xref: /titanic_51/usr/src/uts/i86pc/io/rootnex.c (revision 50200e773f0242e336d032a7b43485e1bcfc9bfe)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * x86 root nexus driver
27  */
28 
29 #include <sys/sysmacros.h>
30 #include <sys/conf.h>
31 #include <sys/autoconf.h>
32 #include <sys/sysmacros.h>
33 #include <sys/debug.h>
34 #include <sys/psw.h>
35 #include <sys/ddidmareq.h>
36 #include <sys/promif.h>
37 #include <sys/devops.h>
38 #include <sys/kmem.h>
39 #include <sys/cmn_err.h>
40 #include <vm/seg.h>
41 #include <vm/seg_kmem.h>
42 #include <vm/seg_dev.h>
43 #include <sys/vmem.h>
44 #include <sys/mman.h>
45 #include <vm/hat.h>
46 #include <vm/as.h>
47 #include <vm/page.h>
48 #include <sys/avintr.h>
49 #include <sys/errno.h>
50 #include <sys/modctl.h>
51 #include <sys/ddi_impldefs.h>
52 #include <sys/sunddi.h>
53 #include <sys/sunndi.h>
54 #include <sys/mach_intr.h>
55 #include <sys/psm.h>
56 #include <sys/ontrap.h>
57 #include <sys/atomic.h>
58 #include <sys/sdt.h>
59 #include <sys/rootnex.h>
60 #include <vm/hat_i86.h>
61 #include <sys/ddifm.h>
62 #include <sys/ddi_isa.h>
63 #include <sys/apic.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 #ifdef DEBUG
138 /* shortcut to rootnex counters */
139 static uint64_t *rootnex_cnt;
140 #endif
141 
142 /*
143  * XXX - does x86 even need these or are they left over from the SPARC days?
144  */
145 /* statically defined integer/boolean properties for the root node */
146 static rootnex_intprop_t rootnex_intprp[] = {
147 	{ "PAGESIZE",			PAGESIZE },
148 	{ "MMU_PAGESIZE",		MMU_PAGESIZE },
149 	{ "MMU_PAGEOFFSET",		MMU_PAGEOFFSET },
150 	{ DDI_RELATIVE_ADDRESSING,	1 },
151 };
152 #define	NROOT_INTPROPS	(sizeof (rootnex_intprp) / sizeof (rootnex_intprop_t))
153 
154 /*
155  * If we're dom0, we're using a real device so we need to load
156  * the cookies with MFNs instead of PFNs.
157  */
158 #ifdef __xpv
159 typedef maddr_t rootnex_addr_t;
160 #define	ROOTNEX_PADDR_TO_RBASE(pa)	\
161 	(DOMAIN_IS_INITDOMAIN(xen_info) ? pa_to_ma(pa) : (pa))
162 #else
163 typedef paddr_t rootnex_addr_t;
164 #define	ROOTNEX_PADDR_TO_RBASE(pa)	(pa)
165 #endif
166 
167 #if !defined(__xpv)
168 char _depends_on[] = "misc/iommulib misc/acpica";
169 #endif
170 
171 static struct cb_ops rootnex_cb_ops = {
172 	nodev,		/* open */
173 	nodev,		/* close */
174 	nodev,		/* strategy */
175 	nodev,		/* print */
176 	nodev,		/* dump */
177 	nodev,		/* read */
178 	nodev,		/* write */
179 	nodev,		/* ioctl */
180 	nodev,		/* devmap */
181 	nodev,		/* mmap */
182 	nodev,		/* segmap */
183 	nochpoll,	/* chpoll */
184 	ddi_prop_op,	/* cb_prop_op */
185 	NULL,		/* struct streamtab */
186 	D_NEW | D_MP | D_HOTPLUG, /* compatibility flags */
187 	CB_REV,		/* Rev */
188 	nodev,		/* cb_aread */
189 	nodev		/* cb_awrite */
190 };
191 
192 static int rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp,
193     off_t offset, off_t len, caddr_t *vaddrp);
194 static int rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip,
195     struct hat *hat, struct seg *seg, caddr_t addr,
196     struct devpage *dp, pfn_t pfn, uint_t prot, uint_t lock);
197 static int rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
198     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep);
199 static int rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip,
200     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
201     ddi_dma_handle_t *handlep);
202 static int rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip,
203     ddi_dma_handle_t handle);
204 static int rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
205     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
206     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
207 static int rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
208     ddi_dma_handle_t handle);
209 static int rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip,
210     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
211 static int rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip,
212     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
213     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
214 static int rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip,
215     ddi_dma_handle_t handle, enum ddi_dma_ctlops request,
216     off_t *offp, size_t *lenp, caddr_t *objp, uint_t cache_flags);
217 static int rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip,
218     ddi_ctl_enum_t ctlop, void *arg, void *result);
219 static int rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
220     ddi_iblock_cookie_t *ibc);
221 static int rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip,
222     ddi_intr_op_t intr_op, ddi_intr_handle_impl_t *hdlp, void *result);
223 static int rootnex_alloc_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *,
224     void *);
225 static int rootnex_free_intr_fixed(dev_info_t *, ddi_intr_handle_impl_t *);
226 
227 static int rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
228     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
229     ddi_dma_handle_t *handlep);
230 static int rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
231     ddi_dma_handle_t handle);
232 static int rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
233     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
234     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
235 static int rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
236     ddi_dma_handle_t handle);
237 #if defined(__amd64) && !defined(__xpv)
238 static void rootnex_coredma_reset_cookies(dev_info_t *dip,
239     ddi_dma_handle_t handle);
240 static int rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
241     ddi_dma_cookie_t **cookiepp, uint_t *ccountp);
242 static int rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
243     ddi_dma_cookie_t *cookiep, uint_t ccount);
244 static int rootnex_coredma_clear_cookies(dev_info_t *dip,
245     ddi_dma_handle_t handle);
246 static int rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle);
247 #endif
248 static int rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip,
249     ddi_dma_handle_t handle, off_t off, size_t len, uint_t cache_flags);
250 static int rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip,
251     ddi_dma_handle_t handle, uint_t win, off_t *offp, size_t *lenp,
252     ddi_dma_cookie_t *cookiep, uint_t *ccountp);
253 
254 static int rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
255     ddi_dma_handle_t handle, void *v);
256 static void *rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
257     ddi_dma_handle_t handle);
258 
259 
260 static struct bus_ops rootnex_bus_ops = {
261 	BUSO_REV,
262 	rootnex_map,
263 	NULL,
264 	NULL,
265 	NULL,
266 	rootnex_map_fault,
267 	rootnex_dma_map,
268 	rootnex_dma_allochdl,
269 	rootnex_dma_freehdl,
270 	rootnex_dma_bindhdl,
271 	rootnex_dma_unbindhdl,
272 	rootnex_dma_sync,
273 	rootnex_dma_win,
274 	rootnex_dma_mctl,
275 	rootnex_ctlops,
276 	ddi_bus_prop_op,
277 	i_ddi_rootnex_get_eventcookie,
278 	i_ddi_rootnex_add_eventcall,
279 	i_ddi_rootnex_remove_eventcall,
280 	i_ddi_rootnex_post_event,
281 	0,			/* bus_intr_ctl */
282 	0,			/* bus_config */
283 	0,			/* bus_unconfig */
284 	rootnex_fm_init,	/* bus_fm_init */
285 	NULL,			/* bus_fm_fini */
286 	NULL,			/* bus_fm_access_enter */
287 	NULL,			/* bus_fm_access_exit */
288 	NULL,			/* bus_powr */
289 	rootnex_intr_ops	/* bus_intr_op */
290 };
291 
292 static int rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
293 static int rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
294 static int rootnex_quiesce(dev_info_t *dip);
295 
296 static struct dev_ops rootnex_ops = {
297 	DEVO_REV,
298 	0,
299 	ddi_no_info,
300 	nulldev,
301 	nulldev,
302 	rootnex_attach,
303 	rootnex_detach,
304 	nulldev,
305 	&rootnex_cb_ops,
306 	&rootnex_bus_ops,
307 	NULL,
308 	rootnex_quiesce,		/* quiesce */
309 };
310 
311 static struct modldrv rootnex_modldrv = {
312 	&mod_driverops,
313 	"i86pc root nexus",
314 	&rootnex_ops
315 };
316 
317 static struct modlinkage rootnex_modlinkage = {
318 	MODREV_1,
319 	(void *)&rootnex_modldrv,
320 	NULL
321 };
322 
323 #if defined(__amd64) && !defined(__xpv)
324 static iommulib_nexops_t iommulib_nexops = {
325 	IOMMU_NEXOPS_VERSION,
326 	"Rootnex IOMMU ops Vers 1.1",
327 	NULL,
328 	rootnex_coredma_allochdl,
329 	rootnex_coredma_freehdl,
330 	rootnex_coredma_bindhdl,
331 	rootnex_coredma_unbindhdl,
332 	rootnex_coredma_reset_cookies,
333 	rootnex_coredma_get_cookies,
334 	rootnex_coredma_set_cookies,
335 	rootnex_coredma_clear_cookies,
336 	rootnex_coredma_get_sleep_flags,
337 	rootnex_coredma_sync,
338 	rootnex_coredma_win,
339 	rootnex_dma_map,
340 	rootnex_dma_mctl,
341 	rootnex_coredma_hdl_setprivate,
342 	rootnex_coredma_hdl_getprivate
343 };
344 #endif
345 
346 /*
347  *  extern hacks
348  */
349 extern struct seg_ops segdev_ops;
350 extern int ignore_hardware_nodes;	/* force flag from ddi_impl.c */
351 #ifdef	DDI_MAP_DEBUG
352 extern int ddi_map_debug_flag;
353 #define	ddi_map_debug	if (ddi_map_debug_flag) prom_printf
354 #endif
355 extern void i86_pp_map(page_t *pp, caddr_t kaddr);
356 extern void i86_va_map(caddr_t vaddr, struct as *asp, caddr_t kaddr);
357 extern int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *,
358     psm_intr_op_t, int *);
359 extern int impl_ddi_sunbus_initchild(dev_info_t *dip);
360 extern void impl_ddi_sunbus_removechild(dev_info_t *dip);
361 
362 /*
363  * Use device arena to use for device control register mappings.
364  * Various kernel memory walkers (debugger, dtrace) need to know
365  * to avoid this address range to prevent undesired device activity.
366  */
367 extern void *device_arena_alloc(size_t size, int vm_flag);
368 extern void device_arena_free(void * vaddr, size_t size);
369 
370 
371 /*
372  *  Internal functions
373  */
374 static int rootnex_dma_init();
375 static void rootnex_add_props(dev_info_t *);
376 static int rootnex_ctl_reportdev(dev_info_t *dip);
377 static struct intrspec *rootnex_get_ispec(dev_info_t *rdip, int inum);
378 static int rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
379 static int rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp);
380 static int rootnex_map_handle(ddi_map_req_t *mp);
381 static void rootnex_clean_dmahdl(ddi_dma_impl_t *hp);
382 static int rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegsize);
383 static int rootnex_valid_bind_parms(ddi_dma_req_t *dmareq,
384     ddi_dma_attr_t *attr);
385 static void rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
386     rootnex_sglinfo_t *sglinfo);
387 static void rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object,
388     ddi_dma_cookie_t *sgl, rootnex_sglinfo_t *sglinfo);
389 static int rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
390     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
391 static int rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
392     rootnex_dma_t *dma, ddi_dma_attr_t *attr);
393 static void rootnex_teardown_copybuf(rootnex_dma_t *dma);
394 static int rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
395     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag);
396 static void rootnex_teardown_windows(rootnex_dma_t *dma);
397 static void rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
398     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset);
399 static void rootnex_setup_cookie(ddi_dma_obj_t *dmar_object,
400     rootnex_dma_t *dma, ddi_dma_cookie_t *cookie, off_t cur_offset,
401     size_t *copybuf_used, page_t **cur_pp);
402 static int rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp,
403     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie,
404     ddi_dma_attr_t *attr, off_t cur_offset);
405 static int rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp,
406     rootnex_dma_t *dma, rootnex_window_t **windowp,
407     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used);
408 static int rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp,
409     rootnex_dma_t *dma, rootnex_window_t **windowp, ddi_dma_cookie_t *cookie);
410 static int rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
411     off_t offset, size_t size, uint_t cache_flags);
412 static int rootnex_verify_buffer(rootnex_dma_t *dma);
413 static int rootnex_dma_check(dev_info_t *dip, const void *handle,
414     const void *comp_addr, const void *not_used);
415 static boolean_t rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object,
416     rootnex_sglinfo_t *sglinfo);
417 static struct as *rootnex_get_as(ddi_dma_obj_t *dmar_object);
418 
419 /*
420  * _init()
421  *
422  */
423 int
424 _init(void)
425 {
426 
427 	rootnex_state = NULL;
428 	return (mod_install(&rootnex_modlinkage));
429 }
430 
431 
432 /*
433  * _info()
434  *
435  */
436 int
437 _info(struct modinfo *modinfop)
438 {
439 	return (mod_info(&rootnex_modlinkage, modinfop));
440 }
441 
442 
443 /*
444  * _fini()
445  *
446  */
447 int
448 _fini(void)
449 {
450 	return (EBUSY);
451 }
452 
453 
454 /*
455  * rootnex_attach()
456  *
457  */
458 static int
459 rootnex_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
460 {
461 	int fmcap;
462 	int e;
463 
464 	switch (cmd) {
465 	case DDI_ATTACH:
466 		break;
467 	case DDI_RESUME:
468 #if defined(__amd64) && !defined(__xpv)
469 		return (immu_unquiesce());
470 #else
471 		return (DDI_SUCCESS);
472 #endif
473 	default:
474 		return (DDI_FAILURE);
475 	}
476 
477 	/*
478 	 * We should only have one instance of rootnex. Save it away since we
479 	 * don't have an easy way to get it back later.
480 	 */
481 	ASSERT(rootnex_state == NULL);
482 	rootnex_state = kmem_zalloc(sizeof (rootnex_state_t), KM_SLEEP);
483 
484 	rootnex_state->r_dip = dip;
485 	rootnex_state->r_err_ibc = (ddi_iblock_cookie_t)ipltospl(15);
486 	rootnex_state->r_reserved_msg_printed = B_FALSE;
487 #ifdef DEBUG
488 	rootnex_cnt = &rootnex_state->r_counters[0];
489 #endif
490 
491 	/*
492 	 * Set minimum fm capability level for i86pc platforms and then
493 	 * initialize error handling. Since we're the rootnex, we don't
494 	 * care what's returned in the fmcap field.
495 	 */
496 	ddi_system_fmcap = DDI_FM_EREPORT_CAPABLE | DDI_FM_ERRCB_CAPABLE |
497 	    DDI_FM_ACCCHK_CAPABLE | DDI_FM_DMACHK_CAPABLE;
498 	fmcap = ddi_system_fmcap;
499 	ddi_fm_init(dip, &fmcap, &rootnex_state->r_err_ibc);
500 
501 	/* initialize DMA related state */
502 	e = rootnex_dma_init();
503 	if (e != DDI_SUCCESS) {
504 		kmem_free(rootnex_state, sizeof (rootnex_state_t));
505 		return (DDI_FAILURE);
506 	}
507 
508 	/* Add static root node properties */
509 	rootnex_add_props(dip);
510 
511 	/* since we can't call ddi_report_dev() */
512 	cmn_err(CE_CONT, "?root nexus = %s\n", ddi_get_name(dip));
513 
514 	/* Initialize rootnex event handle */
515 	i_ddi_rootnex_init_events(dip);
516 
517 #if defined(__amd64) && !defined(__xpv)
518 	e = iommulib_nexus_register(dip, &iommulib_nexops,
519 	    &rootnex_state->r_iommulib_handle);
520 
521 	ASSERT(e == DDI_SUCCESS);
522 #endif
523 
524 	return (DDI_SUCCESS);
525 }
526 
527 
528 /*
529  * rootnex_detach()
530  *
531  */
532 /*ARGSUSED*/
533 static int
534 rootnex_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
535 {
536 	switch (cmd) {
537 	case DDI_SUSPEND:
538 #if defined(__amd64) && !defined(__xpv)
539 		return (immu_quiesce());
540 #else
541 		return (DDI_SUCCESS);
542 #endif
543 	default:
544 		return (DDI_FAILURE);
545 	}
546 	/*NOTREACHED*/
547 
548 }
549 
550 
551 /*
552  * rootnex_dma_init()
553  *
554  */
555 /*ARGSUSED*/
556 static int
557 rootnex_dma_init()
558 {
559 	size_t bufsize;
560 
561 
562 	/*
563 	 * size of our cookie/window/copybuf state needed in dma bind that we
564 	 * pre-alloc in dma_alloc_handle
565 	 */
566 	rootnex_state->r_prealloc_cookies = rootnex_prealloc_cookies;
567 	rootnex_state->r_prealloc_size =
568 	    (rootnex_state->r_prealloc_cookies * sizeof (ddi_dma_cookie_t)) +
569 	    (rootnex_prealloc_windows * sizeof (rootnex_window_t)) +
570 	    (rootnex_prealloc_copybuf * sizeof (rootnex_pgmap_t));
571 
572 	/*
573 	 * setup DDI DMA handle kmem cache, align each handle on 64 bytes,
574 	 * allocate 16 extra bytes for struct pointer alignment
575 	 * (p->dmai_private & dma->dp_prealloc_buffer)
576 	 */
577 	bufsize = sizeof (ddi_dma_impl_t) + sizeof (rootnex_dma_t) +
578 	    rootnex_state->r_prealloc_size + 0x10;
579 	rootnex_state->r_dmahdl_cache = kmem_cache_create("rootnex_dmahdl",
580 	    bufsize, 64, NULL, NULL, NULL, NULL, NULL, 0);
581 	if (rootnex_state->r_dmahdl_cache == NULL) {
582 		return (DDI_FAILURE);
583 	}
584 
585 	/*
586 	 * allocate array to track which major numbers we have printed warnings
587 	 * for.
588 	 */
589 	rootnex_warn_list = kmem_zalloc(devcnt * sizeof (*rootnex_warn_list),
590 	    KM_SLEEP);
591 
592 	return (DDI_SUCCESS);
593 }
594 
595 
596 /*
597  * rootnex_add_props()
598  *
599  */
600 static void
601 rootnex_add_props(dev_info_t *dip)
602 {
603 	rootnex_intprop_t *rpp;
604 	int i;
605 
606 	/* Add static integer/boolean properties to the root node */
607 	rpp = rootnex_intprp;
608 	for (i = 0; i < NROOT_INTPROPS; i++) {
609 		(void) e_ddi_prop_update_int(DDI_DEV_T_NONE, dip,
610 		    rpp[i].prop_name, rpp[i].prop_value);
611 	}
612 }
613 
614 
615 
616 /*
617  * *************************
618  *  ctlops related routines
619  * *************************
620  */
621 
622 /*
623  * rootnex_ctlops()
624  *
625  */
626 /*ARGSUSED*/
627 static int
628 rootnex_ctlops(dev_info_t *dip, dev_info_t *rdip, ddi_ctl_enum_t ctlop,
629     void *arg, void *result)
630 {
631 	int n, *ptr;
632 	struct ddi_parent_private_data *pdp;
633 
634 	switch (ctlop) {
635 	case DDI_CTLOPS_DMAPMAPC:
636 		/*
637 		 * Return 'partial' to indicate that dma mapping
638 		 * has to be done in the main MMU.
639 		 */
640 		return (DDI_DMA_PARTIAL);
641 
642 	case DDI_CTLOPS_BTOP:
643 		/*
644 		 * Convert byte count input to physical page units.
645 		 * (byte counts that are not a page-size multiple
646 		 * are rounded down)
647 		 */
648 		*(ulong_t *)result = btop(*(ulong_t *)arg);
649 		return (DDI_SUCCESS);
650 
651 	case DDI_CTLOPS_PTOB:
652 		/*
653 		 * Convert size in physical pages to bytes
654 		 */
655 		*(ulong_t *)result = ptob(*(ulong_t *)arg);
656 		return (DDI_SUCCESS);
657 
658 	case DDI_CTLOPS_BTOPR:
659 		/*
660 		 * Convert byte count input to physical page units
661 		 * (byte counts that are not a page-size multiple
662 		 * are rounded up)
663 		 */
664 		*(ulong_t *)result = btopr(*(ulong_t *)arg);
665 		return (DDI_SUCCESS);
666 
667 	case DDI_CTLOPS_INITCHILD:
668 		return (impl_ddi_sunbus_initchild(arg));
669 
670 	case DDI_CTLOPS_UNINITCHILD:
671 		impl_ddi_sunbus_removechild(arg);
672 		return (DDI_SUCCESS);
673 
674 	case DDI_CTLOPS_REPORTDEV:
675 		return (rootnex_ctl_reportdev(rdip));
676 
677 	case DDI_CTLOPS_IOMIN:
678 		/*
679 		 * Nothing to do here but reflect back..
680 		 */
681 		return (DDI_SUCCESS);
682 
683 	case DDI_CTLOPS_REGSIZE:
684 	case DDI_CTLOPS_NREGS:
685 		break;
686 
687 	case DDI_CTLOPS_SIDDEV:
688 		if (ndi_dev_is_prom_node(rdip))
689 			return (DDI_SUCCESS);
690 		if (ndi_dev_is_persistent_node(rdip))
691 			return (DDI_SUCCESS);
692 		return (DDI_FAILURE);
693 
694 	case DDI_CTLOPS_POWER:
695 		return ((*pm_platform_power)((power_req_t *)arg));
696 
697 	case DDI_CTLOPS_RESERVED0: /* Was DDI_CTLOPS_NINTRS, obsolete */
698 	case DDI_CTLOPS_RESERVED1: /* Was DDI_CTLOPS_POKE_INIT, obsolete */
699 	case DDI_CTLOPS_RESERVED2: /* Was DDI_CTLOPS_POKE_FLUSH, obsolete */
700 	case DDI_CTLOPS_RESERVED3: /* Was DDI_CTLOPS_POKE_FINI, obsolete */
701 	case DDI_CTLOPS_RESERVED4: /* Was DDI_CTLOPS_INTR_HILEVEL, obsolete */
702 	case DDI_CTLOPS_RESERVED5: /* Was DDI_CTLOPS_XLATE_INTRS, obsolete */
703 		if (!rootnex_state->r_reserved_msg_printed) {
704 			rootnex_state->r_reserved_msg_printed = B_TRUE;
705 			cmn_err(CE_WARN, "Failing ddi_ctlops call(s) for "
706 			    "1 or more reserved/obsolete operations.");
707 		}
708 		return (DDI_FAILURE);
709 
710 	default:
711 		return (DDI_FAILURE);
712 	}
713 	/*
714 	 * The rest are for "hardware" properties
715 	 */
716 	if ((pdp = ddi_get_parent_data(rdip)) == NULL)
717 		return (DDI_FAILURE);
718 
719 	if (ctlop == DDI_CTLOPS_NREGS) {
720 		ptr = (int *)result;
721 		*ptr = pdp->par_nreg;
722 	} else {
723 		off_t *size = (off_t *)result;
724 
725 		ptr = (int *)arg;
726 		n = *ptr;
727 		if (n >= pdp->par_nreg) {
728 			return (DDI_FAILURE);
729 		}
730 		*size = (off_t)pdp->par_reg[n].regspec_size;
731 	}
732 	return (DDI_SUCCESS);
733 }
734 
735 
736 /*
737  * rootnex_ctl_reportdev()
738  *
739  */
740 static int
741 rootnex_ctl_reportdev(dev_info_t *dev)
742 {
743 	int i, n, len, f_len = 0;
744 	char *buf;
745 
746 	buf = kmem_alloc(REPORTDEV_BUFSIZE, KM_SLEEP);
747 	f_len += snprintf(buf, REPORTDEV_BUFSIZE,
748 	    "%s%d at root", ddi_driver_name(dev), ddi_get_instance(dev));
749 	len = strlen(buf);
750 
751 	for (i = 0; i < sparc_pd_getnreg(dev); i++) {
752 
753 		struct regspec *rp = sparc_pd_getreg(dev, i);
754 
755 		if (i == 0)
756 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
757 			    ": ");
758 		else
759 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
760 			    " and ");
761 		len = strlen(buf);
762 
763 		switch (rp->regspec_bustype) {
764 
765 		case BTEISA:
766 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
767 			    "%s 0x%x", DEVI_EISA_NEXNAME, rp->regspec_addr);
768 			break;
769 
770 		case BTISA:
771 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
772 			    "%s 0x%x", DEVI_ISA_NEXNAME, rp->regspec_addr);
773 			break;
774 
775 		default:
776 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
777 			    "space %x offset %x",
778 			    rp->regspec_bustype, rp->regspec_addr);
779 			break;
780 		}
781 		len = strlen(buf);
782 	}
783 	for (i = 0, n = sparc_pd_getnintr(dev); i < n; i++) {
784 		int pri;
785 
786 		if (i != 0) {
787 			f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
788 			    ",");
789 			len = strlen(buf);
790 		}
791 		pri = INT_IPL(sparc_pd_getintr(dev, i)->intrspec_pri);
792 		f_len += snprintf(buf + len, REPORTDEV_BUFSIZE - len,
793 		    " sparc ipl %d", pri);
794 		len = strlen(buf);
795 	}
796 #ifdef DEBUG
797 	if (f_len + 1 >= REPORTDEV_BUFSIZE) {
798 		cmn_err(CE_NOTE, "next message is truncated: "
799 		    "printed length 1024, real length %d", f_len);
800 	}
801 #endif /* DEBUG */
802 	cmn_err(CE_CONT, "?%s\n", buf);
803 	kmem_free(buf, REPORTDEV_BUFSIZE);
804 	return (DDI_SUCCESS);
805 }
806 
807 
808 /*
809  * ******************
810  *  map related code
811  * ******************
812  */
813 
814 /*
815  * rootnex_map()
816  *
817  */
818 static int
819 rootnex_map(dev_info_t *dip, dev_info_t *rdip, ddi_map_req_t *mp, off_t offset,
820     off_t len, caddr_t *vaddrp)
821 {
822 	struct regspec *rp, tmp_reg;
823 	ddi_map_req_t mr = *mp;		/* Get private copy of request */
824 	int error;
825 
826 	mp = &mr;
827 
828 	switch (mp->map_op)  {
829 	case DDI_MO_MAP_LOCKED:
830 	case DDI_MO_UNMAP:
831 	case DDI_MO_MAP_HANDLE:
832 		break;
833 	default:
834 #ifdef	DDI_MAP_DEBUG
835 		cmn_err(CE_WARN, "rootnex_map: unimplemented map op %d.",
836 		    mp->map_op);
837 #endif	/* DDI_MAP_DEBUG */
838 		return (DDI_ME_UNIMPLEMENTED);
839 	}
840 
841 	if (mp->map_flags & DDI_MF_USER_MAPPING)  {
842 #ifdef	DDI_MAP_DEBUG
843 		cmn_err(CE_WARN, "rootnex_map: unimplemented map type: user.");
844 #endif	/* DDI_MAP_DEBUG */
845 		return (DDI_ME_UNIMPLEMENTED);
846 	}
847 
848 	/*
849 	 * First, if given an rnumber, convert it to a regspec...
