xref: /illumos-gate/usr/src/uts/sun4/io/px/px_dma.c (revision 10a4fa49f51ed9ae1c857a626de6ce9ebf41661a)
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 
114 	if (px_child_prefetch(mp->dmai_rdip))
115 		mp->dmai_flags |= (PX_DMAI_FLAGS_MAP_BUFZONE |
116 		    PX_DMAI_FLAGS_REDZONE);
117 
118 	return (mp);
119 }
120 
121 void
122 px_dma_freemp(ddi_dma_impl_t *mp)
123 {
124 	if (mp->dmai_ndvmapages > 1)
125 		px_dma_freepfn(mp);
126 	if (mp->dmai_winlst)
127 		px_dma_freewin(mp);
128 	kmem_free(mp, sizeof (px_dma_hdl_t));
129 }
130 
131 void
132 px_dma_freepfn(ddi_dma_impl_t *mp)
133 {
134 	void *addr = mp->dmai_pfnlst;
135 	if (addr) {
136 		size_t npages = mp->dmai_ndvmapages;
137 		if (npages > 1)
138 			kmem_free(addr, npages * sizeof (px_iopfn_t));
139 		mp->dmai_pfnlst = NULL;
140 	}
141 	mp->dmai_ndvmapages = 0;
142 }
143 
144 /*
145  * px_dma_lmts2hdl - alloate a ddi_dma_impl_t, validate practical limits
146  *			and convert dmareq->dmar_limits to mp->dmai_attr
147  *
148  * ddi_dma_impl_t member modified     input
149  * ------------------------------------------------------------------------
150  * mp->dmai_minxfer		    - dev
151  * mp->dmai_burstsizes		    - dev
152  * mp->dmai_flags		    - no limit? peer-to-peer only?
153  *
154  * ddi_dma_attr member modified       input
155  * ------------------------------------------------------------------------
156  * mp->dmai_attr.dma_attr_addr_lo   - dev lo, sys lo
157  * mp->dmai_attr.dma_attr_addr_hi   - dev hi, sys hi
158  * mp->dmai_attr.dma_attr_count_max - dev count max, dev/sys lo/hi delta
159  * mp->dmai_attr.dma_attr_seg       - 0         (no nocross   restriction)
160  * mp->dmai_attr.dma_attr_align     - 1         (no alignment restriction)
161  *
162  * The dlim_dmaspeed member of dmareq->dmar_limits is ignored.
163  */
164 ddi_dma_impl_t *
165 px_dma_lmts2hdl(dev_info_t *dip, dev_info_t *rdip, px_mmu_t *mmu_p,
166 	ddi_dma_req_t *dmareq)
167 {
168 	ddi_dma_impl_t *mp;
169 	ddi_dma_attr_t *attr_p;
170 	uint64_t syslo		= mmu_p->mmu_dvma_base;
171 	uint64_t syshi		= mmu_p->mmu_dvma_end;
172 	uint64_t fasthi		= mmu_p->mmu_dvma_fast_end;
173 	ddi_dma_lim_t *lim_p	= dmareq->dmar_limits;
174 	uint32_t count_max	= lim_p->dlim_cntr_max;
175 	uint64_t lo		= lim_p->dlim_addr_lo;
176 	uint64_t hi		= lim_p->dlim_addr_hi;
177 	if (hi <= lo) {
178 		DBG(DBG_DMA_MAP, dip, "Bad limits\n");
179 		return ((ddi_dma_impl_t *)DDI_DMA_NOMAPPING);
180 	}
181 	if (!count_max)
182 		count_max--;
183 
184 	if (!(mp = px_dma_allocmp(dip, rdip, dmareq->dmar_fp,
185 		dmareq->dmar_arg)))
186 		return (NULL);
187 
188 	/* store original dev input at the 2nd ddi_dma_attr */
189 	attr_p = PX_DEV_ATTR(mp);
190 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
191 	SET_DMAALIGN(attr_p, 1);
192 
193 	lo = MAX(lo, syslo);
194 	hi = MIN(hi, syshi);
195 	if (hi <= lo)
196 		mp->dmai_flags |= PX_DMAI_FLAGS_PEER_ONLY;
197 	count_max = MIN(count_max, hi - lo);
198 
199 	if (PX_DEV_NOSYSLIMIT(lo, hi, syslo, fasthi, 1))
200 		mp->dmai_flags |= PX_DMAI_FLAGS_NOFASTLIMIT |
201 			PX_DMAI_FLAGS_NOSYSLIMIT;
202 	else {
203 		if (PX_DEV_NOFASTLIMIT(lo, hi, syslo, syshi, 1))
204 			mp->dmai_flags |= PX_DMAI_FLAGS_NOFASTLIMIT;
205 	}
206 	if (PX_DMA_NOCTX(rdip))
207 		mp->dmai_flags |= PX_DMAI_FLAGS_NOCTX;
208 
209 	/* store augumented dev input to mp->dmai_attr */
210 	mp->dmai_minxfer	= lim_p->dlim_minxfer;
211 	mp->dmai_burstsizes	= lim_p->dlim_burstsizes;
212 	attr_p = &mp->dmai_attr;
213 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
214 	SET_DMAALIGN(attr_p, 1);
215 	return (mp);
216 }
217 
218 /*
219  * Called from px_attach to check for bypass dma support and set
220  * flags accordingly.
221  */
222 int
223 px_dma_attach(px_t *px_p)
224 {
225 	uint64_t baddr;
226 
227 	if (px_lib_iommu_getbypass(px_p->px_dip, 0ull,
228 			PCI_MAP_ATTR_WRITE|PCI_MAP_ATTR_READ,
229 			&baddr) != DDI_ENOTSUP)
230 		/* ignore all other errors */
231 		px_p->px_dev_caps |= PX_BYPASS_DMA_ALLOWED;
232 
233 	return (DDI_SUCCESS);
234 }
235 
236 /*
237  * px_dma_attr2hdl
238  *
239  * This routine is called from the alloc handle entry point to sanity check the
240  * dma attribute structure.
241  *
242  * use by: px_dma_allochdl()
243  *
244  * return value:
245  *
246  *	DDI_SUCCESS		- on success
247  *	DDI_DMA_BADATTR		- attribute has invalid version number
248  *				  or address limits exclude dvma space
249  */
250 int
251 px_dma_attr2hdl(px_t *px_p, ddi_dma_impl_t *mp)
252 {
253 	px_mmu_t *mmu_p = px_p->px_mmu_p;
254 	uint64_t syslo, syshi;
255 	int	ret;
256 	ddi_dma_attr_t *attrp		= PX_DEV_ATTR(mp);
257 	uint64_t hi			= attrp->dma_attr_addr_hi;
258 	uint64_t lo			= attrp->dma_attr_addr_lo;
259 	uint64_t align			= attrp->dma_attr_align;
260 	uint64_t nocross		= attrp->dma_attr_seg;
261 	uint64_t count_max		= attrp->dma_attr_count_max;
262 
263 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "attrp=%p cntr_max=%x.%08x\n",
264 		attrp, HI32(count_max), LO32(count_max));
265 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "hi=%x.%08x lo=%x.%08x\n",
266 		HI32(hi), LO32(hi), HI32(lo), LO32(lo));
267 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "seg=%x.%08x align=%x.%08x\n",
268 		HI32(nocross), LO32(nocross), HI32(align), LO32(align));
269 
270 	if (!nocross)
271 		nocross--;
272 	if (attrp->dma_attr_flags & DDI_DMA_FORCE_PHYSICAL) { /* BYPASS */
273 
274 		DBG(DBG_DMA_ALLOCH, px_p->px_dip, "bypass mode\n");
275 		/*
276 		 * If Bypass DMA is not supported, return error so that
277 		 * target driver can fall back to dvma mode of operation
278 		 */
279 		if (!(px_p->px_dev_caps & PX_BYPASS_DMA_ALLOWED))
280 			return (DDI_DMA_BADATTR);
281 		mp->dmai_flags |= PX_DMAI_FLAGS_BYPASSREQ;
282 		if (nocross != UINT64_MAX)
283 			return (DDI_DMA_BADATTR);
284 		if (align && (align > MMU_PAGE_SIZE))
285 			return (DDI_DMA_BADATTR);
286 		align = 1; /* align on 1 page boundary */
287 
288 		/* do a range check and get the limits */
289 		ret = px_lib_dma_bypass_rngchk(px_p->px_dip, attrp,
290 				&syslo, &syshi);
291 		if (ret != DDI_SUCCESS)
292 			return (ret);
