xref: /titanic_50/usr/src/uts/sun4/io/px/px_dma.c (revision 210938ebfcf386cf4819f16c23979173b0d78097)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * PCI Express nexus DVMA and DMA core routines:
30  *	dma_map/dma_bind_handle implementation
31  *	bypass and peer-to-peer support
32  *	fast track DVMA space allocation
33  *	runtime DVMA debug
34  */
35 #include <sys/types.h>
36 #include <sys/kmem.h>
37 #include <sys/async.h>
38 #include <sys/sysmacros.h>
39 #include <sys/sunddi.h>
40 #include <sys/ddi_impldefs.h>
41 #include "px_obj.h"
42 
43 /*LINTLIBRARY*/
44 
45 /*
46  * px_dma_allocmp - Allocate a pci dma implementation structure
47  *
48  * An extra ddi_dma_attr structure is bundled with the usual ddi_dma_impl
49  * to hold unmodified device limits. The ddi_dma_attr inside the
50  * ddi_dma_impl structure is augumented with system limits to enhance
51  * DVMA performance at runtime. The unaugumented device limits saved
52  * right after (accessed through (ddi_dma_attr_t *)(mp + 1)) is used
53  * strictly for peer-to-peer transfers which do not obey system limits.
54  *
55  * return: DDI_SUCCESS DDI_DMA_NORESOURCES
56  */
57 ddi_dma_impl_t *
58 px_dma_allocmp(dev_info_t *dip, dev_info_t *rdip, int (*waitfp)(caddr_t),
59 	caddr_t arg)
60 {
61 	register ddi_dma_impl_t *mp;
62 	int sleep = (waitfp == DDI_DMA_SLEEP) ? KM_SLEEP : KM_NOSLEEP;
63 
64 	/* Caution: we don't use zalloc to enhance performance! */
65 	if ((mp = kmem_alloc(sizeof (px_dma_hdl_t), sleep)) == 0) {
66 		DBG(DBG_DMA_MAP, dip, "can't alloc dma_handle\n");
67 		if (waitfp != DDI_DMA_DONTWAIT) {
68 			DBG(DBG_DMA_MAP, dip, "alloc_mp kmem cb\n");
69 			ddi_set_callback(waitfp, arg, &px_kmem_clid);
70 		}
71 		return (mp);
72 	}
73 
74 	mp->dmai_rdip = rdip;
75 	mp->dmai_flags = 0;
76 	mp->dmai_pfnlst = NULL;
77 	mp->dmai_winlst = NULL;
78 
79 	/*
80 	 * kmem_alloc debug: the following fields are not zero-ed
81 	 * mp->dmai_mapping = 0;
82 	 * mp->dmai_size = 0;
83 	 * mp->dmai_offset = 0;
84 	 * mp->dmai_minxfer = 0;
85 	 * mp->dmai_burstsizes = 0;
86 	 * mp->dmai_ndvmapages = 0;
87 	 * mp->dmai_pool/roffset = 0;
88 	 * mp->dmai_rflags = 0;
89 	 * mp->dmai_inuse/flags
90 	 * mp->dmai_nwin = 0;
91 	 * mp->dmai_winsize = 0;
92 	 * mp->dmai_nexus_private/tte = 0;
93 	 * mp->dmai_iopte/pfnlst
94 	 * mp->dmai_sbi/pfn0 = 0;
95 	 * mp->dmai_minfo/winlst/fdvma
96 	 * mp->dmai_rdip
97 	 * bzero(&mp->dmai_object, sizeof (ddi_dma_obj_t));
98 	 * bzero(&mp->dmai_attr, sizeof (ddi_dma_attr_t));
99 	 * mp->dmai_cookie = 0;
100 	 */
101 
102 	mp->dmai_attr.dma_attr_version = (uint_t)DMA_ATTR_VERSION;
103 	mp->dmai_attr.dma_attr_flags = (uint_t)0;
104 	mp->dmai_fault = 0;
105 	mp->dmai_fault_check = NULL;
106 	mp->dmai_fault_notify = NULL;
107 
108 	mp->dmai_error.err_ena = 0;
109 	mp->dmai_error.err_status = DDI_FM_OK;
110 	mp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
111 	mp->dmai_error.err_ontrap = NULL;
112 	mp->dmai_error.err_fep = NULL;
113 	mp->dmai_error.err_cf = NULL;
114 
115 	return (mp);
116 }
117 
118 void
119 px_dma_freemp(ddi_dma_impl_t *mp)
120 {
121 	if (mp->dmai_ndvmapages > 1)
122 		px_dma_freepfn(mp);
123 	if (mp->dmai_winlst)
124 		px_dma_freewin(mp);
125 	kmem_free(mp, sizeof (px_dma_hdl_t));
126 }
127 
128 void
129 px_dma_freepfn(ddi_dma_impl_t *mp)
130 {
131 	void *addr = mp->dmai_pfnlst;
132 	if (addr) {
133 		size_t npages = mp->dmai_ndvmapages;
134 		if (npages > 1)
135 			kmem_free(addr, npages * sizeof (px_iopfn_t));
136 		mp->dmai_pfnlst = NULL;
137 	}
138 	mp->dmai_ndvmapages = 0;
139 }
140 
141 /*
142  * px_dma_lmts2hdl - alloate a ddi_dma_impl_t, validate practical limits
143  *			and convert dmareq->dmar_limits to mp->dmai_attr
144  *
145  * ddi_dma_impl_t member modified     input
146  * ------------------------------------------------------------------------
147  * mp->dmai_minxfer		    - dev
148  * mp->dmai_burstsizes		    - dev
149  * mp->dmai_flags		    - no limit? peer-to-peer only?
150  *
151  * ddi_dma_attr member modified       input
152  * ------------------------------------------------------------------------
153  * mp->dmai_attr.dma_attr_addr_lo   - dev lo, sys lo
154  * mp->dmai_attr.dma_attr_addr_hi   - dev hi, sys hi
155  * mp->dmai_attr.dma_attr_count_max - dev count max, dev/sys lo/hi delta
156  * mp->dmai_attr.dma_attr_seg       - 0         (no nocross   restriction)
157  * mp->dmai_attr.dma_attr_align     - 1         (no alignment restriction)
158  *
159  * The dlim_dmaspeed member of dmareq->dmar_limits is ignored.
160  */
161 ddi_dma_impl_t *
162 px_dma_lmts2hdl(dev_info_t *dip, dev_info_t *rdip, px_mmu_t *mmu_p,
163 	ddi_dma_req_t *dmareq)
164 {
165 	ddi_dma_impl_t *mp;
166 	ddi_dma_attr_t *attr_p;
167 	uint64_t syslo		= mmu_p->mmu_dvma_base;
168 	uint64_t syshi		= mmu_p->mmu_dvma_end;
169 	uint64_t fasthi		= mmu_p->mmu_dvma_fast_end;
170 	ddi_dma_lim_t *lim_p	= dmareq->dmar_limits;
171 	uint32_t count_max	= lim_p->dlim_cntr_max;
172 	uint64_t lo		= lim_p->dlim_addr_lo;
173 	uint64_t hi		= lim_p->dlim_addr_hi;
174 	if (hi <= lo) {
175 		DBG(DBG_DMA_MAP, dip, "Bad limits\n");
176 		return ((ddi_dma_impl_t *)DDI_DMA_NOMAPPING);
177 	}
178 	if (!count_max)
179 		count_max--;
180 
181 	if (!(mp = px_dma_allocmp(dip, rdip, dmareq->dmar_fp,
182 		dmareq->dmar_arg)))
183 		return (NULL);
184 
185 	/* store original dev input at the 2nd ddi_dma_attr */
186 	attr_p = PX_DEV_ATTR(mp);
187 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
188 	SET_DMAALIGN(attr_p, 1);
189 
190 	lo = MAX(lo, syslo);
191 	hi = MIN(hi, syshi);
192 	if (hi <= lo)
193 		mp->dmai_flags |= PX_DMAI_FLAGS_PEER_ONLY;
194 	count_max = MIN(count_max, hi - lo);
195 
196 	if (PX_DEV_NOSYSLIMIT(lo, hi, syslo, fasthi, 1))
197 		mp->dmai_flags |= PX_DMAI_FLAGS_NOFASTLIMIT |
198 			PX_DMAI_FLAGS_NOSYSLIMIT;
199 	else {
200 		if (PX_DEV_NOFASTLIMIT(lo, hi, syslo, syshi, 1))
201 			mp->dmai_flags |= PX_DMAI_FLAGS_NOFASTLIMIT;
202 	}
203 	if (PX_DMA_NOCTX(rdip))
204 		mp->dmai_flags |= PX_DMAI_FLAGS_NOCTX;
205 
206 	/* store augumented dev input to mp->dmai_attr */
207 	mp->dmai_minxfer	= lim_p->dlim_minxfer;
208 	mp->dmai_burstsizes	= lim_p->dlim_burstsizes;
209 	attr_p = &mp->dmai_attr;
210 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
211 	SET_DMAALIGN(attr_p, 1);
212 	return (mp);
213 }
214 
215 /*
216  * Called from px_attach to check for bypass dma support and set
217  * flags accordingly.