850 	 * (Presumably, this is on behalf of a child of the root node?)
851 	 */
852 
853 	if (mp->map_type == DDI_MT_RNUMBER)  {
854 
855 		int rnumber = mp->map_obj.rnumber;
856 #ifdef	DDI_MAP_DEBUG
857 		static char *out_of_range =
858 		    "rootnex_map: Out of range rnumber <%d>, device <%s>";
859 #endif	/* DDI_MAP_DEBUG */
860 
861 		rp = i_ddi_rnumber_to_regspec(rdip, rnumber);
862 		if (rp == NULL)  {
863 #ifdef	DDI_MAP_DEBUG
864 			cmn_err(CE_WARN, out_of_range, rnumber,
865 			    ddi_get_name(rdip));
866 #endif	/* DDI_MAP_DEBUG */
867 			return (DDI_ME_RNUMBER_RANGE);
868 		}
869 
870 		/*
871 		 * Convert the given ddi_map_req_t from rnumber to regspec...
872 		 */
873 
874 		mp->map_type = DDI_MT_REGSPEC;
875 		mp->map_obj.rp = rp;
876 	}
877 
878 	/*
879 	 * Adjust offset and length correspnding to called values...
880 	 * XXX: A non-zero length means override the one in the regspec
881 	 * XXX: (regardless of what's in the parent's range?)
882 	 */
883 
884 	tmp_reg = *(mp->map_obj.rp);		/* Preserve underlying data */
885 	rp = mp->map_obj.rp = &tmp_reg;		/* Use tmp_reg in request */
886 
887 #ifdef	DDI_MAP_DEBUG
888 	cmn_err(CE_CONT, "rootnex: <%s,%s> <0x%x, 0x%x, 0x%d> offset %d len %d "
889 	    "handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
890 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size, offset,
891 	    len, mp->map_handlep);
892 #endif	/* DDI_MAP_DEBUG */
893 
894 	/*
895 	 * I/O or memory mapping:
896 	 *
897 	 *	<bustype=0, addr=x, len=x>: memory
898 	 *	<bustype=1, addr=x, len=x>: i/o
899 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
900 	 */
901 
902 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
903 		cmn_err(CE_WARN, "<%s,%s> invalid register spec"
904 		    " <0x%x, 0x%x, 0x%x>", ddi_get_name(dip),
905 		    ddi_get_name(rdip), rp->regspec_bustype,
906 		    rp->regspec_addr, rp->regspec_size);
907 		return (DDI_ME_INVAL);
908 	}
909 
910 	if (rp->regspec_bustype > 1 && rp->regspec_addr == 0) {
911 		/*
912 		 * compatibility i/o mapping
913 		 */
914 		rp->regspec_bustype += (uint_t)offset;
915 	} else {
916 		/*
917 		 * Normal memory or i/o mapping
918 		 */
919 		rp->regspec_addr += (uint_t)offset;
920 	}
921 
922 	if (len != 0)
923 		rp->regspec_size = (uint_t)len;
924 
925 #ifdef	DDI_MAP_DEBUG
926 	cmn_err(CE_CONT, "             <%s,%s> <0x%x, 0x%x, 0x%d> offset %d "
927 	    "len %d handle 0x%x\n", ddi_get_name(dip), ddi_get_name(rdip),
928 	    rp->regspec_bustype, rp->regspec_addr, rp->regspec_size,
929 	    offset, len, mp->map_handlep);
930 #endif	/* DDI_MAP_DEBUG */
931 
932 	/*
933 	 * Apply any parent ranges at this level, if applicable.
934 	 * (This is where nexus specific regspec translation takes place.
935 	 * Use of this function is implicit agreement that translation is
936 	 * provided via ddi_apply_range.)
937 	 */
938 
939 #ifdef	DDI_MAP_DEBUG
940 	ddi_map_debug("applying range of parent <%s> to child <%s>...\n",
941 	    ddi_get_name(dip), ddi_get_name(rdip));
942 #endif	/* DDI_MAP_DEBUG */
943 
944 	if ((error = i_ddi_apply_range(dip, rdip, mp->map_obj.rp)) != 0)
945 		return (error);
946 
947 	switch (mp->map_op)  {
948 	case DDI_MO_MAP_LOCKED:
949 
950 		/*
951 		 * Set up the locked down kernel mapping to the regspec...
952 		 */
953 
954 		return (rootnex_map_regspec(mp, vaddrp));
955 
956 	case DDI_MO_UNMAP:
957 
958 		/*
959 		 * Release mapping...
960 		 */
961 
962 		return (rootnex_unmap_regspec(mp, vaddrp));
963 
964 	case DDI_MO_MAP_HANDLE:
965 
966 		return (rootnex_map_handle(mp));
967 
968 	default:
969 		return (DDI_ME_UNIMPLEMENTED);
970 	}
971 }
972 
973 
974 /*
975  * rootnex_map_fault()
976  *
977  *	fault in mappings for requestors
978  */
979 /*ARGSUSED*/
980 static int
981 rootnex_map_fault(dev_info_t *dip, dev_info_t *rdip, struct hat *hat,
982     struct seg *seg, caddr_t addr, struct devpage *dp, pfn_t pfn, uint_t prot,
983     uint_t lock)
984 {
985 
986 #ifdef	DDI_MAP_DEBUG
987 	ddi_map_debug("rootnex_map_fault: address <%x> pfn <%x>", addr, pfn);
988 	ddi_map_debug(" Seg <%s>\n",
989 	    seg->s_ops == &segdev_ops ? "segdev" :
990 	    seg == &kvseg ? "segkmem" : "NONE!");
991 #endif	/* DDI_MAP_DEBUG */
992 
993 	/*
994 	 * This is all terribly broken, but it is a start
995 	 *
996 	 * XXX	Note that this test means that segdev_ops
997 	 *	must be exported from seg_dev.c.
998 	 * XXX	What about devices with their own segment drivers?
999 	 */
1000 	if (seg->s_ops == &segdev_ops) {
1001 		struct segdev_data *sdp = (struct segdev_data *)seg->s_data;
1002 
1003 		if (hat == NULL) {
1004 			/*
1005 			 * This is one plausible interpretation of
1006 			 * a null hat i.e. use the first hat on the
1007 			 * address space hat list which by convention is
1008 			 * the hat of the system MMU.  At alternative
1009 			 * would be to panic .. this might well be better ..
1010 			 */
1011 			ASSERT(AS_READ_HELD(seg->s_as, &seg->s_as->a_lock));
1012 			hat = seg->s_as->a_hat;
1013 			cmn_err(CE_NOTE, "rootnex_map_fault: nil hat");
1014 		}
1015 		hat_devload(hat, addr, MMU_PAGESIZE, pfn, prot | sdp->hat_attr,
1016 		    (lock ? HAT_LOAD_LOCK : HAT_LOAD));
1017 	} else if (seg == &kvseg && dp == NULL) {
1018 		hat_devload(kas.a_hat, addr, MMU_PAGESIZE, pfn, prot,
1019 		    HAT_LOAD_LOCK);
1020 	} else
1021 		return (DDI_FAILURE);
1022 	return (DDI_SUCCESS);
1023 }
1024 
1025 
1026 /*
1027  * rootnex_map_regspec()
1028  *     we don't support mapping of I/O cards above 4Gb
1029  */
1030 static int
1031 rootnex_map_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1032 {
1033 	rootnex_addr_t rbase;
1034 	void *cvaddr;
1035 	uint_t npages, pgoffset;
1036 	struct regspec *rp;
1037 	ddi_acc_hdl_t *hp;
1038 	ddi_acc_impl_t *ap;
1039 	uint_t	hat_acc_flags;
1040 	paddr_t pbase;
1041 
1042 	rp = mp->map_obj.rp;
1043 	hp = mp->map_handlep;
1044 
1045 #ifdef	DDI_MAP_DEBUG
1046 	ddi_map_debug(
1047 	    "rootnex_map_regspec: <0x%x 0x%x 0x%x> handle 0x%x\n",
1048 	    rp->regspec_bustype, rp->regspec_addr,
1049 	    rp->regspec_size, mp->map_handlep);
1050 #endif	/* DDI_MAP_DEBUG */
1051 
1052 	/*
1053 	 * I/O or memory mapping
1054 	 *
1055 	 *	<bustype=0, addr=x, len=x>: memory
1056 	 *	<bustype=1, addr=x, len=x>: i/o
1057 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1058 	 */
1059 
1060 	if (rp->regspec_bustype > 1 && rp->regspec_addr != 0) {
1061 		cmn_err(CE_WARN, "rootnex: invalid register spec"
1062 		    " <0x%x, 0x%x, 0x%x>", rp->regspec_bustype,
1063 		    rp->regspec_addr, rp->regspec_size);
1064 		return (DDI_FAILURE);
1065 	}
1066 
1067 	if (rp->regspec_bustype != 0) {
1068 		/*
1069 		 * I/O space - needs a handle.
1070 		 */
1071 		if (hp == NULL) {
1072 			return (DDI_FAILURE);
1073 		}
1074 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1075 		ap->ahi_acc_attr |= DDI_ACCATTR_IO_SPACE;
1076 		impl_acc_hdl_init(hp);
1077 
1078 		if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1079 #ifdef  DDI_MAP_DEBUG
1080 			ddi_map_debug("rootnex_map_regspec: mmap() "
1081 			    "to I/O space is not supported.\n");
1082 #endif  /* DDI_MAP_DEBUG */
1083 			return (DDI_ME_INVAL);
1084 		} else {
1085 			/*
1086 			 * 1275-compliant vs. compatibility i/o mapping
1087 			 */
1088 			*vaddrp =
1089 			    (rp->regspec_bustype > 1 && rp->regspec_addr == 0) ?
1090 			    ((caddr_t)(uintptr_t)rp->regspec_bustype) :
1091 			    ((caddr_t)(uintptr_t)rp->regspec_addr);
1092 #ifdef __xpv
1093 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1094 				hp->ah_pfn = xen_assign_pfn(
1095 				    mmu_btop((ulong_t)rp->regspec_addr &
1096 				    MMU_PAGEMASK));
1097 			} else {
1098 				hp->ah_pfn = mmu_btop(
1099 				    (ulong_t)rp->regspec_addr & MMU_PAGEMASK);
1100 			}
1101 #else
1102 			hp->ah_pfn = mmu_btop((ulong_t)rp->regspec_addr &
1103 			    MMU_PAGEMASK);
1104 #endif
1105 			hp->ah_pnum = mmu_btopr(rp->regspec_size +
1106 			    (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET);
1107 		}
1108 
1109 #ifdef	DDI_MAP_DEBUG
1110 		ddi_map_debug(
1111 	    "rootnex_map_regspec: \"Mapping\" %d bytes I/O space at 0x%x\n",
1112 		    rp->regspec_size, *vaddrp);
1113 #endif	/* DDI_MAP_DEBUG */
1114 		return (DDI_SUCCESS);
1115 	}
1116 
1117 	/*
1118 	 * Memory space
1119 	 */
1120 
1121 	if (hp != NULL) {
1122 		/*
1123 		 * hat layer ignores
1124 		 * hp->ah_acc.devacc_attr_endian_flags.
1125 		 */
1126 		switch (hp->ah_acc.devacc_attr_dataorder) {
1127 		case DDI_STRICTORDER_ACC:
1128 			hat_acc_flags = HAT_STRICTORDER;
1129 			break;
1130 		case DDI_UNORDERED_OK_ACC:
1131 			hat_acc_flags = HAT_UNORDERED_OK;
1132 			break;
1133 		case DDI_MERGING_OK_ACC:
1134 			hat_acc_flags = HAT_MERGING_OK;
1135 			break;
1136 		case DDI_LOADCACHING_OK_ACC:
1137 			hat_acc_flags = HAT_LOADCACHING_OK;
1138 			break;
1139 		case DDI_STORECACHING_OK_ACC:
1140 			hat_acc_flags = HAT_STORECACHING_OK;
1141 			break;
1142 		}
1143 		ap = (ddi_acc_impl_t *)hp->ah_platform_private;
1144 		ap->ahi_acc_attr |= DDI_ACCATTR_CPU_VADDR;
1145 		impl_acc_hdl_init(hp);
1146 		hp->ah_hat_flags = hat_acc_flags;
1147 	} else {
1148 		hat_acc_flags = HAT_STRICTORDER;
1149 	}
1150 
1151 	rbase = (rootnex_addr_t)(rp->regspec_addr & MMU_PAGEMASK);
1152 #ifdef __xpv
1153 	/*
1154 	 * If we're dom0, we're using a real device so we need to translate
1155 	 * the MA to a PA.
1156 	 */
1157 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1158 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase)));
1159 	} else {
1160 		pbase = rbase;
1161 	}
1162 #else
1163 	pbase = rbase;
1164 #endif
1165 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1166 
1167 	if (rp->regspec_size == 0) {
1168 #ifdef  DDI_MAP_DEBUG
1169 		ddi_map_debug("rootnex_map_regspec: zero regspec_size\n");
1170 #endif  /* DDI_MAP_DEBUG */
1171 		return (DDI_ME_INVAL);
1172 	}
1173 
1174 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING) {
1175 		/* extra cast to make gcc happy */
1176 		*vaddrp = (caddr_t)((uintptr_t)mmu_btop(pbase));
1177 	} else {
1178 		npages = mmu_btopr(rp->regspec_size + pgoffset);
1179 
1180 #ifdef	DDI_MAP_DEBUG
1181 		ddi_map_debug("rootnex_map_regspec: Mapping %d pages "
1182 		    "physical %llx", npages, pbase);
1183 #endif	/* DDI_MAP_DEBUG */
1184 
1185 		cvaddr = device_arena_alloc(ptob(npages), VM_NOSLEEP);
1186 		if (cvaddr == NULL)
1187 			return (DDI_ME_NORESOURCES);
1188 
1189 		/*
1190 		 * Now map in the pages we've allocated...
1191 		 */
1192 		hat_devload(kas.a_hat, cvaddr, mmu_ptob(npages),
1193 		    mmu_btop(pbase), mp->map_prot | hat_acc_flags,
1194 		    HAT_LOAD_LOCK);
1195 		*vaddrp = (caddr_t)cvaddr + pgoffset;
1196 
1197 		/* save away pfn and npages for FMA */
1198 		hp = mp->map_handlep;
1199 		if (hp) {
1200 			hp->ah_pfn = mmu_btop(pbase);
1201 			hp->ah_pnum = npages;
1202 		}
1203 	}
1204 
1205 #ifdef	DDI_MAP_DEBUG
1206 	ddi_map_debug("at virtual 0x%x\n", *vaddrp);
1207 #endif	/* DDI_MAP_DEBUG */
1208 	return (DDI_SUCCESS);
1209 }
1210 
1211 
1212 /*
1213  * rootnex_unmap_regspec()
1214  *
1215  */
1216 static int
1217 rootnex_unmap_regspec(ddi_map_req_t *mp, caddr_t *vaddrp)
1218 {
1219 	caddr_t addr = (caddr_t)*vaddrp;
1220 	uint_t npages, pgoffset;
1221 	struct regspec *rp;
1222 
1223 	if (mp->map_flags & DDI_MF_DEVICE_MAPPING)
1224 		return (0);
1225 
1226 	rp = mp->map_obj.rp;
1227 
1228 	if (rp->regspec_size == 0) {
1229 #ifdef  DDI_MAP_DEBUG
1230 		ddi_map_debug("rootnex_unmap_regspec: zero regspec_size\n");
1231 #endif  /* DDI_MAP_DEBUG */
1232 		return (DDI_ME_INVAL);
1233 	}
1234 
1235 	/*
1236 	 * I/O or memory mapping:
1237 	 *
1238 	 *	<bustype=0, addr=x, len=x>: memory
1239 	 *	<bustype=1, addr=x, len=x>: i/o
1240 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1241 	 */
1242 	if (rp->regspec_bustype != 0) {
1243 		/*
1244 		 * This is I/O space, which requires no particular
1245 		 * processing on unmap since it isn't mapped in the
1246 		 * first place.
1247 		 */
1248 		return (DDI_SUCCESS);
1249 	}
1250 
1251 	/*
1252 	 * Memory space
1253 	 */
1254 	pgoffset = (uintptr_t)addr & MMU_PAGEOFFSET;
1255 	npages = mmu_btopr(rp->regspec_size + pgoffset);
1256 	hat_unload(kas.a_hat, addr - pgoffset, ptob(npages), HAT_UNLOAD_UNLOCK);
1257 	device_arena_free(addr - pgoffset, ptob(npages));
1258 
1259 	/*
1260 	 * Destroy the pointer - the mapping has logically gone
1261 	 */
1262 	*vaddrp = NULL;
1263 
1264 	return (DDI_SUCCESS);
1265 }
1266 
1267 
1268 /*
1269  * rootnex_map_handle()
1270  *
1271  */
1272 static int
1273 rootnex_map_handle(ddi_map_req_t *mp)
1274 {
1275 	rootnex_addr_t rbase;
1276 	ddi_acc_hdl_t *hp;
1277 	uint_t pgoffset;
1278 	struct regspec *rp;
1279 	paddr_t pbase;
1280 
1281 	rp = mp->map_obj.rp;
1282 
1283 #ifdef	DDI_MAP_DEBUG
1284 	ddi_map_debug(
1285 	    "rootnex_map_handle: <0x%x 0x%x 0x%x> handle 0x%x\n",
1286 	    rp->regspec_bustype, rp->regspec_addr,
1287 	    rp->regspec_size, mp->map_handlep);
1288 #endif	/* DDI_MAP_DEBUG */
1289 
1290 	/*
1291 	 * I/O or memory mapping:
1292 	 *
1293 	 *	<bustype=0, addr=x, len=x>: memory
1294 	 *	<bustype=1, addr=x, len=x>: i/o
1295 	 *	<bustype>1, addr=0, len=x>: x86-compatibility i/o
1296 	 */
1297 	if (rp->regspec_bustype != 0) {
1298 		/*
1299 		 * This refers to I/O space, and we don't support "mapping"
1300 		 * I/O space to a user.
1301 		 */
1302 		return (DDI_FAILURE);
1303 	}
1304 
1305 	/*
1306 	 * Set up the hat_flags for the mapping.
1307 	 */
1308 	hp = mp->map_handlep;
1309 
1310 	switch (hp->ah_acc.devacc_attr_endian_flags) {
1311 	case DDI_NEVERSWAP_ACC:
1312 		hp->ah_hat_flags = HAT_NEVERSWAP | HAT_STRICTORDER;
1313 		break;
1314 	case DDI_STRUCTURE_LE_ACC:
1315 		hp->ah_hat_flags = HAT_STRUCTURE_LE;
1316 		break;
1317 	case DDI_STRUCTURE_BE_ACC:
1318 		return (DDI_FAILURE);
1319 	default:
1320 		return (DDI_REGS_ACC_CONFLICT);
1321 	}
1322 
1323 	switch (hp->ah_acc.devacc_attr_dataorder) {
1324 	case DDI_STRICTORDER_ACC:
1325 		break;
1326 	case DDI_UNORDERED_OK_ACC:
1327 		hp->ah_hat_flags |= HAT_UNORDERED_OK;
1328 		break;
1329 	case DDI_MERGING_OK_ACC:
1330 		hp->ah_hat_flags |= HAT_MERGING_OK;
1331 		break;
1332 	case DDI_LOADCACHING_OK_ACC:
1333 		hp->ah_hat_flags |= HAT_LOADCACHING_OK;
1334 		break;
1335 	case DDI_STORECACHING_OK_ACC:
1336 		hp->ah_hat_flags |= HAT_STORECACHING_OK;
1337 		break;
1338 	default:
1339 		return (DDI_FAILURE);
1340 	}
1341 
1342 	rbase = (rootnex_addr_t)rp->regspec_addr &
1343 	    (~(rootnex_addr_t)MMU_PAGEOFFSET);
1344 	pgoffset = (ulong_t)rp->regspec_addr & MMU_PAGEOFFSET;
1345 
1346 	if (rp->regspec_size == 0)
1347 		return (DDI_ME_INVAL);
1348 
1349 #ifdef __xpv
1350 	/*
1351 	 * If we're dom0, we're using a real device so we need to translate
1352 	 * the MA to a PA.
1353 	 */
1354 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1355 		pbase = pfn_to_pa(xen_assign_pfn(mmu_btop(rbase))) |
1356 		    (rbase & MMU_PAGEOFFSET);
1357 	} else {
1358 		pbase = rbase;
1359 	}
1360 #else
1361 	pbase = rbase;
1362 #endif
1363 
1364 	hp->ah_pfn = mmu_btop(pbase);
1365 	hp->ah_pnum = mmu_btopr(rp->regspec_size + pgoffset);
1366 
1367 	return (DDI_SUCCESS);
1368 }
1369 
1370 
1371 
1372 /*
1373  * ************************
1374  *  interrupt related code
1375  * ************************
1376  */
1377 
1378 /*
1379  * rootnex_intr_ops()
1380  *	bus_intr_op() function for interrupt support
1381  */
1382 /* ARGSUSED */
1383 static int
1384 rootnex_intr_ops(dev_info_t *pdip, dev_info_t *rdip, ddi_intr_op_t intr_op,
1385     ddi_intr_handle_impl_t *hdlp, void *result)
1386 {
1387 	struct intrspec			*ispec;
1388 
1389 	DDI_INTR_NEXDBG((CE_CONT,
1390 	    "rootnex_intr_ops: pdip = %p, rdip = %p, intr_op = %x, hdlp = %p\n",
1391 	    (void *)pdip, (void *)rdip, intr_op, (void *)hdlp));
1392 
1393 	/* Process the interrupt operation */
1394 	switch (intr_op) {
1395 	case DDI_INTROP_GETCAP:
1396 		/* First check with pcplusmp */
1397 		if (psm_intr_ops == NULL)
1398 			return (DDI_FAILURE);
1399 
1400 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_CAP, result)) {
1401 			*(int *)result = 0;
1402 			return (DDI_FAILURE);
1403 		}
1404 		break;
1405 	case DDI_INTROP_SETCAP:
1406 		if (psm_intr_ops == NULL)
1407 			return (DDI_FAILURE);
1408 
1409 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_CAP, result))
1410 			return (DDI_FAILURE);
1411 		break;
1412 	case DDI_INTROP_ALLOC:
1413 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1414 		return (rootnex_alloc_intr_fixed(rdip, hdlp, result));
1415 	case DDI_INTROP_FREE:
1416 		ASSERT(hdlp->ih_type == DDI_INTR_TYPE_FIXED);
1417 		return (rootnex_free_intr_fixed(rdip, hdlp));
1418 	case DDI_INTROP_GETPRI:
1419 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1420 			return (DDI_FAILURE);
1421 		*(int *)result = ispec->intrspec_pri;
1422 		break;
1423 	case DDI_INTROP_SETPRI:
1424 		/* Validate the interrupt priority passed to us */
1425 		if (*(int *)result > LOCK_LEVEL)
1426 			return (DDI_FAILURE);
1427 
1428 		/* Ensure that PSM is all initialized and ispec is ok */
1429 		if ((psm_intr_ops == NULL) ||
1430 		    ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL))
1431 			return (DDI_FAILURE);
1432 
1433 		/* Change the priority */
1434 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_PRI, result) ==
1435 		    PSM_FAILURE)
1436 			return (DDI_FAILURE);
1437 
1438 		/* update the ispec with the new priority */
1439 		ispec->intrspec_pri =  *(int *)result;
1440 		break;
1441 	case DDI_INTROP_ADDISR:
1442 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1443 			return (DDI_FAILURE);
1444 		ispec->intrspec_func = hdlp->ih_cb_func;
1445 		break;
1446 	case DDI_INTROP_REMISR:
1447 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1448 			return (DDI_FAILURE);
1449 		ispec->intrspec_func = (uint_t (*)()) 0;
1450 		break;
1451 	case DDI_INTROP_ENABLE:
1452 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1453 			return (DDI_FAILURE);
1454 
1455 		/* Call psmi to translate irq with the dip */
1456 		if (psm_intr_ops == NULL)
1457 			return (DDI_FAILURE);
1458 
1459 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1460 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_XLATE_VECTOR,
1461 		    (int *)&hdlp->ih_vector) == PSM_FAILURE)
1462 			return (DDI_FAILURE);
1463 
1464 		/* Add the interrupt handler */
1465 		if (!add_avintr((void *)hdlp, ispec->intrspec_pri,
1466 		    hdlp->ih_cb_func, DEVI(rdip)->devi_name, hdlp->ih_vector,
1467 		    hdlp->ih_cb_arg1, hdlp->ih_cb_arg2, NULL, rdip))
1468 			return (DDI_FAILURE);
1469 		break;
1470 	case DDI_INTROP_DISABLE:
1471 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1472 			return (DDI_FAILURE);
1473 
1474 		/* Call psm_ops() to translate irq with the dip */
1475 		if (psm_intr_ops == NULL)
1476 			return (DDI_FAILURE);
1477 
1478 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1479 		(void) (*psm_intr_ops)(rdip, hdlp,
1480 		    PSM_INTR_OP_XLATE_VECTOR, (int *)&hdlp->ih_vector);
1481 
1482 		/* Remove the interrupt handler */
1483 		rem_avintr((void *)hdlp, ispec->intrspec_pri,
1484 		    hdlp->ih_cb_func, hdlp->ih_vector);
1485 		break;
1486 	case DDI_INTROP_SETMASK:
1487 		if (psm_intr_ops == NULL)
1488 			return (DDI_FAILURE);
1489 
1490 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_SET_MASK, NULL))
1491 			return (DDI_FAILURE);
1492 		break;
1493 	case DDI_INTROP_CLRMASK:
1494 		if (psm_intr_ops == NULL)
1495 			return (DDI_FAILURE);
1496 
1497 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_CLEAR_MASK, NULL))
1498 			return (DDI_FAILURE);
1499 		break;
1500 	case DDI_INTROP_GETPENDING:
1501 		if (psm_intr_ops == NULL)
1502 			return (DDI_FAILURE);
1503 
1504 		if ((*psm_intr_ops)(rdip, hdlp, PSM_INTR_OP_GET_PENDING,
1505 		    result)) {
1506 			*(int *)result = 0;
1507 			return (DDI_FAILURE);
1508 		}
1509 		break;
1510 	case DDI_INTROP_NAVAIL:
1511 	case DDI_INTROP_NINTRS:
1512 		*(int *)result = i_ddi_get_intx_nintrs(rdip);
1513 		if (*(int *)result == 0) {
1514 			/*
1515 			 * Special case for 'pcic' driver' only. This driver
1516 			 * driver is a child of 'isa' and 'rootnex' drivers.