293 	} else { /* MMU_XLATE or PEER_TO_PEER */
294 		align = MAX(align, MMU_PAGE_SIZE) - 1;
295 		if ((align & nocross) != align) {
296 			dev_info_t *rdip = mp->dmai_rdip;
297 			cmn_err(CE_WARN, "%s%d dma_attr_seg not aligned",
298 				NAMEINST(rdip));
299 			return (DDI_DMA_BADATTR);
300 		}
301 		align = MMU_BTOP(align + 1);
302 		syslo = mmu_p->mmu_dvma_base;
303 		syshi = mmu_p->mmu_dvma_end;
304 	}
305 	if (hi <= lo) {
306 		dev_info_t *rdip = mp->dmai_rdip;
307 		cmn_err(CE_WARN, "%s%d limits out of range", NAMEINST(rdip));
308 		return (DDI_DMA_BADATTR);
309 	}
310 	lo = MAX(lo, syslo);
311 	hi = MIN(hi, syshi);
312 	if (!count_max)
313 		count_max--;
314 
315 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "hi=%x.%08x, lo=%x.%08x\n",
316 		HI32(hi), LO32(hi), HI32(lo), LO32(lo));
317 	if (hi <= lo) { /* peer transfers cannot have alignment & nocross */
318 		dev_info_t *rdip = mp->dmai_rdip;
319 		cmn_err(CE_WARN, "%s%d peer only dev %p", NAMEINST(rdip), mp);
320 		if ((nocross < UINT32_MAX) || (align > 1)) {
321 			cmn_err(CE_WARN, "%s%d peer only device bad attr",
322 				NAMEINST(rdip));
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 
376 	mp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS | DMP_NOSYNC;
377 
378 	switch (dobj_p->dmao_type) {
379 	case DMA_OTYP_BUFVADDR:
380 	case DMA_OTYP_VADDR: {
381 		page_t **pplist = dobj_p->dmao_obj.virt_obj.v_priv;
382 		caddr_t vaddr = dobj_p->dmao_obj.virt_obj.v_addr;
383 
384 		DBG(DBG_DMA_MAP, dip, "vaddr=%p pplist=%p\n", vaddr, pplist);
385 		offset = (ulong_t)vaddr & MMU_PAGE_OFFSET;
386 		if (pplist) {				/* shadow list */
387 			mp->dmai_flags |= PX_DMAI_FLAGS_PGPFN;
388 			pfn0 = page_pptonum(*pplist);
389 		} else {
390 			struct as *as_p = dobj_p->dmao_obj.virt_obj.v_as;
391 			struct hat *hat_p = as_p ? as_p->a_hat : kas.a_hat;
392 			pfn0 = hat_getpfnum(hat_p, vaddr);
393 		}
394 		}
395 		break;
396 
397 	case DMA_OTYP_PAGES:
398 		offset = dobj_p->dmao_obj.pp_obj.pp_offset;
399 		mp->dmai_flags |= PX_DMAI_FLAGS_PGPFN;
400 		pfn0 = page_pptonum(dobj_p->dmao_obj.pp_obj.pp_pp);
401 		break;
402 
403 	case DMA_OTYP_PADDR:
404 	default:
405 		cmn_err(CE_WARN, "%s%d requested unsupported dma type %x",
406 			NAMEINST(mp->dmai_rdip), dobj_p->dmao_type);
407 		return (DDI_DMA_NOMAPPING);
408 	}
409 	if (pfn0 == PFN_INVALID) {
410 		cmn_err(CE_WARN, "%s%d: invalid pfn0 for DMA object %p",
411 			NAMEINST(dip), dobj_p);
412 		return (DDI_DMA_NOMAPPING);
413 	}
414 	if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base32_pfn,
415 			pec_p->pec_last32_pfn)) {
416 		mp->dmai_flags |= PX_DMAI_FLAGS_PTP|PX_DMAI_FLAGS_PTP32;
417 		goto done;	/* leave bypass and dvma flag as 0 */
418 	} else if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base64_pfn,
419 			pec_p->pec_last64_pfn)) {
420 		mp->dmai_flags |= PX_DMAI_FLAGS_PTP|PX_DMAI_FLAGS_PTP64;
421 		goto done;	/* leave bypass and dvma flag as 0 */
422 	}
423 	if (PX_DMA_ISPEERONLY(mp)) {
424 		dev_info_t *rdip = mp->dmai_rdip;
425 		cmn_err(CE_WARN, "Bad peer-to-peer req %s%d", NAMEINST(rdip));
426 		return (DDI_DMA_NOMAPPING);
427 	}
428 	mp->dmai_flags |= (mp->dmai_flags & PX_DMAI_FLAGS_BYPASSREQ) ?
429 	    PX_DMAI_FLAGS_BYPASS : PX_DMAI_FLAGS_DVMA |
430 	    (mp->dmai_rflags & DDI_DMA_REDZONE ? PX_DMAI_FLAGS_REDZONE : 0);
431 done:
432 	mp->dmai_object	 = *dobj_p;			/* whole object    */
433 	mp->dmai_pfn0	 = (void *)pfn0;		/* cache pfn0	   */
434 	mp->dmai_roffset = offset;			/* win0 pg0 offset */
435 	mp->dmai_ndvmapages = MMU_BTOPR(offset + mp->dmai_object.dmao_size);
436 	return (DDI_SUCCESS);
437 }
438 
439 /*
440  * px_dma_pgpfn - set up pfnlst array according to pages
441  *	VA/size pair: <shadow IO, bypass, peer-to-peer>, or OTYP_PAGES
442  */
443 /*ARGSUSED*/
444 static int
445 px_dma_pgpfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
446 {
447 	int i;
448 	dev_info_t *dip = px_p->px_dip;
449 
450 	switch (mp->dmai_object.dmao_type) {
451 	case DMA_OTYP_BUFVADDR:
452 	case DMA_OTYP_VADDR: {
453 		page_t **pplist = mp->dmai_object.dmao_obj.virt_obj.v_priv;
454 		DBG(DBG_DMA_MAP, dip, "shadow pplist=%p, %x pages, pfns=",
455 			pplist, npages);
456 		for (i = 1; i < npages; i++) {
457 			px_iopfn_t pfn = page_pptonum(pplist[i]);
458 			PX_SET_MP_PFN1(mp, i, pfn);
459 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
460 		}
461 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
462 		}
463 		break;
464 
465 	case DMA_OTYP_PAGES: {
466 		page_t *pp = mp->dmai_object.dmao_obj.pp_obj.pp_pp->p_next;
467 		DBG(DBG_DMA_MAP, dip, "pp=%p pfns=", pp);
468 		for (i = 1; i < npages; i++, pp = pp->p_next) {
469 			px_iopfn_t pfn = page_pptonum(pp);
470 			PX_SET_MP_PFN1(mp, i, pfn);
471 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
472 		}
473 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
474 		}
475 		break;
476 
477 	default:	/* check is already done by px_dma_type */
478 		ASSERT(0);
479 		break;
480 	}
481 	return (DDI_SUCCESS);
482 }
483 
484 /*
485  * px_dma_vapfn - set up pfnlst array according to VA
486  *	VA/size pair: <normal, bypass, peer-to-peer>
487  *	pfn0 is skipped as it is already done.
488  *	In this case, the cached pfn0 is used to fill pfnlst[0]
489  */
490 static int
491 px_dma_vapfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
492 {
493 	dev_info_t *dip = px_p->px_dip;
494 	int i;
495 	caddr_t vaddr = (caddr_t)mp->dmai_object.dmao_obj.virt_obj.v_as;
496 	struct hat *hat_p = vaddr ? ((struct as *)vaddr)->a_hat : kas.a_hat;
497 
498 	vaddr = mp->dmai_object.dmao_obj.virt_obj.v_addr + MMU_PAGE_SIZE;
499 	for (i = 1; i < npages; i++, vaddr += MMU_PAGE_SIZE) {
500 		px_iopfn_t pfn = hat_getpfnum(hat_p, vaddr);
501 		if (pfn == PFN_INVALID)
502 			goto err_badpfn;
503 		PX_SET_MP_PFN1(mp, i, pfn);
504 		DBG(DBG_DMA_BINDH, dip, "px_dma_vapfn: mp=%p pfnlst[%x]=%x\n",
505 			mp, i, pfn);
506 	}
507 	return (DDI_SUCCESS);
508 err_badpfn:
509 	cmn_err(CE_WARN, "%s%d: bad page frame vaddr=%p", NAMEINST(dip), vaddr);
510 	return (DDI_DMA_NOMAPPING);
511 }
512 
513 /*
514  * px_dma_pfn - Fills pfn list for all pages being DMA-ed.