218  */
219 int
220 px_dma_attach(px_t *px_p)
221 {
222 	uint64_t baddr;
223 
224 	if (px_lib_iommu_getbypass(px_p->px_dip, 0ull,
225 			PCI_MAP_ATTR_WRITE|PCI_MAP_ATTR_READ,
226 			&baddr) != DDI_ENOTSUP)
227 		/* ignore all other errors */
228 		px_p->px_dev_caps |= PX_BYPASS_DMA_ALLOWED;
229 
230 	return (DDI_SUCCESS);
231 }
232 
233 /*
234  * px_dma_attr2hdl
235  *
236  * This routine is called from the alloc handle entry point to sanity check the
237  * dma attribute structure.
238  *
239  * use by: px_dma_allochdl()
240  *
241  * return value:
242  *
243  *	DDI_SUCCESS		- on success
244  *	DDI_DMA_BADATTR		- attribute has invalid version number
245  *				  or address limits exclude dvma space
246  */
247 int
248 px_dma_attr2hdl(px_t *px_p, ddi_dma_impl_t *mp)
249 {
250 	px_mmu_t *mmu_p = px_p->px_mmu_p;
251 	uint64_t syslo, syshi;
252 	int	ret;
253 	ddi_dma_attr_t *attrp		= PX_DEV_ATTR(mp);
254 	uint64_t hi			= attrp->dma_attr_addr_hi;
255 	uint64_t lo			= attrp->dma_attr_addr_lo;
256 	uint64_t align			= attrp->dma_attr_align;
257 	uint64_t nocross		= attrp->dma_attr_seg;
258 	uint64_t count_max		= attrp->dma_attr_count_max;
259 
260 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "attrp=%p cntr_max=%x.%08x\n",
261 		attrp, HI32(count_max), LO32(count_max));
262 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "hi=%x.%08x lo=%x.%08x\n",
263 		HI32(hi), LO32(hi), HI32(lo), LO32(lo));
264 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "seg=%x.%08x align=%x.%08x\n",
265 		HI32(nocross), LO32(nocross), HI32(align), LO32(align));
266 
267 	if (!nocross)
268 		nocross--;
269 	if (attrp->dma_attr_flags & DDI_DMA_FORCE_PHYSICAL) { /* BYPASS */
270 
271 		DBG(DBG_DMA_ALLOCH, px_p->px_dip, "bypass mode\n");
272 		/*
273 		 * If Bypass DMA is not supported, return error so that
274 		 * target driver can fall back to dvma mode of operation
275 		 */
276 		if (!(px_p->px_dev_caps & PX_BYPASS_DMA_ALLOWED))
277 			return (DDI_DMA_BADATTR);
278 		mp->dmai_flags |= PX_DMAI_FLAGS_BYPASSREQ;
279 		if (nocross != UINT64_MAX)
280 			return (DDI_DMA_BADATTR);
281 		if (align && (align > MMU_PAGE_SIZE))
282 			return (DDI_DMA_BADATTR);
283 		align = 1; /* align on 1 page boundary */
284 
285 		/* do a range check and get the limits */
286 		ret = px_lib_dma_bypass_rngchk(px_p->px_dip, attrp,
287 				&syslo, &syshi);
288 		if (ret != DDI_SUCCESS)
289 			return (ret);
290 	} else { /* MMU_XLATE or PEER_TO_PEER */
291 		align = MAX(align, MMU_PAGE_SIZE) - 1;
292 		if ((align & nocross) != align) {
293 			dev_info_t *rdip = mp->dmai_rdip;
294 			cmn_err(CE_WARN, "%s%d dma_attr_seg not aligned",
295 				NAMEINST(rdip));
296 			return (DDI_DMA_BADATTR);
297 		}
298 		align = MMU_BTOP(align + 1);
299 		syslo = mmu_p->mmu_dvma_base;
300 		syshi = mmu_p->mmu_dvma_end;
301 	}
302 	if (hi <= lo) {
303 		dev_info_t *rdip = mp->dmai_rdip;
304 		cmn_err(CE_WARN, "%s%d limits out of range", NAMEINST(rdip));
305 		return (DDI_DMA_BADATTR);
306 	}
307 	lo = MAX(lo, syslo);
308 	hi = MIN(hi, syshi);
309 	if (!count_max)
310 		count_max--;
311 
312 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "hi=%x.%08x, lo=%x.%08x\n",
313 		HI32(hi), LO32(hi), HI32(lo), LO32(lo));
314 	if (hi <= lo) {
315 		/*
316 		 * If this is an IOMMU bypass access, the caller can't use
317 		 * the required addresses, so fail it.  Otherwise, it's
318 		 * peer-to-peer; ensure that the caller has no alignment or
319 		 * segment size restrictions.
320 		 */
321 		if ((mp->dmai_flags & PX_DMAI_FLAGS_BYPASSREQ) ||
322 		    (nocross < UINT32_MAX) || (align > 1))
323 			return (DDI_DMA_BADATTR);
324 
325 		mp->dmai_flags |= PX_DMAI_FLAGS_PEER_ONLY;
326 	} else /* set practical counter_max value */
327 		count_max = MIN(count_max, hi - lo);
328 
329 	if (PX_DEV_NOSYSLIMIT(lo, hi, syslo, syshi, align))
330 		mp->dmai_flags |= PX_DMAI_FLAGS_NOSYSLIMIT |
331 			PX_DMAI_FLAGS_NOFASTLIMIT;
332 	else {
333 		syshi = mmu_p->mmu_dvma_fast_end;
334 		if (PX_DEV_NOFASTLIMIT(lo, hi, syslo, syshi, align))
335 			mp->dmai_flags |= PX_DMAI_FLAGS_NOFASTLIMIT;
336 	}
337 	if (PX_DMA_NOCTX(mp->dmai_rdip))
338 		mp->dmai_flags |= PX_DMAI_FLAGS_NOCTX;
339 
340 	mp->dmai_minxfer	= attrp->dma_attr_minxfer;
341 	mp->dmai_burstsizes	= attrp->dma_attr_burstsizes;
342 	attrp = &mp->dmai_attr;
343 	SET_DMAATTR(attrp, lo, hi, nocross, count_max);
344 	return (DDI_SUCCESS);
345 }
346 
347 #define	TGT_PFN_INBETWEEN(pfn, bgn, end) ((pfn >= bgn) && (pfn <= end))
348 
349 /*
350  * px_dma_type - determine which of the three types DMA (peer-to-peer,
351  *		mmu bypass, or mmu translate) we are asked to do.
352  *		Also checks pfn0 and rejects any non-peer-to-peer
353  *		requests for peer-only devices.
354  *
355  *	return values:
356  *		DDI_DMA_NOMAPPING - can't get valid pfn0, or bad dma type
357  *		DDI_SUCCESS
358  *
359  *	dma handle members affected (set on exit):
360  *	mp->dmai_object		- dmareq->dmar_object
361  *	mp->dmai_rflags		- consistent?, nosync?, dmareq->dmar_flags
362  *	mp->dmai_flags   	- DMA type
363  *	mp->dmai_pfn0   	- 1st page pfn (if va/size pair and not shadow)
364  *	mp->dmai_roffset 	- initialized to starting MMU page offset
365  *	mp->dmai_ndvmapages	- # of total MMU pages of entire object
366  */
367 int
368 px_dma_type(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
369 {
370 	dev_info_t *dip = px_p->px_dip;
371 	ddi_dma_obj_t *dobj_p = &dmareq->dmar_object;
372 	px_pec_t *pec_p = px_p->px_pec_p;
373 	uint32_t offset;
374 	pfn_t pfn0;
375 	uint_t redzone;
376 
377 	mp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS | DMP_NOSYNC;
378 
379 	switch (dobj_p->dmao_type) {
380 	case DMA_OTYP_BUFVADDR:
381 	case DMA_OTYP_VADDR: {
382 		page_t **pplist = dobj_p->dmao_obj.virt_obj.v_priv;
383 		caddr_t vaddr = dobj_p->dmao_obj.virt_obj.v_addr;
384 
385 		DBG(DBG_DMA_MAP, dip, "vaddr=%p pplist=%p\n", vaddr, pplist);
386 		offset = (ulong_t)vaddr & MMU_PAGE_OFFSET;
387 		if (pplist) {				/* shadow list */
388 			mp->dmai_flags |= PX_DMAI_FLAGS_PGPFN;
389 			pfn0 = page_pptonum(*pplist);
390 		} else {
391 			struct as *as_p = dobj_p->dmao_obj.virt_obj.v_as;
392 			struct hat *hat_p = as_p ? as_p->a_hat : kas.a_hat;
393 			pfn0 = hat_getpfnum(hat_p, vaddr);
394 		}
395 		}
396 		break;
397 
398 	case DMA_OTYP_PAGES:
399 		offset = dobj_p->dmao_obj.pp_obj.pp_offset;
400 		mp->dmai_flags |= PX_DMAI_FLAGS_PGPFN;
401 		pfn0 = page_pptonum(dobj_p->dmao_obj.pp_obj.pp_pp);
402 		break;
403 
404 	case DMA_OTYP_PADDR:
405 	default:
406 		cmn_err(CE_WARN, "%s%d requested unsupported dma type %x",
407 			NAMEINST(mp->dmai_rdip), dobj_p->dmao_type);
408 		return (DDI_DMA_NOMAPPING);
409 	}
410 	if (pfn0 == PFN_INVALID) {
411 		cmn_err(CE_WARN, "%s%d: invalid pfn0 for DMA object %p",
412 			NAMEINST(dip), dobj_p);
413 		return (DDI_DMA_NOMAPPING);
414 	}
415 	if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base32_pfn,
416 			pec_p->pec_last32_pfn)) {
417 		mp->dmai_flags |= PX_DMAI_FLAGS_PTP|PX_DMAI_FLAGS_PTP32;
418 		goto done;	/* leave bypass and dvma flag as 0 */
419 	} else if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base64_pfn,
420 			pec_p->pec_last64_pfn)) {
421 		mp->dmai_flags |= PX_DMAI_FLAGS_PTP|PX_DMAI_FLAGS_PTP64;
422 		goto done;	/* leave bypass and dvma flag as 0 */
423 	}
424 	if (PX_DMA_ISPEERONLY(mp)) {
425 		dev_info_t *rdip = mp->dmai_rdip;
426 		cmn_err(CE_WARN, "Bad peer-to-peer req %s%d", NAMEINST(rdip));
427 		return (DDI_DMA_NOMAPPING);
428 	}
429 
430 	redzone = (mp->dmai_rflags & DDI_DMA_REDZONE) ||
431 	    (mp->dmai_flags & PX_DMAI_FLAGS_MAP_BUFZONE) ?