1517 			 *
1518 			 * See detailed comments on this in the function
1519 			 * rootnex_get_ispec().
1520 			 *
1521 			 * Children of 'pcic' send 'NINITR' request all the
1522 			 * way to rootnex driver. But, the 'pdp->par_nintr'
1523 			 * field may not initialized. So, we fake it here
1524 			 * to return 1 (a la what PCMCIA nexus does).
1525 			 */
1526 			if (strcmp(ddi_get_name(rdip), "pcic") == 0)
1527 				*(int *)result = 1;
1528 			else
1529 				return (DDI_FAILURE);
1530 		}
1531 		break;
1532 	case DDI_INTROP_SUPPORTED_TYPES:
1533 		*(int *)result = DDI_INTR_TYPE_FIXED;	/* Always ... */
1534 		break;
1535 	default:
1536 		return (DDI_FAILURE);
1537 	}
1538 
1539 	return (DDI_SUCCESS);
1540 }
1541 
1542 
1543 /*
1544  * rootnex_get_ispec()
1545  *	convert an interrupt number to an interrupt specification.
1546  *	The interrupt number determines which interrupt spec will be
1547  *	returned if more than one exists.
1548  *
1549  *	Look into the parent private data area of the 'rdip' to find out
1550  *	the interrupt specification.  First check to make sure there is
1551  *	one that matchs "inumber" and then return a pointer to it.
1552  *
1553  *	Return NULL if one could not be found.
1554  *
1555  *	NOTE: This is needed for rootnex_intr_ops()
1556  */
1557 static struct intrspec *
1558 rootnex_get_ispec(dev_info_t *rdip, int inum)
1559 {
1560 	struct ddi_parent_private_data *pdp = ddi_get_parent_data(rdip);
1561 
1562 	/*
1563 	 * Special case handling for drivers that provide their own
1564 	 * intrspec structures instead of relying on the DDI framework.
1565 	 *
1566 	 * A broken hardware driver in ON could potentially provide its
1567 	 * own intrspec structure, instead of relying on the hardware.
1568 	 * If these drivers are children of 'rootnex' then we need to
1569 	 * continue to provide backward compatibility to them here.
1570 	 *
1571 	 * Following check is a special case for 'pcic' driver which
1572 	 * was found to have broken hardwre andby provides its own intrspec.
1573 	 *
1574 	 * Verbatim comments from this driver are shown here:
1575 	 * "Don't use the ddi_add_intr since we don't have a
1576 	 * default intrspec in all cases."
1577 	 *
1578 	 * Since an 'ispec' may not be always created for it,
1579 	 * check for that and create one if so.
1580 	 *
1581 	 * NOTE: Currently 'pcic' is the only driver found to do this.
1582 	 */
1583 	if (!pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1584 		pdp->par_nintr = 1;
1585 		pdp->par_intr = kmem_zalloc(sizeof (struct intrspec) *
1586 		    pdp->par_nintr, KM_SLEEP);
1587 	}
1588 
1589 	/* Validate the interrupt number */
1590 	if (inum >= pdp->par_nintr)
1591 		return (NULL);
1592 
1593 	/* Get the interrupt structure pointer and return that */
1594 	return ((struct intrspec *)&pdp->par_intr[inum]);
1595 }
1596 
1597 /*
1598  * Allocate interrupt vector for FIXED (legacy) type.
1599  */
1600 static int
1601 rootnex_alloc_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp,
1602     void *result)
1603 {
1604 	struct intrspec		*ispec;
1605 	ddi_intr_handle_impl_t	info_hdl;
1606 	int			ret;
1607 	int			free_phdl = 0;
1608 	apic_get_type_t		type_info;
1609 
1610 	if (psm_intr_ops == NULL)
1611 		return (DDI_FAILURE);
1612 
1613 	if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1614 		return (DDI_FAILURE);
1615 
1616 	/*
1617 	 * If the PSM module is "APIX" then pass the request for it
1618 	 * to allocate the vector now.
1619 	 */
1620 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1621 	info_hdl.ih_private = &type_info;
1622 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1623 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1624 		if (hdlp->ih_private == NULL) { /* allocate phdl structure */
1625 			free_phdl = 1;
1626 			i_ddi_alloc_intr_phdl(hdlp);
1627 		}
1628 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1629 		ret = (*psm_intr_ops)(rdip, hdlp,
1630 		    PSM_INTR_OP_ALLOC_VECTORS, result);
1631 		if (free_phdl) { /* free up the phdl structure */
1632 			free_phdl = 0;
1633 			i_ddi_free_intr_phdl(hdlp);
1634 			hdlp->ih_private = NULL;
1635 		}
1636 	} else {
1637 		/*
1638 		 * No APIX module; fall back to the old scheme where the
1639 		 * interrupt vector is allocated during ddi_enable_intr() call.
1640 		 */
1641 		hdlp->ih_pri = ispec->intrspec_pri;
1642 		*(int *)result = hdlp->ih_scratch1;
1643 		ret = DDI_SUCCESS;
1644 	}
1645 
1646 	return (ret);
1647 }
1648 
1649 /*
1650  * Free up interrupt vector for FIXED (legacy) type.
1651  */
1652 static int
1653 rootnex_free_intr_fixed(dev_info_t *rdip, ddi_intr_handle_impl_t *hdlp)
1654 {
1655 	struct intrspec			*ispec;
1656 	struct ddi_parent_private_data	*pdp;
1657 	ddi_intr_handle_impl_t		info_hdl;
1658 	int				ret;
1659 	apic_get_type_t			type_info;
1660 
1661 	if (psm_intr_ops == NULL)
1662 		return (DDI_FAILURE);
1663 
1664 	/*
1665 	 * If the PSM module is "APIX" then pass the request for it
1666 	 * to free up the vector now.
1667 	 */
1668 	bzero(&info_hdl, sizeof (ddi_intr_handle_impl_t));
1669 	info_hdl.ih_private = &type_info;
1670 	if ((*psm_intr_ops)(NULL, &info_hdl, PSM_INTR_OP_APIC_TYPE, NULL) ==
1671 	    PSM_SUCCESS && strcmp(type_info.avgi_type, APIC_APIX_NAME) == 0) {
1672 		if ((ispec = rootnex_get_ispec(rdip, hdlp->ih_inum)) == NULL)
1673 			return (DDI_FAILURE);
1674 		((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp = ispec;
1675 		ret = (*psm_intr_ops)(rdip, hdlp,
1676 		    PSM_INTR_OP_FREE_VECTORS, NULL);
1677 	} else {
1678 		/*
1679 		 * No APIX module; fall back to the old scheme where
1680 		 * the interrupt vector was already freed during
1681 		 * ddi_disable_intr() call.
1682 		 */
1683 		ret = DDI_SUCCESS;
1684 	}
1685 
1686 	pdp = ddi_get_parent_data(rdip);
1687 
1688 	/*
1689 	 * Special case for 'pcic' driver' only.
1690 	 * If an intrspec was created for it, clean it up here
1691 	 * See detailed comments on this in the function
1692 	 * rootnex_get_ispec().
1693 	 */
1694 	if (pdp->par_intr && strcmp(ddi_get_name(rdip), "pcic") == 0) {
1695 		kmem_free(pdp->par_intr, sizeof (struct intrspec) *
1696 		    pdp->par_nintr);
1697 		/*
1698 		 * Set it to zero; so that
1699 		 * DDI framework doesn't free it again
1700 		 */
1701 		pdp->par_intr = NULL;
1702 		pdp->par_nintr = 0;
1703 	}
1704 
1705 	return (ret);
1706 }
1707 
1708 
1709 /*
1710  * ******************
1711  *  dma related code
1712  * ******************
1713  */
1714 
1715 /*ARGSUSED*/
1716 static int
1717 rootnex_coredma_allochdl(dev_info_t *dip, dev_info_t *rdip,
1718     ddi_dma_attr_t *attr, int (*waitfp)(caddr_t), caddr_t arg,
1719     ddi_dma_handle_t *handlep)
1720 {
1721 	uint64_t maxsegmentsize_ll;
1722 	uint_t maxsegmentsize;
1723 	ddi_dma_impl_t *hp;
1724 	rootnex_dma_t *dma;
1725 	uint64_t count_max;
1726 	uint64_t seg;
1727 	int kmflag;
1728 	int e;
1729 
1730 
1731 	/* convert our sleep flags */
1732 	if (waitfp == DDI_DMA_SLEEP) {
1733 		kmflag = KM_SLEEP;
1734 	} else {
1735 		kmflag = KM_NOSLEEP;
1736 	}
1737 
1738 	/*
1739 	 * We try to do only one memory allocation here. We'll do a little
1740 	 * pointer manipulation later. If the bind ends up taking more than
1741 	 * our prealloc's space, we'll have to allocate more memory in the
1742 	 * bind operation. Not great, but much better than before and the
1743 	 * best we can do with the current bind interfaces.
1744 	 */
1745 	hp = kmem_cache_alloc(rootnex_state->r_dmahdl_cache, kmflag);
1746 	if (hp == NULL)
1747 		return (DDI_DMA_NORESOURCES);
1748 
1749 	/* Do our pointer manipulation now, align the structures */
1750 	hp->dmai_private = (void *)(((uintptr_t)hp +
1751 	    (uintptr_t)sizeof (ddi_dma_impl_t) + 0x7) & ~0x7);
1752 	dma = (rootnex_dma_t *)hp->dmai_private;
1753 	dma->dp_prealloc_buffer = (uchar_t *)(((uintptr_t)dma +
1754 	    sizeof (rootnex_dma_t) + 0x7) & ~0x7);
1755 
1756 	/* setup the handle */
1757 	rootnex_clean_dmahdl(hp);
1758 	hp->dmai_error.err_fep = NULL;
1759 	hp->dmai_error.err_cf = NULL;
1760 	dma->dp_dip = rdip;
1761 	dma->dp_sglinfo.si_flags = attr->dma_attr_flags;
1762 	dma->dp_sglinfo.si_min_addr = attr->dma_attr_addr_lo;
1763 
1764 	/*
1765 	 * The BOUNCE_ON_SEG workaround is not needed when an IOMMU
1766 	 * is being used. Set the upper limit to the seg value.
1767 	 * There will be enough DVMA space to always get addresses
1768 	 * that will match the constraints.
1769 	 */
1770 	if (IOMMU_USED(rdip) &&
1771 	    (attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG)) {
1772 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_seg;
1773 		dma->dp_sglinfo.si_flags &= ~_DDI_DMA_BOUNCE_ON_SEG;
1774 	} else
1775 		dma->dp_sglinfo.si_max_addr = attr->dma_attr_addr_hi;
1776 
1777 	hp->dmai_minxfer = attr->dma_attr_minxfer;
1778 	hp->dmai_burstsizes = attr->dma_attr_burstsizes;
1779 	hp->dmai_rdip = rdip;
1780 	hp->dmai_attr = *attr;
1781 
1782 	if (attr->dma_attr_seg >= dma->dp_sglinfo.si_max_addr)
1783 		dma->dp_sglinfo.si_cancross = B_FALSE;
1784 	else
1785 		dma->dp_sglinfo.si_cancross = B_TRUE;
1786 
1787 	/* we don't need to worry about the SPL since we do a tryenter */
1788 	mutex_init(&dma->dp_mutex, NULL, MUTEX_DRIVER, NULL);
1789 
1790 	/*
1791 	 * Figure out our maximum segment size. If the segment size is greater
1792 	 * than 4G, we will limit it to (4G - 1) since the max size of a dma
1793 	 * object (ddi_dma_obj_t.dmao_size) is 32 bits. dma_attr_seg and
1794 	 * dma_attr_count_max are size-1 type values.
1795 	 *
1796 	 * Maximum segment size is the largest physically contiguous chunk of
1797 	 * memory that we can return from a bind (i.e. the maximum size of a
1798 	 * single cookie).
1799 	 */
1800 
1801 	/* handle the rollover cases */
1802 	seg = attr->dma_attr_seg + 1;
1803 	if (seg < attr->dma_attr_seg) {
1804 		seg = attr->dma_attr_seg;
1805 	}
1806 	count_max = attr->dma_attr_count_max + 1;
1807 	if (count_max < attr->dma_attr_count_max) {
1808 		count_max = attr->dma_attr_count_max;
1809 	}
1810 
1811 	/*
1812 	 * granularity may or may not be a power of two. If it isn't, we can't
1813 	 * use a simple mask.
1814 	 */
1815 	if (attr->dma_attr_granular & (attr->dma_attr_granular - 1)) {
1816 		dma->dp_granularity_power_2 = B_FALSE;
1817 	} else {
1818 		dma->dp_granularity_power_2 = B_TRUE;
1819 	}
1820 
1821 	/*
1822 	 * maxxfer should be a whole multiple of granularity. If we're going to
1823 	 * break up a window because we're greater than maxxfer, we might as
1824 	 * well make sure it's maxxfer is a whole multiple so we don't have to
1825 	 * worry about triming the window later on for this case.
1826 	 */
1827 	if (attr->dma_attr_granular > 1) {
1828 		if (dma->dp_granularity_power_2) {
1829 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1830 			    (attr->dma_attr_maxxfer &
1831 			    (attr->dma_attr_granular - 1));
1832 		} else {
1833 			dma->dp_maxxfer = attr->dma_attr_maxxfer -
1834 			    (attr->dma_attr_maxxfer % attr->dma_attr_granular);
1835 		}
1836 	} else {
1837 		dma->dp_maxxfer = attr->dma_attr_maxxfer;
1838 	}
1839 
1840 	maxsegmentsize_ll = MIN(seg, dma->dp_maxxfer);
1841 	maxsegmentsize_ll = MIN(maxsegmentsize_ll, count_max);
1842 	if (maxsegmentsize_ll == 0 || (maxsegmentsize_ll > 0xFFFFFFFF)) {
1843 		maxsegmentsize = 0xFFFFFFFF;
1844 	} else {
1845 		maxsegmentsize = maxsegmentsize_ll;
1846 	}
1847 	dma->dp_sglinfo.si_max_cookie_size = maxsegmentsize;
1848 	dma->dp_sglinfo.si_segmask = attr->dma_attr_seg;
1849 
1850 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
1851 	if (rootnex_alloc_check_parms) {
1852 		e = rootnex_valid_alloc_parms(attr, maxsegmentsize);
1853 		if (e != DDI_SUCCESS) {
1854 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ALLOC_FAIL]);
1855 			(void) rootnex_dma_freehdl(dip, rdip,
1856 			    (ddi_dma_handle_t)hp);
1857 			return (e);
1858 		}
1859 	}
1860 
1861 	*handlep = (ddi_dma_handle_t)hp;
1862 
1863 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1864 	ROOTNEX_DPROBE1(rootnex__alloc__handle, uint64_t,
1865 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1866 
1867 	return (DDI_SUCCESS);
1868 }
1869 
1870 
1871 /*
1872  * rootnex_dma_allochdl()
1873  *    called from ddi_dma_alloc_handle().
1874  */
1875 static int
1876 rootnex_dma_allochdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_attr_t *attr,
1877     int (*waitfp)(caddr_t), caddr_t arg, ddi_dma_handle_t *handlep)
1878 {
1879 	int retval = DDI_SUCCESS;
1880 #if defined(__amd64) && !defined(__xpv)
1881 
1882 	if (IOMMU_UNITIALIZED(rdip)) {
1883 		retval = iommulib_nex_open(dip, rdip);
1884 
1885 		if (retval != DDI_SUCCESS && retval != DDI_ENOTSUP)
1886 			return (retval);
1887 	}
1888 
1889 	if (IOMMU_UNUSED(rdip)) {
1890 		retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1891 		    handlep);
1892 	} else {
1893 		retval = iommulib_nexdma_allochdl(dip, rdip, attr,
1894 		    waitfp, arg, handlep);
1895 	}
1896 #else
1897 	retval = rootnex_coredma_allochdl(dip, rdip, attr, waitfp, arg,
1898 	    handlep);
1899 #endif
1900 	switch (retval) {
1901 	case DDI_DMA_NORESOURCES:
1902 		if (waitfp != DDI_DMA_DONTWAIT) {
1903 			ddi_set_callback(waitfp, arg,
1904 			    &rootnex_state->r_dvma_call_list_id);
1905 		}
1906 		break;
1907 	case DDI_SUCCESS:
1908 		ndi_fmc_insert(rdip, DMA_HANDLE, *handlep, NULL);
1909 		break;
1910 	default:
1911 		break;
1912 	}
1913 	return (retval);
1914 }
1915 
1916 /*ARGSUSED*/
1917 static int
1918 rootnex_coredma_freehdl(dev_info_t *dip, dev_info_t *rdip,
1919     ddi_dma_handle_t handle)
1920 {
1921 	ddi_dma_impl_t *hp;
1922 	rootnex_dma_t *dma;
1923 
1924 
1925 	hp = (ddi_dma_impl_t *)handle;
1926 	dma = (rootnex_dma_t *)hp->dmai_private;
1927 
1928 	/* unbind should have been called first */
1929 	ASSERT(!dma->dp_inuse);
1930 
1931 	mutex_destroy(&dma->dp_mutex);
1932 	kmem_cache_free(rootnex_state->r_dmahdl_cache, hp);
1933 
1934 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1935 	ROOTNEX_DPROBE1(rootnex__free__handle, uint64_t,
1936 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_HDLS]);
1937 
1938 	return (DDI_SUCCESS);
1939 }
1940 
1941 /*
1942  * rootnex_dma_freehdl()
1943  *    called from ddi_dma_free_handle().
1944  */
1945 static int
1946 rootnex_dma_freehdl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle)
1947 {
1948 	int ret;
1949 
1950 	ndi_fmc_remove(rdip, DMA_HANDLE, handle);
1951 #if defined(__amd64) && !defined(__xpv)
1952 	if (IOMMU_USED(rdip))
1953 		ret = iommulib_nexdma_freehdl(dip, rdip, handle);
1954 	else
1955 #endif
1956 	ret = rootnex_coredma_freehdl(dip, rdip, handle);
1957 
1958 	if (rootnex_state->r_dvma_call_list_id)
1959 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
1960 
1961 	return (ret);
1962 }
1963 
1964 /*ARGSUSED*/
1965 static int
1966 rootnex_coredma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
1967     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
1968     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
1969 {
1970 	rootnex_sglinfo_t *sinfo;
1971 	ddi_dma_obj_t *dmao;
1972 	struct dvmaseg *dvs;
1973 	ddi_dma_cookie_t *cookie;
1974 	ddi_dma_attr_t *attr;
1975 	ddi_dma_impl_t *hp;
1976 	rootnex_dma_t *dma;
1977 	int kmflag;
1978 	int e;
1979 	uint_t ncookies;
1980 
1981 	hp = (ddi_dma_impl_t *)handle;
1982 	dma = (rootnex_dma_t *)hp->dmai_private;
1983 	sinfo = &dma->dp_sglinfo;
1984 	attr = &hp->dmai_attr;
1985 
1986 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
1987 		dma->dp_sleep_flags = KM_SLEEP;
1988 	} else {
1989 		dma->dp_sleep_flags = KM_NOSLEEP;
1990 	}
1991 
1992 	hp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS;
1993 
1994 	/*
1995 	 * This is useful for debugging a driver. Not as useful in a production
1996 	 * system. The only time this will fail is if you have a driver bug.
1997 	 */
1998 	if (rootnex_bind_check_inuse) {
1999 		/*
2000 		 * No one else should ever have this lock unless someone else
2001 		 * is trying to use this handle. So contention on the lock
2002 		 * is the same as inuse being set.
2003 		 */
2004 		e = mutex_tryenter(&dma->dp_mutex);
2005 		if (e == 0) {
2006 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2007 			return (DDI_DMA_INUSE);
2008 		}
2009 		if (dma->dp_inuse) {
2010 			mutex_exit(&dma->dp_mutex);
2011 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2012 			return (DDI_DMA_INUSE);
2013 		}
2014 		dma->dp_inuse = B_TRUE;
2015 		mutex_exit(&dma->dp_mutex);
2016 	}
2017 
2018 	/* check the ddi_dma_attr arg to make sure it makes a little sense */
2019 	if (rootnex_bind_check_parms) {
2020 		e = rootnex_valid_bind_parms(dmareq, attr);
2021 		if (e != DDI_SUCCESS) {
2022 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2023 			rootnex_clean_dmahdl(hp);
2024 			return (e);
2025 		}
2026 	}
2027 
2028 	/* save away the original bind info */
2029 	dma->dp_dma = dmareq->dmar_object;
2030 
2031 #if defined(__amd64) && !defined(__xpv)
2032 	if (IOMMU_USED(rdip)) {
2033 		dmao = &dma->dp_dvma;
2034 		e = iommulib_nexdma_mapobject(dip, rdip, handle, dmareq, dmao);
2035 		switch (e) {
2036 		case DDI_SUCCESS:
2037 			if (sinfo->si_cancross ||
2038 			    dmao->dmao_obj.dvma_obj.dv_nseg != 1 ||
2039 			    dmao->dmao_size > sinfo->si_max_cookie_size) {
2040 				dma->dp_dvma_used = B_TRUE;
2041 				break;
2042 			}
2043 			sinfo->si_sgl_size = 1;
2044 			hp->dmai_rflags |= DMP_NOSYNC;
2045 
2046 			dma->dp_dvma_used = B_TRUE;
2047 			dma->dp_need_to_free_cookie = B_FALSE;
2048 
2049 			dvs = &dmao->dmao_obj.dvma_obj.dv_seg[0];
2050 			cookie = hp->dmai_cookie = dma->dp_cookies =
2051 			    (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2052 			cookie->dmac_laddress = dvs->dvs_start +
2053 			    dmao->dmao_obj.dvma_obj.dv_off;
2054 			cookie->dmac_size = dvs->dvs_len;
2055 			cookie->dmac_type = 0;
2056 
2057 			ROOTNEX_DPROBE1(rootnex__bind__dvmafast, dev_info_t *,
2058 			    rdip);
2059 			goto fast;
2060 		case DDI_ENOTSUP:
2061 			break;
2062 		default:
2063 			rootnex_clean_dmahdl(hp);
2064 			return (e);
2065 		}
2066 	}
2067 
2068 #endif
2069 
2070 	/*
2071 	 * Figure out a rough estimate of what maximum number of pages
2072 	 * this buffer could use (a high estimate of course).
2073 	 */
2074 	sinfo->si_max_pages = mmu_btopr(dma->dp_dma.dmao_size) + 1;
2075 
2076 	if (dma->dp_dvma_used) {
2077 		/*
2078 		 * The number of physical pages is the worst case.
2079 		 *
2080 		 * For DVMA, the worst case is the length divided
2081 		 * by the maximum cookie length, plus 1. Add to that
2082 		 * the number of segment boundaries potentially crossed, and
2083 		 * the additional number of DVMA segments that was returned.
2084 		 *
2085 		 * In the normal case, for modern devices, si_cancross will
2086 		 * be false, and dv_nseg will be 1, and the fast path will
2087 		 * have been taken above.
2088 		 */
2089 		ncookies = (dma->dp_dma.dmao_size / sinfo->si_max_cookie_size)
2090 		    + 1;
2091 		if (sinfo->si_cancross)
2092 			ncookies +=
2093 			    (dma->dp_dma.dmao_size / attr->dma_attr_seg) + 1;
2094 		ncookies += (dmao->dmao_obj.dvma_obj.dv_nseg - 1);
2095 
2096 		sinfo->si_max_pages = MIN(sinfo->si_max_pages, ncookies);
2097 	} else {
2098 		dmao = &dma->dp_dma;
2099 	}
2100 
2101 	/*
2102 	 * We'll use the pre-allocated cookies for any bind that will *always*
2103 	 * fit (more important to be consistent, we don't want to create
2104 	 * additional degenerate cases).
2105 	 */
2106 	if (sinfo->si_max_pages <= rootnex_state->r_prealloc_cookies) {
2107 		dma->dp_cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer;
2108 		dma->dp_need_to_free_cookie = B_FALSE;
2109 		ROOTNEX_DPROBE2(rootnex__bind__prealloc, dev_info_t *, rdip,
2110 		    uint_t, sinfo->si_max_pages);
2111 
2112 	/*
2113 	 * For anything larger than that, we'll go ahead and allocate the
2114 	 * maximum number of pages we expect to see. Hopefuly, we won't be
2115 	 * seeing this path in the fast path for high performance devices very
2116 	 * frequently.
2117 	 *
2118 	 * a ddi bind interface that allowed the driver to provide storage to
2119 	 * the bind interface would speed this case up.
2120 	 */
2121 	} else {
2122 		/* convert the sleep flags */
2123 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
2124 			kmflag =  KM_SLEEP;
2125 		} else {
2126 			kmflag =  KM_NOSLEEP;
2127 		}
2128 
2129 		/*
2130 		 * Save away how much memory we allocated. If we're doing a
2131 		 * nosleep, the alloc could fail...
