xref: /titanic_41/usr/src/uts/sun4/io/px/px_dma.c (revision 6fec3791b5a9a5621db93bfef3a6514bc0511b5d)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * PCI Express nexus DVMA and DMA core routines:
31  *	dma_map/dma_bind_handle implementation
32  *	bypass and peer-to-peer support
33  *	fast track DVMA space allocation
34  *	runtime DVMA debug
35  */
36 #include <sys/types.h>
37 #include <sys/kmem.h>
38 #include <sys/async.h>
39 #include <sys/sysmacros.h>
40 #include <sys/sunddi.h>
41 #include <sys/ddi_impldefs.h>
42 #include "px_obj.h"
43 
44 /*LINTLIBRARY*/
45 
46 /*
47  * px_dma_allocmp - Allocate a pci dma implementation structure
48  *
49  * An extra ddi_dma_attr structure is bundled with the usual ddi_dma_impl
50  * to hold unmodified device limits. The ddi_dma_attr inside the
51  * ddi_dma_impl structure is augumented with system limits to enhance
52  * DVMA performance at runtime. The unaugumented device limits saved
53  * right after (accessed through (ddi_dma_attr_t *)(mp + 1)) is used
54  * strictly for peer-to-peer transfers which do not obey system limits.
55  *
56  * return: DDI_SUCCESS DDI_DMA_NORESOURCES
57  */
58 ddi_dma_impl_t *
59 px_dma_allocmp(dev_info_t *dip, dev_info_t *rdip, int (*waitfp)(caddr_t),
60 	caddr_t arg)
61 {
62 	register ddi_dma_impl_t *mp;
63 	int sleep = (waitfp == DDI_DMA_SLEEP) ? KM_SLEEP : KM_NOSLEEP;
64 
65 	/* Caution: we don't use zalloc to enhance performance! */
66 	if ((mp = kmem_alloc(sizeof (px_dma_hdl_t), sleep)) == 0) {
67 		DBG(DBG_DMA_MAP, dip, "can't alloc dma_handle\n");
68 		if (waitfp != DDI_DMA_DONTWAIT) {
69 			DBG(DBG_DMA_MAP, dip, "alloc_mp kmem cb\n");
70 			ddi_set_callback(waitfp, arg, &px_kmem_clid);
71 		}
72 		return (mp);
73 	}
74 
75 	mp->dmai_rdip = rdip;
76 	mp->dmai_flags = 0;
77 	mp->dmai_pfnlst = NULL;
78 	mp->dmai_winlst = NULL;
79 
80 	/*
81 	 * kmem_alloc debug: the following fields are not zero-ed
82 	 * mp->dmai_mapping = 0;
83 	 * mp->dmai_size = 0;
84 	 * mp->dmai_offset = 0;
85 	 * mp->dmai_minxfer = 0;
86 	 * mp->dmai_burstsizes = 0;
87 	 * mp->dmai_ndvmapages = 0;
88 	 * mp->dmai_pool/roffset = 0;
89 	 * mp->dmai_rflags = 0;
90 	 * mp->dmai_inuse/flags
91 	 * mp->dmai_nwin = 0;
92 	 * mp->dmai_winsize = 0;
93 	 * mp->dmai_nexus_private/tte = 0;
94 	 * mp->dmai_iopte/pfnlst
95 	 * mp->dmai_sbi/pfn0 = 0;
96 	 * mp->dmai_minfo/winlst/fdvma
97 	 * mp->dmai_rdip
98 	 * bzero(&mp->dmai_object, sizeof (ddi_dma_obj_t));
99 	 * bzero(&mp->dmai_attr, sizeof (ddi_dma_attr_t));
100 	 * mp->dmai_cookie = 0;
101 	 */
102 
103 	mp->dmai_attr.dma_attr_version = (uint_t)DMA_ATTR_VERSION;
104 	mp->dmai_attr.dma_attr_flags = (uint_t)0;
105 	mp->dmai_fault = 0;
106 	mp->dmai_fault_check = NULL;
107 	mp->dmai_fault_notify = NULL;
108 
109 	mp->dmai_error.err_ena = 0;
110 	mp->dmai_error.err_status = DDI_FM_OK;
111 	mp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
112 	mp->dmai_error.err_ontrap = NULL;
113 	mp->dmai_error.err_fep = 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 = 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 |= DMAI_FLAGS_PEER_ONLY;
