cookiep = hp->dmai_cookie[0];
4598 	hp->dmai_cookie++;
4599 
4600 #if !defined(__amd64)
4601 	/* re-map copybuf if required for this window */
4602 	if (dma->dp_cb_remaping) {
4603 		/*
4604 		 * calculate the page index into the buffer where this
4605 		 * window starts.
4606 		 */
4607 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4608 		    MMU_PAGESHIFT;
4609 		ASSERT(pidx < sinfo->si_max_pages);
4610 
4611 		/*
4612 		 * the first page can get unmapped if it's shared with the
4613 		 * previous window. Even if the rest of this window is already
4614 		 * mapped in, we need to still check this one.
4615 		 */
4616 		pmap = &dma->dp_pgmap[pidx];
4617 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4618 			if (pmap->pm_pp != NULL) {
4619 				pmap->pm_mapped = B_TRUE;
4620 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4621 			} else if (pmap->pm_vaddr != NULL) {
4622 				pmap->pm_mapped = B_TRUE;
4623 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4624 				    pmap->pm_kaddr);
4625 			}
4626 		}
4627 		pidx++;
4628 
4629 		/* map in the rest of the pages if required */
4630 		if (window->wd_remap_copybuf) {
4631 			window->wd_remap_copybuf = B_FALSE;
4632 
4633 			/* figure out many pages this window takes up */
4634 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4635 			    MMU_PAGEOFFSET;
4636 			pcnt = mmu_btopr(window->wd_size + poff);
4637 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4638 
4639 			/* map pages which require it */
4640 			for (i = 1; i < pcnt; i++) {
4641 				pmap = &dma->dp_pgmap[pidx];
4642 				if (pmap->pm_uses_copybuf) {
4643 					ASSERT(pmap->pm_mapped == B_FALSE);
4644 					if (pmap->pm_pp != NULL) {
4645 						pmap->pm_mapped = B_TRUE;
4646 						i86_pp_map(pmap->pm_pp,
4647 						    pmap->pm_kaddr);
4648 					} else if (pmap->pm_vaddr != NULL) {
4649 						pmap->pm_mapped = B_TRUE;
4650 						i86_va_map(pmap->pm_vaddr,
4651 						    sinfo->si_asp,
4652 						    pmap->pm_kaddr);
4653 					}
4654 				}
4655 				pidx++;
4656 			}
4657 		}
4658 	}
4659 #endif
4660 
4661 	/* if the new window uses the copy buffer, sync it for the device */
4662 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4663 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4664 		    DDI_DMA_SYNC_FORDEV);
4665 	}
4666 
4667 	return (DDI_SUCCESS);
4668 }
4669 
4670 /*
4671  * rootnex_dma_win()
4672  *    called from ddi_dma_getwin()
4673  */
4674 /*ARGSUSED*/
4675 static int
4676 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4677     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4678     uint_t *ccountp)
4679 {
4680 #if defined(__amd64) && !defined(__xpv)
4681 	if (IOMMU_USED(rdip)) {
4682 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4683 		    cookiep, ccountp));
4684 	}
4685 #endif
4686 
4687 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4688 	    cookiep, ccountp));
4689 }
4690 
4691 /*
4692  * ************************
4693  *  obsoleted dma routines
4694  * ************************
4695  */
4696 
4697 /*
4698  * rootnex_dma_map()
4699  *    called from ddi_dma_setup()
4700  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
4701  */
4702 /* ARGSUSED */
4703 static int
4704 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
4705     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
4706 {
4707 #if defined(__amd64)
4708 	/*
4709 	 * this interface is not supported in 64-bit x86 kernel. See comment in
4710 	 * rootnex_dma_mctl()
4711 	 */
4712 	return (DDI_DMA_NORESOURCES);
4713 
4714 #else /* 32-bit x86 kernel */
4715 	ddi_dma_handle_t *lhandlep;
4716 	ddi_dma_handle_t lhandle;
4717 	ddi_dma_cookie_t cookie;
4718 	ddi_dma_attr_t dma_attr;
4719 	ddi_dma_lim_t *dma_lim;
4720 	uint_t ccnt;
4721 	int e;
4722 
4723 
4724 	/*
4725 	 * if the driver is just testing to see if it's possible to do the bind,
4726 	 * we'll use local state. Otherwise, use the handle pointer passed in.