432 	    PX_DMAI_FLAGS_REDZONE : 0;
433 
434 	mp->dmai_flags |= (mp->dmai_flags & PX_DMAI_FLAGS_BYPASSREQ) ?
435 	    PX_DMAI_FLAGS_BYPASS : (PX_DMAI_FLAGS_DVMA | redzone);
436 done:
437 	mp->dmai_object	 = *dobj_p;			/* whole object    */
438 	mp->dmai_pfn0	 = (void *)pfn0;		/* cache pfn0	   */
439 	mp->dmai_roffset = offset;			/* win0 pg0 offset */
440 	mp->dmai_ndvmapages = MMU_BTOPR(offset + mp->dmai_object.dmao_size);
441 	return (DDI_SUCCESS);
442 }
443 
444 /*
445  * px_dma_pgpfn - set up pfnlst array according to pages
446  *	VA/size pair: <shadow IO, bypass, peer-to-peer>, or OTYP_PAGES
447  */
448 /*ARGSUSED*/
449 static int
450 px_dma_pgpfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
451 {
452 	int i;
453 	dev_info_t *dip = px_p->px_dip;
454 
455 	switch (mp->dmai_object.dmao_type) {
456 	case DMA_OTYP_BUFVADDR:
457 	case DMA_OTYP_VADDR: {
458 		page_t **pplist = mp->dmai_object.dmao_obj.virt_obj.v_priv;
459 		DBG(DBG_DMA_MAP, dip, "shadow pplist=%p, %x pages, pfns=",
460 			pplist, npages);
461 		for (i = 1; i < npages; i++) {
462 			px_iopfn_t pfn = page_pptonum(pplist[i]);
463 			PX_SET_MP_PFN1(mp, i, pfn);
464 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
465 		}
466 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
467 		}
468 		break;
469 
470 	case DMA_OTYP_PAGES: {
471 		page_t *pp = mp->dmai_object.dmao_obj.pp_obj.pp_pp->p_next;
472 		DBG(DBG_DMA_MAP, dip, "pp=%p pfns=", pp);
473 		for (i = 1; i < npages; i++, pp = pp->p_next) {
474 			px_iopfn_t pfn = page_pptonum(pp);
475 			PX_SET_MP_PFN1(mp, i, pfn);
476 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
477 		}
478 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
479 		}
480 		break;
481 
482 	default:	/* check is already done by px_dma_type */
483 		ASSERT(0);
484 		break;
485 	}
486 	return (DDI_SUCCESS);
487 }
488 
489 /*
490  * px_dma_vapfn - set up pfnlst array according to VA
491  *	VA/size pair: <normal, bypass, peer-to-peer>
492  *	pfn0 is skipped as it is already done.
493  *	In this case, the cached pfn0 is used to fill pfnlst[0]
494  */
495 static int
496 px_dma_vapfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
497 {
498 	dev_info_t *dip = px_p->px_dip;
499 	int i;
500 	caddr_t vaddr = (caddr_t)mp->dmai_object.dmao_obj.virt_obj.v_as;
501 	struct hat *hat_p = vaddr ? ((struct as *)vaddr)->a_hat : kas.a_hat;
502 
503 	vaddr = mp->dmai_object.dmao_obj.virt_obj.v_addr + MMU_PAGE_SIZE;
504 	for (i = 1; i < npages; i++, vaddr += MMU_PAGE_SIZE) {
505 		px_iopfn_t pfn = hat_getpfnum(hat_p, vaddr);
506 		if (pfn == PFN_INVALID)
507 			goto err_badpfn;
508 		PX_SET_MP_PFN1(mp, i, pfn);
509 		DBG(DBG_DMA_BINDH, dip, "px_dma_vapfn: mp=%p pfnlst[%x]=%x\n",
510 			mp, i, pfn);
511 	}
512 	return (DDI_SUCCESS);
513 err_badpfn:
514 	cmn_err(CE_WARN, "%s%d: bad page frame vaddr=%p", NAMEINST(dip), vaddr);
515 	return (DDI_DMA_NOMAPPING);
516 }
517 
518 /*
519  * px_dma_pfn - Fills pfn list for all pages being DMA-ed.
520  *
521  * dependencies:
522  *	mp->dmai_ndvmapages	- set to total # of dma pages
523  *
524  * return value:
525  *	DDI_SUCCESS
526  *	DDI_DMA_NOMAPPING
527  */
528 int
529 px_dma_pfn(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
530 {
531 	uint32_t npages = mp->dmai_ndvmapages;
532 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
533 	int i, ret, peer = PX_DMA_ISPTP(mp);
534 	int peer32 = PX_DMA_ISPTP32(mp);
535 	dev_info_t *dip = px_p->px_dip;
536 
537 	px_pec_t *pec_p = px_p->px_pec_p;
538 	px_iopfn_t pfn_base = peer32 ? pec_p->pec_base32_pfn :
539 					pec_p->pec_base64_pfn;
540 	px_iopfn_t pfn_last = peer32 ? pec_p->pec_last32_pfn :
541 					pec_p->pec_last64_pfn;
542 	px_iopfn_t pfn_adj = peer ? pfn_base : 0;
543 
544 	DBG(DBG_DMA_BINDH, dip, "px_dma_pfn: mp=%p pfn0=%x\n",
545 		mp, PX_MP_PFN0(mp) - pfn_adj);
546 	/* 1 page: no array alloc/fill, no mixed mode check */
547 	if (npages == 1) {
548 		PX_SET_MP_PFN(mp, 0, PX_MP_PFN0(mp) - pfn_adj);
549 		return (DDI_SUCCESS);
550 	}
551 	/* allocate pfn array */
552 	if (!(mp->dmai_pfnlst = kmem_alloc(npages * sizeof (px_iopfn_t),
553 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP))) {
554 		if (waitfp != DDI_DMA_DONTWAIT)
555 			ddi_set_callback(waitfp, dmareq->dmar_arg,
556 				&px_kmem_clid);
557 		return (DDI_DMA_NORESOURCES);
558 	}
559 	/* fill pfn array */
560 	PX_SET_MP_PFN(mp, 0, PX_MP_PFN0(mp) - pfn_adj);	/* pfnlst[0] */
561 	if ((ret = PX_DMA_ISPGPFN(mp) ? px_dma_pgpfn(px_p, mp, npages) :
562 		px_dma_vapfn(px_p, mp, npages)) != DDI_SUCCESS)
563 		goto err;
564 
565 	/* skip pfn0, check mixed mode and adjust peer to peer pfn */
566 	for (i = 1; i < npages; i++) {
567 		px_iopfn_t pfn = PX_GET_MP_PFN1(mp, i);
568 		if (peer ^ TGT_PFN_INBETWEEN(pfn, pfn_base, pfn_last)) {
569 			cmn_err(CE_WARN, "%s%d mixed mode DMA %lx %lx",
570 				NAMEINST(mp->dmai_rdip), PX_MP_PFN0(mp), pfn);
571 			ret = DDI_DMA_NOMAPPING;	/* mixed mode */
572 			goto err;
573 		}
574 		DBG(DBG_DMA_MAP, dip,
575 			"px_dma_pfn: pfnlst[%x]=%x-%x\n", i, pfn, pfn_adj);
576 		if (pfn_adj)
577 			PX_SET_MP_PFN1(mp, i, pfn - pfn_adj);
578 	}
579 	return (DDI_SUCCESS);
580 err:
581 	px_dma_freepfn(mp);
582 	return (ret);
583 }
584 
585 /*
586  * px_dvma_win() - trim requested DVMA size down to window size
587  *	The 1st window starts from offset and ends at page-aligned boundary.
588  *	From the 2nd window on, each window starts and ends at page-aligned
589  *	boundary except the last window ends at wherever requested.