515  *
516  * dependencies:
517  *	mp->dmai_ndvmapages	- set to total # of dma pages
518  *
519  * return value:
520  *	DDI_SUCCESS
521  *	DDI_DMA_NOMAPPING
522  */
523 int
524 px_dma_pfn(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
525 {
526 	uint32_t npages = mp->dmai_ndvmapages;
527 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
528 	int i, ret, peer = PX_DMA_ISPTP(mp);
529 	int peer32 = PX_DMA_ISPTP32(mp);
530 	dev_info_t *dip = px_p->px_dip;
531 
532 	px_pec_t *pec_p = px_p->px_pec_p;
533 	px_iopfn_t pfn_base = peer32 ? pec_p->pec_base32_pfn :
534 					pec_p->pec_base64_pfn;
535 	px_iopfn_t pfn_last = peer32 ? pec_p->pec_last32_pfn :
536 					pec_p->pec_last64_pfn;
537 	px_iopfn_t pfn_adj = peer ? pfn_base : 0;
538 
539 	DBG(DBG_DMA_BINDH, dip, "px_dma_pfn: mp=%p pfn0=%x\n",
540 		mp, PX_MP_PFN0(mp) - pfn_adj);
541 	/* 1 page: no array alloc/fill, no mixed mode check */
542 	if (npages == 1) {
543 		PX_SET_MP_PFN(mp, 0, PX_MP_PFN0(mp) - pfn_adj);
544 		return (DDI_SUCCESS);
545 	}
546 	/* allocate pfn array */
547 	if (!(mp->dmai_pfnlst = kmem_alloc(npages * sizeof (px_iopfn_t),
548 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP))) {
549 		if (waitfp != DDI_DMA_DONTWAIT)
550 			ddi_set_callback(waitfp, dmareq->dmar_arg,
551 				&px_kmem_clid);
552 		return (DDI_DMA_NORESOURCES);
553 	}
554 	/* fill pfn array */
555 	PX_SET_MP_PFN(mp, 0, PX_MP_PFN0(mp) - pfn_adj);	/* pfnlst[0] */
556 	if ((ret = PX_DMA_ISPGPFN(mp) ? px_dma_pgpfn(px_p, mp, npages) :
557 		px_dma_vapfn(px_p, mp, npages)) != DDI_SUCCESS)
558 		goto err;
559 
560 	/* skip pfn0, check mixed mode and adjust peer to peer pfn */
561 	for (i = 1; i < npages; i++) {
562 		px_iopfn_t pfn = PX_GET_MP_PFN1(mp, i);
563 		if (peer ^ TGT_PFN_INBETWEEN(pfn, pfn_base, pfn_last)) {
564 			cmn_err(CE_WARN, "%s%d mixed mode DMA %lx %lx",
565 				NAMEINST(mp->dmai_rdip), PX_MP_PFN0(mp), pfn);
566 			ret = DDI_DMA_NOMAPPING;	/* mixed mode */
567 			goto err;
568 		}
569 		DBG(DBG_DMA_MAP, dip,
570 			"px_dma_pfn: pfnlst[%x]=%x-%x\n", i, pfn, pfn_adj);
571 		if (pfn_adj)
572 			PX_SET_MP_PFN1(mp, i, pfn - pfn_adj);
573 	}
574 	return (DDI_SUCCESS);
575 err:
576 	px_dma_freepfn(mp);
577 	return (ret);
578 }
579 
580 /*
581  * px_dvma_win() - trim requested DVMA size down to window size
582  *	The 1st window starts from offset and ends at page-aligned boundary.
583  *	From the 2nd window on, each window starts and ends at page-aligned
584  *	boundary except the last window ends at wherever requested.
585  *
586  *	accesses the following mp-> members:
587  *	mp->dmai_attr.dma_attr_count_max
588  *	mp->dmai_attr.dma_attr_seg
589  *	mp->dmai_roffset   - start offset of 1st window
590  *	mp->dmai_rflags (redzone)
591  *	mp->dmai_ndvmapages (for 1 page fast path)
592  *
593  *	sets the following mp-> members:
594  *	mp->dmai_size	   - xfer size, != winsize if 1st/last win  (not fixed)
595  *	mp->dmai_winsize   - window size (no redzone), n * page size    (fixed)
596  *	mp->dmai_nwin	   - # of DMA windows of entire object		(fixed)
597  *	mp->dmai_rflags	   - remove partial flag if nwin == 1		(fixed)
598  *	mp->dmai_winlst	   - NULL, window objects not used for DVMA	(fixed)
599  *
600  *	fixed - not changed across different DMA windows
601  */
602 /*ARGSUSED*/
603 int
604 px_dvma_win(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
605 {
606 	uint32_t redzone_sz	= PX_HAS_REDZONE(mp) ? MMU_PAGE_SIZE : 0;
607 	size_t obj_sz		= mp->dmai_object.dmao_size;
608 	size_t xfer_sz;
609 	ulong_t pg_off;
610 
611 	if ((mp->dmai_ndvmapages == 1) && !redzone_sz) {
612 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
613 		mp->dmai_size = obj_sz;
614 		mp->dmai_winsize = MMU_PAGE_SIZE;
615 		mp->dmai_nwin = 1;
616 		goto done;
617 	}
618 
619 	pg_off	= mp->dmai_roffset;
620 	xfer_sz	= obj_sz + redzone_sz;
621 
622 	/* include redzone in nocross check */ {
623 		uint64_t nocross = mp->dmai_attr.dma_attr_seg;
624 		if (xfer_sz + pg_off - 1 > nocross)
625 			xfer_sz = nocross - pg_off + 1;
626 		if (redzone_sz && (xfer_sz <= redzone_sz)) {
627 			DBG(DBG_DMA_MAP, px_p->px_dip,
628 			    "nocross too small: "
629 			    "%lx(%lx)+%lx+%lx < %llx\n",
630 			    xfer_sz, obj_sz, pg_off, redzone_sz, nocross);
631 			return (DDI_DMA_TOOBIG);
632 		}
633 	}
634 	xfer_sz -= redzone_sz;		/* restore transfer size  */
635 	/* check counter max */ {
636 		uint32_t count_max = mp->dmai_attr.dma_attr_count_max;
637 		if (xfer_sz - 1 > count_max)
638 			xfer_sz = count_max + 1;
639 	}
640 	if (xfer_sz >= obj_sz) {
641 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
642 		mp->dmai_size = xfer_sz;
643 		mp->dmai_winsize = P2ROUNDUP(xfer_sz + pg_off, MMU_PAGE_SIZE);
644 		mp->dmai_nwin = 1;
645 		goto done;
646 	}
647 	if (!(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
648 		DBG(DBG_DMA_MAP, px_p->px_dip, "too big: %lx+%lx+%lx > %lx\n",
649 			obj_sz, pg_off, redzone_sz, xfer_sz);
650 		return (DDI_DMA_TOOBIG);
651 	}
652 
653 	xfer_sz = MMU_PTOB(MMU_BTOP(xfer_sz + pg_off)); /* page align */
654 	mp->dmai_size = xfer_sz - pg_off;	/* 1st window xferrable size */
655 	mp->dmai_winsize = xfer_sz;		/* redzone not in winsize */
656 	mp->dmai_nwin = (obj_sz + pg_off + xfer_sz - 1) / xfer_sz;
657 done:
658 	mp->dmai_winlst = NULL;
659 	px_dump_dma_handle(DBG_DMA_MAP, px_p->px_dip, mp);
660 	return (DDI_SUCCESS);
661 }
662 
663 /*
664  * fast track cache entry to mmu context, inserts 3 0 bits between
665  * upper 6-bits and lower 3-bits of the 9-bit cache entry
666  */
667 #define	MMU_FCE_TO_CTX(i)	(((i) << 3) | ((i) & 0x7) | 0x38)
668 
669 /*
670  * px_dvma_map_fast - attempts to map fast trackable DVMA
671  */
672 /*ARGSUSED*/
673 int
674 px_dvma_map_fast(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
675 {
676 	uint_t clustsz = px_dvma_page_cache_clustsz;
677 	uint_t entries = px_dvma_page_cache_entries;
678 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
679 	    mp->dmai_attr.dma_attr_flags);
680 	int i = mmu_p->mmu_dvma_addr_scan_start;