4727 	 */
4728 	if (handlep == NULL) {
4729 		lhandlep = &lhandle;
4730 	} else {
4731 		lhandlep = handlep;
4732 	}
4733 
4734 	/* convert the limit structure to a dma_attr one */
4735 	dma_lim = dmareq->dmar_limits;
4736 	dma_attr.dma_attr_version = DMA_ATTR_V0;
4737 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
4738 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
4739 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
4740 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
4741 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
4742 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
4743 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
4744 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
4745 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
4746 	dma_attr.dma_attr_align = MMU_PAGESIZE;
4747 	dma_attr.dma_attr_flags = 0;
4748 
4749 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
4750 	    dmareq->dmar_arg, lhandlep);
4751 	if (e != DDI_SUCCESS) {
4752 		return (e);
4753 	}
4754 
4755 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
4756 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
4757 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4758 		return (e);
4759 	}
4760 
4761 	/*
4762 	 * if the driver is just testing to see if it's possible to do the bind,
4763 	 * free up the local state and return the result.
4764 	 */
4765 	if (handlep == NULL) {
4766 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
4767 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
4768 		if (e == DDI_DMA_MAPPED) {
4769 			return (DDI_DMA_MAPOK);
4770 		} else {
4771 			return (DDI_DMA_NOMAPPING);
4772 		}
4773 	}
4774 
4775 	return (e);
4776 #endif /* defined(__amd64) */
4777 }
4778 
4779 /*
4780  * rootnex_dma_mctl()
4781  *
4782  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
4783  */
4784 /* ARGSUSED */
4785 static int
4786 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4787     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
4788     uint_t cache_flags)
4789 {
4790 #if defined(__amd64)
4791 	/*
4792 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
4793 	 * common implementation in genunix, so they no longer have x86
4794 	 * specific functionality which called into dma_ctl.
4795 	 *
4796 	 * The rest of the obsoleted interfaces were never supported in the
4797 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
4798 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
4799 	 * implementation issues.
4800 	 *
4801 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
4802 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
4803 	 * reflect that now too...
4804 	 *
4805 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
4806 	 * not going to put this functionality into the 64-bit x86 kernel now.
4807 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
4808 	 * that in a future release.
4809 	 */
4810 	return (DDI_FAILURE);
4811 
4812 #else /* 32-bit x86 kernel */
4813 	ddi_dma_cookie_t lcookie;
4814 	ddi_dma_cookie_t *cookie;
4815 	rootnex_window_t *window;
4816 	ddi_dma_impl_t *hp;
4817 	rootnex_dma_t *dma;
4818 	uint_t nwin;
4819 	uint_t ccnt;
4820 	size_t len;
4821 	off_t off;
4822 	int e;
4823 
4824 
4825 	/*
4826 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
4827 	 * hacky since were optimizing for the current interfaces and so we can
4828 	 * cleanup the mess in genunix. Hopefully we will remove the this
4829 	 * obsoleted routines someday soon.
4830 	 */
4831 
4832 	switch (request) {
4833 
4834 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
4835 		hp = (ddi_dma_impl_t *)handle;
4836 		cookie = (ddi_dma_cookie_t *)objpp;
4837 
4838 		/*
4839 		 * convert segment to cookie. We don't distinguish between the
4840 		 * two :-)
4841 		 */
4842 		*cookie = *hp->dmai_cookie;
4843 		*lenp = cookie->dmac_size;
4844 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
4845 		return (DDI_SUCCESS);
4846 
4847 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
4848 		hp = (ddi_dma_impl_t *)handle;
4849 		dma = (rootnex_dma_t *)hp->dmai_private;
4850 
4851 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
4852 			return (DDI_DMA_STALE);
4853 		}
4854 
4855 		/* handle the case where we don't have any windows */
4856 		if (dma->dp_window == NULL) {
4857 			/*
4858 			 * if seg == NULL, and we don't have any windows,
4859 			 * return the first cookie in the sgl.