2132 		 */
2133 		dma->dp_cookie_size = sinfo->si_max_pages *
2134 		    sizeof (ddi_dma_cookie_t);
2135 		dma->dp_cookies = kmem_alloc(dma->dp_cookie_size, kmflag);
2136 		if (dma->dp_cookies == NULL) {
2137 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2138 			rootnex_clean_dmahdl(hp);
2139 			return (DDI_DMA_NORESOURCES);
2140 		}
2141 		dma->dp_need_to_free_cookie = B_TRUE;
2142 		ROOTNEX_DPROBE2(rootnex__bind__alloc, dev_info_t *, rdip,
2143 		    uint_t, sinfo->si_max_pages);
2144 	}
2145 	hp->dmai_cookie = dma->dp_cookies;
2146 
2147 	/*
2148 	 * Get the real sgl. rootnex_get_sgl will fill in cookie array while
2149 	 * looking at the constraints in the dma structure. It will then put
2150 	 * some additional state about the sgl in the dma struct (i.e. is
2151 	 * the sgl clean, or do we need to do some munging; how many pages
2152 	 * need to be copied, etc.)
2153 	 */
2154 	if (dma->dp_dvma_used)
2155 		rootnex_dvma_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2156 	else
2157 		rootnex_get_sgl(dmao, dma->dp_cookies, &dma->dp_sglinfo);
2158 
2159 out:
2160 	ASSERT(sinfo->si_sgl_size <= sinfo->si_max_pages);
2161 	/* if we don't need a copy buffer, we don't need to sync */
2162 	if (sinfo->si_copybuf_req == 0) {
2163 		hp->dmai_rflags |= DMP_NOSYNC;
2164 	}
2165 
2166 	/*
2167 	 * if we don't need the copybuf and we don't need to do a partial,  we
2168 	 * hit the fast path. All the high performance devices should be trying
2169 	 * to hit this path. To hit this path, a device should be able to reach
2170 	 * all of memory, shouldn't try to bind more than it can transfer, and
2171 	 * the buffer shouldn't require more cookies than the driver/device can
2172 	 * handle [sgllen]).
2173 	 */
2174 	if ((sinfo->si_copybuf_req == 0) &&
2175 	    (sinfo->si_sgl_size <= attr->dma_attr_sgllen) &&
2176 	    (dmao->dmao_size < dma->dp_maxxfer)) {
2177 fast:
2178 		/*
2179 		 * If the driver supports FMA, insert the handle in the FMA DMA
2180 		 * handle cache.
2181 		 */
2182 		if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2183 			hp->dmai_error.err_cf = rootnex_dma_check;
2184 
2185 		/*
2186 		 * copy out the first cookie and ccountp, set the cookie
2187 		 * pointer to the second cookie. The first cookie is passed
2188 		 * back on the stack. Additional cookies are accessed via
2189 		 * ddi_dma_nextcookie()
2190 		 */
2191 		*cookiep = dma->dp_cookies[0];
2192 		*ccountp = sinfo->si_sgl_size;
2193 		hp->dmai_cookie++;
2194 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2195 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2196 		ROOTNEX_DPROBE4(rootnex__bind__fast, dev_info_t *, rdip,
2197 		    uint64_t, rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS],
2198 		    uint_t, dmao->dmao_size, uint_t, *ccountp);
2199 
2200 
2201 		return (DDI_DMA_MAPPED);
2202 	}
2203 
2204 	/*
2205 	 * go to the slow path, we may need to alloc more memory, create
2206 	 * multiple windows, and munge up a sgl to make the device happy.
2207 	 */
2208 
2209 	/*
2210 	 * With the IOMMU mapobject method used, we should never hit
2211 	 * the slow path. If we do, something is seriously wrong.
2212 	 * Clean up and return an error.
2213 	 */
2214 
2215 	if (dma->dp_dvma_used) {
2216 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2217 		    &dma->dp_dvma);
2218 		e = DDI_DMA_NOMAPPING;
2219 	} else {
2220 		e = rootnex_bind_slowpath(hp, dmareq, dma, attr, &dma->dp_dma,
2221 		    kmflag);
2222 	}
2223 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
2224 		if (dma->dp_need_to_free_cookie) {
2225 			kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2226 		}
2227 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_BIND_FAIL]);
2228 		rootnex_clean_dmahdl(hp); /* must be after free cookie */
2229 		return (e);
2230 	}
2231 
2232 	/*
2233 	 * If the driver supports FMA, insert the handle in the FMA DMA handle
2234 	 * cache.
2235 	 */
2236 	if (attr->dma_attr_flags & DDI_DMA_FLAGERR)
2237 		hp->dmai_error.err_cf = rootnex_dma_check;
2238 
2239 	/* if the first window uses the copy buffer, sync it for the device */
2240 	if ((dma->dp_window[dma->dp_current_win].wd_dosync) &&
2241 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
2242 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2243 		    DDI_DMA_SYNC_FORDEV);
2244 	}
2245 
2246 	/*
2247 	 * copy out the first cookie and ccountp, set the cookie pointer to the
2248 	 * second cookie. Make sure the partial flag is set/cleared correctly.
2249 	 * If we have a partial map (i.e. multiple windows), the number of
2250 	 * cookies we return is the number of cookies in the first window.
2251 	 */
2252 	if (e == DDI_DMA_MAPPED) {
2253 		hp->dmai_rflags &= ~DDI_DMA_PARTIAL;
2254 		*ccountp = sinfo->si_sgl_size;
2255 		hp->dmai_nwin = 1;
2256 	} else {
2257 		hp->dmai_rflags |= DDI_DMA_PARTIAL;
2258 		*ccountp = dma->dp_window[dma->dp_current_win].wd_cookie_cnt;
2259 		ASSERT(hp->dmai_nwin <= dma->dp_max_win);
2260 	}
2261 	*cookiep = dma->dp_cookies[0];
2262 	hp->dmai_cookie++;
2263 
2264 	ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2265 	ROOTNEX_DPROBE4(rootnex__bind__slow, dev_info_t *, rdip, uint64_t,
2266 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS], uint_t,
2267 	    dmao->dmao_size, uint_t, *ccountp);
2268 	return (e);
2269 }
2270 
2271 /*
2272  * rootnex_dma_bindhdl()
2273  *    called from ddi_dma_addr_bind_handle() and ddi_dma_buf_bind_handle().
2274  */
2275 static int
2276 rootnex_dma_bindhdl(dev_info_t *dip, dev_info_t *rdip,
2277     ddi_dma_handle_t handle, struct ddi_dma_req *dmareq,
2278     ddi_dma_cookie_t *cookiep, uint_t *ccountp)
2279 {
2280 	int ret;
2281 #if defined(__amd64) && !defined(__xpv)
2282 	if (IOMMU_USED(rdip))
2283 		ret = iommulib_nexdma_bindhdl(dip, rdip, handle, dmareq,
2284 		    cookiep, ccountp);
2285 	else
2286 #endif
2287 	ret = rootnex_coredma_bindhdl(dip, rdip, handle, dmareq,
2288 	    cookiep, ccountp);
2289 
2290 	if (ret == DDI_DMA_NORESOURCES && dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
2291 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
2292 		    &rootnex_state->r_dvma_call_list_id);
2293 	}
2294 
2295 	return (ret);
2296 }
2297 
2298 
2299 
2300 /*ARGSUSED*/
2301 static int
2302 rootnex_coredma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2303     ddi_dma_handle_t handle)
2304 {
2305 	ddi_dma_impl_t *hp;
2306 	rootnex_dma_t *dma;
2307 	int e;
2308 
2309 	hp = (ddi_dma_impl_t *)handle;
2310 	dma = (rootnex_dma_t *)hp->dmai_private;
2311 
2312 	/* make sure the buffer wasn't free'd before calling unbind */
2313 	if (rootnex_unbind_verify_buffer) {
2314 		e = rootnex_verify_buffer(dma);
2315 		if (e != DDI_SUCCESS) {
2316 			ASSERT(0);
2317 			return (DDI_FAILURE);
2318 		}
2319 	}
2320 
2321 	/* sync the current window before unbinding the buffer */
2322 	if (dma->dp_window && dma->dp_window[dma->dp_current_win].wd_dosync &&
2323 	    (hp->dmai_rflags & DDI_DMA_READ)) {
2324 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
2325 		    DDI_DMA_SYNC_FORCPU);
2326 	}
2327 
2328 	/*
2329 	 * cleanup and copy buffer or window state. if we didn't use the copy
2330 	 * buffer or windows, there won't be much to do :-)
2331 	 */
2332 	rootnex_teardown_copybuf(dma);
2333 	rootnex_teardown_windows(dma);
2334 
2335 	if (IOMMU_USED(rdip))
2336 		(void) iommulib_nexdma_unmapobject(dip, rdip, handle,
2337 		    &dma->dp_dvma);
2338 
2339 	/*
2340 	 * If we had to allocate space to for the worse case sgl (it didn't
2341 	 * fit into our pre-allocate buffer), free that up now
2342 	 */
2343 	if (dma->dp_need_to_free_cookie) {
2344 		kmem_free(dma->dp_cookies, dma->dp_cookie_size);
2345 	}
2346 
2347 	/*
2348 	 * clean up the handle so it's ready for the next bind (i.e. if the
2349 	 * handle is reused).
2350 	 */
2351 	rootnex_clean_dmahdl(hp);
2352 	hp->dmai_error.err_cf = NULL;
2353 
2354 	ROOTNEX_DPROF_DEC(&rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2355 	ROOTNEX_DPROBE1(rootnex__unbind, uint64_t,
2356 	    rootnex_cnt[ROOTNEX_CNT_ACTIVE_BINDS]);
2357 
2358 	return (DDI_SUCCESS);
2359 }
2360 
2361 /*
2362  * rootnex_dma_unbindhdl()
2363  *    called from ddi_dma_unbind_handle()
2364  */
2365 /*ARGSUSED*/
2366 static int
2367 rootnex_dma_unbindhdl(dev_info_t *dip, dev_info_t *rdip,
2368     ddi_dma_handle_t handle)
2369 {
2370 	int ret;
2371 
2372 #if defined(__amd64) && !defined(__xpv)
2373 	if (IOMMU_USED(rdip))
2374 		ret = iommulib_nexdma_unbindhdl(dip, rdip, handle);
2375 	else
2376 #endif
2377 	ret = rootnex_coredma_unbindhdl(dip, rdip, handle);
2378 
2379 	if (rootnex_state->r_dvma_call_list_id)
2380 		ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
2381 
2382 	return (ret);
2383 }
2384 
2385 #if defined(__amd64) && !defined(__xpv)
2386 
2387 static int
2388 rootnex_coredma_get_sleep_flags(ddi_dma_handle_t handle)
2389 {
2390 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2391 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2392 
2393 	if (dma->dp_sleep_flags != KM_SLEEP &&
2394 	    dma->dp_sleep_flags != KM_NOSLEEP)
2395 		cmn_err(CE_PANIC, "kmem sleep flags not set in DMA handle");
2396 	return (dma->dp_sleep_flags);
2397 }
2398 /*ARGSUSED*/
2399 static void
2400 rootnex_coredma_reset_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2401 {
2402 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2403 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2404 	rootnex_window_t *window;
2405 
2406 	if (dma->dp_window) {
2407 		window = &dma->dp_window[dma->dp_current_win];
2408 		hp->dmai_cookie = window->wd_first_cookie;
2409 	} else {
2410 		hp->dmai_cookie = dma->dp_cookies;
2411 	}
2412 	hp->dmai_cookie++;
2413 }
2414 
2415 /*ARGSUSED*/
2416 static int
2417 rootnex_coredma_get_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2418     ddi_dma_cookie_t **cookiepp, uint_t *ccountp)
2419 {
2420 	int i;
2421 	int km_flags;
2422 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2423 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2424 	rootnex_window_t *window;
2425 	ddi_dma_cookie_t *cp;
2426 	ddi_dma_cookie_t *cookie;
2427 
2428 	ASSERT(*cookiepp == NULL);
2429 	ASSERT(*ccountp == 0);
2430 
2431 	if (dma->dp_window) {
2432 		window = &dma->dp_window[dma->dp_current_win];
2433 		cp = window->wd_first_cookie;
2434 		*ccountp = window->wd_cookie_cnt;
2435 	} else {
2436 		cp = dma->dp_cookies;
2437 		*ccountp = dma->dp_sglinfo.si_sgl_size;
2438 	}
2439 
2440 	km_flags = rootnex_coredma_get_sleep_flags(handle);
2441 	cookie = kmem_zalloc(sizeof (ddi_dma_cookie_t) * (*ccountp), km_flags);
2442 	if (cookie == NULL) {
2443 		return (DDI_DMA_NORESOURCES);
2444 	}
2445 
2446 	for (i = 0; i < *ccountp; i++) {
2447 		cookie[i].dmac_notused = cp[i].dmac_notused;
2448 		cookie[i].dmac_type = cp[i].dmac_type;
2449 		cookie[i].dmac_address = cp[i].dmac_address;
2450 		cookie[i].dmac_size = cp[i].dmac_size;
2451 	}
2452 
2453 	*cookiepp = cookie;
2454 
2455 	return (DDI_SUCCESS);
2456 }
2457 
2458 /*ARGSUSED*/
2459 static int
2460 rootnex_coredma_set_cookies(dev_info_t *dip, ddi_dma_handle_t handle,
2461     ddi_dma_cookie_t *cookiep, uint_t ccount)
2462 {
2463 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2464 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2465 	rootnex_window_t *window;
2466 	ddi_dma_cookie_t *cur_cookiep;
2467 
2468 	ASSERT(cookiep);
2469 	ASSERT(ccount != 0);
2470 	ASSERT(dma->dp_need_to_switch_cookies == B_FALSE);
2471 
2472 	if (dma->dp_window) {
2473 		window = &dma->dp_window[dma->dp_current_win];
2474 		dma->dp_saved_cookies = window->wd_first_cookie;
2475 		window->wd_first_cookie = cookiep;
2476 		ASSERT(ccount == window->wd_cookie_cnt);
2477 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2478 		    + window->wd_first_cookie;
2479 	} else {
2480 		dma->dp_saved_cookies = dma->dp_cookies;
2481 		dma->dp_cookies = cookiep;
2482 		ASSERT(ccount == dma->dp_sglinfo.si_sgl_size);
2483 		cur_cookiep = (hp->dmai_cookie - dma->dp_saved_cookies)
2484 		    + dma->dp_cookies;
2485 	}
2486 
2487 	dma->dp_need_to_switch_cookies = B_TRUE;
2488 	hp->dmai_cookie = cur_cookiep;
2489 
2490 	return (DDI_SUCCESS);
2491 }
2492 
2493 /*ARGSUSED*/
2494 static int
2495 rootnex_coredma_clear_cookies(dev_info_t *dip, ddi_dma_handle_t handle)
2496 {
2497 	ddi_dma_impl_t *hp = (ddi_dma_impl_t *)handle;
2498 	rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private;
2499 	rootnex_window_t *window;
2500 	ddi_dma_cookie_t *cur_cookiep;
2501 	ddi_dma_cookie_t *cookie_array;
2502 	uint_t ccount;
2503 
2504 	/* check if cookies have not been switched */
2505 	if (dma->dp_need_to_switch_cookies == B_FALSE)
2506 		return (DDI_SUCCESS);
2507 
2508 	ASSERT(dma->dp_saved_cookies);
2509 
2510 	if (dma->dp_window) {
2511 		window = &dma->dp_window[dma->dp_current_win];
2512 		cookie_array = window->wd_first_cookie;
2513 		window->wd_first_cookie = dma->dp_saved_cookies;
2514 		dma->dp_saved_cookies = NULL;
2515 		ccount = window->wd_cookie_cnt;
2516 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2517 		    + window->wd_first_cookie;
2518 	} else {
2519 		cookie_array = dma->dp_cookies;
2520 		dma->dp_cookies = dma->dp_saved_cookies;
2521 		dma->dp_saved_cookies = NULL;
2522 		ccount = dma->dp_sglinfo.si_sgl_size;
2523 		cur_cookiep = (hp->dmai_cookie - cookie_array)
2524 		    + dma->dp_cookies;
2525 	}
2526 
2527 	kmem_free(cookie_array, sizeof (ddi_dma_cookie_t) * ccount);
2528 
2529 	hp->dmai_cookie = cur_cookiep;
2530 
2531 	dma->dp_need_to_switch_cookies = B_FALSE;
2532 
2533 	return (DDI_SUCCESS);
2534 }
2535 
2536 #endif
2537 
2538 static struct as *
2539 rootnex_get_as(ddi_dma_obj_t *dmao)
2540 {
2541 	struct as *asp;
2542 
2543 	switch (dmao->dmao_type) {
2544 	case DMA_OTYP_VADDR:
2545 	case DMA_OTYP_BUFVADDR:
2546 		asp = dmao->dmao_obj.virt_obj.v_as;
2547 		if (asp == NULL)
2548 			asp = &kas;
2549 		break;
2550 	default:
2551 		asp = NULL;
2552 		break;
2553 	}
2554 	return (asp);
2555 }
2556 
2557 /*
2558  * rootnex_verify_buffer()
2559  *   verify buffer wasn't free'd
2560  */
2561 static int
2562 rootnex_verify_buffer(rootnex_dma_t *dma)
2563 {
2564 	page_t **pplist;
2565 	caddr_t vaddr;
2566 	uint_t pcnt;
2567 	uint_t poff;
2568 	page_t *pp;
2569 	char b;
2570 	int i;
2571 
2572 	/* Figure out how many pages this buffer occupies */
2573 	if (dma->dp_dma.dmao_type == DMA_OTYP_PAGES) {
2574 		poff = dma->dp_dma.dmao_obj.pp_obj.pp_offset & MMU_PAGEOFFSET;
2575 	} else {
2576 		vaddr = dma->dp_dma.dmao_obj.virt_obj.v_addr;
2577 		poff = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2578 	}
2579 	pcnt = mmu_btopr(dma->dp_dma.dmao_size + poff);
2580 
2581 	switch (dma->dp_dma.dmao_type) {
2582 	case DMA_OTYP_PAGES:
2583 		/*
2584 		 * for a linked list of pp's walk through them to make sure
2585 		 * they're locked and not free.
2586 		 */
2587 		pp = dma->dp_dma.dmao_obj.pp_obj.pp_pp;
2588 		for (i = 0; i < pcnt; i++) {
2589 			if (PP_ISFREE(pp) || !PAGE_LOCKED(pp)) {
2590 				return (DDI_FAILURE);
2591 			}
2592 			pp = pp->p_next;
2593 		}
2594 		break;
2595 
2596 	case DMA_OTYP_VADDR:
2597 	case DMA_OTYP_BUFVADDR:
2598 		pplist = dma->dp_dma.dmao_obj.virt_obj.v_priv;
2599 		/*
2600 		 * for an array of pp's walk through them to make sure they're
2601 		 * not free. It's possible that they may not be locked.
2602 		 */
2603 		if (pplist) {
2604 			for (i = 0; i < pcnt; i++) {
2605 				if (PP_ISFREE(pplist[i])) {
2606 					return (DDI_FAILURE);
2607 				}
2608 			}
2609 
2610 		/* For a virtual address, try to peek at each page */
2611 		} else {
2612 			if (rootnex_get_as(&dma->dp_dma) == &kas) {
2613 				for (i = 0; i < pcnt; i++) {
2614 					if (ddi_peek8(NULL, vaddr, &b) ==
2615 					    DDI_FAILURE)
2616 						return (DDI_FAILURE);
2617 					vaddr += MMU_PAGESIZE;
2618 				}
2619 			}
2620 		}
2621 		break;
2622 
2623 	default:
2624 		cmn_err(CE_PANIC, "rootnex_verify_buffer: bad DMA object");
2625 		break;
2626 	}
2627 
2628 	return (DDI_SUCCESS);
2629 }
2630 
2631 
2632 /*
2633  * rootnex_clean_dmahdl()
2634  *    Clean the dma handle. This should be called on a handle alloc and an
2635  *    unbind handle. Set the handle state to the default settings.
2636  */
2637 static void
2638 rootnex_clean_dmahdl(ddi_dma_impl_t *hp)
2639 {
2640 	rootnex_dma_t *dma;
2641 
2642 
2643 	dma = (rootnex_dma_t *)hp->dmai_private;
2644 
2645 	hp->dmai_nwin = 0;
2646 	dma->dp_current_cookie = 0;
2647 	dma->dp_copybuf_size = 0;
2648 	dma->dp_window = NULL;
2649 	dma->dp_cbaddr = NULL;
2650 	dma->dp_inuse = B_FALSE;
2651 	dma->dp_dvma_used = B_FALSE;
2652 	dma->dp_need_to_free_cookie = B_FALSE;
2653 	dma->dp_need_to_switch_cookies = B_FALSE;
2654 	dma->dp_saved_cookies = NULL;
2655 	dma->dp_sleep_flags = KM_PANIC;
2656 	dma->dp_need_to_free_window = B_FALSE;
2657 	dma->dp_partial_required = B_FALSE;
2658 	dma->dp_trim_required = B_FALSE;
2659 	dma->dp_sglinfo.si_copybuf_req = 0;
2660 #if !defined(__amd64)
2661 	dma->dp_cb_remaping = B_FALSE;
2662 	dma->dp_kva = NULL;
2663 #endif
2664 
2665 	/* FMA related initialization */
2666 	hp->dmai_fault = 0;
2667 	hp->dmai_fault_check = NULL;
2668 	hp->dmai_fault_notify = NULL;
2669 	hp->dmai_error.err_ena = 0;
2670 	hp->dmai_error.err_status = DDI_FM_OK;
2671 	hp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
2672 	hp->dmai_error.err_ontrap = NULL;
2673 }
2674 
2675 
2676 /*
2677  * rootnex_valid_alloc_parms()
2678  *    Called in ddi_dma_alloc_handle path to validate its parameters.
2679  */
2680 static int
2681 rootnex_valid_alloc_parms(ddi_dma_attr_t *attr, uint_t maxsegmentsize)
2682 {
2683 	if ((attr->dma_attr_seg < MMU_PAGEOFFSET) ||
2684 	    (attr->dma_attr_count_max < MMU_PAGEOFFSET) ||
2685 	    (attr->dma_attr_granular > MMU_PAGESIZE) ||
2686 	    (attr->dma_attr_maxxfer < MMU_PAGESIZE)) {
2687 		return (DDI_DMA_BADATTR);
2688 	}
2689 
2690 	if (attr->dma_attr_addr_hi <= attr->dma_attr_addr_lo) {
2691 		return (DDI_DMA_BADATTR);
2692 	}
2693 
2694 	if ((attr->dma_attr_seg & MMU_PAGEOFFSET) != MMU_PAGEOFFSET ||
2695 	    MMU_PAGESIZE & (attr->dma_attr_granular - 1) ||
2696 	    attr->dma_attr_sgllen <= 0) {
2697 		return (DDI_DMA_BADATTR);
2698 	}
2699 
2700 	/* We should be able to DMA into every byte offset in a page */
2701 	if (maxsegmentsize < MMU_PAGESIZE) {
2702 		return (DDI_DMA_BADATTR);
2703 	}
2704 
2705 	/* if we're bouncing on seg, seg must be <= addr_hi */
2706 	if ((attr->dma_attr_flags & _DDI_DMA_BOUNCE_ON_SEG) &&
2707 	    (attr->dma_attr_seg > attr->dma_attr_addr_hi)) {
2708 		return (DDI_DMA_BADATTR);
2709 	}
2710 	return (DDI_SUCCESS);
2711 }
2712 
2713 /*
2714  * rootnex_valid_bind_parms()
2715  *    Called in ddi_dma_*_bind_handle path to validate its parameters.
2716  */
2717 /* ARGSUSED */
2718 static int
2719 rootnex_valid_bind_parms(ddi_dma_req_t *dmareq, ddi_dma_attr_t *attr)
2720 {
2721 #if !defined(__amd64)
2722 	/*
2723 	 * we only support up to a 2G-1 transfer size on 32-bit kernels so
2724 	 * we can track the offset for the obsoleted interfaces.
2725 	 */
2726 	if (dmareq->dmar_object.dmao_size > 0x7FFFFFFF) {
2727 		return (DDI_DMA_TOOBIG);
2728 	}
2729 #endif
2730 
2731 	return (DDI_SUCCESS);
2732 }
2733 
2734 
2735 /*
2736  * rootnex_need_bounce_seg()
2737  *    check to see if the buffer lives on both side of the seg.
2738  */
2739 static boolean_t
2740 rootnex_need_bounce_seg(ddi_dma_obj_t *dmar_object, rootnex_sglinfo_t *sglinfo)
2741 {
2742 	ddi_dma_atyp_t buftype;
2743 	rootnex_addr_t raddr;
2744 	boolean_t lower_addr;
2745 	boolean_t upper_addr;
2746 	uint64_t offset;
2747 	page_t **pplist;
2748 	uint64_t paddr;
2749 	uint32_t psize;
2750 	uint32_t size;
2751 	caddr_t vaddr;
2752 	uint_t pcnt;
2753 	page_t *pp;
2754 
2755 
2756 	/* shortcuts */
2757 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2758 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2759 	buftype = dmar_object->dmao_type;
2760 	size = dmar_object->dmao_size;
2761 
2762 	lower_addr = B_FALSE;
2763 	upper_addr = B_FALSE;
2764 	pcnt = 0;
2765 
2766 	/*
2767 	 * Process the first page to handle the initial offset of the buffer.
2768 	 * We'll use the base address we get later when we loop through all
2769 	 * the pages.
2770 	 */
2771 	if (buftype == DMA_OTYP_PAGES) {
2772 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2773 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2774 		    MMU_PAGEOFFSET;
2775 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2776 		psize = MIN(size, (MMU_PAGESIZE - offset));
2777 		pp = pp->p_next;
2778 		sglinfo->si_asp = NULL;
2779 	} else if (pplist != NULL) {
2780 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2781 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2782 		if (sglinfo->si_asp == NULL) {
2783 			sglinfo->si_asp = &kas;
2784 		}
2785 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2786 		paddr += offset;
2787 		psize = MIN(size, (MMU_PAGESIZE - offset));
2788 		pcnt++;
2789 	} else {
2790 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2791 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2792 		if (sglinfo->si_asp == NULL) {
2793 			sglinfo->si_asp = &kas;
2794 		}
2795 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2796 		paddr += offset;
2797 		psize = MIN(size, (MMU_PAGESIZE - offset));
2798 		vaddr += psize;
2799 	}
2800 
2801 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2802 
2803 	if ((raddr + psize) > sglinfo->si_segmask) {
2804 		upper_addr = B_TRUE;
2805 	} else {
2806 		lower_addr = B_TRUE;
2807 	}
2808 	size -= psize;
2809 
2810 	/*
2811 	 * Walk through the rest of the pages in the buffer. Track to see
2812 	 * if we have pages on both sides of the segment boundary.