194 	count_max = MIN(count_max, hi - lo);
195 
196 	if (DEV_NOSYSLIMIT(lo, hi, syslo, fasthi, 1))
197 		mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT |
198 			DMAI_FLAGS_NOSYSLIMIT;
199 	else {
200 		if (DEV_NOFASTLIMIT(lo, hi, syslo, syshi, 1))
201 			mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT;
202 	}
203 	if (PX_DMA_NOCTX(rdip))
204 		mp->dmai_flags |= 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_soft_state |= 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		= 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_soft_state & PX_BYPASS_DMA_ALLOWED))
277 			return (DDI_DMA_BADATTR);
278 		mp->dmai_flags |= 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(attrp, &syslo, &syshi);
287 		if (ret != DDI_SUCCESS)
288 			return (ret);
289 	} else { /* MMU_XLATE or PEER_TO_PEER */
290 		align = MAX(align, MMU_PAGE_SIZE) - 1;
291 		if ((align & nocross) != align) {
292 			dev_info_t *rdip = mp->dmai_rdip;
293 			cmn_err(CE_WARN, "%s%d dma_attr_seg not aligned",
294 				NAMEINST(rdip));
295 			return (DDI_DMA_BADATTR);
296 		}
297 		align = MMU_BTOP(align + 1);
298 		syslo = mmu_p->mmu_dvma_base;
299 		syshi = mmu_p->mmu_dvma_end;
300 	}
301 	if (hi <= lo) {
302 		dev_info_t *rdip = mp->dmai_rdip;
303 		cmn_err(CE_WARN, "%s%d limits out of range", NAMEINST(rdip));
304 		return (DDI_DMA_BADATTR);
305 	}
306 	lo = MAX(lo, syslo);
307 	hi = MIN(hi, syshi);
308 	if (!count_max)
309 		count_max--;
310 
311 	DBG(DBG_DMA_ALLOCH, px_p->px_dip, "hi=%x.%08x, lo=%x.%08x\n",
312 		HI32(hi), LO32(hi), HI32(lo), LO32(lo));
313 	if (hi <= lo) { /* peer transfers cannot have alignment & nocross */
314 		dev_info_t *rdip = mp->dmai_rdip;
315 		cmn_err(CE_WARN, "%s%d peer only dev %p", NAMEINST(rdip), mp);
316 		if ((nocross < UINT32_MAX) || (align > 1)) {
317 			cmn_err(CE_WARN, "%s%d peer only device bad attr",
318 				NAMEINST(rdip));
319 			return (DDI_DMA_BADATTR);
320 		}
321 		mp->dmai_flags |= DMAI_FLAGS_PEER_ONLY;
322 	} else /* set practical counter_max value */
323 		count_max = MIN(count_max, hi - lo);
324 
325 	if (DEV_NOSYSLIMIT(lo, hi, syslo, syshi, align))
326 		mp->dmai_flags |= DMAI_FLAGS_NOSYSLIMIT |
327 			DMAI_FLAGS_NOFASTLIMIT;
328 	else {
329 		syshi = mmu_p->mmu_dvma_fast_end;
330 		if (DEV_NOFASTLIMIT(lo, hi, syslo, syshi, align))
331 			mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT;
332 	}
333 	if (PX_DMA_NOCTX(mp->dmai_rdip))
334 		mp->dmai_flags |= DMAI_FLAGS_NOCTX;
335 
336 	mp->dmai_minxfer	= attrp->dma_attr_minxfer;
337 	mp->dmai_burstsizes	= attrp->dma_attr_burstsizes;
338 	attrp = &mp->dmai_attr;
339 	SET_DMAATTR(attrp, lo, hi, nocross, count_max);
340 	return (DDI_SUCCESS);
341 }
342 
343 #define	TGT_PFN_INBETWEEN(pfn, bgn, end) ((pfn >= bgn) && (pfn <= end))
344 
345 /*
346  * px_dma_type - determine which of the three types DMA (peer-to-peer,
347  *		mmu bypass, or mmu translate) we are asked to do.
348  *		Also checks pfn0 and rejects any non-peer-to-peer
349  *		requests for peer-only devices.