4860 			 */
4861 			if (*lenp == NULL) {
4862 				dma->dp_current_cookie = 0;
4863 				hp->dmai_cookie = dma->dp_cookies;
4864 				*objpp = (caddr_t)handle;
4865 				return (DDI_SUCCESS);
4866 
4867 			/* if we have more cookies, go to the next cookie */
4868 			} else {
4869 				if ((dma->dp_current_cookie + 1) >=
4870 				    dma->dp_sglinfo.si_sgl_size) {
4871 					return (DDI_DMA_DONE);
4872 				}
4873 				dma->dp_current_cookie++;
4874 				hp->dmai_cookie++;
4875 				return (DDI_SUCCESS);
4876 			}
4877 		}
4878 
4879 		/* We have one or more windows */
4880 		window = &dma->dp_window[dma->dp_current_win];
4881 
4882 		/*
4883 		 * if seg == NULL, return the first cookie in the current
4884 		 * window
4885 		 */
4886 		if (*lenp == NULL) {
4887 			dma->dp_current_cookie = 0;
4888 			hp->dmai_cookie = window->wd_first_cookie;
4889 
4890 		/*
4891 		 * go to the next cookie in the window then see if we done with
4892 		 * this window.
4893 		 */
4894 		} else {
4895 			if ((dma->dp_current_cookie + 1) >=
4896 			    window->wd_cookie_cnt) {
4897 				return (DDI_DMA_DONE);
4898 			}
4899 			dma->dp_current_cookie++;
4900 			hp->dmai_cookie++;
4901 		}
4902 		*objpp = (caddr_t)handle;
4903 		return (DDI_SUCCESS);
4904 
4905 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
4906 		hp = (ddi_dma_impl_t *)handle;
4907 		dma = (rootnex_dma_t *)hp->dmai_private;
4908 
4909 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
4910 			return (DDI_DMA_STALE);
4911 		}
4912 
4913 		/* if win == NULL, return the first window in the bind */
4914 		if (*offp == NULL) {
4915 			nwin = 0;
4916 
4917 		/*
4918 		 * else, go to the next window then see if we're done with all
4919 		 * the windows.
4920 		 */
4921 		} else {
4922 			nwin = dma->dp_current_win + 1;
4923 			if (nwin >= hp->dmai_nwin) {
4924 				return (DDI_DMA_DONE);
4925 			}
4926 		}
4927 
4928 		/* switch to the next window */
4929 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
4930 		    &lcookie, &ccnt);
4931 		ASSERT(e == DDI_SUCCESS);
4932 		if (e != DDI_SUCCESS) {
4933 			return (DDI_DMA_STALE);
4934 		}
4935 
4936 		/* reset the cookie back to the first cookie in the window */
4937 		if (dma->dp_window != NULL) {
4938 			window = &dma->dp_window[dma->dp_current_win];
4939 			hp->dmai_cookie = window->wd_first_cookie;
4940 		} else {
4941 			hp->dmai_cookie = dma->dp_cookies;
4942 		}
4943 
4944 		*objpp = (caddr_t)handle;
4945 		return (DDI_SUCCESS);
4946 
4947 	case DDI_DMA_FREE: /* ddi_dma_free() */
4948 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
4949 		(void) rootnex_dma_freehdl(dip, rdip, handle);
4950 		if (rootnex_state->r_dvma_call_list_id) {
4951 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
4952 		}
4953 		return (DDI_SUCCESS);
4954 
4955 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
4956 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
4957 		/* should never get here, handled in genunix */
4958 		ASSERT(0);
4959 		return (DDI_FAILURE);
4960 
4961 	case DDI_DMA_KVADDR:
4962 	case DDI_DMA_GETERR:
4963 	case DDI_DMA_COFF:
4964 		return (DDI_FAILURE);
4965 	}
4966 
4967 	return (DDI_FAILURE);
4968 #endif /* defined(__amd64) */
4969 }
4970 
4971 /*
4972  * *********
4973  *  FMA Code
4974  * *********
4975  */
4976 
4977 /*
4978  * rootnex_fm_init()
4979  *    FMA init busop
4980  */
4981 /* ARGSUSED */
4982 static int
4983 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
4984     ddi_iblock_cookie_t *ibc)
4985 {
4986 	*ibc = rootnex_state->r_err_ibc;
4987 
4988 	return (ddi_system_fmcap);
4989 }
4990 
4991 /*
4992  * rootnex_dma_check()
4993  *    Function called after a dma fault occurred to find out whether the
4994  *    fault address is associated with a driver that is able to handle faults
4995  *    and recover from faults.