2813 	 */
2814 	while (size > 0) {
2815 		/* partial or full page */
2816 		psize = MIN(size, MMU_PAGESIZE);
2817 
2818 		if (buftype == DMA_OTYP_PAGES) {
2819 			/* get the paddr from the page_t */
2820 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2821 			paddr = pfn_to_pa(pp->p_pagenum);
2822 			pp = pp->p_next;
2823 		} else if (pplist != NULL) {
2824 			/* index into the array of page_t's to get the paddr */
2825 			ASSERT(!PP_ISFREE(pplist[pcnt]));
2826 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2827 			pcnt++;
2828 		} else {
2829 			/* call into the VM to get the paddr */
2830 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
2831 			    vaddr));
2832 			vaddr += psize;
2833 		}
2834 
2835 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2836 
2837 		if ((raddr + psize) > sglinfo->si_segmask) {
2838 			upper_addr = B_TRUE;
2839 		} else {
2840 			lower_addr = B_TRUE;
2841 		}
2842 		/*
2843 		 * if the buffer lives both above and below the segment
2844 		 * boundary, or the current page is the page immediately
2845 		 * after the segment, we will use a copy/bounce buffer for
2846 		 * all pages > seg.
2847 		 */
2848 		if ((lower_addr && upper_addr) ||
2849 		    (raddr == (sglinfo->si_segmask + 1))) {
2850 			return (B_TRUE);
2851 		}
2852 
2853 		size -= psize;
2854 	}
2855 
2856 	return (B_FALSE);
2857 }
2858 
2859 /*
2860  * rootnex_get_sgl()
2861  *    Called in bind fastpath to get the sgl. Most of this will be replaced
2862  *    with a call to the vm layer when vm2.0 comes around...
2863  */
2864 static void
2865 rootnex_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
2866     rootnex_sglinfo_t *sglinfo)
2867 {
2868 	ddi_dma_atyp_t buftype;
2869 	rootnex_addr_t raddr;
2870 	uint64_t last_page;
2871 	uint64_t offset;
2872 	uint64_t addrhi;
2873 	uint64_t addrlo;
2874 	uint64_t maxseg;
2875 	page_t **pplist;
2876 	uint64_t paddr;
2877 	uint32_t psize;
2878 	uint32_t size;
2879 	caddr_t vaddr;
2880 	uint_t pcnt;
2881 	page_t *pp;
2882 	uint_t cnt;
2883 
2884 
2885 	/* shortcuts */
2886 	pplist = dmar_object->dmao_obj.virt_obj.v_priv;
2887 	vaddr = dmar_object->dmao_obj.virt_obj.v_addr;
2888 	maxseg = sglinfo->si_max_cookie_size;
2889 	buftype = dmar_object->dmao_type;
2890 	addrhi = sglinfo->si_max_addr;
2891 	addrlo = sglinfo->si_min_addr;
2892 	size = dmar_object->dmao_size;
2893 
2894 	pcnt = 0;
2895 	cnt = 0;
2896 
2897 
2898 	/*
2899 	 * check to see if we need to use the copy buffer for pages over
2900 	 * the segment attr.
2901 	 */
2902 	sglinfo->si_bounce_on_seg = B_FALSE;
2903 	if (sglinfo->si_flags & _DDI_DMA_BOUNCE_ON_SEG) {
2904 		sglinfo->si_bounce_on_seg = rootnex_need_bounce_seg(
2905 		    dmar_object, sglinfo);
2906 	}
2907 
2908 	/*
2909 	 * if we were passed down a linked list of pages, i.e. pointer to
2910 	 * page_t, use this to get our physical address and buf offset.
2911 	 */
2912 	if (buftype == DMA_OTYP_PAGES) {
2913 		pp = dmar_object->dmao_obj.pp_obj.pp_pp;
2914 		ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
2915 		offset =  dmar_object->dmao_obj.pp_obj.pp_offset &
2916 		    MMU_PAGEOFFSET;
2917 		paddr = pfn_to_pa(pp->p_pagenum) + offset;
2918 		psize = MIN(size, (MMU_PAGESIZE - offset));
2919 		pp = pp->p_next;
2920 		sglinfo->si_asp = NULL;
2921 
2922 	/*
2923 	 * We weren't passed down a linked list of pages, but if we were passed
2924 	 * down an array of pages, use this to get our physical address and buf
2925 	 * offset.
2926 	 */
2927 	} else if (pplist != NULL) {
2928 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2929 		    (buftype == DMA_OTYP_BUFVADDR));
2930 
2931 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2932 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2933 		if (sglinfo->si_asp == NULL) {
2934 			sglinfo->si_asp = &kas;
2935 		}
2936 
2937 		ASSERT(!PP_ISFREE(pplist[pcnt]));
2938 		paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
2939 		paddr += offset;
2940 		psize = MIN(size, (MMU_PAGESIZE - offset));
2941 		pcnt++;
2942 
2943 	/*
2944 	 * All we have is a virtual address, we'll need to call into the VM
2945 	 * to get the physical address.
2946 	 */
2947 	} else {
2948 		ASSERT((buftype == DMA_OTYP_VADDR) ||
2949 		    (buftype == DMA_OTYP_BUFVADDR));
2950 
2951 		offset = (uintptr_t)vaddr & MMU_PAGEOFFSET;
2952 		sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as;
2953 		if (sglinfo->si_asp == NULL) {
2954 			sglinfo->si_asp = &kas;
2955 		}
2956 
2957 		paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, vaddr));
2958 		paddr += offset;
2959 		psize = MIN(size, (MMU_PAGESIZE - offset));
2960 		vaddr += psize;
2961 	}
2962 
2963 	raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
2964 
2965 	/*
2966 	 * Setup the first cookie with the physical address of the page and the
2967 	 * size of the page (which takes into account the initial offset into
2968 	 * the page.
2969 	 */
2970 	sgl[cnt].dmac_laddress = raddr;
2971 	sgl[cnt].dmac_size = psize;
2972 	sgl[cnt].dmac_type = 0;
2973 
2974 	/*
2975 	 * Save away the buffer offset into the page. We'll need this later in
2976 	 * the copy buffer code to help figure out the page index within the
2977 	 * buffer and the offset into the current page.
2978 	 */
2979 	sglinfo->si_buf_offset = offset;
2980 
2981 	/*
2982 	 * If we are using the copy buffer for anything over the segment
2983 	 * boundary, and this page is over the segment boundary.
2984 	 *   OR
2985 	 * if the DMA engine can't reach the physical address.
2986 	 */
2987 	if (((sglinfo->si_bounce_on_seg) &&
2988 	    ((raddr + psize) > sglinfo->si_segmask)) ||
2989 	    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
2990 		/*
2991 		 * Increase how much copy buffer we use. We always increase by
2992 		 * pagesize so we don't have to worry about converting offsets.
2993 		 * Set a flag in the cookies dmac_type to indicate that it uses
2994 		 * the copy buffer. If this isn't the last cookie, go to the
2995 		 * next cookie (since we separate each page which uses the copy
2996 		 * buffer in case the copy buffer is not physically contiguous.
2997 		 */
2998 		sglinfo->si_copybuf_req += MMU_PAGESIZE;
2999 		sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3000 		if ((cnt + 1) < sglinfo->si_max_pages) {
3001 			cnt++;
3002 			sgl[cnt].dmac_laddress = 0;
3003 			sgl[cnt].dmac_size = 0;
3004 			sgl[cnt].dmac_type = 0;
3005 		}
3006 	}
3007 
3008 	/*
3009 	 * save this page's physical address so we can figure out if the next
3010 	 * page is physically contiguous. Keep decrementing size until we are
3011 	 * done with the buffer.
3012 	 */
3013 	last_page = raddr & MMU_PAGEMASK;
3014 	size -= psize;
3015 
3016 	while (size > 0) {
3017 		/* Get the size for this page (i.e. partial or full page) */
3018 		psize = MIN(size, MMU_PAGESIZE);
3019 
3020 		if (buftype == DMA_OTYP_PAGES) {
3021 			/* get the paddr from the page_t */
3022 			ASSERT(!PP_ISFREE(pp) && PAGE_LOCKED(pp));
3023 			paddr = pfn_to_pa(pp->p_pagenum);
3024 			pp = pp->p_next;
3025 		} else if (pplist != NULL) {
3026 			/* index into the array of page_t's to get the paddr */
3027 			ASSERT(!PP_ISFREE(pplist[pcnt]));
3028 			paddr = pfn_to_pa(pplist[pcnt]->p_pagenum);
3029 			pcnt++;
3030 		} else {
3031 			/* call into the VM to get the paddr */
3032 			paddr =  pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat,
3033 			    vaddr));
3034 			vaddr += psize;
3035 		}
3036 
3037 		raddr = ROOTNEX_PADDR_TO_RBASE(paddr);
3038 
3039 		/*
3040 		 * If we are using the copy buffer for anything over the
3041 		 * segment boundary, and this page is over the segment
3042 		 * boundary.
3043 		 *   OR
3044 		 * if the DMA engine can't reach the physical address.
3045 		 */
3046 		if (((sglinfo->si_bounce_on_seg) &&
3047 		    ((raddr + psize) > sglinfo->si_segmask)) ||
3048 		    ((raddr < addrlo) || ((raddr + psize) > addrhi))) {
3049 
3050 			sglinfo->si_copybuf_req += MMU_PAGESIZE;
3051 
3052 			/*
3053 			 * if there is something in the current cookie, go to
3054 			 * the next one. We only want one page in a cookie which
3055 			 * uses the copybuf since the copybuf doesn't have to
3056 			 * be physically contiguous.
3057 			 */
3058 			if (sgl[cnt].dmac_size != 0) {
3059 				cnt++;
3060 			}
3061 			sgl[cnt].dmac_laddress = raddr;
3062 			sgl[cnt].dmac_size = psize;
3063 #if defined(__amd64)
3064 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF;
3065 #else
3066 			/*
3067 			 * save the buf offset for 32-bit kernel. used in the
3068 			 * obsoleted interfaces.
3069 			 */
3070 			sgl[cnt].dmac_type = ROOTNEX_USES_COPYBUF |
3071 			    (dmar_object->dmao_size - size);
3072 #endif
3073 			/* if this isn't the last cookie, go to the next one */
3074 			if ((cnt + 1) < sglinfo->si_max_pages) {
3075 				cnt++;
3076 				sgl[cnt].dmac_laddress = 0;
3077 				sgl[cnt].dmac_size = 0;
3078 				sgl[cnt].dmac_type = 0;
3079 			}
3080 
3081 		/*
3082 		 * this page didn't need the copy buffer, if it's not physically
3083 		 * contiguous, or it would put us over a segment boundary, or it
3084 		 * puts us over the max cookie size, or the current sgl doesn't
3085 		 * have anything in it.
3086 		 */
3087 		} else if (((last_page + MMU_PAGESIZE) != raddr) ||
3088 		    !(raddr & sglinfo->si_segmask) ||
3089 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3090 		    (sgl[cnt].dmac_size == 0)) {
3091 			/*
3092 			 * if we're not already in a new cookie, go to the next
3093 			 * cookie.
3094 			 */
3095 			if (sgl[cnt].dmac_size != 0) {
3096 				cnt++;
3097 			}
3098 
3099 			/* save the cookie information */
3100 			sgl[cnt].dmac_laddress = raddr;
3101 			sgl[cnt].dmac_size = psize;
3102 #if defined(__amd64)
3103 			sgl[cnt].dmac_type = 0;
3104 #else
3105 			/*
3106 			 * save the buf offset for 32-bit kernel. used in the
3107 			 * obsoleted interfaces.
3108 			 */
3109 			sgl[cnt].dmac_type = dmar_object->dmao_size - size;
3110 #endif
3111 
3112 		/*
3113 		 * this page didn't need the copy buffer, it is physically
3114 		 * contiguous with the last page, and it's <= the max cookie
3115 		 * size.
3116 		 */
3117 		} else {
3118 			sgl[cnt].dmac_size += psize;
3119 
3120 			/*
3121 			 * if this exactly ==  the maximum cookie size, and
3122 			 * it isn't the last cookie, go to the next cookie.
3123 			 */
3124 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3125 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3126 				cnt++;
3127 				sgl[cnt].dmac_laddress = 0;
3128 				sgl[cnt].dmac_size = 0;
3129 				sgl[cnt].dmac_type = 0;
3130 			}
3131 		}
3132 
3133 		/*
3134 		 * save this page's physical address so we can figure out if the
3135 		 * next page is physically contiguous. Keep decrementing size
3136 		 * until we are done with the buffer.
3137 		 */
3138 		last_page = raddr;
3139 		size -= psize;
3140 	}
3141 
3142 	/* we're done, save away how many cookies the sgl has */
3143 	if (sgl[cnt].dmac_size == 0) {
3144 		ASSERT(cnt < sglinfo->si_max_pages);
3145 		sglinfo->si_sgl_size = cnt;
3146 	} else {
3147 		sglinfo->si_sgl_size = cnt + 1;
3148 	}
3149 }
3150 
3151 static void
3152 rootnex_dvma_get_sgl(ddi_dma_obj_t *dmar_object, ddi_dma_cookie_t *sgl,
3153     rootnex_sglinfo_t *sglinfo)
3154 {
3155 	uint64_t offset;
3156 	uint64_t maxseg;
3157 	uint64_t dvaddr;
3158 	struct dvmaseg *dvs;
3159 	uint64_t paddr;
3160 	uint32_t psize, ssize;
3161 	uint32_t size;
3162 	uint_t cnt;
3163 	int physcontig;
3164 
3165 	ASSERT(dmar_object->dmao_type == DMA_OTYP_DVADDR);
3166 
3167 	/* shortcuts */
3168 	maxseg = sglinfo->si_max_cookie_size;
3169 	size = dmar_object->dmao_size;
3170 
3171 	cnt = 0;
3172 	sglinfo->si_bounce_on_seg = B_FALSE;
3173 
3174 	dvs = dmar_object->dmao_obj.dvma_obj.dv_seg;
3175 	offset = dmar_object->dmao_obj.dvma_obj.dv_off;
3176 	ssize = dvs->dvs_len;
3177 	paddr = dvs->dvs_start;
3178 	paddr += offset;
3179 	psize = MIN(ssize, (maxseg - offset));
3180 	dvaddr = paddr + psize;
3181 	ssize -= psize;
3182 
3183 	sgl[cnt].dmac_laddress = paddr;
3184 	sgl[cnt].dmac_size = psize;
3185 	sgl[cnt].dmac_type = 0;
3186 
3187 	size -= psize;
3188 	while (size > 0) {
3189 		if (ssize == 0) {
3190 			dvs++;
3191 			ssize = dvs->dvs_len;
3192 			dvaddr = dvs->dvs_start;
3193 			physcontig = 0;
3194 		} else
3195 			physcontig = 1;
3196 
3197 		paddr = dvaddr;
3198 		psize = MIN(ssize, maxseg);
3199 		dvaddr += psize;
3200 		ssize -= psize;
3201 
3202 		if (!physcontig || !(paddr & sglinfo->si_segmask) ||
3203 		    ((sgl[cnt].dmac_size + psize) > maxseg) ||
3204 		    (sgl[cnt].dmac_size == 0)) {
3205 			/*
3206 			 * if we're not already in a new cookie, go to the next
3207 			 * cookie.
3208 			 */
3209 			if (sgl[cnt].dmac_size != 0) {
3210 				cnt++;
3211 			}
3212 
3213 			/* save the cookie information */
3214 			sgl[cnt].dmac_laddress = paddr;
3215 			sgl[cnt].dmac_size = psize;
3216 			sgl[cnt].dmac_type = 0;
3217 		} else {
3218 			sgl[cnt].dmac_size += psize;
3219 
3220 			/*
3221 			 * if this exactly ==  the maximum cookie size, and
3222 			 * it isn't the last cookie, go to the next cookie.
3223 			 */
3224 			if (((sgl[cnt].dmac_size + psize) == maxseg) &&
3225 			    ((cnt + 1) < sglinfo->si_max_pages)) {
3226 				cnt++;
3227 				sgl[cnt].dmac_laddress = 0;
3228 				sgl[cnt].dmac_size = 0;
3229 				sgl[cnt].dmac_type = 0;
3230 			}
3231 		}
3232 		size -= psize;
3233 	}
3234 
3235 	/* we're done, save away how many cookies the sgl has */
3236 	if (sgl[cnt].dmac_size == 0) {
3237 		sglinfo->si_sgl_size = cnt;
3238 	} else {
3239 		sglinfo->si_sgl_size = cnt + 1;
3240 	}
3241 }
3242 
3243 /*
3244  * rootnex_bind_slowpath()
3245  *    Call in the bind path if the calling driver can't use the sgl without
3246  *    modifying it. We either need to use the copy buffer and/or we will end up
3247  *    with a partial bind.
3248  */
3249 static int
3250 rootnex_bind_slowpath(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3251     rootnex_dma_t *dma, ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3252 {
3253 	rootnex_sglinfo_t *sinfo;
3254 	rootnex_window_t *window;
3255 	ddi_dma_cookie_t *cookie;
3256 	size_t copybuf_used;
3257 	size_t dmac_size;
3258 	boolean_t partial;
3259 	off_t cur_offset;
3260 	page_t *cur_pp;
3261 	major_t mnum;
3262 	int e;
3263 	int i;
3264 
3265 
3266 	sinfo = &dma->dp_sglinfo;
3267 	copybuf_used = 0;
3268 	partial = B_FALSE;
3269 
3270 	/*
3271 	 * If we're using the copybuf, set the copybuf state in dma struct.
3272 	 * Needs to be first since it sets the copy buffer size.
3273 	 */
3274 	if (sinfo->si_copybuf_req != 0) {
3275 		e = rootnex_setup_copybuf(hp, dmareq, dma, attr);
3276 		if (e != DDI_SUCCESS) {
3277 			return (e);
3278 		}
3279 	} else {
3280 		dma->dp_copybuf_size = 0;
3281 	}
3282 
3283 	/*
3284 	 * Figure out if we need to do a partial mapping. If so, figure out
3285 	 * if we need to trim the buffers when we munge the sgl.
3286 	 */
3287 	if ((dma->dp_copybuf_size < sinfo->si_copybuf_req) ||
3288 	    (dmao->dmao_size > dma->dp_maxxfer) ||
3289 	    (attr->dma_attr_sgllen < sinfo->si_sgl_size)) {
3290 		dma->dp_partial_required = B_TRUE;
3291 		if (attr->dma_attr_granular != 1) {
3292 			dma->dp_trim_required = B_TRUE;
3293 		}
3294 	} else {
3295 		dma->dp_partial_required = B_FALSE;
3296 		dma->dp_trim_required = B_FALSE;
3297 	}
3298 
3299 	/* If we need to do a partial bind, make sure the driver supports it */
3300 	if (dma->dp_partial_required &&
3301 	    !(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
3302 
3303 		mnum = ddi_driver_major(dma->dp_dip);
3304 		/*
3305 		 * patchable which allows us to print one warning per major
3306 		 * number.
3307 		 */
3308 		if ((rootnex_bind_warn) &&
3309 		    ((rootnex_warn_list[mnum] & ROOTNEX_BIND_WARNING) == 0)) {
3310 			rootnex_warn_list[mnum] |= ROOTNEX_BIND_WARNING;
3311 			cmn_err(CE_WARN, "!%s: coding error detected, the "
3312 			    "driver is using ddi_dma_attr(9S) incorrectly. "
3313 			    "There is a small risk of data corruption in "
3314 			    "particular with large I/Os. The driver should be "
3315 			    "replaced with a corrected version for proper "
3316 			    "system operation. To disable this warning, add "
3317 			    "'set rootnex:rootnex_bind_warn=0' to "
3318 			    "/etc/system(4).", ddi_driver_name(dma->dp_dip));
3319 		}
3320 		return (DDI_DMA_TOOBIG);
3321 	}
3322 
3323 	/*
3324 	 * we might need multiple windows, setup state to handle them. In this
3325 	 * code path, we will have at least one window.
3326 	 */
3327 	e = rootnex_setup_windows(hp, dma, attr, dmao, kmflag);
3328 	if (e != DDI_SUCCESS) {
3329 		rootnex_teardown_copybuf(dma);
3330 		return (e);
3331 	}
3332 
3333 	window = &dma->dp_window[0];
3334 	cookie = &dma->dp_cookies[0];
3335 	cur_offset = 0;
3336 	rootnex_init_win(hp, dma, window, cookie, cur_offset);
3337 	if (dmao->dmao_type == DMA_OTYP_PAGES) {
3338 		cur_pp = dmareq->dmar_object.dmao_obj.pp_obj.pp_pp;
3339 	}
3340 
3341 	/* loop though all the cookies we got back from get_sgl() */
3342 	for (i = 0; i < sinfo->si_sgl_size; i++) {
3343 		/*
3344 		 * If we're using the copy buffer, check this cookie and setup
3345 		 * its associated copy buffer state. If this cookie uses the
3346 		 * copy buffer, make sure we sync this window during dma_sync.
3347 		 */
3348 		if (dma->dp_copybuf_size > 0) {
3349 			rootnex_setup_cookie(dmao, dma, cookie,
3350 			    cur_offset, &copybuf_used, &cur_pp);
3351 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3352 				window->wd_dosync = B_TRUE;
3353 			}
3354 		}
3355 
3356 		/*
3357 		 * save away the cookie size, since it could be modified in
3358 		 * the windowing code.
3359 		 */
3360 		dmac_size = cookie->dmac_size;
3361 
3362 		/* if we went over max copybuf size */
3363 		if (dma->dp_copybuf_size &&
3364 		    (copybuf_used > dma->dp_copybuf_size)) {
3365 			partial = B_TRUE;
3366 			e = rootnex_copybuf_window_boundary(hp, dma, &window,
3367 			    cookie, cur_offset, &copybuf_used);
3368 			if (e != DDI_SUCCESS) {
3369 				rootnex_teardown_copybuf(dma);
3370 				rootnex_teardown_windows(dma);
3371 				return (e);
3372 			}
3373 
3374 			/*
3375 			 * if the coookie uses the copy buffer, make sure the
3376 			 * new window we just moved to is set to sync.
3377 			 */
3378 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3379 				window->wd_dosync = B_TRUE;
3380 			}
3381 			ROOTNEX_DPROBE1(rootnex__copybuf__window, dev_info_t *,
3382 			    dma->dp_dip);
3383 
3384 		/* if the cookie cnt == max sgllen, move to the next window */
3385 		} else if (window->wd_cookie_cnt >= attr->dma_attr_sgllen) {
3386 			partial = B_TRUE;
3387 			ASSERT(window->wd_cookie_cnt == attr->dma_attr_sgllen);
3388 			e = rootnex_sgllen_window_boundary(hp, dma, &window,
3389 			    cookie, attr, cur_offset);
3390 			if (e != DDI_SUCCESS) {
3391 				rootnex_teardown_copybuf(dma);
3392 				rootnex_teardown_windows(dma);
3393 				return (e);
3394 			}
3395 
3396 			/*
3397 			 * if the coookie uses the copy buffer, make sure the
3398 			 * new window we just moved to is set to sync.
3399 			 */
3400 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3401 				window->wd_dosync = B_TRUE;
3402 			}
3403 			ROOTNEX_DPROBE1(rootnex__sgllen__window, dev_info_t *,
3404 			    dma->dp_dip);
3405 
3406 		/* else if we will be over maxxfer */
3407 		} else if ((window->wd_size + dmac_size) >
3408 		    dma->dp_maxxfer) {
3409 			partial = B_TRUE;
3410 			e = rootnex_maxxfer_window_boundary(hp, dma, &window,
3411 			    cookie);
3412 			if (e != DDI_SUCCESS) {
3413 				rootnex_teardown_copybuf(dma);
3414 				rootnex_teardown_windows(dma);
3415 				return (e);
3416 			}
3417 
3418 			/*
3419 			 * if the coookie uses the copy buffer, make sure the
3420 			 * new window we just moved to is set to sync.
3421 			 */
3422 			if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3423 				window->wd_dosync = B_TRUE;
3424 			}
3425 			ROOTNEX_DPROBE1(rootnex__maxxfer__window, dev_info_t *,
3426 			    dma->dp_dip);
3427 
3428 		/* else this cookie fits in the current window */
3429 		} else {
3430 			window->wd_cookie_cnt++;
3431 			window->wd_size += dmac_size;
3432 		}
3433 
3434 		/* track our offset into the buffer, go to the next cookie */
3435 		ASSERT(dmac_size <= dmao->dmao_size);
3436 		ASSERT(cookie->dmac_size <= dmac_size);
3437 		cur_offset += dmac_size;
3438 		cookie++;
3439 	}
3440 
3441 	/* if we ended up with a zero sized window in the end, clean it up */
3442 	if (window->wd_size == 0) {
3443 		hp->dmai_nwin--;
3444 		window--;
3445 	}
3446 
3447 	ASSERT(window->wd_trim.tr_trim_last == B_FALSE);
3448 
3449 	if (!partial) {
3450 		return (DDI_DMA_MAPPED);
3451 	}
3452 
3453 	ASSERT(dma->dp_partial_required);
3454 	return (DDI_DMA_PARTIAL_MAP);
3455 }
3456 
3457 /*
3458  * rootnex_setup_copybuf()
3459  *    Called in bind slowpath. Figures out if we're going to use the copy
3460  *    buffer, and if we do, sets up the basic state to handle it.