590  *
591  *	accesses the following mp-> members:
592  *	mp->dmai_attr.dma_attr_count_max
593  *	mp->dmai_attr.dma_attr_seg
594  *	mp->dmai_roffset   - start offset of 1st window
595  *	mp->dmai_rflags (redzone)
596  *	mp->dmai_ndvmapages (for 1 page fast path)
597  *
598  *	sets the following mp-> members:
599  *	mp->dmai_size	   - xfer size, != winsize if 1st/last win  (not fixed)
600  *	mp->dmai_winsize   - window size (no redzone), n * page size    (fixed)
601  *	mp->dmai_nwin	   - # of DMA windows of entire object		(fixed)
602  *	mp->dmai_rflags	   - remove partial flag if nwin == 1		(fixed)
603  *	mp->dmai_winlst	   - NULL, window objects not used for DVMA	(fixed)
604  *
605  *	fixed - not changed across different DMA windows
606  */
607 /*ARGSUSED*/
608 int
609 px_dvma_win(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
610 {
611 	uint32_t redzone_sz	= PX_HAS_REDZONE(mp) ? MMU_PAGE_SIZE : 0;
612 	size_t obj_sz		= mp->dmai_object.dmao_size;
613 	size_t xfer_sz;
614 	ulong_t pg_off;
615 
616 	if ((mp->dmai_ndvmapages == 1) && !redzone_sz) {
617 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
618 		mp->dmai_size = obj_sz;
619 		mp->dmai_winsize = MMU_PAGE_SIZE;
620 		mp->dmai_nwin = 1;
621 		goto done;
622 	}
623 
624 	pg_off	= mp->dmai_roffset;
625 	xfer_sz	= obj_sz + redzone_sz;
626 
627 	/* include redzone in nocross check */ {
628 		uint64_t nocross = mp->dmai_attr.dma_attr_seg;
629 		if (xfer_sz + pg_off - 1 > nocross)
630 			xfer_sz = nocross - pg_off + 1;
631 		if (redzone_sz && (xfer_sz <= redzone_sz)) {
632 			DBG(DBG_DMA_MAP, px_p->px_dip,
633 			    "nocross too small: "
634 			    "%lx(%lx)+%lx+%lx < %llx\n",
635 			    xfer_sz, obj_sz, pg_off, redzone_sz, nocross);
636 			return (DDI_DMA_TOOBIG);
637 		}
638 	}
639 	xfer_sz -= redzone_sz;		/* restore transfer size  */
640 	/* check counter max */ {
641 		uint32_t count_max = mp->dmai_attr.dma_attr_count_max;
642 		if (xfer_sz - 1 > count_max)
643 			xfer_sz = count_max + 1;
644 	}
645 	if (xfer_sz >= obj_sz) {
646 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
647 		mp->dmai_size = xfer_sz;
648 		mp->dmai_winsize = P2ROUNDUP(xfer_sz + pg_off, MMU_PAGE_SIZE);
649 		mp->dmai_nwin = 1;
650 		goto done;
651 	}
652 	if (!(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
653 		DBG(DBG_DMA_MAP, px_p->px_dip, "too big: %lx+%lx+%lx > %lx\n",
654 			obj_sz, pg_off, redzone_sz, xfer_sz);
655 		return (DDI_DMA_TOOBIG);
656 	}
657 
658 	xfer_sz = MMU_PTOB(MMU_BTOP(xfer_sz + pg_off)); /* page align */
659 	mp->dmai_size = xfer_sz - pg_off;	/* 1st window xferrable size */
660 	mp->dmai_winsize = xfer_sz;		/* redzone not in winsize */
661 	mp->dmai_nwin = (obj_sz + pg_off + xfer_sz - 1) / xfer_sz;
662 done:
663 	mp->dmai_winlst = NULL;
664 	px_dump_dma_handle(DBG_DMA_MAP, px_p->px_dip, mp);
665 	return (DDI_SUCCESS);
666 }
667 
668 /*
669  * fast track cache entry to mmu context, inserts 3 0 bits between
670  * upper 6-bits and lower 3-bits of the 9-bit cache entry
671  */
672 #define	MMU_FCE_TO_CTX(i)	(((i) << 3) | ((i) & 0x7) | 0x38)
673 
674 /*
675  * px_dvma_map_fast - attempts to map fast trackable DVMA
676  */
677 /*ARGSUSED*/
678 int
679 px_dvma_map_fast(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
680 {
681 	uint_t clustsz = px_dvma_page_cache_clustsz;
682 	uint_t entries = px_dvma_page_cache_entries;
683 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
684 	    mp->dmai_attr.dma_attr_flags);
685 	int i = mmu_p->mmu_dvma_addr_scan_start;
686 	uint8_t *lock_addr = mmu_p->mmu_dvma_cache_locks + i;
687 	px_dvma_addr_t dvma_pg;
688 	size_t npages = MMU_BTOP(mp->dmai_winsize);
689 	dev_info_t *dip = mmu_p->mmu_px_p->px_dip;
690 
691 	extern uint8_t ldstub(uint8_t *);
692 	ASSERT(MMU_PTOB(npages) == mp->dmai_winsize);
693 	ASSERT(npages + PX_HAS_REDZONE(mp) <= clustsz);
694 
695 	for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
696 	if (i >= entries) {
697 		lock_addr = mmu_p->mmu_dvma_cache_locks;
698 		i = 0;
699 		for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
700 		if (i >= entries) {
701 #ifdef	PX_DMA_PROF
702 			px_dvmaft_exhaust++;
703 #endif	/* PX_DMA_PROF */
704 			return (DDI_DMA_NORESOURCES);
705 		}
706 	}
707 	mmu_p->mmu_dvma_addr_scan_start = (i + 1) & (entries - 1);
708 
709 	i *= clustsz;
710 	dvma_pg = mmu_p->dvma_base_pg + i;
711 
712 	if (px_lib_iommu_map(dip, PCI_TSBID(0, i), npages, attr,
713 	    (void *)mp, 0, MMU_MAP_PFN) != DDI_SUCCESS) {
714 		DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: "
715 		    "px_lib_iommu_map failed\n");
716 
717 		return (DDI_FAILURE);
718 	}
719 
720 	if (!PX_MAP_BUFZONE(mp))
721 		goto done;
722 
723 	DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: redzone pg=%x\n", i + npages);
724 
725 	ASSERT(PX_HAS_REDZONE(mp));
726 
727 	if (px_lib_iommu_map(dip, PCI_TSBID(0, i + npages), 1, attr,
728 	    (void *)mp, npages - 1, MMU_MAP_PFN) != DDI_SUCCESS) {
729 		DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: "
730 		    "mapping REDZONE page failed\n");
731 
732 		(void) px_lib_iommu_demap(dip, PCI_TSBID(0, i), npages);
733 		return (DDI_FAILURE);
734 	}
735 
736 done:
737 #ifdef PX_DMA_PROF
738 	px_dvmaft_success++;
739 #endif
740 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
741 	mp->dmai_offset = 0;
742 	mp->dmai_flags |= PX_DMAI_FLAGS_FASTTRACK;
743 	PX_SAVE_MP_TTE(mp, attr);	/* save TTE template for unmapping */
744 	if (PX_DVMA_DBG_ON(mmu_p))
745 		px_dvma_alloc_debug(mmu_p, (char *)mp->dmai_mapping,
746 			mp->dmai_size, mp);
747 	return (DDI_SUCCESS);
748 }
749 
750 /*
751  * px_dvma_map: map non-fasttrack DMA
752  *		Use quantum cache if single page DMA.
753  */
754 int
755 px_dvma_map(ddi_dma_impl_t *mp, ddi_dma_req_t *dmareq, px_mmu_t *mmu_p)
756 {
757 	uint_t npages = PX_DMA_WINNPGS(mp);
758 	px_dvma_addr_t dvma_pg, dvma_pg_index;
759 	void *dvma_addr;
760 	uint64_t tte = PX_GET_TTE_ATTR(mp->dmai_rflags,
761 	    mp->dmai_attr.dma_attr_flags);
762 	int sleep = dmareq->dmar_fp == DDI_DMA_SLEEP ? VM_SLEEP : VM_NOSLEEP;
763 	dev_info_t *dip = mp->dmai_rdip;
764 	int	ret = DDI_SUCCESS;
765 
766 	/*
767 	 * allocate dvma space resource and map in the first window.