681 	uint8_t *lock_addr = mmu_p->mmu_dvma_cache_locks + i;
682 	px_dvma_addr_t dvma_pg;
683 	size_t npages = MMU_BTOP(mp->dmai_winsize);
684 	dev_info_t *dip = mmu_p->mmu_px_p->px_dip;
685 
686 	extern uint8_t ldstub(uint8_t *);
687 	ASSERT(MMU_PTOB(npages) == mp->dmai_winsize);
688 	ASSERT(npages + PX_HAS_REDZONE(mp) <= clustsz);
689 
690 	for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
691 	if (i >= entries) {
692 		lock_addr = mmu_p->mmu_dvma_cache_locks;
693 		i = 0;
694 		for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
695 		if (i >= entries) {
696 #ifdef	PX_DMA_PROF
697 			px_dvmaft_exhaust++;
698 #endif	/* PX_DMA_PROF */
699 			return (DDI_DMA_NORESOURCES);
700 		}
701 	}
702 	mmu_p->mmu_dvma_addr_scan_start = (i + 1) & (entries - 1);
703 
704 	i *= clustsz;
705 	dvma_pg = mmu_p->dvma_base_pg + i;
706 
707 	if (px_lib_iommu_map(dip, PCI_TSBID(0, i), npages, attr,
708 	    (void *)mp, 0, MMU_MAP_PFN) != DDI_SUCCESS) {
709 		DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: "
710 		    "px_lib_iommu_map failed\n");
711 
712 		return (DDI_FAILURE);
713 	}
714 
715 	if (!PX_MAP_BUFZONE(mp))
716 		goto done;
717 
718 	DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: redzone pg=%x\n", i + npages);
719 
720 	ASSERT(PX_HAS_REDZONE(mp));
721 
722 	if (px_lib_iommu_map(dip, PCI_TSBID(0, i + npages), 1, attr,
723 	    (void *)mp, npages - 1, MMU_MAP_PFN) != DDI_SUCCESS) {
724 		DBG(DBG_MAP_WIN, dip, "px_dvma_map_fast: "
725 		    "mapping REDZONE page failed\n");
726 
727 		(void) px_lib_iommu_demap(dip, PCI_TSBID(0, i), npages);
728 		return (DDI_FAILURE);
729 	}
730 
731 done:
732 #ifdef PX_DMA_PROF
733 	px_dvmaft_success++;
734 #endif
735 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
736 	mp->dmai_offset = 0;
737 	mp->dmai_flags |= PX_DMAI_FLAGS_FASTTRACK;
738 	PX_SAVE_MP_TTE(mp, attr);	/* save TTE template for unmapping */
739 	if (PX_DVMA_DBG_ON(mmu_p))
740 		px_dvma_alloc_debug(mmu_p, (char *)mp->dmai_mapping,
741 			mp->dmai_size, mp);
742 	return (DDI_SUCCESS);
743 }
744 
745 /*
746  * px_dvma_map: map non-fasttrack DMA
747  *		Use quantum cache if single page DMA.
748  */
749 int
750 px_dvma_map(ddi_dma_impl_t *mp, ddi_dma_req_t *dmareq, px_mmu_t *mmu_p)
751 {
752 	uint_t npages = PX_DMA_WINNPGS(mp);
753 	px_dvma_addr_t dvma_pg, dvma_pg_index;
754 	void *dvma_addr;
755 	uint64_t tte = PX_GET_TTE_ATTR(mp->dmai_rflags,
756 	    mp->dmai_attr.dma_attr_flags);
757 	int sleep = dmareq->dmar_fp == DDI_DMA_SLEEP ? VM_SLEEP : VM_NOSLEEP;
758 	dev_info_t *dip = mp->dmai_rdip;
759 	int	ret = DDI_SUCCESS;
760 
761 	/*
762 	 * allocate dvma space resource and map in the first window.
763 	 * (vmem_t *vmp, size_t size,
764 	 *	size_t align, size_t phase, size_t nocross,
765 	 *	void *minaddr, void *maxaddr, int vmflag)
766 	 */
767 	if ((npages == 1) && !PX_HAS_REDZONE(mp) && PX_HAS_NOSYSLIMIT(mp)) {
768 		dvma_addr = vmem_alloc(mmu_p->mmu_dvma_map,
769 			MMU_PAGE_SIZE, sleep);
770 		mp->dmai_flags |= PX_DMAI_FLAGS_VMEMCACHE;
771 #ifdef	PX_DMA_PROF
772 		px_dvma_vmem_alloc++;
773 #endif	/* PX_DMA_PROF */
774 	} else {
775 		dvma_addr = vmem_xalloc(mmu_p->mmu_dvma_map,
776 			MMU_PTOB(npages + PX_HAS_REDZONE(mp)),
777 			MAX(mp->dmai_attr.dma_attr_align, MMU_PAGE_SIZE),
778 			0,
779 			mp->dmai_attr.dma_attr_seg + 1,
780 			(void *)mp->dmai_attr.dma_attr_addr_lo,
781 			(void *)(mp->dmai_attr.dma_attr_addr_hi + 1),
782 			sleep);
783 #ifdef	PX_DMA_PROF
784 		px_dvma_vmem_xalloc++;
785 #endif	/* PX_DMA_PROF */
786 	}
787 	dvma_pg = MMU_BTOP((ulong_t)dvma_addr);
788 	dvma_pg_index = dvma_pg - mmu_p->dvma_base_pg;
789 	DBG(DBG_DMA_MAP, dip, "fallback dvma_pages: dvma_pg=%x index=%x\n",
790 		dvma_pg, dvma_pg_index);
791 	if (dvma_pg == 0)
792 		goto noresource;
793 
794 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
795 	mp->dmai_offset = 0;
796 	PX_SAVE_MP_TTE(mp, tte);	/* mp->dmai_tte = tte */
797 
798 	if ((ret = px_mmu_map_pages(mmu_p,
799 	    mp, dvma_pg, npages, 0)) != DDI_SUCCESS) {
800 		if (mp->dmai_flags & PX_DMAI_FLAGS_VMEMCACHE) {
801 			vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
802 			    MMU_PAGE_SIZE);
803 #ifdef PX_DMA_PROF
804 			px_dvma_vmem_free++;
805 #endif /* PX_DMA_PROF */
806 		} else {
807 			vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
808 			    MMU_PTOB(npages + PX_HAS_REDZONE(mp)));
809 #ifdef PX_DMA_PROF
810 			px_dvma_vmem_xfree++;
811 #endif /* PX_DMA_PROF */
812 		}
813 	}
814 
815 	return (ret);
816 noresource:
817 	if (dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
818 		DBG(DBG_DMA_MAP, dip, "dvma_pg 0 - set callback\n");
819 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
820 			&mmu_p->mmu_dvma_clid);
821 	}
822 	DBG(DBG_DMA_MAP, dip, "vmem_xalloc - DDI_DMA_NORESOURCES\n");
823 	return (DDI_DMA_NORESOURCES);
824 }
825 
826 void
827 px_dvma_unmap(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
828 {
829 	px_dvma_addr_t dvma_addr = (px_dvma_addr_t)mp->dmai_mapping;
830 	px_dvma_addr_t dvma_pg = MMU_BTOP(dvma_addr);
831 	dvma_addr = MMU_PTOB(dvma_pg);
832 
833 	if (mp->dmai_flags & PX_DMAI_FLAGS_FASTTRACK) {
834 		px_iopfn_t index = dvma_pg - mmu_p->dvma_base_pg;
835 		ASSERT(index % px_dvma_page_cache_clustsz == 0);
836 		index /= px_dvma_page_cache_clustsz;
837 		ASSERT(index < px_dvma_page_cache_entries);
838 		mmu_p->mmu_dvma_cache_locks[index] = 0;
839 #ifdef	PX_DMA_PROF
840 		px_dvmaft_free++;
841 #endif	/* PX_DMA_PROF */
842 		return;
843 	}
844 
845 	if (mp->dmai_flags & PX_DMAI_FLAGS_VMEMCACHE) {
846 		vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
847 			MMU_PAGE_SIZE);
848 #ifdef PX_DMA_PROF
849 		px_dvma_vmem_free++;
850 #endif /* PX_DMA_PROF */
851 	} else {
852 		size_t npages = MMU_BTOP(mp->dmai_winsize) + PX_HAS_REDZONE(mp);
853 		vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
854 			MMU_PTOB(npages));
855 #ifdef PX_DMA_PROF
856 		px_dvma_vmem_xfree++;
857 #endif /* PX_DMA_PROF */
858 	}
859 }
860 
861 /*
862  * DVMA mappings may have multiple windows, but each window always have
863  * one segment.