3461  */
3462 static int
3463 rootnex_setup_copybuf(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq,
3464     rootnex_dma_t *dma, ddi_dma_attr_t *attr)
3465 {
3466 	rootnex_sglinfo_t *sinfo;
3467 	ddi_dma_attr_t lattr;
3468 	size_t max_copybuf;
3469 	int cansleep;
3470 	int e;
3471 #if !defined(__amd64)
3472 	int vmflag;
3473 #endif
3474 
3475 	ASSERT(!dma->dp_dvma_used);
3476 
3477 	sinfo = &dma->dp_sglinfo;
3478 
3479 	/* read this first so it's consistent through the routine  */
3480 	max_copybuf = i_ddi_copybuf_size() & MMU_PAGEMASK;
3481 
3482 	/* We need to call into the rootnex on ddi_dma_sync() */
3483 	hp->dmai_rflags &= ~DMP_NOSYNC;
3484 
3485 	/* make sure the copybuf size <= the max size */
3486 	dma->dp_copybuf_size = MIN(sinfo->si_copybuf_req, max_copybuf);
3487 	ASSERT((dma->dp_copybuf_size & MMU_PAGEOFFSET) == 0);
3488 
3489 #if !defined(__amd64)
3490 	/*
3491 	 * if we don't have kva space to copy to/from, allocate the KVA space
3492 	 * now. We only do this for the 32-bit kernel. We use seg kpm space for
3493 	 * the 64-bit kernel.
3494 	 */
3495 	if ((dmareq->dmar_object.dmao_type == DMA_OTYP_PAGES) ||
3496 	    (dmareq->dmar_object.dmao_obj.virt_obj.v_as != NULL)) {
3497 
3498 		/* convert the sleep flags */
3499 		if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3500 			vmflag = VM_SLEEP;
3501 		} else {
3502 			vmflag = VM_NOSLEEP;
3503 		}
3504 
3505 		/* allocate Kernel VA space that we can bcopy to/from */
3506 		dma->dp_kva = vmem_alloc(heap_arena, dma->dp_copybuf_size,
3507 		    vmflag);
3508 		if (dma->dp_kva == NULL) {
3509 			return (DDI_DMA_NORESOURCES);
3510 		}
3511 	}
3512 #endif
3513 
3514 	/* convert the sleep flags */
3515 	if (dmareq->dmar_fp == DDI_DMA_SLEEP) {
3516 		cansleep = 1;
3517 	} else {
3518 		cansleep = 0;
3519 	}
3520 
3521 	/*
3522 	 * Allocate the actual copy buffer. This needs to fit within the DMA
3523 	 * engine limits, so we can't use kmem_alloc... We don't need
3524 	 * contiguous memory (sgllen) since we will be forcing windows on
3525 	 * sgllen anyway.
3526 	 */
3527 	lattr = *attr;
3528 	lattr.dma_attr_align = MMU_PAGESIZE;
3529 	/*
3530 	 * this should be < 0 to indicate no limit, but due to a bug in
3531 	 * the rootnex, we'll set it to the maximum positive int.
3532 	 */
3533 	lattr.dma_attr_sgllen = 0x7fffffff;
3534 	/*
3535 	 * if we're using the copy buffer because of seg, use that for our
3536 	 * upper address limit.
3537 	 */
3538 	if (sinfo->si_bounce_on_seg) {
3539 		lattr.dma_attr_addr_hi = lattr.dma_attr_seg;
3540 	}
3541 	e = i_ddi_mem_alloc(dma->dp_dip, &lattr, dma->dp_copybuf_size, cansleep,
3542 	    0, NULL, &dma->dp_cbaddr, &dma->dp_cbsize, NULL);
3543 	if (e != DDI_SUCCESS) {
3544 #if !defined(__amd64)
3545 		if (dma->dp_kva != NULL) {
3546 			vmem_free(heap_arena, dma->dp_kva,
3547 			    dma->dp_copybuf_size);
3548 		}
3549 #endif
3550 		return (DDI_DMA_NORESOURCES);
3551 	}
3552 
3553 	ROOTNEX_DPROBE2(rootnex__alloc__copybuf, dev_info_t *, dma->dp_dip,
3554 	    size_t, dma->dp_copybuf_size);
3555 
3556 	return (DDI_SUCCESS);
3557 }
3558 
3559 
3560 /*
3561  * rootnex_setup_windows()
3562  *    Called in bind slowpath to setup the window state. We always have windows
3563  *    in the slowpath. Even if the window count = 1.
3564  */
3565 static int
3566 rootnex_setup_windows(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3567     ddi_dma_attr_t *attr, ddi_dma_obj_t *dmao, int kmflag)
3568 {
3569 	rootnex_window_t *windowp;
3570 	rootnex_sglinfo_t *sinfo;
3571 	size_t copy_state_size;
3572 	size_t win_state_size;
3573 	size_t state_available;
3574 	size_t space_needed;
3575 	uint_t copybuf_win;
3576 	uint_t maxxfer_win;
3577 	size_t space_used;
3578 	uint_t sglwin;
3579 
3580 
3581 	sinfo = &dma->dp_sglinfo;
3582 
3583 	dma->dp_current_win = 0;
3584 	hp->dmai_nwin = 0;
3585 
3586 	/* If we don't need to do a partial, we only have one window */
3587 	if (!dma->dp_partial_required) {
3588 		dma->dp_max_win = 1;
3589 
3590 	/*
3591 	 * we need multiple windows, need to figure out the worse case number
3592 	 * of windows.
3593 	 */
3594 	} else {
3595 		/*
3596 		 * if we need windows because we need more copy buffer that
3597 		 * we allow, the worse case number of windows we could need
3598 		 * here would be (copybuf space required / copybuf space that
3599 		 * we have) plus one for remainder, and plus 2 to handle the
3600 		 * extra pages on the trim for the first and last pages of the
3601 		 * buffer (a page is the minimum window size so under the right
3602 		 * attr settings, you could have a window for each page).
3603 		 * The last page will only be hit here if the size is not a
3604 		 * multiple of the granularity (which theoretically shouldn't
3605 		 * be the case but never has been enforced, so we could have
3606 		 * broken things without it).
3607 		 */
3608 		if (sinfo->si_copybuf_req > dma->dp_copybuf_size) {
3609 			ASSERT(dma->dp_copybuf_size > 0);
3610 			copybuf_win = (sinfo->si_copybuf_req /
3611 			    dma->dp_copybuf_size) + 1 + 2;
3612 		} else {
3613 			copybuf_win = 0;
3614 		}
3615 
3616 		/*
3617 		 * if we need windows because we have more cookies than the H/W
3618 		 * can handle, the number of windows we would need here would
3619 		 * be (cookie count / cookies count H/W supports minus 1[for
3620 		 * trim]) plus one for remainder.
3621 		 */
3622 		if (attr->dma_attr_sgllen < sinfo->si_sgl_size) {
3623 			sglwin = (sinfo->si_sgl_size /
3624 			    (attr->dma_attr_sgllen - 1)) + 1;
3625 		} else {
3626 			sglwin = 0;
3627 		}
3628 
3629 		/*
3630 		 * if we need windows because we're binding more memory than the
3631 		 * H/W can transfer at once, the number of windows we would need
3632 		 * here would be (xfer count / max xfer H/W supports) plus one
3633 		 * for remainder, and plus 2 to handle the extra pages on the
3634 		 * trim (see above comment about trim)
3635 		 */
3636 		if (dmao->dmao_size > dma->dp_maxxfer) {
3637 			maxxfer_win = (dmao->dmao_size /
3638 			    dma->dp_maxxfer) + 1 + 2;
3639 		} else {
3640 			maxxfer_win = 0;
3641 		}
3642 		dma->dp_max_win =  copybuf_win + sglwin + maxxfer_win;
3643 		ASSERT(dma->dp_max_win > 0);
3644 	}
3645 	win_state_size = dma->dp_max_win * sizeof (rootnex_window_t);
3646 
3647 	/*
3648 	 * Get space for window and potential copy buffer state. Before we
3649 	 * go and allocate memory, see if we can get away with using what's
3650 	 * left in the pre-allocted state or the dynamically allocated sgl.
3651 	 */
3652 	space_used = (uintptr_t)(sinfo->si_sgl_size *
3653 	    sizeof (ddi_dma_cookie_t));
3654 
3655 	/* if we dynamically allocated space for the cookies */
3656 	if (dma->dp_need_to_free_cookie) {
3657 		/* if we have more space in the pre-allocted buffer, use it */
3658 		ASSERT(space_used <= dma->dp_cookie_size);
3659 		if ((dma->dp_cookie_size - space_used) <=
3660 		    rootnex_state->r_prealloc_size) {
3661 			state_available = rootnex_state->r_prealloc_size;
3662 			windowp = (rootnex_window_t *)dma->dp_prealloc_buffer;
3663 
3664 		/*
3665 		 * else, we have more free space in the dynamically allocated
3666 		 * buffer, i.e. the buffer wasn't worse case fragmented so we
3667 		 * didn't need a lot of cookies.
3668 		 */
3669 		} else {
3670 			state_available = dma->dp_cookie_size - space_used;
3671 			windowp = (rootnex_window_t *)
3672 			    &dma->dp_cookies[sinfo->si_sgl_size];
3673 		}
3674 
3675 	/* we used the pre-alloced buffer */
3676 	} else {
3677 		ASSERT(space_used <= rootnex_state->r_prealloc_size);
3678 		state_available = rootnex_state->r_prealloc_size - space_used;
3679 		windowp = (rootnex_window_t *)
3680 		    &dma->dp_cookies[sinfo->si_sgl_size];
3681 	}
3682 
3683 	/*
3684 	 * figure out how much state we need to track the copy buffer. Add an
3685 	 * addition 8 bytes for pointer alignemnt later.
3686 	 */
3687 	if (dma->dp_copybuf_size > 0) {
3688 		copy_state_size = sinfo->si_max_pages *
3689 		    sizeof (rootnex_pgmap_t);
3690 	} else {
3691 		copy_state_size = 0;
3692 	}
3693 	/* add an additional 8 bytes for pointer alignment */
3694 	space_needed = win_state_size + copy_state_size + 0x8;
3695 
3696 	/* if we have enough space already, use it */
3697 	if (state_available >= space_needed) {
3698 		dma->dp_window = windowp;
3699 		dma->dp_need_to_free_window = B_FALSE;
3700 
3701 	/* not enough space, need to allocate more. */
3702 	} else {
3703 		dma->dp_window = kmem_alloc(space_needed, kmflag);
3704 		if (dma->dp_window == NULL) {
3705 			return (DDI_DMA_NORESOURCES);
3706 		}
3707 		dma->dp_need_to_free_window = B_TRUE;
3708 		dma->dp_window_size = space_needed;
3709 		ROOTNEX_DPROBE2(rootnex__bind__sp__alloc, dev_info_t *,
3710 		    dma->dp_dip, size_t, space_needed);
3711 	}
3712 
3713 	/*
3714 	 * we allocate copy buffer state and window state at the same time.
3715 	 * setup our copy buffer state pointers. Make sure it's aligned.
3716 	 */
3717 	if (dma->dp_copybuf_size > 0) {
3718 		dma->dp_pgmap = (rootnex_pgmap_t *)(((uintptr_t)
3719 		    &dma->dp_window[dma->dp_max_win] + 0x7) & ~0x7);
3720 
3721 #if !defined(__amd64)
3722 		/*
3723 		 * make sure all pm_mapped, pm_vaddr, and pm_pp are set to
3724 		 * false/NULL. Should be quicker to bzero vs loop and set.
3725 		 */
3726 		bzero(dma->dp_pgmap, copy_state_size);
3727 #endif
3728 	} else {
3729 		dma->dp_pgmap = NULL;
3730 	}
3731 
3732 	return (DDI_SUCCESS);
3733 }
3734 
3735 
3736 /*
3737  * rootnex_teardown_copybuf()
3738  *    cleans up after rootnex_setup_copybuf()
3739  */
3740 static void
3741 rootnex_teardown_copybuf(rootnex_dma_t *dma)
3742 {
3743 #if !defined(__amd64)
3744 	int i;
3745 
3746 	/*
3747 	 * if we allocated kernel heap VMEM space, go through all the pages and
3748 	 * map out any of the ones that we're mapped into the kernel heap VMEM
3749 	 * arena. Then free the VMEM space.
3750 	 */
3751 	if (dma->dp_kva != NULL) {
3752 		for (i = 0; i < dma->dp_sglinfo.si_max_pages; i++) {
3753 			if (dma->dp_pgmap[i].pm_mapped) {
3754 				hat_unload(kas.a_hat, dma->dp_pgmap[i].pm_kaddr,
3755 				    MMU_PAGESIZE, HAT_UNLOAD);
3756 				dma->dp_pgmap[i].pm_mapped = B_FALSE;
3757 			}
3758 		}
3759 
3760 		vmem_free(heap_arena, dma->dp_kva, dma->dp_copybuf_size);
3761 	}
3762 
3763 #endif
3764 
3765 	/* if we allocated a copy buffer, free it */
3766 	if (dma->dp_cbaddr != NULL) {
3767 		i_ddi_mem_free(dma->dp_cbaddr, NULL);
3768 	}
3769 }
3770 
3771 
3772 /*
3773  * rootnex_teardown_windows()
3774  *    cleans up after rootnex_setup_windows()
3775  */
3776 static void
3777 rootnex_teardown_windows(rootnex_dma_t *dma)
3778 {
3779 	/*
3780 	 * if we had to allocate window state on the last bind (because we
3781 	 * didn't have enough pre-allocated space in the handle), free it.
3782 	 */
3783 	if (dma->dp_need_to_free_window) {
3784 		kmem_free(dma->dp_window, dma->dp_window_size);
3785 	}
3786 }
3787 
3788 
3789 /*
3790  * rootnex_init_win()
3791  *    Called in bind slow path during creation of a new window. Initializes
3792  *    window state to default values.
3793  */
3794 /*ARGSUSED*/
3795 static void
3796 rootnex_init_win(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
3797     rootnex_window_t *window, ddi_dma_cookie_t *cookie, off_t cur_offset)
3798 {
3799 	hp->dmai_nwin++;
3800 	window->wd_dosync = B_FALSE;
3801 	window->wd_offset = cur_offset;
3802 	window->wd_size = 0;
3803 	window->wd_first_cookie = cookie;
3804 	window->wd_cookie_cnt = 0;
3805 	window->wd_trim.tr_trim_first = B_FALSE;
3806 	window->wd_trim.tr_trim_last = B_FALSE;
3807 	window->wd_trim.tr_first_copybuf_win = B_FALSE;
3808 	window->wd_trim.tr_last_copybuf_win = B_FALSE;
3809 #if !defined(__amd64)
3810 	window->wd_remap_copybuf = dma->dp_cb_remaping;
3811 #endif
3812 }
3813 
3814 
3815 /*
3816  * rootnex_setup_cookie()
3817  *    Called in the bind slow path when the sgl uses the copy buffer. If any of
3818  *    the sgl uses the copy buffer, we need to go through each cookie, figure
3819  *    out if it uses the copy buffer, and if it does, save away everything we'll
3820  *    need during sync.
3821  */
3822 static void
3823 rootnex_setup_cookie(ddi_dma_obj_t *dmar_object, rootnex_dma_t *dma,
3824     ddi_dma_cookie_t *cookie, off_t cur_offset, size_t *copybuf_used,
3825     page_t **cur_pp)
3826 {
3827 	boolean_t copybuf_sz_power_2;
3828 	rootnex_sglinfo_t *sinfo;
3829 	paddr_t paddr;
3830 	uint_t pidx;
3831 	uint_t pcnt;
3832 	off_t poff;
3833 #if defined(__amd64)
3834 	pfn_t pfn;
3835 #else
3836 	page_t **pplist;
3837 #endif
3838 
3839 	ASSERT(dmar_object->dmao_type != DMA_OTYP_DVADDR);
3840 
3841 	sinfo = &dma->dp_sglinfo;
3842 
3843 	/*
3844 	 * Calculate the page index relative to the start of the buffer. The
3845 	 * index to the current page for our buffer is the offset into the
3846 	 * first page of the buffer plus our current offset into the buffer
3847 	 * itself, shifted of course...
3848 	 */
3849 	pidx = (sinfo->si_buf_offset + cur_offset) >> MMU_PAGESHIFT;
3850 	ASSERT(pidx < sinfo->si_max_pages);
3851 
3852 	/* if this cookie uses the copy buffer */
3853 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
3854 		/*
3855 		 * NOTE: we know that since this cookie uses the copy buffer, it
3856 		 * is <= MMU_PAGESIZE.
3857 		 */
3858 
3859 		/*
3860 		 * get the offset into the page. For the 64-bit kernel, get the
3861 		 * pfn which we'll use with seg kpm.
3862 		 */
3863 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
3864 #if defined(__amd64)
3865 		/* mfn_to_pfn() is a NOP on i86pc */
3866 		pfn = mfn_to_pfn(cookie->dmac_laddress >> MMU_PAGESHIFT);
3867 #endif /* __amd64 */
3868 
3869 		/* figure out if the copybuf size is a power of 2 */
3870 		if (dma->dp_copybuf_size & (dma->dp_copybuf_size - 1)) {
3871 			copybuf_sz_power_2 = B_FALSE;
3872 		} else {
3873 			copybuf_sz_power_2 = B_TRUE;
3874 		}
3875 
3876 		/* This page uses the copy buffer */
3877 		dma->dp_pgmap[pidx].pm_uses_copybuf = B_TRUE;
3878 
3879 		/*
3880 		 * save the copy buffer KVA that we'll use with this page.
3881 		 * if we still fit within the copybuf, it's a simple add.
3882 		 * otherwise, we need to wrap over using & or % accordingly.
3883 		 */
3884 		if ((*copybuf_used + MMU_PAGESIZE) <= dma->dp_copybuf_size) {
3885 			dma->dp_pgmap[pidx].pm_cbaddr = dma->dp_cbaddr +
3886 			    *copybuf_used;
3887 		} else {
3888 			if (copybuf_sz_power_2) {
3889 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3890 				    (uintptr_t)dma->dp_cbaddr +
3891 				    (*copybuf_used &
3892 				    (dma->dp_copybuf_size - 1)));
3893 			} else {
3894 				dma->dp_pgmap[pidx].pm_cbaddr = (caddr_t)(
3895 				    (uintptr_t)dma->dp_cbaddr +
3896 				    (*copybuf_used % dma->dp_copybuf_size));
3897 			}
3898 		}
3899 
3900 		/*
3901 		 * over write the cookie physical address with the address of
3902 		 * the physical address of the copy buffer page that we will
3903 		 * use.
3904 		 */
3905 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
3906 		    dma->dp_pgmap[pidx].pm_cbaddr)) + poff;
3907 
3908 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
3909 
3910 		/* if we have a kernel VA, it's easy, just save that address */
3911 		if ((dmar_object->dmao_type != DMA_OTYP_PAGES) &&
3912 		    (sinfo->si_asp == &kas)) {
3913 			/*
3914 			 * save away the page aligned virtual address of the
3915 			 * driver buffer. Offsets are handled in the sync code.
3916 			 */
3917 			dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)(((uintptr_t)
3918 			    dmar_object->dmao_obj.virt_obj.v_addr + cur_offset)
3919 			    & MMU_PAGEMASK);
3920 #if !defined(__amd64)
3921 			/*
3922 			 * we didn't need to, and will never need to map this
3923 			 * page.
3924 			 */
3925 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
3926 #endif
3927 
3928 		/* we don't have a kernel VA. We need one for the bcopy. */
3929 		} else {
3930 #if defined(__amd64)
3931 			/*
3932 			 * for the 64-bit kernel, it's easy. We use seg kpm to
3933 			 * get a Kernel VA for the corresponding pfn.
3934 			 */
3935 			dma->dp_pgmap[pidx].pm_kaddr = hat_kpm_pfn2va(pfn);
3936 #else
3937 			/*
3938 			 * for the 32-bit kernel, this is a pain. First we'll
3939 			 * save away the page_t or user VA for this page. This
3940 			 * is needed in rootnex_dma_win() when we switch to a
3941 			 * new window which requires us to re-map the copy
3942 			 * buffer.
3943 			 */
3944 			pplist = dmar_object->dmao_obj.virt_obj.v_priv;
3945 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
3946 				dma->dp_pgmap[pidx].pm_pp = *cur_pp;
3947 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3948 			} else if (pplist != NULL) {
3949 				dma->dp_pgmap[pidx].pm_pp = pplist[pidx];
3950 				dma->dp_pgmap[pidx].pm_vaddr = NULL;
3951 			} else {
3952 				dma->dp_pgmap[pidx].pm_pp = NULL;
3953 				dma->dp_pgmap[pidx].pm_vaddr = (caddr_t)
3954 				    (((uintptr_t)
3955 				    dmar_object->dmao_obj.virt_obj.v_addr +
3956 				    cur_offset) & MMU_PAGEMASK);
3957 			}
3958 
3959 			/*
3960 			 * save away the page aligned virtual address which was
3961 			 * allocated from the kernel heap arena (taking into
3962 			 * account if we need more copy buffer than we alloced
3963 			 * and use multiple windows to handle this, i.e. &,%).
3964 			 * NOTE: there isn't and physical memory backing up this
3965 			 * virtual address space currently.
3966 			 */
3967 			if ((*copybuf_used + MMU_PAGESIZE) <=
3968 			    dma->dp_copybuf_size) {
3969 				dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3970 				    (((uintptr_t)dma->dp_kva + *copybuf_used) &
3971 				    MMU_PAGEMASK);
3972 			} else {
3973 				if (copybuf_sz_power_2) {
3974 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3975 					    (((uintptr_t)dma->dp_kva +
3976 					    (*copybuf_used &
3977 					    (dma->dp_copybuf_size - 1))) &
3978 					    MMU_PAGEMASK);
3979 				} else {
3980 					dma->dp_pgmap[pidx].pm_kaddr = (caddr_t)
3981 					    (((uintptr_t)dma->dp_kva +
3982 					    (*copybuf_used %
3983 					    dma->dp_copybuf_size)) &
3984 					    MMU_PAGEMASK);
3985 				}
3986 			}
3987 
3988 			/*
3989 			 * if we haven't used up the available copy buffer yet,
3990 			 * map the kva to the physical page.
3991 			 */
3992 			if (!dma->dp_cb_remaping && ((*copybuf_used +
3993 			    MMU_PAGESIZE) <= dma->dp_copybuf_size)) {
3994 				dma->dp_pgmap[pidx].pm_mapped = B_TRUE;
3995 				if (dma->dp_pgmap[pidx].pm_pp != NULL) {
3996 					i86_pp_map(dma->dp_pgmap[pidx].pm_pp,
3997 					    dma->dp_pgmap[pidx].pm_kaddr);
3998 				} else {
3999 					i86_va_map(dma->dp_pgmap[pidx].pm_vaddr,
4000 					    sinfo->si_asp,
4001 					    dma->dp_pgmap[pidx].pm_kaddr);
4002 				}
4003 
4004 			/*
4005 			 * we've used up the available copy buffer, this page
4006 			 * will have to be mapped during rootnex_dma_win() when
4007 			 * we switch to a new window which requires a re-map
4008 			 * the copy buffer. (32-bit kernel only)
4009 			 */
4010 			} else {
4011 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4012 			}
4013 #endif
4014 			/* go to the next page_t */
4015 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4016 				*cur_pp = (*cur_pp)->p_next;
4017 			}
4018 		}
4019 
4020 		/* add to the copy buffer count */
4021 		*copybuf_used += MMU_PAGESIZE;
4022 
4023 	/*
4024 	 * This cookie doesn't use the copy buffer. Walk through the pages this
4025 	 * cookie occupies to reflect this.
4026 	 */
4027 	} else {
4028 		/*
4029 		 * figure out how many pages the cookie occupies. We need to
4030 		 * use the original page offset of the buffer and the cookies
4031 		 * offset in the buffer to do this.
4032 		 */
4033 		poff = (sinfo->si_buf_offset + cur_offset) & MMU_PAGEOFFSET;
4034 		pcnt = mmu_btopr(cookie->dmac_size + poff);
4035 
4036 		while (pcnt > 0) {
4037 #if !defined(__amd64)
4038 			/*
4039 			 * the 32-bit kernel doesn't have seg kpm, so we need
4040 			 * to map in the driver buffer (if it didn't come down
4041 			 * with a kernel VA) on the fly. Since this page doesn't
4042 			 * use the copy buffer, it's not, or will it ever, have
4043 			 * to be mapped in.
4044 			 */
4045 			dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4046 #endif
4047 			dma->dp_pgmap[pidx].pm_uses_copybuf = B_FALSE;
4048 
4049 			/*
4050 			 * we need to update pidx and cur_pp or we'll loose
4051 			 * track of where we are.
4052 			 */
4053 			if (dmar_object->dmao_type == DMA_OTYP_PAGES) {
4054 				*cur_pp = (*cur_pp)->p_next;
4055 			}
4056 			pidx++;
4057 			pcnt--;
4058 		}
4059 	}
4060 }
4061 
4062 
4063 /*
4064  * rootnex_sgllen_window_boundary()
4065  *    Called in the bind slow path when the next cookie causes us to exceed (in
4066  *    this case == since we start at 0 and sgllen starts at 1) the maximum sgl
4067  *    length supported by the DMA H/W.
4068  */
4069 static int
4070 rootnex_sgllen_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4071     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, ddi_dma_attr_t *attr,
4072     off_t cur_offset)
4073 {
4074 	off_t new_offset;
4075 	size_t trim_sz;
4076 	off_t coffset;
4077 
4078 
4079 	/*
4080 	 * if we know we'll never have to trim, it's pretty easy. Just move to
4081 	 * the next window and init it. We're done.
4082 	 */
4083 	if (!dma->dp_trim_required) {
4084 		(*windowp)++;
4085 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4086 		(*windowp)->wd_cookie_cnt++;
4087 		(*windowp)->wd_size = cookie->dmac_size;
4088 		return (DDI_SUCCESS);
4089 	}
4090 
4091 	/* figure out how much we need to trim from the window */
4092 	ASSERT(attr->dma_attr_granular != 0);
4093 	if (dma->dp_granularity_power_2) {
4094 		trim_sz = (*windowp)->wd_size & (attr->dma_attr_granular - 1);
4095 	} else {
4096 		trim_sz = (*windowp)->wd_size % attr->dma_attr_granular;
4097 	}
4098 
4099 	/* The window's a whole multiple of granularity. We're done */
4100 	if (trim_sz == 0) {
4101 		(*windowp)++;
4102 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4103 		(*windowp)->wd_cookie_cnt++;
4104 		(*windowp)->wd_size = cookie->dmac_size;
4105 		return (DDI_SUCCESS);
4106 	}
4107 
4108 	/*
4109 	 * The window's not a whole multiple of granularity, since we know this
4110 	 * is due to the sgllen, we need to go back to the last cookie and trim
4111 	 * that one, add the left over part of the old cookie into the new
4112 	 * window, and then add in the new cookie into the new window.