768 	 * (vmem_t *vmp, size_t size,
769 	 *	size_t align, size_t phase, size_t nocross,
770 	 *	void *minaddr, void *maxaddr, int vmflag)
771 	 */
772 	if ((npages == 1) && !PX_HAS_REDZONE(mp) && PX_HAS_NOSYSLIMIT(mp)) {
773 		dvma_addr = vmem_alloc(mmu_p->mmu_dvma_map,
774 			MMU_PAGE_SIZE, sleep);
775 		mp->dmai_flags |= PX_DMAI_FLAGS_VMEMCACHE;
776 #ifdef	PX_DMA_PROF
777 		px_dvma_vmem_alloc++;
778 #endif	/* PX_DMA_PROF */
779 	} else {
780 		dvma_addr = vmem_xalloc(mmu_p->mmu_dvma_map,
781 			MMU_PTOB(npages + PX_HAS_REDZONE(mp)),
782 			MAX(mp->dmai_attr.dma_attr_align, MMU_PAGE_SIZE),
783 			0,
784 			mp->dmai_attr.dma_attr_seg + 1,
785 			(void *)mp->dmai_attr.dma_attr_addr_lo,
786 			(void *)(mp->dmai_attr.dma_attr_addr_hi + 1),
787 			sleep);
788 #ifdef	PX_DMA_PROF
789 		px_dvma_vmem_xalloc++;
790 #endif	/* PX_DMA_PROF */
791 	}
792 	dvma_pg = MMU_BTOP((ulong_t)dvma_addr);
793 	dvma_pg_index = dvma_pg - mmu_p->dvma_base_pg;
794 	DBG(DBG_DMA_MAP, dip, "fallback dvma_pages: dvma_pg=%x index=%x\n",
795 		dvma_pg, dvma_pg_index);
796 	if (dvma_pg == 0)
797 		goto noresource;
798 
799 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
800 	mp->dmai_offset = 0;
801 	PX_SAVE_MP_TTE(mp, tte);	/* mp->dmai_tte = tte */
802 
803 	if ((ret = px_mmu_map_pages(mmu_p,
804 	    mp, dvma_pg, npages, 0)) != DDI_SUCCESS) {
805 		if (mp->dmai_flags & PX_DMAI_FLAGS_VMEMCACHE) {
806 			vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
807 			    MMU_PAGE_SIZE);
808 #ifdef PX_DMA_PROF
809 			px_dvma_vmem_free++;
810 #endif /* PX_DMA_PROF */
811 		} else {
812 			vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
813 			    MMU_PTOB(npages + PX_HAS_REDZONE(mp)));
814 #ifdef PX_DMA_PROF
815 			px_dvma_vmem_xfree++;
816 #endif /* PX_DMA_PROF */
817 		}
818 	}
819 
820 	return (ret);
821 noresource:
822 	if (dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
823 		DBG(DBG_DMA_MAP, dip, "dvma_pg 0 - set callback\n");
824 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
825 			&mmu_p->mmu_dvma_clid);
826 	}
827 	DBG(DBG_DMA_MAP, dip, "vmem_xalloc - DDI_DMA_NORESOURCES\n");
828 	return (DDI_DMA_NORESOURCES);
829 }
830 
831 void
832 px_dvma_unmap(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
833 {
834 	px_dvma_addr_t dvma_addr = (px_dvma_addr_t)mp->dmai_mapping;
835 	px_dvma_addr_t dvma_pg = MMU_BTOP(dvma_addr);
836 	dvma_addr = MMU_PTOB(dvma_pg);
837 
838 	if (mp->dmai_flags & PX_DMAI_FLAGS_FASTTRACK) {
839 		px_iopfn_t index = dvma_pg - mmu_p->dvma_base_pg;
840 		ASSERT(index % px_dvma_page_cache_clustsz == 0);
841 		index /= px_dvma_page_cache_clustsz;
842 		ASSERT(index < px_dvma_page_cache_entries);
843 		mmu_p->mmu_dvma_cache_locks[index] = 0;
844 #ifdef	PX_DMA_PROF
845 		px_dvmaft_free++;
846 #endif	/* PX_DMA_PROF */
847 		return;
848 	}
849 
850 	if (mp->dmai_flags & PX_DMAI_FLAGS_VMEMCACHE) {
851 		vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
852 			MMU_PAGE_SIZE);
853 #ifdef PX_DMA_PROF
854 		px_dvma_vmem_free++;
855 #endif /* PX_DMA_PROF */
856 	} else {
857 		size_t npages = MMU_BTOP(mp->dmai_winsize) + PX_HAS_REDZONE(mp);
858 		vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
859 			MMU_PTOB(npages));
860 #ifdef PX_DMA_PROF
861 		px_dvma_vmem_xfree++;
862 #endif /* PX_DMA_PROF */
863 	}
864 }
865 
866 /*
867  * DVMA mappings may have multiple windows, but each window always have
868  * one segment.
869  */
870 int
871 px_dvma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
872 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
873 	uint_t cache_flags)
874 {
875 	switch (cmd) {
876 	case DDI_DMA_SYNC:
877 		return (px_lib_dma_sync(dip, rdip, (ddi_dma_handle_t)mp,
878 		    *offp, *lenp, cache_flags));
879 
880 	case DDI_DMA_HTOC: {
881 		int ret;
882 		off_t wo_off, off = *offp;	/* wo_off: wnd's obj offset */
883 		uint_t win_size = mp->dmai_winsize;
884 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)objp;
885 
886 		if (off >= mp->dmai_object.dmao_size) {
887 			cmn_err(CE_WARN, "%s%d invalid dma_htoc offset %lx",
888 				NAMEINST(mp->dmai_rdip), off);
889 			return (DDI_FAILURE);
890 		}
891 		off += mp->dmai_roffset;
892 		ret = px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
893 		    off / win_size, &wo_off, NULL, cp, NULL); /* lenp == NULL */
894 		if (ret)
895 			return (ret);
896 		DBG(DBG_DMA_CTL, dip, "HTOC:cookie=%x+%lx off=%lx,%lx\n",
897 			cp->dmac_address, cp->dmac_size, off, *offp);
898 
899 		/* adjust cookie addr/len if we are not on window boundary */
900 		ASSERT((off % win_size) == (off -
901 			(PX_DMA_CURWIN(mp) ? mp->dmai_roffset : 0) - wo_off));
902 		off = PX_DMA_CURWIN(mp) ? off % win_size : *offp;
903 		ASSERT(cp->dmac_size > off);
904 		cp->dmac_laddress += off;
905 		cp->dmac_size -= off;
906 		DBG(DBG_DMA_CTL, dip, "HTOC:mp=%p cookie=%x+%lx off=%lx,%lx\n",
907 			mp, cp->dmac_address, cp->dmac_size, off, wo_off);
908 		}
909 		return (DDI_SUCCESS);
910 
911 	case DDI_DMA_REPWIN:
912 		*offp = mp->dmai_offset;
913 		*lenp = mp->dmai_size;
914 		return (DDI_SUCCESS);
915 
916 	case DDI_DMA_MOVWIN: {
917 		off_t off = *offp;
918 		if (off >= mp->dmai_object.dmao_size)
919 			return (DDI_FAILURE);
920 		off += mp->dmai_roffset;
921 		return (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
922 		    off / mp->dmai_winsize, offp, lenp,
923 		    (ddi_dma_cookie_t *)objp, NULL));
924 		}
925 
926 	case DDI_DMA_NEXTWIN: {
927 		px_window_t win = PX_DMA_CURWIN(mp);
928 		if (offp) {
929 			if (*(px_window_t *)offp != win) {
930 				/* window not active */
931 				*(px_window_t *)objp = win; /* return cur win */
932 				return (DDI_DMA_STALE);
933 			}
934 			win++;
935 		} else	/* map win 0 */
936 			win = 0;
937 		if (win >= mp->dmai_nwin) {
938 			*(px_window_t *)objp = win - 1;
939 			return (DDI_DMA_DONE);
940 		}
941 		if (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
942 		    win, 0, 0, 0, 0)) {
943 			*(px_window_t *)objp = win - 1;
944 			return (DDI_FAILURE);
945 		}
946 		*(px_window_t *)objp = win;
947 		}
948 		return (DDI_SUCCESS);
949 
950 	case DDI_DMA_NEXTSEG:
951 		if (*(px_window_t *)offp != PX_DMA_CURWIN(mp))
952 			return (DDI_DMA_STALE);
953 		if (lenp)				/* only 1 seg allowed */
954 			return (DDI_DMA_DONE);
955 
956 		/* return mp as seg 0 */
957 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
958 		return (DDI_SUCCESS);
959 
960 	case DDI_DMA_SEGTOC:
961 		MAKE_DMA_COOKIE((ddi_dma_cookie_t *)objp, mp->dmai_mapping,
962 			mp->dmai_size);
963 		*offp = mp->dmai_offset;
964 		*lenp = mp->dmai_size;
965 		return (DDI_SUCCESS);
966 
967 	case DDI_DMA_COFF: {
968 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)offp;
969 		if (cp->dmac_address < mp->dmai_mapping ||
970 			(cp->dmac_address + cp->dmac_size) >
971 			(mp->dmai_mapping + mp->dmai_size))
972 			return (DDI_FAILURE);
973 		*objp = (caddr_t)(cp->dmac_address - mp->dmai_mapping +
974 			mp->dmai_offset);
975 		}
976 		return (DDI_SUCCESS);
977 	default:
978 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
979 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
980 		break;
981 	}
982 	return (DDI_FAILURE);
983 }
984 
985 void
986 px_dma_freewin(ddi_dma_impl_t *mp)
987 {
988 	px_dma_win_t *win_p = mp->dmai_winlst, *win2_p;
989 	for (win2_p = win_p; win_p; win2_p = win_p) {
990 		win_p = win2_p->win_next;
991 		kmem_free(win2_p, sizeof (px_dma_win_t) +
992 			sizeof (ddi_dma_cookie_t) * win2_p->win_ncookies);
993 	}
994 	mp->dmai_nwin = 0;
995 	mp->dmai_winlst = NULL;
996 }
997 
998 /*
999  * px_dma_newwin - create a dma window object and cookies
1000  *
1001  *	After the initial scan in px_dma_physwin(), which identifies
1002  *	a portion of the pfn array that belongs to a dma window,
1003  *	we are called to allocate and initialize representing memory
1004  *	resources. We know from the 1st scan the number of cookies
1005  *	or dma segment in this window so we can allocate a contiguous
1006  *	memory array for the dma cookies (The implementation of
1007  *	ddi_dma_nextcookie(9f) dictates dma cookies be contiguous).
1008  *
1009  *	A second round scan is done on the pfn array to identify
1010  *	each dma segment and initialize its corresponding dma cookie.
1011  *	We don't need to do all the safety checking and we know they
1012  *	all belong to the same dma window.