4996  */
4997 /* ARGSUSED */
4998 static int
4999 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5000     const void *not_used)
5001 {
5002 	rootnex_window_t *window;
5003 	uint64_t start_addr;
5004 	uint64_t fault_addr;
5005 	ddi_dma_impl_t *hp;
5006 	rootnex_dma_t *dma;
5007 	uint64_t end_addr;
5008 	size_t csize;
5009 	int i;
5010 	int j;
5011 
5012 
5013 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5014 	hp = (ddi_dma_impl_t *)handle;
5015 	ASSERT(hp);
5016 
5017 	dma = (rootnex_dma_t *)hp->dmai_private;
5018 
5019 	/* Get the address that we need to search for */
5020 	fault_addr = *(uint64_t *)addr;
5021 
5022 	/*
5023 	 * if we don't have any windows, we can just walk through all the
5024 	 * cookies.
5025 	 */
5026 	if (dma->dp_window == NULL) {
5027 		/* for each cookie */
5028 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5029 			/*
5030 			 * if the faulted address is within the physical address
5031 			 * range of the cookie, return DDI_FM_NONFATAL.
5032 			 */
5033 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5034 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5035 			    dma->dp_cookies[i].dmac_size))) {
5036 				return (DDI_FM_NONFATAL);
5037 			}
5038 		}
5039 
5040 		/* fault_addr not within this DMA handle */
5041 		return (DDI_FM_UNKNOWN);
5042 	}
5043 
5044 	/* we have mutiple windows, walk through each window */
5045 	for (i = 0; i < hp->dmai_nwin; i++) {
5046 		window = &dma->dp_window[i];
5047 
5048 		/* Go through all the cookies in the window */
5049 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5050 
5051 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5052 			csize = window->wd_first_cookie[j].dmac_size;
5053 
5054 			/*
5055 			 * if we are trimming the first cookie in the window,
5056 			 * and this is the first cookie, adjust the start
5057 			 * address and size of the cookie to account for the
5058 			 * trim.
5059 			 */
5060 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5061 				start_addr = window->wd_trim.tr_first_paddr;
5062 				csize = window->wd_trim.tr_first_size;
5063 			}
5064 
5065 			/*
5066 			 * if we are trimming the last cookie in the window,
5067 			 * and this is the last cookie, adjust the start
5068 			 * address and size of the cookie to account for the
5069 			 * trim.
5070 			 */
5071 			if (window->wd_trim.tr_trim_last &&
5072 			    (j == (window->wd_cookie_cnt - 1))) {
5073 				start_addr = window->wd_trim.tr_last_paddr;
5074 				csize = window->wd_trim.tr_last_size;
5075 			}
5076 
5077 			end_addr = start_addr + csize;
5078 
5079 			/*
5080 			 * if the faulted address is within the physical
5081 			 * address of the cookie, return DDI_FM_NONFATAL.
5082 			 */
5083 			if ((fault_addr >= start_addr) &&
5084 			    (fault_addr <= end_addr)) {
5085 				return (DDI_FM_NONFATAL);
5086 			}
5087 		}
5088 	}
5089 
5090 	/* fault_addr not within this DMA handle */
5091 	return (DDI_FM_UNKNOWN);
5092 }
5093 
5094 /*ARGSUSED*/
5095 static int
5096 rootnex_quiesce(dev_info_t *dip)
5097 {
5098 #if defined(__amd64) && !defined(__xpv)
5099 	return (immu_quiesce());
5100 #else
5101 	return (DDI_SUCCESS);
5102 #endif
5103 }
5104 
5105 #if defined(__xpv)
5106 void
5107 immu_init(void)
5108 {
5109 	;
5110 }
5111 
5112 void
5113 immu_startup(void)
5114 {
5115 	;
5116 }
5117 /*ARGSUSED*/
5118 void
5119 immu_physmem_update(uint64_t addr, uint64_t size)
5120 {
5121 	;
5122 }
5123 #endif
5124