4113 	 */
4114 
4115 	/*
4116 	 * make sure the driver isn't making us do something bad... Trimming and
4117 	 * sgllen == 1 don't go together.
4118 	 */
4119 	if (attr->dma_attr_sgllen == 1) {
4120 		return (DDI_DMA_NOMAPPING);
4121 	}
4122 
4123 	/*
4124 	 * first, setup the current window to account for the trim. Need to go
4125 	 * back to the last cookie for this.
4126 	 */
4127 	cookie--;
4128 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4129 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4130 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4131 	ASSERT(cookie->dmac_size > trim_sz);
4132 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4133 	(*windowp)->wd_size -= trim_sz;
4134 
4135 	/* save the buffer offsets for the next window */
4136 	coffset = cookie->dmac_size - trim_sz;
4137 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4138 
4139 	/*
4140 	 * set this now in case this is the first window. all other cases are
4141 	 * set in dma_win()
4142 	 */
4143 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4144 
4145 	/*
4146 	 * initialize the next window using what's left over in the previous
4147 	 * cookie.
4148 	 */
4149 	(*windowp)++;
4150 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4151 	(*windowp)->wd_cookie_cnt++;
4152 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4153 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4154 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4155 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4156 		(*windowp)->wd_dosync = B_TRUE;
4157 	}
4158 
4159 	/*
4160 	 * now go back to the current cookie and add it to the new window. set
4161 	 * the new window size to the what was left over from the previous
4162 	 * cookie and what's in the current cookie.
4163 	 */
4164 	cookie++;
4165 	(*windowp)->wd_cookie_cnt++;
4166 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4167 
4168 	/*
4169 	 * trim plus the next cookie could put us over maxxfer (a cookie can be
4170 	 * a max size of maxxfer). Handle that case.
4171 	 */
4172 	if ((*windowp)->wd_size > dma->dp_maxxfer) {
4173 		/*
4174 		 * maxxfer is already a whole multiple of granularity, and this
4175 		 * trim will be <= the previous trim (since a cookie can't be
4176 		 * larger than maxxfer). Make things simple here.
4177 		 */
4178 		trim_sz = (*windowp)->wd_size - dma->dp_maxxfer;
4179 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4180 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4181 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4182 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4183 		(*windowp)->wd_size -= trim_sz;
4184 		ASSERT((*windowp)->wd_size == dma->dp_maxxfer);
4185 
4186 		/* save the buffer offsets for the next window */
4187 		coffset = cookie->dmac_size - trim_sz;
4188 		new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4189 
4190 		/* setup the next window */
4191 		(*windowp)++;
4192 		rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4193 		(*windowp)->wd_cookie_cnt++;
4194 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4195 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4196 		    coffset;
4197 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4198 	}
4199 
4200 	return (DDI_SUCCESS);
4201 }
4202 
4203 
4204 /*
4205  * rootnex_copybuf_window_boundary()
4206  *    Called in bind slowpath when we get to a window boundary because we used
4207  *    up all the copy buffer that we have.
4208  */
4209 static int
4210 rootnex_copybuf_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4211     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie, off_t cur_offset,
4212     size_t *copybuf_used)
4213 {
4214 	rootnex_sglinfo_t *sinfo;
4215 	off_t new_offset;
4216 	size_t trim_sz;
4217 	paddr_t paddr;
4218 	off_t coffset;
4219 	uint_t pidx;
4220 	off_t poff;
4221 
4222 
4223 	sinfo = &dma->dp_sglinfo;
4224 
4225 	/*
4226 	 * the copy buffer should be a whole multiple of page size. We know that
4227 	 * this cookie is <= MMU_PAGESIZE.
4228 	 */
4229 	ASSERT(cookie->dmac_size <= MMU_PAGESIZE);
4230 
4231 	/*
4232 	 * from now on, all new windows in this bind need to be re-mapped during
4233 	 * ddi_dma_getwin() (32-bit kernel only). i.e. we ran out out copybuf
4234 	 * space...
4235 	 */
4236 #if !defined(__amd64)
4237 	dma->dp_cb_remaping = B_TRUE;
4238 #endif
4239 
4240 	/* reset copybuf used */
4241 	*copybuf_used = 0;
4242 
4243 	/*
4244 	 * if we don't have to trim (since granularity is set to 1), go to the
4245 	 * next window and add the current cookie to it. We know the current
4246 	 * cookie uses the copy buffer since we're in this code path.
4247 	 */
4248 	if (!dma->dp_trim_required) {
4249 		(*windowp)++;
4250 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4251 
4252 		/* Add this cookie to the new window */
4253 		(*windowp)->wd_cookie_cnt++;
4254 		(*windowp)->wd_size += cookie->dmac_size;
4255 		*copybuf_used += MMU_PAGESIZE;
4256 		return (DDI_SUCCESS);
4257 	}
4258 
4259 	/*
4260 	 * *** may need to trim, figure it out.
4261 	 */
4262 
4263 	/* figure out how much we need to trim from the window */
4264 	if (dma->dp_granularity_power_2) {
4265 		trim_sz = (*windowp)->wd_size &
4266 		    (hp->dmai_attr.dma_attr_granular - 1);
4267 	} else {
4268 		trim_sz = (*windowp)->wd_size % hp->dmai_attr.dma_attr_granular;
4269 	}
4270 
4271 	/*
4272 	 * if the window's a whole multiple of granularity, go to the next
4273 	 * window, init it, then add in the current cookie. We know the current
4274 	 * cookie uses the copy buffer since we're in this code path.
4275 	 */
4276 	if (trim_sz == 0) {
4277 		(*windowp)++;
4278 		rootnex_init_win(hp, dma, *windowp, cookie, cur_offset);
4279 
4280 		/* Add this cookie to the new window */
4281 		(*windowp)->wd_cookie_cnt++;
4282 		(*windowp)->wd_size += cookie->dmac_size;
4283 		*copybuf_used += MMU_PAGESIZE;
4284 		return (DDI_SUCCESS);
4285 	}
4286 
4287 	/*
4288 	 * *** We figured it out, we definitly need to trim
4289 	 */
4290 
4291 	/*
4292 	 * make sure the driver isn't making us do something bad...
4293 	 * Trimming and sgllen == 1 don't go together.
4294 	 */
4295 	if (hp->dmai_attr.dma_attr_sgllen == 1) {
4296 		return (DDI_DMA_NOMAPPING);
4297 	}
4298 
4299 	/*
4300 	 * first, setup the current window to account for the trim. Need to go
4301 	 * back to the last cookie for this. Some of the last cookie will be in
4302 	 * the current window, and some of the last cookie will be in the new
4303 	 * window. All of the current cookie will be in the new window.
4304 	 */
4305 	cookie--;
4306 	(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4307 	(*windowp)->wd_trim.tr_last_cookie = cookie;
4308 	(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4309 	ASSERT(cookie->dmac_size > trim_sz);
4310 	(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4311 	(*windowp)->wd_size -= trim_sz;
4312 
4313 	/*
4314 	 * we're trimming the last cookie (not the current cookie). So that
4315 	 * last cookie may have or may not have been using the copy buffer (
4316 	 * we know the cookie passed in uses the copy buffer since we're in
4317 	 * this code path).
4318 	 *
4319 	 * If the last cookie doesn't use the copy buffer, nothing special to
4320 	 * do. However, if it does uses the copy buffer, it will be both the
4321 	 * last page in the current window and the first page in the next
4322 	 * window. Since we are reusing the copy buffer (and KVA space on the
4323 	 * 32-bit kernel), this page will use the end of the copy buffer in the
4324 	 * current window, and the start of the copy buffer in the next window.
4325 	 * Track that info... The cookie physical address was already set to
4326 	 * the copy buffer physical address in setup_cookie..
4327 	 */
4328 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4329 		pidx = (sinfo->si_buf_offset + (*windowp)->wd_offset +
4330 		    (*windowp)->wd_size) >> MMU_PAGESHIFT;
4331 		(*windowp)->wd_trim.tr_last_copybuf_win = B_TRUE;
4332 		(*windowp)->wd_trim.tr_last_pidx = pidx;
4333 		(*windowp)->wd_trim.tr_last_cbaddr =
4334 		    dma->dp_pgmap[pidx].pm_cbaddr;
4335 #if !defined(__amd64)
4336 		(*windowp)->wd_trim.tr_last_kaddr =
4337 		    dma->dp_pgmap[pidx].pm_kaddr;
4338 #endif
4339 	}
4340 
4341 	/* save the buffer offsets for the next window */
4342 	coffset = cookie->dmac_size - trim_sz;
4343 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4344 
4345 	/*
4346 	 * set this now in case this is the first window. all other cases are
4347 	 * set in dma_win()
4348 	 */
4349 	cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4350 
4351 	/*
4352 	 * initialize the next window using what's left over in the previous
4353 	 * cookie.
4354 	 */
4355 	(*windowp)++;
4356 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4357 	(*windowp)->wd_cookie_cnt++;
4358 	(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4359 	(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress + coffset;
4360 	(*windowp)->wd_trim.tr_first_size = trim_sz;
4361 
4362 	/*
4363 	 * again, we're tracking if the last cookie uses the copy buffer.
4364 	 * read the comment above for more info on why we need to track
4365 	 * additional state.
4366 	 *
4367 	 * For the first cookie in the new window, we need reset the physical
4368 	 * address to DMA into to the start of the copy buffer plus any
4369 	 * initial page offset which may be present.
4370 	 */
4371 	if (cookie->dmac_type & ROOTNEX_USES_COPYBUF) {
4372 		(*windowp)->wd_dosync = B_TRUE;
4373 		(*windowp)->wd_trim.tr_first_copybuf_win = B_TRUE;
4374 		(*windowp)->wd_trim.tr_first_pidx = pidx;
4375 		(*windowp)->wd_trim.tr_first_cbaddr = dma->dp_cbaddr;
4376 		poff = (*windowp)->wd_trim.tr_first_paddr & MMU_PAGEOFFSET;
4377 
4378 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, dma->dp_cbaddr)) +
4379 		    poff;
4380 		(*windowp)->wd_trim.tr_first_paddr =
4381 		    ROOTNEX_PADDR_TO_RBASE(paddr);
4382 
4383 #if !defined(__amd64)
4384 		(*windowp)->wd_trim.tr_first_kaddr = dma->dp_kva;
4385 #endif
4386 		/* account for the cookie copybuf usage in the new window */
4387 		*copybuf_used += MMU_PAGESIZE;
4388 
4389 		/*
4390 		 * every piece of code has to have a hack, and here is this
4391 		 * ones :-)
4392 		 *
4393 		 * There is a complex interaction between setup_cookie and the
4394 		 * copybuf window boundary. The complexity had to be in either
4395 		 * the maxxfer window, or the copybuf window, and I chose the
4396 		 * copybuf code.
4397 		 *
4398 		 * So in this code path, we have taken the last cookie,
4399 		 * virtually broken it in half due to the trim, and it happens
4400 		 * to use the copybuf which further complicates life. At the
4401 		 * same time, we have already setup the current cookie, which
4402 		 * is now wrong. More background info: the current cookie uses
4403 		 * the copybuf, so it is only a page long max. So we need to
4404 		 * fix the current cookies copy buffer address, physical
4405 		 * address, and kva for the 32-bit kernel. We due this by
4406 		 * bumping them by page size (of course, we can't due this on
4407 		 * the physical address since the copy buffer may not be
4408 		 * physically contiguous).
4409 		 */
4410 		cookie++;
4411 		dma->dp_pgmap[pidx + 1].pm_cbaddr += MMU_PAGESIZE;
4412 		poff = cookie->dmac_laddress & MMU_PAGEOFFSET;
4413 
4414 		paddr = pfn_to_pa(hat_getpfnum(kas.a_hat,
4415 		    dma->dp_pgmap[pidx + 1].pm_cbaddr)) + poff;
4416 		cookie->dmac_laddress = ROOTNEX_PADDR_TO_RBASE(paddr);
4417 
4418 #if !defined(__amd64)
4419 		ASSERT(dma->dp_pgmap[pidx + 1].pm_mapped == B_FALSE);
4420 		dma->dp_pgmap[pidx + 1].pm_kaddr += MMU_PAGESIZE;
4421 #endif
4422 	} else {
4423 		/* go back to the current cookie */
4424 		cookie++;
4425 	}
4426 
4427 	/*
4428 	 * add the current cookie to the new window. set the new window size to
4429 	 * the what was left over from the previous cookie and what's in the
4430 	 * current cookie.
4431 	 */
4432 	(*windowp)->wd_cookie_cnt++;
4433 	(*windowp)->wd_size = trim_sz + cookie->dmac_size;
4434 	ASSERT((*windowp)->wd_size < dma->dp_maxxfer);
4435 
4436 	/*
4437 	 * we know that the cookie passed in always uses the copy buffer. We
4438 	 * wouldn't be here if it didn't.
4439 	 */
4440 	*copybuf_used += MMU_PAGESIZE;
4441 
4442 	return (DDI_SUCCESS);
4443 }
4444 
4445 
4446 /*
4447  * rootnex_maxxfer_window_boundary()
4448  *    Called in bind slowpath when we get to a window boundary because we will
4449  *    go over maxxfer.
4450  */
4451 static int
4452 rootnex_maxxfer_window_boundary(ddi_dma_impl_t *hp, rootnex_dma_t *dma,
4453     rootnex_window_t **windowp, ddi_dma_cookie_t *cookie)
4454 {
4455 	size_t dmac_size;
4456 	off_t new_offset;
4457 	size_t trim_sz;
4458 	off_t coffset;
4459 
4460 
4461 	/*
4462 	 * calculate how much we have to trim off of the current cookie to equal
4463 	 * maxxfer. We don't have to account for granularity here since our
4464 	 * maxxfer already takes that into account.
4465 	 */
4466 	trim_sz = ((*windowp)->wd_size + cookie->dmac_size) - dma->dp_maxxfer;
4467 	ASSERT(trim_sz <= cookie->dmac_size);
4468 	ASSERT(trim_sz <= dma->dp_maxxfer);
4469 
4470 	/* save cookie size since we need it later and we might change it */
4471 	dmac_size = cookie->dmac_size;
4472 
4473 	/*
4474 	 * if we're not trimming the entire cookie, setup the current window to
4475 	 * account for the trim.
4476 	 */
4477 	if (trim_sz < cookie->dmac_size) {
4478 		(*windowp)->wd_cookie_cnt++;
4479 		(*windowp)->wd_trim.tr_trim_last = B_TRUE;
4480 		(*windowp)->wd_trim.tr_last_cookie = cookie;
4481 		(*windowp)->wd_trim.tr_last_paddr = cookie->dmac_laddress;
4482 		(*windowp)->wd_trim.tr_last_size = cookie->dmac_size - trim_sz;
4483 		(*windowp)->wd_size = dma->dp_maxxfer;
4484 
4485 		/*
4486 		 * set the adjusted cookie size now in case this is the first
4487 		 * window. All other windows are taken care of in get win
4488 		 */
4489 		cookie->dmac_size = (*windowp)->wd_trim.tr_last_size;
4490 	}
4491 
4492 	/*
4493 	 * coffset is the current offset within the cookie, new_offset is the
4494 	 * current offset with the entire buffer.
4495 	 */
4496 	coffset = dmac_size - trim_sz;
4497 	new_offset = (*windowp)->wd_offset + (*windowp)->wd_size;
4498 
4499 	/* initialize the next window */
4500 	(*windowp)++;
4501 	rootnex_init_win(hp, dma, *windowp, cookie, new_offset);
4502 	(*windowp)->wd_cookie_cnt++;
4503 	(*windowp)->wd_size = trim_sz;
4504 	if (trim_sz < dmac_size) {
4505 		(*windowp)->wd_trim.tr_trim_first = B_TRUE;
4506 		(*windowp)->wd_trim.tr_first_paddr = cookie->dmac_laddress +
4507 		    coffset;
4508 		(*windowp)->wd_trim.tr_first_size = trim_sz;
4509 	}
4510 
4511 	return (DDI_SUCCESS);
4512 }
4513 
4514 
4515 /*ARGSUSED*/
4516 static int
4517 rootnex_coredma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4518     off_t off, size_t len, uint_t cache_flags)
4519 {
4520 	rootnex_sglinfo_t *sinfo;
4521 	rootnex_pgmap_t *cbpage;
4522 	rootnex_window_t *win;
4523 	ddi_dma_impl_t *hp;
4524 	rootnex_dma_t *dma;
4525 	caddr_t fromaddr;
4526 	caddr_t toaddr;
4527 	uint_t psize;
4528 	off_t offset;
4529 	uint_t pidx;
4530 	size_t size;
4531 	off_t poff;
4532 	int e;
4533 
4534 
4535 	hp = (ddi_dma_impl_t *)handle;
4536 	dma = (rootnex_dma_t *)hp->dmai_private;
4537 	sinfo = &dma->dp_sglinfo;
4538 
4539 	/*
4540 	 * if we don't have any windows, we don't need to sync. A copybuf
4541 	 * will cause us to have at least one window.
4542 	 */
4543 	if (dma->dp_window == NULL) {
4544 		return (DDI_SUCCESS);
4545 	}
4546 
4547 	/* This window may not need to be sync'd */
4548 	win = &dma->dp_window[dma->dp_current_win];
4549 	if (!win->wd_dosync) {
4550 		return (DDI_SUCCESS);
4551 	}
4552 
4553 	/* handle off and len special cases */
4554 	if ((off == 0) || (rootnex_sync_ignore_params)) {
4555 		offset = win->wd_offset;
4556 	} else {
4557 		offset = off;
4558 	}
4559 	if ((len == 0) || (rootnex_sync_ignore_params)) {
4560 		size = win->wd_size;
4561 	} else {
4562 		size = len;
4563 	}
4564 
4565 	/* check the sync args to make sure they make a little sense */
4566 	if (rootnex_sync_check_parms) {
4567 		e = rootnex_valid_sync_parms(hp, win, offset, size,
4568 		    cache_flags);
4569 		if (e != DDI_SUCCESS) {
4570 			ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_SYNC_FAIL]);
4571 			return (DDI_FAILURE);
4572 		}
4573 	}
4574 
4575 	/*
4576 	 * special case the first page to handle the offset into the page. The
4577 	 * offset to the current page for our buffer is the offset into the
4578 	 * first page of the buffer plus our current offset into the buffer
4579 	 * itself, masked of course.
4580 	 */
4581 	poff = (sinfo->si_buf_offset + offset) & MMU_PAGEOFFSET;
4582 	psize = MIN((MMU_PAGESIZE - poff), size);
4583 
4584 	/* go through all the pages that we want to sync */
4585 	while (size > 0) {
4586 		/*
4587 		 * Calculate the page index relative to the start of the buffer.
4588 		 * The index to the current page for our buffer is the offset
4589 		 * into the first page of the buffer plus our current offset
4590 		 * into the buffer itself, shifted of course...
4591 		 */
4592 		pidx = (sinfo->si_buf_offset + offset) >> MMU_PAGESHIFT;
4593 		ASSERT(pidx < sinfo->si_max_pages);
4594 
4595 		/*
4596 		 * if this page uses the copy buffer, we need to sync it,
4597 		 * otherwise, go on to the next page.
4598 		 */
4599 		cbpage = &dma->dp_pgmap[pidx];
4600 		ASSERT((cbpage->pm_uses_copybuf == B_TRUE) ||
4601 		    (cbpage->pm_uses_copybuf == B_FALSE));
4602 		if (cbpage->pm_uses_copybuf) {
4603 			/* cbaddr and kaddr should be page aligned */
4604 			ASSERT(((uintptr_t)cbpage->pm_cbaddr &
4605 			    MMU_PAGEOFFSET) == 0);
4606 			ASSERT(((uintptr_t)cbpage->pm_kaddr &
4607 			    MMU_PAGEOFFSET) == 0);
4608 
4609 			/*
4610 			 * if we're copying for the device, we are going to
4611 			 * copy from the drivers buffer and to the rootnex
4612 			 * allocated copy buffer.
4613 			 */
4614 			if (cache_flags == DDI_DMA_SYNC_FORDEV) {
4615 				fromaddr = cbpage->pm_kaddr + poff;
4616 				toaddr = cbpage->pm_cbaddr + poff;
4617 				ROOTNEX_DPROBE2(rootnex__sync__dev,
4618 				    dev_info_t *, dma->dp_dip, size_t, psize);
4619 
4620 			/*
4621 			 * if we're copying for the cpu/kernel, we are going to
4622 			 * copy from the rootnex allocated copy buffer to the
4623 			 * drivers buffer.
4624 			 */
4625 			} else {
4626 				fromaddr = cbpage->pm_cbaddr + poff;
4627 				toaddr = cbpage->pm_kaddr + poff;
4628 				ROOTNEX_DPROBE2(rootnex__sync__cpu,
4629 				    dev_info_t *, dma->dp_dip, size_t, psize);
4630 			}
4631 
4632 			bcopy(fromaddr, toaddr, psize);
4633 		}
4634 
4635 		/*
4636 		 * decrement size until we're done, update our offset into the
4637 		 * buffer, and get the next page size.
4638 		 */
4639 		size -= psize;
4640 		offset += psize;
4641 		psize = MIN(MMU_PAGESIZE, size);
4642 
4643 		/* page offset is zero for the rest of this loop */
4644 		poff = 0;
4645 	}
4646 
4647 	return (DDI_SUCCESS);
4648 }
4649 
4650 /*
4651  * rootnex_dma_sync()
4652  *    called from ddi_dma_sync() if DMP_NOSYNC is not set in hp->dmai_rflags.
4653  *    We set DMP_NOSYNC if we're not using the copy buffer. If DMP_NOSYNC
4654  *    is set, ddi_dma_sync() returns immediately passing back success.
4655  */
4656 /*ARGSUSED*/
4657 static int
4658 rootnex_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4659     off_t off, size_t len, uint_t cache_flags)
4660 {
4661 #if defined(__amd64) && !defined(__xpv)
4662 	if (IOMMU_USED(rdip)) {
4663 		return (iommulib_nexdma_sync(dip, rdip, handle, off, len,
4664 		    cache_flags));
4665 	}
4666 #endif
4667 	return (rootnex_coredma_sync(dip, rdip, handle, off, len,
4668 	    cache_flags));
4669 }
4670 
4671 /*
4672  * rootnex_valid_sync_parms()
4673  *    checks the parameters passed to sync to verify they are correct.
4674  */
4675 static int
4676 rootnex_valid_sync_parms(ddi_dma_impl_t *hp, rootnex_window_t *win,
4677     off_t offset, size_t size, uint_t cache_flags)
4678 {
4679 	off_t woffset;
4680 
4681 
4682 	/*
4683 	 * the first part of the test to make sure the offset passed in is
4684 	 * within the window.
4685 	 */
4686 	if (offset < win->wd_offset) {
4687 		return (DDI_FAILURE);
4688 	}
4689 
4690 	/*
4691 	 * second and last part of the test to make sure the offset and length
4692 	 * passed in is within the window.
4693 	 */
4694 	woffset = offset - win->wd_offset;
4695 	if ((woffset + size) > win->wd_size) {
4696 		return (DDI_FAILURE);
4697 	}
4698 
4699 	/*
4700 	 * if we are sync'ing for the device, the DDI_DMA_WRITE flag should
4701 	 * be set too.
4702 	 */
4703 	if ((cache_flags == DDI_DMA_SYNC_FORDEV) &&
4704 	    (hp->dmai_rflags & DDI_DMA_WRITE)) {
4705 		return (DDI_SUCCESS);
4706 	}
4707 
4708 	/*
4709 	 * at this point, either DDI_DMA_SYNC_FORCPU or DDI_DMA_SYNC_FORKERNEL
4710 	 * should be set. Also DDI_DMA_READ should be set in the flags.
4711 	 */
4712 	if (((cache_flags == DDI_DMA_SYNC_FORCPU) ||
4713 	    (cache_flags == DDI_DMA_SYNC_FORKERNEL)) &&
4714 	    (hp->dmai_rflags & DDI_DMA_READ)) {
4715 		return (DDI_SUCCESS);
4716 	}
4717 
4718 	return (DDI_FAILURE);
4719 }
4720 
4721 
4722 /*ARGSUSED*/
4723 static int
4724 rootnex_coredma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4725     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4726     uint_t *ccountp)
4727 {
4728 	rootnex_window_t *window;
4729 	rootnex_trim_t *trim;
4730 	ddi_dma_impl_t *hp;
4731 	rootnex_dma_t *dma;
4732 	ddi_dma_obj_t *dmao;
4733 #if !defined(__amd64)
4734 	rootnex_sglinfo_t *sinfo;
4735 	rootnex_pgmap_t *pmap;
4736 	uint_t pidx;
4737 	uint_t pcnt;
4738 	off_t poff;
4739 	int i;
4740 #endif
4741 
4742 
4743 	hp = (ddi_dma_impl_t *)handle;
4744 	dma = (rootnex_dma_t *)hp->dmai_private;
4745 #if !defined(__amd64)
4746 	sinfo = &dma->dp_sglinfo;
4747 #endif
4748 
4749 	/* If we try and get a window which doesn't exist, return failure */
4750 	if (win >= hp->dmai_nwin) {
4751 		ROOTNEX_DPROF_INC(&rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4752 		return (DDI_FAILURE);
4753 	}
4754 
4755 	dmao = dma->dp_dvma_used ? &dma->dp_dma : &dma->dp_dvma;
4756 
4757 	/*
4758 	 * if we don't have any windows, and they're asking for the first
4759 	 * window, setup the cookie pointer to the first cookie in the bind.
4760 	 * setup our return values, then increment the cookie since we return
4761 	 * the first cookie on the stack.