1013  *
1014  *	Input:	cookie_no - # of cookies identified by the 1st scan
1015  *		start_idx - subscript of the pfn array for the starting pfn
1016  *		end_idx   - subscript of the last pfn in dma window
1017  *		win_pp    - pointer to win_next member of previous window
1018  *	Return:	DDI_SUCCESS - with **win_pp as newly created window object
1019  *		DDI_DMA_NORESROUCE - caller frees all previous window objs
1020  *	Note:	Each cookie and window size are all initialized on page
1021  *		boundary. This is not true for the 1st cookie of the 1st
1022  *		window and the last cookie of the last window.
1023  *		We fix that later in upper layer which has access to size
1024  *		and offset info.
1025  *
1026  */
1027 /*ARGSUSED*/
1028 static int
1029 px_dma_newwin(dev_info_t *dip, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp,
1030 	uint32_t cookie_no, uint32_t start_idx, uint32_t end_idx,
1031 	px_dma_win_t **win_pp, uint64_t count_max, uint64_t bypass)
1032 {
1033 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
1034 	ddi_dma_cookie_t *cookie_p;
1035 	uint32_t pfn_no = 1;
1036 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, start_idx);
1037 	px_iopfn_t prev_pfn = pfn;
1038 	uint64_t baddr, seg_pfn0 = pfn;
1039 	size_t sz = cookie_no * sizeof (ddi_dma_cookie_t);
1040 	px_dma_win_t *win_p = kmem_zalloc(sizeof (px_dma_win_t) + sz,
1041 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP);
1042 	io_attributes_t	attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
1043 	    mp->dmai_attr.dma_attr_flags);
1044 
1045 	if (!win_p)
1046 		goto noresource;
1047 
1048 	win_p->win_next = NULL;
1049 	win_p->win_ncookies = cookie_no;
1050 	win_p->win_curseg = 0;	/* start from segment 0 */
1051 	win_p->win_size = MMU_PTOB(end_idx - start_idx + 1);
1052 	/* win_p->win_offset is left uninitialized */
1053 
1054 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1055 	start_idx++;
1056 	for (; start_idx <= end_idx; start_idx++, prev_pfn = pfn, pfn_no++) {
1057 		pfn = PX_GET_MP_PFN1(mp, start_idx);
1058 		if ((pfn == prev_pfn + 1) &&
1059 			(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1060 			continue;
1061 
1062 		/* close up the cookie up to (including) prev_pfn */
1063 		baddr = MMU_PTOB(seg_pfn0);
1064 		if (bypass && (px_lib_iommu_getbypass(dip,
1065 				baddr, attr, &baddr) != DDI_SUCCESS))
1066 			return (DDI_FAILURE);
1067 
1068 		MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1069 		DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages)\n",
1070 			MMU_PTOB(seg_pfn0), pfn_no);
1071 
1072 		cookie_p++;	/* advance to next available cookie cell */
1073 		pfn_no = 0;
1074 		seg_pfn0 = pfn;	/* start a new segment from current pfn */
1075 	}
1076 
1077 	baddr = MMU_PTOB(seg_pfn0);
1078 	if (bypass && (px_lib_iommu_getbypass(dip,
1079 			baddr, attr, &baddr) != DDI_SUCCESS))
1080 		return (DDI_FAILURE);
1081 
1082 	MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1083 	DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages) of total %x\n",
1084 		MMU_PTOB(seg_pfn0), pfn_no, cookie_no);
1085 #ifdef	DEBUG
1086 	cookie_p++;
1087 	ASSERT((cookie_p - (ddi_dma_cookie_t *)(win_p + 1)) == cookie_no);
1088 #endif	/* DEBUG */
1089 	*win_pp = win_p;
1090 	return (DDI_SUCCESS);
1091 noresource:
1092 	if (waitfp != DDI_DMA_DONTWAIT)
1093 		ddi_set_callback(waitfp, dmareq->dmar_arg, &px_kmem_clid);
1094 	return (DDI_DMA_NORESOURCES);
1095 }
1096 
1097 /*
1098  * px_dma_adjust - adjust 1st and last cookie and window sizes
1099  *	remove initial dma page offset from 1st cookie and window size
1100  *	remove last dma page remainder from last cookie and window size
1101  *	fill win_offset of each dma window according to just fixed up
1102  *		each window sizes
1103  *	px_dma_win_t members modified:
1104  *	win_p->win_offset - this window's offset within entire DMA object
1105  *	win_p->win_size	  - xferrable size (in bytes) for this window
1106  *
1107  *	ddi_dma_impl_t members modified:
1108  *	mp->dmai_size	  - 1st window xferrable size
1109  *	mp->dmai_offset   - 0, which is the dma offset of the 1st window
1110  *
1111  *	ddi_dma_cookie_t members modified:
1112  *	cookie_p->dmac_size - 1st and last cookie remove offset or remainder
1113  *	cookie_p->dmac_laddress - 1st cookie add page offset
1114  */
1115 static void
1116 px_dma_adjust(ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp, px_dma_win_t *win_p)
1117 {
1118 	ddi_dma_cookie_t *cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1119 	size_t pg_offset = mp->dmai_roffset;
1120 	size_t win_offset = 0;
1121 
1122 	cookie_p->dmac_size -= pg_offset;
1123 	cookie_p->dmac_laddress |= pg_offset;
1124 	win_p->win_size -= pg_offset;
1125 	DBG(DBG_BYPASS, mp->dmai_rdip, "pg0 adjust %lx\n", pg_offset);
1126 
1127 	mp->dmai_size = win_p->win_size;
1128 	mp->dmai_offset = 0;
1129 
1130 	pg_offset += mp->dmai_object.dmao_size;
1131 	pg_offset &= MMU_PAGE_OFFSET;
1132 	if (pg_offset)
1133 		pg_offset = MMU_PAGE_SIZE - pg_offset;
1134 	DBG(DBG_BYPASS, mp->dmai_rdip, "last pg adjust %lx\n", pg_offset);
1135 
1136 	for (; win_p->win_next; win_p = win_p->win_next) {
1137 		DBG(DBG_BYPASS, mp->dmai_rdip, "win off %p\n", win_offset);
1138 		win_p->win_offset = win_offset;
1139 		win_offset += win_p->win_size;
1140 	}
1141 	/* last window */
1142 	win_p->win_offset = win_offset;
1143 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1144 	cookie_p[win_p->win_ncookies - 1].dmac_size -= pg_offset;
1145 	win_p->win_size -= pg_offset;
1146 	ASSERT((win_offset + win_p->win_size) == mp->dmai_object.dmao_size);
1147 }
1148 
1149 /*
1150  * px_dma_physwin() - carve up dma windows using physical addresses.
1151  *	Called to handle mmu bypass and pci peer-to-peer transfers.
1152  *	Calls px_dma_newwin() to allocate window objects.
1153  *
1154  * Dependency: mp->dmai_pfnlst points to an array of pfns
1155  *
1156  * 1. Each dma window is represented by a px_dma_win_t object.
1157  *	The object will be casted to ddi_dma_win_t and returned
1158  *	to leaf driver through the DDI interface.
1159  * 2. Each dma window can have several dma segments with each
1160  *	segment representing a physically contiguous either memory
1161  *	space (if we are doing an mmu bypass transfer) or pci address
1162  *	space (if we are doing a peer-to-peer transfer).
1163  * 3. Each segment has a DMA cookie to program the DMA engine.
1164  *	The cookies within each DMA window must be located in a
1165  *	contiguous array per ddi_dma_nextcookie(9f).
1166  * 4. The number of DMA segments within each DMA window cannot exceed
1167  *	mp->dmai_attr.dma_attr_sgllen. If the transfer size is
1168  *	too large to fit in the sgllen, the rest needs to be
1169  *	relocated to the next dma window.
1170  * 5. Peer-to-peer DMA segment follows device hi, lo, count_max,
1171  *	and nocross restrictions while bypass DMA follows the set of
1172  *	restrictions with system limits factored in.
1173  *
1174  * Return:
1175  *	mp->dmai_winlst	 - points to a link list of px_dma_win_t objects.
1176  *		Each px_dma_win_t object on the link list contains
1177  *		infomation such as its window size (# of pages),
1178  *		starting offset (also see Restriction), an array of
1179  *		DMA cookies, and # of cookies in the array.
1180  *	mp->dmai_pfnlst	 - NULL, the pfn list is freed to conserve memory.
1181  *	mp->dmai_nwin	 - # of total DMA windows on mp->dmai_winlst.
1182  *	mp->dmai_mapping - starting cookie address
1183  *	mp->dmai_rflags	 - consistent, nosync, no redzone
1184  *	mp->dmai_cookie	 - start of cookie table of the 1st DMA window
1185  *
1186  * Restriction:
1187  *	Each px_dma_win_t object can theoratically start from any offset
1188  *	since the mmu is not involved. However, this implementation
1189  *	always make windows start from page aligned offset (except
1190  *	the 1st window, which follows the requested offset) due to the
1191  *	fact that we are handed a pfn list. This does require device's
1192  *	count_max and attr_seg to be at least MMU_PAGE_SIZE aligned.