350  *
351  *	return values:
352  *		DDI_DMA_NOMAPPING - can't get valid pfn0, or bad dma type
353  *		DDI_SUCCESS
354  *
355  *	dma handle members affected (set on exit):
356  *	mp->dmai_object		- dmareq->dmar_object
357  *	mp->dmai_rflags		- consistent?, nosync?, dmareq->dmar_flags
358  *	mp->dmai_flags   	- DMA type
359  *	mp->dmai_pfn0   	- 1st page pfn (if va/size pair and not shadow)
360  *	mp->dmai_roffset 	- initialized to starting MMU page offset
361  *	mp->dmai_ndvmapages	- # of total MMU pages of entire object
362  */
363 int
364 px_dma_type(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
365 {
366 	dev_info_t *dip = px_p->px_dip;
367 	ddi_dma_obj_t *dobj_p = &dmareq->dmar_object;
368 	px_pec_t *pec_p = px_p->px_pec_p;
369 	uint32_t offset;
370 	pfn_t pfn0;
371 
372 	mp->dmai_rflags = dmareq->dmar_flags & DMP_DDIFLAGS | DMP_NOSYNC;
373 
374 	switch (dobj_p->dmao_type) {
375 	case DMA_OTYP_BUFVADDR:
376 	case DMA_OTYP_VADDR: {
377 		page_t **pplist = dobj_p->dmao_obj.virt_obj.v_priv;
378 		caddr_t vaddr = dobj_p->dmao_obj.virt_obj.v_addr;
379 
380 		DBG(DBG_DMA_MAP, dip, "vaddr=%p pplist=%p\n", vaddr, pplist);
381 		offset = (ulong_t)vaddr & MMU_PAGE_OFFSET;
382 		if (pplist) {				/* shadow list */
383 			mp->dmai_flags |= DMAI_FLAGS_PGPFN;
384 			pfn0 = page_pptonum(*pplist);
385 		} else {
386 			struct as *as_p = dobj_p->dmao_obj.virt_obj.v_as;
387 			struct hat *hat_p = as_p ? as_p->a_hat : kas.a_hat;
388 			pfn0 = hat_getpfnum(hat_p, vaddr);
389 		}
390 		}
391 		break;
392 
393 	case DMA_OTYP_PAGES:
394 		offset = dobj_p->dmao_obj.pp_obj.pp_offset;
395 		mp->dmai_flags |= DMAI_FLAGS_PGPFN;
396 		pfn0 = page_pptonum(dobj_p->dmao_obj.pp_obj.pp_pp);
397 		break;
398 
399 	case DMA_OTYP_PADDR:
400 	default:
401 		cmn_err(CE_WARN, "%s%d requested unsupported dma type %x",
402 			NAMEINST(mp->dmai_rdip), dobj_p->dmao_type);
403 		return (DDI_DMA_NOMAPPING);
404 	}
405 	if (pfn0 == PFN_INVALID) {
406 		cmn_err(CE_WARN, "%s%d: invalid pfn0 for DMA object %p",
407 			NAMEINST(dip), dobj_p);
408 		return (DDI_DMA_NOMAPPING);
409 	}
410 	if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base32_pfn,
411 			pec_p->pec_last32_pfn)) {
412 		mp->dmai_flags |= DMAI_FLAGS_PTP|DMAI_FLAGS_PTP32;
413 		goto done;	/* leave bypass and dvma flag as 0 */
414 	} else if (TGT_PFN_INBETWEEN(pfn0, pec_p->pec_base64_pfn,
415 			pec_p->pec_last64_pfn)) {
416 		mp->dmai_flags |= DMAI_FLAGS_PTP|DMAI_FLAGS_PTP64;
417 		goto done;	/* leave bypass and dvma flag as 0 */
418 	}
419 	if (PX_DMA_ISPEERONLY(mp)) {
420 		dev_info_t *rdip = mp->dmai_rdip;
421 		cmn_err(CE_WARN, "Bad peer-to-peer req %s%d", NAMEINST(rdip));
422 		return (DDI_DMA_NOMAPPING);
423 	}
424 	mp->dmai_flags |= (mp->dmai_flags & DMAI_FLAGS_BYPASSREQ) ?