4762 	 */
4763 	if (dma->dp_window == NULL) {
4764 		if (win != 0) {
4765 			ROOTNEX_DPROF_INC(
4766 			    &rootnex_cnt[ROOTNEX_CNT_GETWIN_FAIL]);
4767 			return (DDI_FAILURE);
4768 		}
4769 		hp->dmai_cookie = dma->dp_cookies;
4770 		*offp = 0;
4771 		*lenp = dmao->dmao_size;
4772 		*ccountp = dma->dp_sglinfo.si_sgl_size;
4773 		*cookiep = hp->dmai_cookie[0];
4774 		hp->dmai_cookie++;
4775 		return (DDI_SUCCESS);
4776 	}
4777 
4778 	/* sync the old window before moving on to the new one */
4779 	window = &dma->dp_window[dma->dp_current_win];
4780 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_READ)) {
4781 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4782 		    DDI_DMA_SYNC_FORCPU);
4783 	}
4784 
4785 #if !defined(__amd64)
4786 	/*
4787 	 * before we move to the next window, if we need to re-map, unmap all
4788 	 * the pages in this window.
4789 	 */
4790 	if (dma->dp_cb_remaping) {
4791 		/*
4792 		 * If we switch to this window again, we'll need to map in
4793 		 * on the fly next time.
4794 		 */
4795 		window->wd_remap_copybuf = B_TRUE;
4796 
4797 		/*
4798 		 * calculate the page index into the buffer where this window
4799 		 * starts, and the number of pages this window takes up.
4800 		 */
4801 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4802 		    MMU_PAGESHIFT;
4803 		poff = (sinfo->si_buf_offset + window->wd_offset) &
4804 		    MMU_PAGEOFFSET;
4805 		pcnt = mmu_btopr(window->wd_size + poff);
4806 		ASSERT((pidx + pcnt) <= sinfo->si_max_pages);
4807 
4808 		/* unmap pages which are currently mapped in this window */
4809 		for (i = 0; i < pcnt; i++) {
4810 			if (dma->dp_pgmap[pidx].pm_mapped) {
4811 				hat_unload(kas.a_hat,
4812 				    dma->dp_pgmap[pidx].pm_kaddr, MMU_PAGESIZE,
4813 				    HAT_UNLOAD);
4814 				dma->dp_pgmap[pidx].pm_mapped = B_FALSE;
4815 			}
4816 			pidx++;
4817 		}
4818 	}
4819 #endif
4820 
4821 	/*
4822 	 * Move to the new window.
4823 	 * NOTE: current_win must be set for sync to work right
4824 	 */
4825 	dma->dp_current_win = win;
4826 	window = &dma->dp_window[win];
4827 
4828 	/* if needed, adjust the first and/or last cookies for trim */
4829 	trim = &window->wd_trim;
4830 	if (trim->tr_trim_first) {
4831 		window->wd_first_cookie->dmac_laddress = trim->tr_first_paddr;
4832 		window->wd_first_cookie->dmac_size = trim->tr_first_size;
4833 #if !defined(__amd64)
4834 		window->wd_first_cookie->dmac_type =
4835 		    (window->wd_first_cookie->dmac_type &
4836 		    ROOTNEX_USES_COPYBUF) + window->wd_offset;
4837 #endif
4838 		if (trim->tr_first_copybuf_win) {
4839 			dma->dp_pgmap[trim->tr_first_pidx].pm_cbaddr =
4840 			    trim->tr_first_cbaddr;
4841 #if !defined(__amd64)
4842 			dma->dp_pgmap[trim->tr_first_pidx].pm_kaddr =
4843 			    trim->tr_first_kaddr;
4844 #endif
4845 		}
4846 	}
4847 	if (trim->tr_trim_last) {
4848 		trim->tr_last_cookie->dmac_laddress = trim->tr_last_paddr;
4849 		trim->tr_last_cookie->dmac_size = trim->tr_last_size;
4850 		if (trim->tr_last_copybuf_win) {
4851 			dma->dp_pgmap[trim->tr_last_pidx].pm_cbaddr =
4852 			    trim->tr_last_cbaddr;
4853 #if !defined(__amd64)
4854 			dma->dp_pgmap[trim->tr_last_pidx].pm_kaddr =
4855 			    trim->tr_last_kaddr;
4856 #endif
4857 		}
4858 	}
4859 
4860 	/*
4861 	 * setup the cookie pointer to the first cookie in the window. setup
4862 	 * our return values, then increment the cookie since we return the
4863 	 * first cookie on the stack.
4864 	 */
4865 	hp->dmai_cookie = window->wd_first_cookie;
4866 	*offp = window->wd_offset;
4867 	*lenp = window->wd_size;
4868 	*ccountp = window->wd_cookie_cnt;
4869 	*cookiep = hp->dmai_cookie[0];
4870 	hp->dmai_cookie++;
4871 
4872 #if !defined(__amd64)
4873 	/* re-map copybuf if required for this window */
4874 	if (dma->dp_cb_remaping) {
4875 		/*
4876 		 * calculate the page index into the buffer where this
4877 		 * window starts.
4878 		 */
4879 		pidx = (sinfo->si_buf_offset + window->wd_offset) >>
4880 		    MMU_PAGESHIFT;
4881 		ASSERT(pidx < sinfo->si_max_pages);
4882 
4883 		/*
4884 		 * the first page can get unmapped if it's shared with the
4885 		 * previous window. Even if the rest of this window is already
4886 		 * mapped in, we need to still check this one.
4887 		 */
4888 		pmap = &dma->dp_pgmap[pidx];
4889 		if ((pmap->pm_uses_copybuf) && (pmap->pm_mapped == B_FALSE)) {
4890 			if (pmap->pm_pp != NULL) {
4891 				pmap->pm_mapped = B_TRUE;
4892 				i86_pp_map(pmap->pm_pp, pmap->pm_kaddr);
4893 			} else if (pmap->pm_vaddr != NULL) {
4894 				pmap->pm_mapped = B_TRUE;
4895 				i86_va_map(pmap->pm_vaddr, sinfo->si_asp,
4896 				    pmap->pm_kaddr);
4897 			}
4898 		}
4899 		pidx++;
4900 
4901 		/* map in the rest of the pages if required */
4902 		if (window->wd_remap_copybuf) {
4903 			window->wd_remap_copybuf = B_FALSE;
4904 
4905 			/* figure out many pages this window takes up */
4906 			poff = (sinfo->si_buf_offset + window->wd_offset) &
4907 			    MMU_PAGEOFFSET;
4908 			pcnt = mmu_btopr(window->wd_size + poff);
4909 			ASSERT(((pidx - 1) + pcnt) <= sinfo->si_max_pages);
4910 
4911 			/* map pages which require it */
4912 			for (i = 1; i < pcnt; i++) {
4913 				pmap = &dma->dp_pgmap[pidx];
4914 				if (pmap->pm_uses_copybuf) {
4915 					ASSERT(pmap->pm_mapped == B_FALSE);
4916 					if (pmap->pm_pp != NULL) {
4917 						pmap->pm_mapped = B_TRUE;
4918 						i86_pp_map(pmap->pm_pp,
4919 						    pmap->pm_kaddr);
4920 					} else if (pmap->pm_vaddr != NULL) {
4921 						pmap->pm_mapped = B_TRUE;
4922 						i86_va_map(pmap->pm_vaddr,
4923 						    sinfo->si_asp,
4924 						    pmap->pm_kaddr);
4925 					}
4926 				}
4927 				pidx++;
4928 			}
4929 		}
4930 	}
4931 #endif
4932 
4933 	/* if the new window uses the copy buffer, sync it for the device */
4934 	if ((window->wd_dosync) && (hp->dmai_rflags & DDI_DMA_WRITE)) {
4935 		(void) rootnex_coredma_sync(dip, rdip, handle, 0, 0,
4936 		    DDI_DMA_SYNC_FORDEV);
4937 	}
4938 
4939 	return (DDI_SUCCESS);
4940 }
4941 
4942 /*
4943  * rootnex_dma_win()
4944  *    called from ddi_dma_getwin()
4945  */
4946 /*ARGSUSED*/
4947 static int
4948 rootnex_dma_win(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
4949     uint_t win, off_t *offp, size_t *lenp, ddi_dma_cookie_t *cookiep,
4950     uint_t *ccountp)
4951 {
4952 #if defined(__amd64) && !defined(__xpv)
4953 	if (IOMMU_USED(rdip)) {
4954 		return (iommulib_nexdma_win(dip, rdip, handle, win, offp, lenp,
4955 		    cookiep, ccountp));
4956 	}
4957 #endif
4958 
4959 	return (rootnex_coredma_win(dip, rdip, handle, win, offp, lenp,
4960 	    cookiep, ccountp));
4961 }
4962 
4963 /*ARGSUSED*/
4964 static int
4965 rootnex_coredma_hdl_setprivate(dev_info_t *dip, dev_info_t *rdip,
4966     ddi_dma_handle_t handle, void *v)
4967 {
4968 	ddi_dma_impl_t *hp;
4969 	rootnex_dma_t *dma;
4970 
4971 	hp = (ddi_dma_impl_t *)handle;
4972 	dma = (rootnex_dma_t *)hp->dmai_private;
4973 	dma->dp_iommu_private = v;
4974 
4975 	return (DDI_SUCCESS);
4976 }
4977 
4978 /*ARGSUSED*/
4979 static void *
4980 rootnex_coredma_hdl_getprivate(dev_info_t *dip, dev_info_t *rdip,
4981     ddi_dma_handle_t handle)
4982 {
4983 	ddi_dma_impl_t *hp;
4984 	rootnex_dma_t *dma;
4985 
4986 	hp = (ddi_dma_impl_t *)handle;
4987 	dma = (rootnex_dma_t *)hp->dmai_private;
4988 
4989 	return (dma->dp_iommu_private);
4990 }
4991 
4992 /*
4993  * ************************
4994  *  obsoleted dma routines
4995  * ************************
4996  */
4997 
4998 /*
4999  * rootnex_dma_map()
5000  *    called from ddi_dma_setup()
5001  * NO IOMMU in 32 bit mode. The below routines doesn't work in 64 bit mode.
5002  */
5003 /* ARGSUSED */
5004 static int
5005 rootnex_dma_map(dev_info_t *dip, dev_info_t *rdip,
5006     struct ddi_dma_req *dmareq, ddi_dma_handle_t *handlep)
5007 {
5008 #if defined(__amd64)
5009 	/*
5010 	 * this interface is not supported in 64-bit x86 kernel. See comment in
5011 	 * rootnex_dma_mctl()
5012 	 */
5013 	return (DDI_DMA_NORESOURCES);
5014 
5015 #else /* 32-bit x86 kernel */
5016 	ddi_dma_handle_t *lhandlep;
5017 	ddi_dma_handle_t lhandle;
5018 	ddi_dma_cookie_t cookie;
5019 	ddi_dma_attr_t dma_attr;
5020 	ddi_dma_lim_t *dma_lim;
5021 	uint_t ccnt;
5022 	int e;
5023 
5024 
5025 	/*
5026 	 * if the driver is just testing to see if it's possible to do the bind,
5027 	 * we'll use local state. Otherwise, use the handle pointer passed in.
5028 	 */
5029 	if (handlep == NULL) {
5030 		lhandlep = &lhandle;
5031 	} else {
5032 		lhandlep = handlep;
5033 	}
5034 
5035 	/* convert the limit structure to a dma_attr one */
5036 	dma_lim = dmareq->dmar_limits;
5037 	dma_attr.dma_attr_version = DMA_ATTR_V0;
5038 	dma_attr.dma_attr_addr_lo = dma_lim->dlim_addr_lo;
5039 	dma_attr.dma_attr_addr_hi = dma_lim->dlim_addr_hi;
5040 	dma_attr.dma_attr_minxfer = dma_lim->dlim_minxfer;
5041 	dma_attr.dma_attr_seg = dma_lim->dlim_adreg_max;
5042 	dma_attr.dma_attr_count_max = dma_lim->dlim_ctreg_max;
5043 	dma_attr.dma_attr_granular = dma_lim->dlim_granular;
5044 	dma_attr.dma_attr_sgllen = dma_lim->dlim_sgllen;
5045 	dma_attr.dma_attr_maxxfer = dma_lim->dlim_reqsize;
5046 	dma_attr.dma_attr_burstsizes = dma_lim->dlim_burstsizes;
5047 	dma_attr.dma_attr_align = MMU_PAGESIZE;
5048 	dma_attr.dma_attr_flags = 0;
5049 
5050 	e = rootnex_dma_allochdl(dip, rdip, &dma_attr, dmareq->dmar_fp,
5051 	    dmareq->dmar_arg, lhandlep);
5052 	if (e != DDI_SUCCESS) {
5053 		return (e);
5054 	}
5055 
5056 	e = rootnex_dma_bindhdl(dip, rdip, *lhandlep, dmareq, &cookie, &ccnt);
5057 	if ((e != DDI_DMA_MAPPED) && (e != DDI_DMA_PARTIAL_MAP)) {
5058 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5059 		return (e);
5060 	}
5061 
5062 	/*
5063 	 * if the driver is just testing to see if it's possible to do the bind,
5064 	 * free up the local state and return the result.
5065 	 */
5066 	if (handlep == NULL) {
5067 		(void) rootnex_dma_unbindhdl(dip, rdip, *lhandlep);
5068 		(void) rootnex_dma_freehdl(dip, rdip, *lhandlep);
5069 		if (e == DDI_DMA_MAPPED) {
5070 			return (DDI_DMA_MAPOK);
5071 		} else {
5072 			return (DDI_DMA_NOMAPPING);
5073 		}
5074 	}
5075 
5076 	return (e);
5077 #endif /* defined(__amd64) */
5078 }
5079 
5080 /*
5081  * rootnex_dma_mctl()
5082  *
5083  * No IOMMU in 32 bit mode. The below routine doesn't work in 64 bit mode.
5084  */
5085 /* ARGSUSED */
5086 static int
5087 rootnex_dma_mctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
5088     enum ddi_dma_ctlops request, off_t *offp, size_t *lenp, caddr_t *objpp,
5089     uint_t cache_flags)
5090 {
5091 #if defined(__amd64)
5092 	/*
5093 	 * DDI_DMA_SMEM_ALLOC & DDI_DMA_IOPB_ALLOC we're changed to have a
5094 	 * common implementation in genunix, so they no longer have x86
5095 	 * specific functionality which called into dma_ctl.
5096 	 *
5097 	 * The rest of the obsoleted interfaces were never supported in the
5098 	 * 64-bit x86 kernel. For s10, the obsoleted DDI_DMA_SEGTOC interface
5099 	 * was not ported to the x86 64-bit kernel do to serious x86 rootnex
5100 	 * implementation issues.
5101 	 *
5102 	 * If you can't use DDI_DMA_SEGTOC; DDI_DMA_NEXTSEG, DDI_DMA_FREE, and
5103 	 * DDI_DMA_NEXTWIN are useless since you can get to the cookie, so we
5104 	 * reflect that now too...
5105 	 *
5106 	 * Even though we fixed the pointer problem in DDI_DMA_SEGTOC, we are
5107 	 * not going to put this functionality into the 64-bit x86 kernel now.
5108 	 * It wasn't ported to the 64-bit kernel for s10, no reason to change
5109 	 * that in a future release.
5110 	 */
5111 	return (DDI_FAILURE);
5112 
5113 #else /* 32-bit x86 kernel */
5114 	ddi_dma_cookie_t lcookie;
5115 	ddi_dma_cookie_t *cookie;
5116 	rootnex_window_t *window;
5117 	ddi_dma_impl_t *hp;
5118 	rootnex_dma_t *dma;
5119 	uint_t nwin;
5120 	uint_t ccnt;
5121 	size_t len;
5122 	off_t off;
5123 	int e;
5124 
5125 
5126 	/*
5127 	 * DDI_DMA_SEGTOC, DDI_DMA_NEXTSEG, and DDI_DMA_NEXTWIN are a little
5128 	 * hacky since were optimizing for the current interfaces and so we can
5129 	 * cleanup the mess in genunix. Hopefully we will remove the this
5130 	 * obsoleted routines someday soon.
5131 	 */
5132 
5133 	switch (request) {
5134 
5135 	case DDI_DMA_SEGTOC: /* ddi_dma_segtocookie() */
5136 		hp = (ddi_dma_impl_t *)handle;
5137 		cookie = (ddi_dma_cookie_t *)objpp;
5138 
5139 		/*
5140 		 * convert segment to cookie. We don't distinguish between the
5141 		 * two :-)
5142 		 */
5143 		*cookie = *hp->dmai_cookie;
5144 		*lenp = cookie->dmac_size;
5145 		*offp = cookie->dmac_type & ~ROOTNEX_USES_COPYBUF;
5146 		return (DDI_SUCCESS);
5147 
5148 	case DDI_DMA_NEXTSEG: /* ddi_dma_nextseg() */
5149 		hp = (ddi_dma_impl_t *)handle;
5150 		dma = (rootnex_dma_t *)hp->dmai_private;
5151 
5152 		if ((*lenp != NULL) && ((uintptr_t)*lenp != (uintptr_t)hp)) {
5153 			return (DDI_DMA_STALE);
5154 		}
5155 
5156 		/* handle the case where we don't have any windows */
5157 		if (dma->dp_window == NULL) {
5158 			/*
5159 			 * if seg == NULL, and we don't have any windows,
5160 			 * return the first cookie in the sgl.
5161 			 */
5162 			if (*lenp == NULL) {
5163 				dma->dp_current_cookie = 0;
5164 				hp->dmai_cookie = dma->dp_cookies;
5165 				*objpp = (caddr_t)handle;
5166 				return (DDI_SUCCESS);
5167 
5168 			/* if we have more cookies, go to the next cookie */
5169 			} else {
5170 				if ((dma->dp_current_cookie + 1) >=
5171 				    dma->dp_sglinfo.si_sgl_size) {
5172 					return (DDI_DMA_DONE);
5173 				}
5174 				dma->dp_current_cookie++;
5175 				hp->dmai_cookie++;
5176 				return (DDI_SUCCESS);
5177 			}
5178 		}
5179 
5180 		/* We have one or more windows */
5181 		window = &dma->dp_window[dma->dp_current_win];
5182 
5183 		/*
5184 		 * if seg == NULL, return the first cookie in the current
5185 		 * window
5186 		 */
5187 		if (*lenp == NULL) {
5188 			dma->dp_current_cookie = 0;
5189 			hp->dmai_cookie = window->wd_first_cookie;
5190 
5191 		/*
5192 		 * go to the next cookie in the window then see if we done with
5193 		 * this window.
5194 		 */
5195 		} else {
5196 			if ((dma->dp_current_cookie + 1) >=
5197 			    window->wd_cookie_cnt) {
5198 				return (DDI_DMA_DONE);
5199 			}
5200 			dma->dp_current_cookie++;
5201 			hp->dmai_cookie++;
5202 		}
5203 		*objpp = (caddr_t)handle;
5204 		return (DDI_SUCCESS);
5205 
5206 	case DDI_DMA_NEXTWIN: /* ddi_dma_nextwin() */
5207 		hp = (ddi_dma_impl_t *)handle;
5208 		dma = (rootnex_dma_t *)hp->dmai_private;
5209 
5210 		if ((*offp != NULL) && ((uintptr_t)*offp != (uintptr_t)hp)) {
5211 			return (DDI_DMA_STALE);
5212 		}
5213 
5214 		/* if win == NULL, return the first window in the bind */
5215 		if (*offp == NULL) {
5216 			nwin = 0;
5217 
5218 		/*
5219 		 * else, go to the next window then see if we're done with all
5220 		 * the windows.
5221 		 */
5222 		} else {
5223 			nwin = dma->dp_current_win + 1;
5224 			if (nwin >= hp->dmai_nwin) {
5225 				return (DDI_DMA_DONE);
5226 			}
5227 		}
5228 
5229 		/* switch to the next window */
5230 		e = rootnex_dma_win(dip, rdip, handle, nwin, &off, &len,
5231 		    &lcookie, &ccnt);
5232 		ASSERT(e == DDI_SUCCESS);
5233 		if (e != DDI_SUCCESS) {
5234 			return (DDI_DMA_STALE);
5235 		}
5236 
5237 		/* reset the cookie back to the first cookie in the window */
5238 		if (dma->dp_window != NULL) {
5239 			window = &dma->dp_window[dma->dp_current_win];
5240 			hp->dmai_cookie = window->wd_first_cookie;
5241 		} else {
5242 			hp->dmai_cookie = dma->dp_cookies;
5243 		}
5244 
5245 		*objpp = (caddr_t)handle;
5246 		return (DDI_SUCCESS);
5247 
5248 	case DDI_DMA_FREE: /* ddi_dma_free() */
5249 		(void) rootnex_dma_unbindhdl(dip, rdip, handle);
5250 		(void) rootnex_dma_freehdl(dip, rdip, handle);
5251 		if (rootnex_state->r_dvma_call_list_id) {
5252 			ddi_run_callback(&rootnex_state->r_dvma_call_list_id);
5253 		}
5254 		return (DDI_SUCCESS);
5255 
5256 	case DDI_DMA_IOPB_ALLOC:	/* get contiguous DMA-able memory */
5257 	case DDI_DMA_SMEM_ALLOC:	/* get contiguous DMA-able memory */
5258 		/* should never get here, handled in genunix */
5259 		ASSERT(0);
5260 		return (DDI_FAILURE);
5261 
5262 	case DDI_DMA_KVADDR:
5263 	case DDI_DMA_GETERR:
5264 	case DDI_DMA_COFF:
5265 		return (DDI_FAILURE);
5266 	}
5267 
5268 	return (DDI_FAILURE);
5269 #endif /* defined(__amd64) */
5270 }
5271 
5272 /*
5273  * *********
5274  *  FMA Code
5275  * *********
5276  */
5277 
5278 /*
5279  * rootnex_fm_init()
5280  *    FMA init busop
5281  */
5282 /* ARGSUSED */
5283 static int
5284 rootnex_fm_init(dev_info_t *dip, dev_info_t *tdip, int tcap,
5285     ddi_iblock_cookie_t *ibc)
5286 {
5287 	*ibc = rootnex_state->r_err_ibc;
5288 
5289 	return (ddi_system_fmcap);
5290 }
5291 
5292 /*
5293  * rootnex_dma_check()
5294  *    Function called after a dma fault occurred to find out whether the
5295  *    fault address is associated with a driver that is able to handle faults
5296  *    and recover from faults.
5297  */
5298 /* ARGSUSED */
5299 static int
5300 rootnex_dma_check(dev_info_t *dip, const void *handle, const void *addr,
5301     const void *not_used)
5302 {
5303 	rootnex_window_t *window;
5304 	uint64_t start_addr;
5305 	uint64_t fault_addr;
5306 	ddi_dma_impl_t *hp;
5307 	rootnex_dma_t *dma;
5308 	uint64_t end_addr;
5309 	size_t csize;
5310 	int i;
5311 	int j;
5312 
5313 
5314 	/* The driver has to set DDI_DMA_FLAGERR to recover from dma faults */
5315 	hp = (ddi_dma_impl_t *)handle;
5316 	ASSERT(hp);
5317 
5318 	dma = (rootnex_dma_t *)hp->dmai_private;
5319 
5320 	/* Get the address that we need to search for */
5321 	fault_addr = *(uint64_t *)addr;
5322 
5323 	/*
5324 	 * if we don't have any windows, we can just walk through all the
5325 	 * cookies.
5326 	 */
5327 	if (dma->dp_window == NULL) {
5328 		/* for each cookie */
5329 		for (i = 0; i < dma->dp_sglinfo.si_sgl_size; i++) {
5330 			/*
5331 			 * if the faulted address is within the physical address
5332 			 * range of the cookie, return DDI_FM_NONFATAL.
5333 			 */
5334 			if ((fault_addr >= dma->dp_cookies[i].dmac_laddress) &&
5335 			    (fault_addr <= (dma->dp_cookies[i].dmac_laddress +
5336 			    dma->dp_cookies[i].dmac_size))) {
5337 				return (DDI_FM_NONFATAL);
5338 			}
5339 		}
5340 
5341 		/* fault_addr not within this DMA handle */
5342 		return (DDI_FM_UNKNOWN);
5343 	}
5344 
5345 	/* we have mutiple windows, walk through each window */
5346 	for (i = 0; i < hp->dmai_nwin; i++) {
5347 		window = &dma->dp_window[i];
5348 
5349 		/* Go through all the cookies in the window */
5350 		for (j = 0; j < window->wd_cookie_cnt; j++) {
5351 
5352 			start_addr = window->wd_first_cookie[j].dmac_laddress;
5353 			csize = window->wd_first_cookie[j].dmac_size;
5354 
5355 			/*
5356 			 * if we are trimming the first cookie in the window,
5357 			 * and this is the first cookie, adjust the start
5358 			 * address and size of the cookie to account for the
5359 			 * trim.
5360 			 */
5361 			if (window->wd_trim.tr_trim_first && (j == 0)) {
5362 				start_addr = window->wd_trim.tr_first_paddr;
5363 				csize = window->wd_trim.tr_first_size;
5364 			}
5365 
5366 			/*
5367 			 * if we are trimming the last cookie in the window,
5368 			 * and this is the last cookie, adjust the start
5369 			 * address and size of the cookie to account for the
5370 			 * trim.
5371 			 */
5372 			if (window->wd_trim.tr_trim_last &&
5373 			    (j == (window->wd_cookie_cnt - 1))) {
5374 				start_addr = window->wd_trim.tr_last_paddr;
5375 				csize = window->wd_trim.tr_last_size;
5376 			}
5377 
5378 			end_addr = start_addr + csize;
5379 
5380 			/*
5381 			 * if the faulted address is within the physical
5382 			 * address of the cookie, return DDI_FM_NONFATAL.
5383 			 */
5384 			if ((fault_addr >= start_addr) &&
5385 			    (fault_addr <= end_addr)) {
5386 				return (DDI_FM_NONFATAL);
5387 			}
5388 		}
5389 	}
5390 
5391 	/* fault_addr not within this DMA handle */
5392 	return (DDI_FM_UNKNOWN);
5393 }
5394 
5395 /*ARGSUSED*/
5396 static int
5397 rootnex_quiesce(dev_info_t *dip)
5398 {
5399 #if defined(__amd64) && !defined(__xpv)
5400 	return (immu_quiesce());
5401 #else
5402 	return (DDI_SUCCESS);
5403 #endif
5404 }
5405 
5406 #if defined(__xpv)
5407 void
5408 immu_init(void)
5409 {
5410 	;
5411 }
5412 
5413 void
5414 immu_startup(void)
5415 {
5416 	;
5417 }
5418 /*ARGSUSED*/
5419 void
5420 immu_physmem_update(uint64_t addr, uint64_t size)
5421 {
5422 	;
5423 }
5424 #endif
5425