1193  */
1194 int
1195 px_dma_physwin(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
1196 {
1197 	uint_t npages = mp->dmai_ndvmapages;
1198 	int ret, sgllen = mp->dmai_attr.dma_attr_sgllen;
1199 	px_iopfn_t pfn_lo, pfn_hi, prev_pfn;
1200 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, 0);
1201 	uint32_t i, win_no = 0, pfn_no = 1, win_pfn0_index = 0, cookie_no = 0;
1202 	uint64_t count_max, bypass_addr = 0;
1203 	px_dma_win_t **win_pp = (px_dma_win_t **)&mp->dmai_winlst;
1204 	ddi_dma_cookie_t *cookie0_p;
1205 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
1206 	    mp->dmai_attr.dma_attr_flags);
1207 	dev_info_t *dip = px_p->px_dip;
1208 
1209 	ASSERT(PX_DMA_ISPTP(mp) || PX_DMA_ISBYPASS(mp));
1210 	if (PX_DMA_ISPTP(mp)) { /* ignore sys limits for peer-to-peer */
1211 		ddi_dma_attr_t *dev_attr_p = PX_DEV_ATTR(mp);
1212 		uint64_t nocross = dev_attr_p->dma_attr_seg;
1213 		px_pec_t *pec_p = px_p->px_pec_p;
1214 		px_iopfn_t pfn_last = PX_DMA_ISPTP32(mp) ?
1215 				pec_p->pec_last32_pfn - pec_p->pec_base32_pfn :
1216 				pec_p->pec_last64_pfn - pec_p->pec_base64_pfn;
1217 
1218 		if (nocross && (nocross < UINT32_MAX))
1219 			return (DDI_DMA_NOMAPPING);
1220 		if (dev_attr_p->dma_attr_align > MMU_PAGE_SIZE)
1221 			return (DDI_DMA_NOMAPPING);
1222 		pfn_lo = MMU_BTOP(dev_attr_p->dma_attr_addr_lo);
1223 		pfn_hi = MMU_BTOP(dev_attr_p->dma_attr_addr_hi);
1224 		pfn_hi = MIN(pfn_hi, pfn_last);
1225 		if ((pfn_lo > pfn_hi) || (pfn < pfn_lo))
1226 			return (DDI_DMA_NOMAPPING);
1227 
1228 		count_max = dev_attr_p->dma_attr_count_max;
1229 		count_max = MIN(count_max, nocross);
1230 		/*
1231 		 * the following count_max trim is not done because we are
1232 		 * making sure pfn_lo <= pfn <= pfn_hi inside the loop
1233 		 * count_max=MIN(count_max, MMU_PTOB(pfn_hi - pfn_lo + 1)-1);
1234 		 */
1235 	} else { /* bypass hi/lo/count_max have been processed by attr2hdl() */
1236 		count_max = mp->dmai_attr.dma_attr_count_max;
1237 		pfn_lo = MMU_BTOP(mp->dmai_attr.dma_attr_addr_lo);
1238 		pfn_hi = MMU_BTOP(mp->dmai_attr.dma_attr_addr_hi);
1239 
1240 		if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn),
1241 				attr, &bypass_addr) != DDI_SUCCESS) {
1242 			cmn_err(CE_WARN, "bypass cookie failure %lx\n", pfn);
1243 			return (DDI_DMA_NOMAPPING);
1244 		}
1245 		pfn = MMU_BTOP(bypass_addr);
1246 	}
1247 
1248 	/* pfn: absolute (bypass mode) or relative (p2p mode) */
1249 	for (prev_pfn = pfn, i = 1; i < npages;
1250 	    i++, prev_pfn = pfn, pfn_no++) {
1251 		pfn = PX_GET_MP_PFN1(mp, i);
1252 		if (bypass_addr) {
1253 			if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn), attr,
1254 					&bypass_addr) != DDI_SUCCESS) {
1255 				ret = DDI_DMA_NOMAPPING;
1256 				goto err;
1257 			}
1258 			pfn = MMU_BTOP(bypass_addr);
1259 		}
1260 		if ((pfn == prev_pfn + 1) &&
1261 				(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1262 			continue;
1263 		if ((pfn < pfn_lo) || (prev_pfn > pfn_hi)) {
1264 			ret = DDI_DMA_NOMAPPING;
1265 			goto err;
1266 		}
1267 		cookie_no++;
1268 		pfn_no = 0;
1269 		if (cookie_no < sgllen)
1270 			continue;
1271 
1272 		DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1273 			win_pfn0_index, i - 1, cookie_no);
1274 		if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no,
1275 			win_pfn0_index, i - 1, win_pp, count_max, bypass_addr))
1276 			goto err;
1277 
1278 		win_pp = &(*win_pp)->win_next;	/* win_pp = *(win_pp) */
1279 		win_no++;
1280 		win_pfn0_index = i;
1281 		cookie_no = 0;
1282 	}
1283 	if (pfn > pfn_hi) {
1284 		ret = DDI_DMA_NOMAPPING;
1285 		goto err;
1286 	}
1287 	cookie_no++;
1288 	DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1289 		win_pfn0_index, i - 1, cookie_no);
1290 	if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no, win_pfn0_index,
1291 		i - 1, win_pp, count_max, bypass_addr))
1292 		goto err;
1293 	win_no++;
1294 	px_dma_adjust(dmareq, mp, mp->dmai_winlst);
1295 	mp->dmai_nwin = win_no;
1296 	mp->dmai_rflags |= DDI_DMA_CONSISTENT | DMP_NOSYNC;
1297 	mp->dmai_rflags &= ~DDI_DMA_REDZONE;
1298 	mp->dmai_flags |= PX_DMAI_FLAGS_NOSYNC;
1299 	cookie0_p = (ddi_dma_cookie_t *)(PX_WINLST(mp) + 1);
1300 	mp->dmai_cookie = PX_WINLST(mp)->win_ncookies > 1 ? cookie0_p + 1 : 0;
1301 	mp->dmai_mapping = cookie0_p->dmac_laddress;
1302 
1303 	px_dma_freepfn(mp);
1304 	return (DDI_DMA_MAPPED);
1305 err:
1306 	px_dma_freewin(mp);
1307 	return (ret);
1308 }
1309 
1310 int
1311 px_dma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
1312 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
1313 	uint_t cache_flags)
1314 {
1315 	switch (cmd) {
1316 	case DDI_DMA_SYNC:
1317 		return (DDI_SUCCESS);
1318 
1319 	case DDI_DMA_HTOC: {
1320 		off_t off = *offp;
1321 		ddi_dma_cookie_t *loop_cp, *cp;
1322 		px_dma_win_t *win_p = mp->dmai_winlst;
1323 
1324 		if (off >= mp->dmai_object.dmao_size)
1325 			return (DDI_FAILURE);
1326 
1327 		/* locate window */
1328 		while (win_p->win_offset + win_p->win_size <= off)
1329 			win_p = win_p->win_next;
1330 
1331 		loop_cp = cp = (ddi_dma_cookie_t *)(win_p + 1);
1332 		mp->dmai_offset = win_p->win_offset;
1333 		mp->dmai_size   = win_p->win_size;
1334 		mp->dmai_mapping = cp->dmac_laddress; /* cookie0 start addr */
1335 
1336 		/* adjust cookie addr/len if we are not on cookie boundary */
1337 		off -= win_p->win_offset;	   /* offset within window */
1338 		for (; off >= loop_cp->dmac_size; loop_cp++)
1339 			off -= loop_cp->dmac_size; /* offset within cookie */
1340 
1341 		mp->dmai_cookie = loop_cp + 1;
1342 		win_p->win_curseg = loop_cp - cp;
1343 		cp = (ddi_dma_cookie_t *)objp;
1344 		MAKE_DMA_COOKIE(cp, loop_cp->dmac_laddress + off,
1345 			loop_cp->dmac_size - off);
1346 
1347 		DBG(DBG_DMA_CTL, dip,
1348 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1349 			cp->dmac_laddress, cp->dmac_size);
1350 		}
1351 		return (DDI_SUCCESS);
1352 
1353 	case DDI_DMA_REPWIN:
1354 		*offp = mp->dmai_offset;
1355 		*lenp = mp->dmai_size;
1356 		return (DDI_SUCCESS);
1357 
1358 	case DDI_DMA_MOVWIN: {
1359 		off_t off = *offp;
1360 		ddi_dma_cookie_t *cp;
1361 		px_dma_win_t *win_p = mp->dmai_winlst;
1362 
1363 		if (off >= mp->dmai_object.dmao_size)
1364 			return (DDI_FAILURE);
1365 
1366 		/* locate window */
1367 		while (win_p->win_offset + win_p->win_size <= off)
1368 			win_p = win_p->win_next;
1369 
1370 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1371 		mp->dmai_offset = win_p->win_offset;
1372 		mp->dmai_size   = win_p->win_size;
1373 		mp->dmai_mapping = cp->dmac_laddress;	/* cookie0 star addr */
1374 		mp->dmai_cookie = cp + 1;
1375 		win_p->win_curseg = 0;
1376 
1377 		*(ddi_dma_cookie_t *)objp = *cp;
1378 		*offp = win_p->win_offset;
1379 		*lenp = win_p->win_size;
1380 		DBG(DBG_DMA_CTL, dip,
1381 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1382 			cp->dmac_laddress, cp->dmac_size);
1383 		}
1384 		return (DDI_SUCCESS);
1385 
1386 	case DDI_DMA_NEXTWIN: {