864  */
865 int
866 px_dvma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
867 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
868 	uint_t cache_flags)
869 {
870 	switch (cmd) {
871 	case DDI_DMA_SYNC:
872 		return (px_lib_dma_sync(dip, rdip, (ddi_dma_handle_t)mp,
873 		    *offp, *lenp, cache_flags));
874 
875 	case DDI_DMA_HTOC: {
876 		int ret;
877 		off_t wo_off, off = *offp;	/* wo_off: wnd's obj offset */
878 		uint_t win_size = mp->dmai_winsize;
879 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)objp;
880 
881 		if (off >= mp->dmai_object.dmao_size) {
882 			cmn_err(CE_WARN, "%s%d invalid dma_htoc offset %lx",
883 				NAMEINST(mp->dmai_rdip), off);
884 			return (DDI_FAILURE);
885 		}
886 		off += mp->dmai_roffset;
887 		ret = px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
888 		    off / win_size, &wo_off, NULL, cp, NULL); /* lenp == NULL */
889 		if (ret)
890 			return (ret);
891 		DBG(DBG_DMA_CTL, dip, "HTOC:cookie=%x+%lx off=%lx,%lx\n",
892 			cp->dmac_address, cp->dmac_size, off, *offp);
893 
894 		/* adjust cookie addr/len if we are not on window boundary */
895 		ASSERT((off % win_size) == (off -
896 			(PX_DMA_CURWIN(mp) ? mp->dmai_roffset : 0) - wo_off));
897 		off = PX_DMA_CURWIN(mp) ? off % win_size : *offp;
898 		ASSERT(cp->dmac_size > off);
899 		cp->dmac_laddress += off;
900 		cp->dmac_size -= off;
901 		DBG(DBG_DMA_CTL, dip, "HTOC:mp=%p cookie=%x+%lx off=%lx,%lx\n",
902 			mp, cp->dmac_address, cp->dmac_size, off, wo_off);
903 		}
904 		return (DDI_SUCCESS);
905 
906 	case DDI_DMA_REPWIN:
907 		*offp = mp->dmai_offset;
908 		*lenp = mp->dmai_size;
909 		return (DDI_SUCCESS);
910 
911 	case DDI_DMA_MOVWIN: {
912 		off_t off = *offp;
913 		if (off >= mp->dmai_object.dmao_size)
914 			return (DDI_FAILURE);
915 		off += mp->dmai_roffset;
916 		return (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
917 		    off / mp->dmai_winsize, offp, lenp,
918 		    (ddi_dma_cookie_t *)objp, NULL));
919 		}
920 
921 	case DDI_DMA_NEXTWIN: {
922 		px_window_t win = PX_DMA_CURWIN(mp);
923 		if (offp) {
924 			if (*(px_window_t *)offp != win) {
925 				/* window not active */
926 				*(px_window_t *)objp = win; /* return cur win */
927 				return (DDI_DMA_STALE);
928 			}
929 			win++;
930 		} else	/* map win 0 */
931 			win = 0;
932 		if (win >= mp->dmai_nwin) {
933 			*(px_window_t *)objp = win - 1;
934 			return (DDI_DMA_DONE);
935 		}
936 		if (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
937 		    win, 0, 0, 0, 0)) {
938 			*(px_window_t *)objp = win - 1;
939 			return (DDI_FAILURE);
940 		}
941 		*(px_window_t *)objp = win;
942 		}
943 		return (DDI_SUCCESS);
944 
945 	case DDI_DMA_NEXTSEG:
946 		if (*(px_window_t *)offp != PX_DMA_CURWIN(mp))
947 			return (DDI_DMA_STALE);
948 		if (lenp)				/* only 1 seg allowed */
949 			return (DDI_DMA_DONE);
950 
951 		/* return mp as seg 0 */
952 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
953 		return (DDI_SUCCESS);
954 
955 	case DDI_DMA_SEGTOC:
956 		MAKE_DMA_COOKIE((ddi_dma_cookie_t *)objp, mp->dmai_mapping,
957 			mp->dmai_size);
958 		*offp = mp->dmai_offset;
959 		*lenp = mp->dmai_size;
960 		return (DDI_SUCCESS);
961 
962 	case DDI_DMA_COFF: {
963 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)offp;
964 		if (cp->dmac_address < mp->dmai_mapping ||
965 			(cp->dmac_address + cp->dmac_size) >
966 			(mp->dmai_mapping + mp->dmai_size))
967 			return (DDI_FAILURE);
968 		*objp = (caddr_t)(cp->dmac_address - mp->dmai_mapping +
969 			mp->dmai_offset);
970 		}
971 		return (DDI_SUCCESS);
972 	default:
973 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
974 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
975 		break;
976 	}
977 	return (DDI_FAILURE);
978 }
979 
980 void
981 px_dma_freewin(ddi_dma_impl_t *mp)
982 {
983 	px_dma_win_t *win_p = mp->dmai_winlst, *win2_p;
984 	for (win2_p = win_p; win_p; win2_p = win_p) {
985 		win_p = win2_p->win_next;
986 		kmem_free(win2_p, sizeof (px_dma_win_t) +
987 			sizeof (ddi_dma_cookie_t) * win2_p->win_ncookies);
988 	}
989 	mp->dmai_nwin = 0;
990 	mp->dmai_winlst = NULL;
991 }
992 
993 /*
994  * px_dma_newwin - create a dma window object and cookies
995  *
996  *	After the initial scan in px_dma_physwin(), which identifies
997  *	a portion of the pfn array that belongs to a dma window,
998  *	we are called to allocate and initialize representing memory
999  *	resources. We know from the 1st scan the number of cookies
1000  *	or dma segment in this window so we can allocate a contiguous
1001  *	memory array for the dma cookies (The implementation of
1002  *	ddi_dma_nextcookie(9f) dictates dma cookies be contiguous).
1003  *
1004  *	A second round scan is done on the pfn array to identify
1005  *	each dma segment and initialize its corresponding dma cookie.
1006  *	We don't need to do all the safety checking and we know they
1007  *	all belong to the same dma window.
1008  *
1009  *	Input:	cookie_no - # of cookies identified by the 1st scan
1010  *		start_idx - subscript of the pfn array for the starting pfn
1011  *		end_idx   - subscript of the last pfn in dma window
1012  *		win_pp    - pointer to win_next member of previous window
1013  *	Return:	DDI_SUCCESS - with **win_pp as newly created window object
1014  *		DDI_DMA_NORESROUCE - caller frees all previous window objs
1015  *	Note:	Each cookie and window size are all initialized on page
1016  *		boundary. This is not true for the 1st cookie of the 1st
1017  *		window and the last cookie of the last window.
1018  *		We fix that later in upper layer which has access to size
1019  *		and offset info.
1020  *
1021  */
1022 /*ARGSUSED*/
1023 static int
1024 px_dma_newwin(dev_info_t *dip, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp,
1025 	uint32_t cookie_no, uint32_t start_idx, uint32_t end_idx,
1026 	px_dma_win_t **win_pp, uint64_t count_max, uint64_t bypass)
1027 {
1028 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
1029 	ddi_dma_cookie_t *cookie_p;
1030 	uint32_t pfn_no = 1;
1031 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, start_idx);
1032 	px_iopfn_t prev_pfn = pfn;
1033 	uint64_t baddr, seg_pfn0 = pfn;
1034 	size_t sz = cookie_no * sizeof (ddi_dma_cookie_t);
1035 	px_dma_win_t *win_p = kmem_zalloc(sizeof (px_dma_win_t) + sz,
1036 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP);
1037 	io_attributes_t	attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
1038 	    mp->dmai_attr.dma_attr_flags);
1039 
1040 	if (!win_p)
1041 		goto noresource;
1042 
1043 	win_p->win_next = NULL;
1044 	win_p->win_ncookies = cookie_no;
1045 	win_p->win_curseg = 0;	/* start from segment 0 */
1046 	win_p->win_size = MMU_PTOB(end_idx - start_idx + 1);
1047 	/* win_p->win_offset is left uninitialized */
1048 
1049 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1050 	start_idx++;
1051 	for (; start_idx <= end_idx; start_idx++, prev_pfn = pfn, pfn_no++) {
1052 		pfn = PX_GET_MP_PFN1(mp, start_idx);
1053 		if ((pfn == prev_pfn + 1) &&
1054 			(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1055 			continue;
1056 
1057 		/* close up the cookie up to (including) prev_pfn */
1058 		baddr = MMU_PTOB(seg_pfn0);
1059 		if (bypass && (px_lib_iommu_getbypass(dip,
1060 				baddr, attr, &baddr) != DDI_SUCCESS))
1061 			return (DDI_FAILURE);
1062 
1063 		MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1064 		DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages)\n",
1065 			MMU_PTOB(seg_pfn0), pfn_no);
1066 
1067 		cookie_p++;	/* advance to next available cookie cell */
1068 		pfn_no = 0;
1069 		seg_pfn0 = pfn;	/* start a new segment from current pfn */
1070 	}
1071 
1072 	baddr = MMU_PTOB(seg_pfn0);
1073 	if (bypass && (px_lib_iommu_getbypass(dip,
1074 			baddr, attr, &baddr) != DDI_SUCCESS))
1075 		return (DDI_FAILURE);
1076 
1077 	MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1078 	DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages) of total %x\n",
1079 		MMU_PTOB(seg_pfn0), pfn_no, cookie_no);
1080 #ifdef	DEBUG
1081 	cookie_p++;
1082 	ASSERT((cookie_p - (ddi_dma_cookie_t *)(win_p + 1)) == cookie_no);
1083 #endif	/* DEBUG */
1084 	*win_pp = win_p;
1085 	return (DDI_SUCCESS);
1086 noresource:
1087 	if (waitfp != DDI_DMA_DONTWAIT)
1088 		ddi_set_callback(waitfp, dmareq->dmar_arg, &px_kmem_clid);
1089 	return (DDI_DMA_NORESOURCES);
1090 }
1091 
1092 /*
1093  * px_dma_adjust - adjust 1st and last cookie and window sizes
1094  *	remove initial dma page offset from 1st cookie and window size
1095  *	remove last dma page remainder from last cookie and window size
1096  *	fill win_offset of each dma window according to just fixed up
1097  *		each window sizes
1098  *	px_dma_win_t members modified:
1099  *	win_p->win_offset - this window's offset within entire DMA object
1100  *	win_p->win_size	  - xferrable size (in bytes) for this window
1101  *
1102  *	ddi_dma_impl_t members modified:
1103  *	mp->dmai_size	  - 1st window xferrable size
1104  *	mp->dmai_offset   - 0, which is the dma offset of the 1st window
1105  *
1106  *	ddi_dma_cookie_t members modified:
1107  *	cookie_p->dmac_size - 1st and last cookie remove offset or remainder
1108  *	cookie_p->dmac_laddress - 1st cookie add page offset
1109  */
1110 static void
1111 px_dma_adjust(ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp, px_dma_win_t *win_p)
1112 {
1113 	ddi_dma_cookie_t *cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1114 	size_t pg_offset = mp->dmai_roffset;
1115 	size_t win_offset = 0;
1116 
1117 	cookie_p->dmac_size -= pg_offset;
1118 	cookie_p->dmac_laddress |= pg_offset;
1119 	win_p->win_size -= pg_offset;
1120 	DBG(DBG_BYPASS, mp->dmai_rdip, "pg0 adjust %lx\n", pg_offset);
1121 
1122 	mp->dmai_size = win_p->win_size;
1123 	mp->dmai_offset = 0;
1124 
1125 	pg_offset += mp->dmai_object.dmao_size;
1126 	pg_offset &= MMU_PAGE_OFFSET;
1127 	if (pg_offset)
1128 		pg_offset = MMU_PAGE_SIZE - pg_offset;
1129 	DBG(DBG_BYPASS, mp->dmai_rdip, "last pg adjust %lx\n", pg_offset);
1130 
1131 	for (; win_p->win_next; win_p = win_p->win_next) {
1132 		DBG(DBG_BYPASS, mp->dmai_rdip, "win off %p\n", win_offset);
1133 		win_p->win_offset = win_offset;
1134 		win_offset += win_p->win_size;
1135 	}
1136 	/* last window */
1137 	win_p->win_offset = win_offset;
1138 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1139 	cookie_p[win_p->win_ncookies - 1].dmac_size -= pg_offset;
1140 	win_p->win_size -= pg_offset;
1141 	ASSERT((win_offset + win_p->win_size) == mp->dmai_object.dmao_size);
1142 }
1143 
1144 /*
1145  * px_dma_physwin() - carve up dma windows using physical addresses.
1146  *	Called to handle mmu bypass and pci peer-to-peer transfers.
1147  *	Calls px_dma_newwin() to allocate window objects.
1148  *
1149  * Dependency: mp->dmai_pfnlst points to an array of pfns
1150  *
1151  * 1. Each dma window is represented by a px_dma_win_t object.
1152  *	The object will be casted to ddi_dma_win_t and returned
1153  *	to leaf driver through the DDI interface.
1154  * 2. Each dma window can have several dma segments with each
1155  *	segment representing a physically contiguous either memory
1156  *	space (if we are doing an mmu bypass transfer) or pci address
1157  *	space (if we are doing a peer-to-peer transfer).
1158  * 3. Each segment has a DMA cookie to program the DMA engine.
1159  *	The cookies within each DMA window must be located in a
1160  *	contiguous array per ddi_dma_nextcookie(9f).
1161  * 4. The number of DMA segments within each DMA window cannot exceed
1162  *	mp->dmai_attr.dma_attr_sgllen. If the transfer size is
1163  *	too large to fit in the sgllen, the rest needs to be
1164  *	relocated to the next dma window.
1165  * 5. Peer-to-peer DMA segment follows device hi, lo, count_max,
1166  *	and nocross restrictions while bypass DMA follows the set of
1167  *	restrictions with system limits factored in.
1168  *
1169  * Return:
1170  *	mp->dmai_winlst	 - points to a link list of px_dma_win_t objects.
1171  *		Each px_dma_win_t object on the link list contains
1172  *		infomation such as its window size (# of pages),
1173  *		starting offset (also see Restriction), an array of
1174  *		DMA cookies, and # of cookies in the array.
1175  *	mp->dmai_pfnlst	 - NULL, the pfn list is freed to conserve memory.
1176  *	mp->dmai_nwin	 - # of total DMA windows on mp->dmai_winlst.
1177  *	mp->dmai_mapping - starting cookie address
1178  *	mp->dmai_rflags	 - consistent, nosync, no redzone
1179  *	mp->dmai_cookie	 - start of cookie table of the 1st DMA window
1180  *
1181  * Restriction:
1182  *	Each px_dma_win_t object can theoratically start from any offset
1183  *	since the mmu is not involved. However, this implementation
1184  *	always make windows start from page aligned offset (except
1185  *	the 1st window, which follows the requested offset) due to the
1186  *	fact that we are handed a pfn list. This does require device's
1187  *	count_max and attr_seg to be at least MMU_PAGE_SIZE aligned.
1188  */
1189 int
1190 px_dma_physwin(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
1191 {
1192 	uint_t npages = mp->dmai_ndvmapages;
1193 	int ret, sgllen = mp->dmai_attr.dma_attr_sgllen;
1194 	px_iopfn_t pfn_lo, pfn_hi, prev_pfn;
1195 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, 0);
1196 	uint32_t i, win_no = 0, pfn_no = 1, win_pfn0_index = 0, cookie_no = 0;
1197 	uint64_t count_max, bypass_addr = 0;
1198 	px_dma_win_t **win_pp = (px_dma_win_t **)&mp->dmai_winlst;
1199 	ddi_dma_cookie_t *cookie0_p;
1200 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags,
1201 	    mp->dmai_attr.dma_attr_flags);
1202 	dev_info_t *dip = px_p->px_dip;
1203 
1204 	ASSERT(PX_DMA_ISPTP(mp) || PX_DMA_ISBYPASS(mp));
1205 	if (PX_DMA_ISPTP(mp)) { /* ignore sys limits for peer-to-peer */
1206 		ddi_dma_attr_t *dev_attr_p = PX_DEV_ATTR(mp);
1207 		uint64_t nocross = dev_attr_p->dma_attr_seg;
1208 		px_pec_t *pec_p = px_p->px_pec_p;
1209 		px_iopfn_t pfn_last = PX_DMA_ISPTP32(mp) ?