1387 		px_dma_win_t *win_p = *(px_dma_win_t **)offp;
1388 		px_dma_win_t **nw_pp = (px_dma_win_t **)objp;
1389 		ddi_dma_cookie_t *cp;
1390 		if (!win_p) {
1391 			*nw_pp = mp->dmai_winlst;
1392 			return (DDI_SUCCESS);
1393 		}
1394 
1395 		if (win_p->win_offset != mp->dmai_offset)
1396 			return (DDI_DMA_STALE);
1397 		if (!win_p->win_next)
1398 			return (DDI_DMA_DONE);
1399 		win_p = win_p->win_next;
1400 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1401 		mp->dmai_offset = win_p->win_offset;
1402 		mp->dmai_size   = win_p->win_size;
1403 		mp->dmai_mapping = cp->dmac_laddress;   /* cookie0 star addr */
1404 		mp->dmai_cookie = cp + 1;
1405 		win_p->win_curseg = 0;
1406 		*nw_pp = win_p;
1407 		}
1408 		return (DDI_SUCCESS);
1409 
1410 	case DDI_DMA_NEXTSEG: {
1411 		px_dma_win_t *w_p = *(px_dma_win_t **)offp;
1412 		if (w_p->win_offset != mp->dmai_offset)
1413 			return (DDI_DMA_STALE);
1414 		if (w_p->win_curseg + 1 >= w_p->win_ncookies)
1415 			return (DDI_DMA_DONE);
1416 		w_p->win_curseg++;
1417 		}
1418 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
1419 		return (DDI_SUCCESS);
1420 
1421 	case DDI_DMA_SEGTOC: {
1422 		px_dma_win_t *win_p = mp->dmai_winlst;
1423 		off_t off = mp->dmai_offset;
1424 		ddi_dma_cookie_t *cp;
1425 		int i;
1426 
1427 		/* locate active window */
1428 		for (; win_p->win_offset != off; win_p = win_p->win_next);
1429 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1430 		for (i = 0; i < win_p->win_curseg; i++, cp++)
1431 			off += cp->dmac_size;
1432 		*offp = off;
1433 		*lenp = cp->dmac_size;
1434 		*(ddi_dma_cookie_t *)objp = *cp;	/* copy cookie */
1435 		}
1436 		return (DDI_SUCCESS);
1437 
1438 	case DDI_DMA_COFF: {
1439 		px_dma_win_t *win_p;
1440 		ddi_dma_cookie_t *cp;
1441 		uint64_t addr, key = ((ddi_dma_cookie_t *)offp)->dmac_laddress;
1442 		size_t win_off;
1443 
1444 		for (win_p = mp->dmai_winlst; win_p; win_p = win_p->win_next) {
1445 			int i;
1446 			win_off = 0;
1447 			cp = (ddi_dma_cookie_t *)(win_p + 1);
1448 			for (i = 0; i < win_p->win_ncookies; i++, cp++) {
1449 				size_t sz = cp->dmac_size;
1450 
1451 				addr = cp->dmac_laddress;
1452 				if ((addr <= key) && (addr + sz >= key))
1453 					goto found;
1454 				win_off += sz;
1455 			}
1456 		}
1457 		return (DDI_FAILURE);
1458 found:
1459 		*objp = (caddr_t)(win_p->win_offset + win_off + (key - addr));
1460 		return (DDI_SUCCESS);
1461 		}
1462 	default:
1463 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
1464 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
1465 		break;
1466 	}
1467 	return (DDI_FAILURE);
1468 }
1469 
1470 static void
1471 px_dvma_debug_init(px_mmu_t *mmu_p)
1472 {
1473 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1474 	ASSERT(MUTEX_HELD(&mmu_p->dvma_debug_lock));
1475 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat ON", mmu_p);
1476 
1477 	mmu_p->dvma_alloc_rec = kmem_alloc(sz, KM_SLEEP);
1478 	mmu_p->dvma_free_rec = kmem_alloc(sz, KM_SLEEP);
1479 
1480 	mmu_p->dvma_active_list = NULL;
1481 	mmu_p->dvma_alloc_rec_index = 0;
1482 	mmu_p->dvma_free_rec_index = 0;
1483 	mmu_p->dvma_active_count = 0;
1484 }
1485 
1486 void
1487 px_dvma_debug_fini(px_mmu_t *mmu_p)
1488 {
1489 	struct px_dvma_rec *prev, *ptr;
1490 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1491 	uint64_t mask = ~(1ull << mmu_p->mmu_inst);
1492 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat OFF", mmu_p);
1493 
1494 	kmem_free(mmu_p->dvma_alloc_rec, sz);
1495 	kmem_free(mmu_p->dvma_free_rec, sz);
1496 	mmu_p->dvma_alloc_rec = mmu_p->dvma_free_rec = NULL;
1497 
1498 	prev = mmu_p->dvma_active_list;
1499 	if (!prev)
1500 		return;
1501 	for (ptr = prev->next; ptr; prev = ptr, ptr = ptr->next)
1502 		kmem_free(prev, sizeof (struct px_dvma_rec));
1503 	kmem_free(prev, sizeof (struct px_dvma_rec));
1504 
1505 	mmu_p->dvma_active_list = NULL;
1506 	mmu_p->dvma_alloc_rec_index = 0;
1507 	mmu_p->dvma_free_rec_index = 0;
1508 	mmu_p->dvma_active_count = 0;
1509 
1510 	px_dvma_debug_off &= mask;
1511 	px_dvma_debug_on &= mask;
1512 }
1513 
1514 void
1515 px_dvma_alloc_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1516 	ddi_dma_impl_t *mp)
1517 {
1518 	struct px_dvma_rec *ptr;
1519 	mutex_enter(&mmu_p->dvma_debug_lock);
1520 
1521 	if (!mmu_p->dvma_alloc_rec)
1522 		px_dvma_debug_init(mmu_p);
1523 	if (PX_DVMA_DBG_OFF(mmu_p)) {
1524 		px_dvma_debug_fini(mmu_p);
1525 		goto done;
1526 	}
1527 
1528 	ptr = &mmu_p->dvma_alloc_rec[mmu_p->dvma_alloc_rec_index];
1529 	ptr->dvma_addr = address;
1530 	ptr->len = len;
1531 	ptr->mp = mp;
1532 	if (++mmu_p->dvma_alloc_rec_index == px_dvma_debug_rec)
1533 		mmu_p->dvma_alloc_rec_index = 0;
1534 
1535 	ptr = kmem_alloc(sizeof (struct px_dvma_rec), KM_SLEEP);
1536 	ptr->dvma_addr = address;
1537 	ptr->len = len;
1538 	ptr->mp = mp;
1539 
1540 	ptr->next = mmu_p->dvma_active_list;
1541 	mmu_p->dvma_active_list = ptr;
1542 	mmu_p->dvma_active_count++;
1543 done:
1544 	mutex_exit(&mmu_p->dvma_debug_lock);
1545 }
1546 
1547 void
1548 px_dvma_free_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1549     ddi_dma_impl_t *mp)
1550 {
1551 	struct px_dvma_rec *ptr, *ptr_save;
1552 	mutex_enter(&mmu_p->dvma_debug_lock);
1553 
1554 	if (!mmu_p->dvma_alloc_rec)
1555 		px_dvma_debug_init(mmu_p);
1556 	if (PX_DVMA_DBG_OFF(mmu_p)) {
1557 		px_dvma_debug_fini(mmu_p);
1558 		goto done;
1559 	}
1560 
1561 	ptr = &mmu_p->dvma_free_rec[mmu_p->dvma_free_rec_index];
1562 	ptr->dvma_addr = address;
1563 	ptr->len = len;
1564 	ptr->mp = mp;
1565 	if (++mmu_p->dvma_free_rec_index == px_dvma_debug_rec)
1566 		mmu_p->dvma_free_rec_index = 0;
1567 
1568 	ptr_save = mmu_p->dvma_active_list;
1569 	for (ptr = ptr_save; ptr; ptr = ptr->next) {
1570 		if ((ptr->dvma_addr == address) && (ptr->len = len))
1571 			break;
1572 		ptr_save = ptr;
1573 	}
1574 	if (!ptr) {
1575 		cmn_err(CE_WARN, "bad dvma free addr=%lx len=%x",
1576 			(long)address, len);
1577 		goto done;
1578 	}
1579 	if (ptr == mmu_p->dvma_active_list)
1580 		mmu_p->dvma_active_list = ptr->next;
1581 	else
1582 		ptr_save->next = ptr->next;
1583 	kmem_free(ptr, sizeof (struct px_dvma_rec));
1584 	mmu_p->dvma_active_count--;
1585 done:
1586 	mutex_exit(&mmu_p->dvma_debug_lock);
1587 }
1588 
1589 #ifdef	DEBUG
1590 void
1591 px_dump_dma_handle(uint64_t flag, dev_info_t *dip, ddi_dma_impl_t *hp)
1592 {
1593 	DBG(flag, dip, "mp(%p): flags=%x mapping=%lx xfer_size=%x\n",
1594 		hp, hp->dmai_inuse, hp->dmai_mapping, hp->dmai_size);
1595 	DBG(flag|DBG_CONT, dip, "\tnpages=%x roffset=%x rflags=%x nwin=%x\n",
1596 		hp->dmai_ndvmapages, hp->dmai_roffset, hp->dmai_rflags,
1597 		hp->dmai_nwin);
1598 	DBG(flag|DBG_CONT, dip, "\twinsize=%x tte=%p pfnlst=%p pfn0=%p\n",
1599 		hp->dmai_winsize, hp->dmai_tte, hp->dmai_pfnlst, hp->dmai_pfn0);
1600 	DBG(flag|DBG_CONT, dip, "\twinlst=%x obj=%p attr=%p ckp=%p\n",
1601 		hp->dmai_winlst, &hp->dmai_object, &hp->dmai_attr,
1602 		hp->dmai_cookie);
1603 }
1604 #endif	/* DEBUG */
1605