1210 				pec_p->pec_last32_pfn - pec_p->pec_base32_pfn :
1211 				pec_p->pec_last64_pfn - pec_p->pec_base64_pfn;
1212 
1213 		if (nocross && (nocross < UINT32_MAX))
1214 			return (DDI_DMA_NOMAPPING);
1215 		if (dev_attr_p->dma_attr_align > MMU_PAGE_SIZE)
1216 			return (DDI_DMA_NOMAPPING);
1217 		pfn_lo = MMU_BTOP(dev_attr_p->dma_attr_addr_lo);
1218 		pfn_hi = MMU_BTOP(dev_attr_p->dma_attr_addr_hi);
1219 		pfn_hi = MIN(pfn_hi, pfn_last);
1220 		if ((pfn_lo > pfn_hi) || (pfn < pfn_lo))
1221 			return (DDI_DMA_NOMAPPING);
1222 
1223 		count_max = dev_attr_p->dma_attr_count_max;
1224 		count_max = MIN(count_max, nocross);
1225 		/*
1226 		 * the following count_max trim is not done because we are
1227 		 * making sure pfn_lo <= pfn <= pfn_hi inside the loop
1228 		 * count_max=MIN(count_max, MMU_PTOB(pfn_hi - pfn_lo + 1)-1);
1229 		 */
1230 	} else { /* bypass hi/lo/count_max have been processed by attr2hdl() */
1231 		count_max = mp->dmai_attr.dma_attr_count_max;
1232 		pfn_lo = MMU_BTOP(mp->dmai_attr.dma_attr_addr_lo);
1233 		pfn_hi = MMU_BTOP(mp->dmai_attr.dma_attr_addr_hi);
1234 
1235 		if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn),
1236 				attr, &bypass_addr) != DDI_SUCCESS) {
1237 			cmn_err(CE_WARN, "bypass cookie failure %lx\n", pfn);
1238 			return (DDI_DMA_NOMAPPING);
1239 		}
1240 		pfn = MMU_BTOP(bypass_addr);
1241 	}
1242 
1243 	/* pfn: absolute (bypass mode) or relative (p2p mode) */
1244 	for (prev_pfn = pfn, i = 1; i < npages;
1245 	    i++, prev_pfn = pfn, pfn_no++) {
1246 		pfn = PX_GET_MP_PFN1(mp, i);
1247 		if (bypass_addr) {
1248 			if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn), attr,
1249 					&bypass_addr) != DDI_SUCCESS) {
1250 				ret = DDI_DMA_NOMAPPING;
1251 				goto err;
1252 			}
1253 			pfn = MMU_BTOP(bypass_addr);
1254 		}
1255 		if ((pfn == prev_pfn + 1) &&
1256 				(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1257 			continue;
1258 		if ((pfn < pfn_lo) || (prev_pfn > pfn_hi)) {
1259 			ret = DDI_DMA_NOMAPPING;
1260 			goto err;
1261 		}
1262 		cookie_no++;
1263 		pfn_no = 0;
1264 		if (cookie_no < sgllen)
1265 			continue;
1266 
1267 		DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1268 			win_pfn0_index, i - 1, cookie_no);
1269 		if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no,
1270 			win_pfn0_index, i - 1, win_pp, count_max, bypass_addr))
1271 			goto err;
1272 
1273 		win_pp = &(*win_pp)->win_next;	/* win_pp = *(win_pp) */
1274 		win_no++;
1275 		win_pfn0_index = i;
1276 		cookie_no = 0;
1277 	}
1278 	if (pfn > pfn_hi) {
1279 		ret = DDI_DMA_NOMAPPING;
1280 		goto err;
1281 	}
1282 	cookie_no++;
1283 	DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1284 		win_pfn0_index, i - 1, cookie_no);
1285 	if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no, win_pfn0_index,
1286 		i - 1, win_pp, count_max, bypass_addr))
1287 		goto err;
1288 	win_no++;
1289 	px_dma_adjust(dmareq, mp, mp->dmai_winlst);
1290 	mp->dmai_nwin = win_no;
1291 	mp->dmai_rflags |= DDI_DMA_CONSISTENT | DMP_NOSYNC;
1292 	mp->dmai_rflags &= ~DDI_DMA_REDZONE;
1293 	mp->dmai_flags |= PX_DMAI_FLAGS_NOSYNC;
1294 	cookie0_p = (ddi_dma_cookie_t *)(PX_WINLST(mp) + 1);
1295 	mp->dmai_cookie = PX_WINLST(mp)->win_ncookies > 1 ? cookie0_p + 1 : 0;
1296 	mp->dmai_mapping = cookie0_p->dmac_laddress;
1297 
1298 	px_dma_freepfn(mp);
1299 	return (DDI_DMA_MAPPED);
1300 err:
1301 	px_dma_freewin(mp);
1302 	return (ret);
1303 }
1304 
1305 int
1306 px_dma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
1307 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
1308 	uint_t cache_flags)
1309 {
1310 	switch (cmd) {
1311 	case DDI_DMA_SYNC:
1312 		return (DDI_SUCCESS);
1313 
1314 	case DDI_DMA_HTOC: {
1315 		off_t off = *offp;
1316 		ddi_dma_cookie_t *loop_cp, *cp;
1317 		px_dma_win_t *win_p = mp->dmai_winlst;
1318 
1319 		if (off >= mp->dmai_object.dmao_size)
1320 			return (DDI_FAILURE);
1321 
1322 		/* locate window */
1323 		while (win_p->win_offset + win_p->win_size <= off)
1324 			win_p = win_p->win_next;
1325 
1326 		loop_cp = cp = (ddi_dma_cookie_t *)(win_p + 1);
1327 		mp->dmai_offset = win_p->win_offset;
1328 		mp->dmai_size   = win_p->win_size;
1329 		mp->dmai_mapping = cp->dmac_laddress; /* cookie0 start addr */
1330 
1331 		/* adjust cookie addr/len if we are not on cookie boundary */
1332 		off -= win_p->win_offset;	   /* offset within window */
1333 		for (; off >= loop_cp->dmac_size; loop_cp++)
1334 			off -= loop_cp->dmac_size; /* offset within cookie */
1335 
1336 		mp->dmai_cookie = loop_cp + 1;
1337 		win_p->win_curseg = loop_cp - cp;
1338 		cp = (ddi_dma_cookie_t *)objp;
1339 		MAKE_DMA_COOKIE(cp, loop_cp->dmac_laddress + off,
1340 			loop_cp->dmac_size - off);
1341 
1342 		DBG(DBG_DMA_CTL, dip,
1343 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1344 			cp->dmac_laddress, cp->dmac_size);
1345 		}
1346 		return (DDI_SUCCESS);
1347 
1348 	case DDI_DMA_REPWIN:
1349 		*offp = mp->dmai_offset;
1350 		*lenp = mp->dmai_size;
1351 		return (DDI_SUCCESS);
1352 
1353 	case DDI_DMA_MOVWIN: {
1354 		off_t off = *offp;
1355 		ddi_dma_cookie_t *cp;
1356 		px_dma_win_t *win_p = mp->dmai_winlst;
1357 
1358 		if (off >= mp->dmai_object.dmao_size)
1359 			return (DDI_FAILURE);
1360 
1361 		/* locate window */
1362 		while (win_p->win_offset + win_p->win_size <= off)
1363 			win_p = win_p->win_next;
1364 
1365 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1366 		mp->dmai_offset = win_p->win_offset;
1367 		mp->dmai_size   = win_p->win_size;
1368 		mp->dmai_mapping = cp->dmac_laddress;	/* cookie0 star addr */
1369 		mp->dmai_cookie = cp + 1;
1370 		win_p->win_curseg = 0;
1371 
1372 		*(ddi_dma_cookie_t *)objp = *cp;
1373 		*offp = win_p->win_offset;
1374 		*lenp = win_p->win_size;
1375 		DBG(DBG_DMA_CTL, dip,
1376 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1377 			cp->dmac_laddress, cp->dmac_size);
1378 		}
1379 		return (DDI_SUCCESS);
1380 
1381 	case DDI_DMA_NEXTWIN: {
1382 		px_dma_win_t *win_p = *(px_dma_win_t **)offp;
1383 		px_dma_win_t **nw_pp = (px_dma_win_t **)objp;
1384 		ddi_dma_cookie_t *cp;
1385 		if (!win_p) {
1386 			*nw_pp = mp->dmai_winlst;
1387 			return (DDI_SUCCESS);
1388 		}
1389 
1390 		if (win_p->win_offset != mp->dmai_offset)
1391 			return (DDI_DMA_STALE);
1392 		if (!win_p->win_next)
1393 			return (DDI_DMA_DONE);
1394 		win_p = win_p->win_next;
1395 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1396 		mp->dmai_offset = win_p->win_offset;
1397 		mp->dmai_size   = win_p->win_size;
1398 		mp->dmai_mapping = cp->dmac_laddress;   /* cookie0 star addr */
1399 		mp->dmai_cookie = cp + 1;
1400 		win_p->win_curseg = 0;
1401 		*nw_pp = win_p;
1402 		}
1403 		return (DDI_SUCCESS);
1404 
1405 	case DDI_DMA_NEXTSEG: {
1406 		px_dma_win_t *w_p = *(px_dma_win_t **)offp;
1407 		if (w_p->win_offset != mp->dmai_offset)
1408 			return (DDI_DMA_STALE);
1409 		if (w_p->win_curseg + 1 >= w_p->win_ncookies)
1410 			return (DDI_DMA_DONE);
1411 		w_p->win_curseg++;
1412 		}
1413 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
1414 		return (DDI_SUCCESS);
1415 
1416 	case DDI_DMA_SEGTOC: {
1417 		px_dma_win_t *win_p = mp->dmai_winlst;
1418 		off_t off = mp->dmai_offset;
1419 		ddi_dma_cookie_t *cp;
1420 		int i;
1421 
1422 		/* locate active window */
1423 		for (; win_p->win_offset != off; win_p = win_p->win_next);
1424 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1425 		for (i = 0; i < win_p->win_curseg; i++, cp++)
1426 			off += cp->dmac_size;
1427 		*offp = off;
1428 		*lenp = cp->dmac_size;
1429 		*(ddi_dma_cookie_t *)objp = *cp;	/* copy cookie */
1430 		}
1431 		return (DDI_SUCCESS);
1432 
1433 	case DDI_DMA_COFF: {
1434 		px_dma_win_t *win_p;
1435 		ddi_dma_cookie_t *cp;
1436 		uint64_t addr, key = ((ddi_dma_cookie_t *)offp)->dmac_laddress;
1437 		size_t win_off;
1438 
1439 		for (win_p = mp->dmai_winlst; win_p; win_p = win_p->win_next) {
1440 			int i;
1441 			win_off = 0;
1442 			cp = (ddi_dma_cookie_t *)(win_p + 1);
1443 			for (i = 0; i < win_p->win_ncookies; i++, cp++) {
1444 				size_t sz = cp->dmac_size;
1445 
1446 				addr = cp->dmac_laddress;
1447 				if ((addr <= key) && (addr + sz >= key))
1448 					goto found;
1449 				win_off += sz;
1450 			}
1451 		}
1452 		return (DDI_FAILURE);
1453 found:
1454 		*objp = (caddr_t)(win_p->win_offset + win_off + (key - addr));
1455 		return (DDI_SUCCESS);
1456 		}
1457 	default:
1458 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
1459 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
1460 		break;
1461 	}
1462 	return (DDI_FAILURE);
1463 }
1464 
1465 static void
1466 px_dvma_debug_init(px_mmu_t *mmu_p)
1467 {
1468 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1469 	ASSERT(MUTEX_HELD(&mmu_p->dvma_debug_lock));
1470 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat ON", mmu_p);
1471 
1472 	mmu_p->dvma_alloc_rec = kmem_alloc(sz, KM_SLEEP);
1473 	mmu_p->dvma_free_rec = kmem_alloc(sz, KM_SLEEP);
1474 
1475 	mmu_p->dvma_active_list = NULL;
1476 	mmu_p->dvma_alloc_rec_index = 0;
1477 	mmu_p->dvma_free_rec_index = 0;
1478 	mmu_p->dvma_active_count = 0;
1479 }
1480 
1481 void
1482 px_dvma_debug_fini(px_mmu_t *mmu_p)
1483 {
1484 	struct px_dvma_rec *prev, *ptr;
1485 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1486 	uint64_t mask = ~(1ull << mmu_p->mmu_inst);
1487 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat OFF", mmu_p);
1488 
1489 	kmem_free(mmu_p->dvma_alloc_rec, sz);
1490 	kmem_free(mmu_p->dvma_free_rec, sz);
1491 	mmu_p->dvma_alloc_rec = mmu_p->dvma_free_rec = NULL;
1492 
1493 	prev = mmu_p->dvma_active_list;
1494 	if (!prev)
1495 		return;
1496 	for (ptr = prev->next; ptr; prev = ptr, ptr = ptr->next)
1497 		kmem_free(prev, sizeof (struct px_dvma_rec));
1498 	kmem_free(prev, sizeof (struct px_dvma_rec));
1499 
1500 	mmu_p->dvma_active_list = NULL;
1501 	mmu_p->dvma_alloc_rec_index = 0;
1502 	mmu_p->dvma_free_rec_index = 0;
1503 	mmu_p->dvma_active_count = 0;
1504 
1505 	px_dvma_debug_off &= mask;
1506 	px_dvma_debug_on &= mask;
1507 }
1508 
1509 void
1510 px_dvma_alloc_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1511 	ddi_dma_impl_t *mp)
1512 {
1513 	struct px_dvma_rec *ptr;
1514 	mutex_enter(&mmu_p->dvma_debug_lock);
1515 
1516 	if (!mmu_p->dvma_alloc_rec)
1517 		px_dvma_debug_init(mmu_p);
1518 	if (PX_DVMA_DBG_OFF(mmu_p)) {
1519 		px_dvma_debug_fini(mmu_p);
1520 		goto done;
1521 	}
1522 
1523 	ptr = &mmu_p->dvma_alloc_rec[mmu_p->dvma_alloc_rec_index];
1524 	ptr->dvma_addr = address;
1525 	ptr->len = len;
1526 	ptr->mp = mp;
1527 	if (++mmu_p->dvma_alloc_rec_index == px_dvma_debug_rec)
1528 		mmu_p->dvma_alloc_rec_index = 0;
1529 
1530 	ptr = kmem_alloc(sizeof (struct px_dvma_rec), KM_SLEEP);
1531 	ptr->dvma_addr = address;
1532 	ptr->len = len;
1533 	ptr->mp = mp;
1534 
1535 	ptr->next = mmu_p->dvma_active_list;
1536 	mmu_p->dvma_active_list = ptr;
1537 	mmu_p->dvma_active_count++;
1538 done:
1539 	mutex_exit(&mmu_p->dvma_debug_lock);
1540 }
1541 
1542 void
1543 px_dvma_free_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1544     ddi_dma_impl_t *mp)
1545 {
1546 	struct px_dvma_rec *ptr, *ptr_save;
1547 	mutex_enter(&mmu_p->dvma_debug_lock);
1548 
1549 	if (!mmu_p->dvma_alloc_rec)
1550 		px_dvma_debug_init(mmu_p);
1551 	if (PX_DVMA_DBG_OFF(mmu_p)) {
1552 		px_dvma_debug_fini(mmu_p);
1553 		goto done;
1554 	}
1555 
1556 	ptr = &mmu_p->dvma_free_rec[mmu_p->dvma_free_rec_index];
1557 	ptr->dvma_addr = address;
1558 	ptr->len = len;
1559 	ptr->mp = mp;
1560 	if (++mmu_p->dvma_free_rec_index == px_dvma_debug_rec)
1561 		mmu_p->dvma_free_rec_index = 0;
1562 
1563 	ptr_save = mmu_p->dvma_active_list;
1564 	for (ptr = ptr_save; ptr; ptr = ptr->next) {
1565 		if ((ptr->dvma_addr == address) && (ptr->len = len))
1566 			break;
1567 		ptr_save = ptr;
1568 	}
1569 	if (!ptr) {
1570 		cmn_err(CE_WARN, "bad dvma free addr=%lx len=%x",
1571 			(long)address, len);
1572 		goto done;
1573 	}
1574 	if (ptr == mmu_p->dvma_active_list)
1575 		mmu_p->dvma_active_list = ptr->next;
1576 	else
1577 		ptr_save->next = ptr->next;
1578 	kmem_free(ptr, sizeof (struct px_dvma_rec));
1579 	mmu_p->dvma_active_count--;
1580 done:
1581 	mutex_exit(&mmu_p->dvma_debug_lock);
1582 }
1583 
1584 #ifdef	DEBUG
1585 void
1586 px_dump_dma_handle(uint64_t flag, dev_info_t *dip, ddi_dma_impl_t *hp)
1587 {
1588 	DBG(flag, dip, "mp(%p): flags=%x mapping=%lx xfer_size=%x\n",
1589 		hp, hp->dmai_inuse, hp->dmai_mapping, hp->dmai_size);
1590 	DBG(flag|DBG_CONT, dip, "\tnpages=%x roffset=%x rflags=%x nwin=%x\n",
1591 		hp->dmai_ndvmapages, hp->dmai_roffset, hp->dmai_rflags,
1592 		hp->dmai_nwin);
1593 	DBG(flag|DBG_CONT, dip, "\twinsize=%x tte=%p pfnlst=%p pfn0=%p\n",
1594 		hp->dmai_winsize, hp->dmai_tte, hp->dmai_pfnlst, hp->dmai_pfn0);
1595 	DBG(flag|DBG_CONT, dip, "\twinlst=%x obj=%p attr=%p ckp=%p\n",
1596 		hp->dmai_winlst, &hp->dmai_object, &hp->dmai_attr,
1597 		hp->dmai_cookie);
1598 }
1599 #endif	/* DEBUG */
1600