425 		DMAI_FLAGS_BYPASS : DMAI_FLAGS_DVMA;
426 done:
427 	mp->dmai_object	 = *dobj_p;			/* whole object    */
428 	mp->dmai_pfn0	 = (void *)pfn0;		/* cache pfn0	   */
429 	mp->dmai_roffset = offset;			/* win0 pg0 offset */
430 	mp->dmai_ndvmapages = MMU_BTOPR(offset + mp->dmai_object.dmao_size);
431 	return (DDI_SUCCESS);
432 }
433 
434 /*
435  * px_dma_pgpfn - set up pfnlst array according to pages
436  *	VA/size pair: <shadow IO, bypass, peer-to-peer>, or OTYP_PAGES
437  */
438 /*ARGSUSED*/
439 static int
440 px_dma_pgpfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
441 {
442 	int i;
443 	dev_info_t *dip = px_p->px_dip;
444 
445 	switch (mp->dmai_object.dmao_type) {
446 	case DMA_OTYP_BUFVADDR:
447 	case DMA_OTYP_VADDR: {
448 		page_t **pplist = mp->dmai_object.dmao_obj.virt_obj.v_priv;
449 		DBG(DBG_DMA_MAP, dip, "shadow pplist=%p, %x pages, pfns=",
450 			pplist, npages);
451 		for (i = 1; i < npages; i++) {
452 			px_iopfn_t pfn = page_pptonum(pplist[i]);
453 			PX_SET_MP_PFN1(mp, i, pfn);
454 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
455 		}
456 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
457 		}
458 		break;
459 
460 	case DMA_OTYP_PAGES: {
461 		page_t *pp = mp->dmai_object.dmao_obj.pp_obj.pp_pp->p_next;
462 		DBG(DBG_DMA_MAP, dip, "pp=%p pfns=", pp);
463 		for (i = 1; i < npages; i++, pp = pp->p_next) {
464 			px_iopfn_t pfn = page_pptonum(pp);
465 			PX_SET_MP_PFN1(mp, i, pfn);
466 			DBG(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
467 		}
468 		DBG(DBG_DMA_MAP|DBG_CONT, dip, "\n");
469 		}
470 		break;
471 
472 	default:	/* check is already done by px_dma_type */
473 		ASSERT(0);
474 		break;
475 	}
476 	return (DDI_SUCCESS);
477 }
478 
479 /*
480  * px_dma_vapfn - set up pfnlst array according to VA
481  *	VA/size pair: <normal, bypass, peer-to-peer>
482  *	pfn0 is skipped as it is already done.
483  *	In this case, the cached pfn0 is used to fill pfnlst[0]
484  */
485 static int
486 px_dma_vapfn(px_t *px_p, ddi_dma_impl_t *mp, uint_t npages)
487 {
488 	dev_info_t *dip = px_p->px_dip;
489 	int i;
490 	caddr_t vaddr = (caddr_t)mp->dmai_object.dmao_obj.virt_obj.v_as;
491 	struct hat *hat_p = vaddr ? ((struct as *)vaddr)->a_hat : kas.a_hat;
492 
493 	vaddr = mp->dmai_object.dmao_obj.virt_obj.v_addr + MMU_PAGE_SIZE;
494 	for (i = 1; i < npages; i++, vaddr += MMU_PAGE_SIZE) {
495 		px_iopfn_t pfn = hat_getpfnum(hat_p, vaddr);
496 		if (pfn == PFN_INVALID)
497 			goto err_badpfn;
498 		PX_SET_MP_PFN1(mp, i, pfn);
499 		DBG(DBG_DMA_BINDH, dip, "px_dma_vapfn: mp=%p pfnlst[%x]=%x\n",
500 			mp, i, pfn);
501 	}
502 	return (DDI_SUCCESS);
503 err_badpfn:
504 	cmn_err(CE_WARN, "%s%d: bad page frame vaddr=%p", NAMEINST(dip), vaddr);
505 	return (DDI_DMA_NOMAPPING);
506 }
507 
508 /*
509  * px_dma_pfn - Fills pfn list for all pages being DMA-ed.
510  *
511  * dependencies:
512  *	mp->dmai_ndvmapages	- set to total # of dma pages
513  *
514  * return value:
515  *	DDI_SUCCESS
516  *	DDI_DMA_NOMAPPING
517  */
518 int
519 px_dma_pfn(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
520 {
521 	uint32_t npages = mp->dmai_ndvmapages;
522 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
523 	int i, ret, peer = PX_DMA_ISPTP(mp);
524 	int peer32 = PX_DMA_ISPTP32(mp);
525 	dev_info_t *dip = px_p->px_dip;
526 
527 	px_pec_t *pec_p = px_p->px_pec_p;
528 	px_iopfn_t pfn_base = peer32 ? pec_p->pec_base32_pfn :
529 					pec_p->pec_base64_pfn;
530 	px_iopfn_t pfn_last = peer32 ? pec_p->pec_last32_pfn :
531 					pec_p->pec_last64_pfn;
532 	px_iopfn_t pfn_adj = peer ? pfn_base : 0;
533 
534 	DBG(DBG_DMA_BINDH, dip, "px_dma_pfn: mp=%p pfn0=%x\n",
535 		mp, MP_PFN0(mp) - pfn_adj);
536 	/* 1 page: no array alloc/fill, no mixed mode check */
537 	if (npages == 1) {
538 		PX_SET_MP_PFN(mp, 0, MP_PFN0(mp) - pfn_adj);
539 		return (DDI_SUCCESS);
540 	}
541 	/* allocate pfn array */
542 	if (!(mp->dmai_pfnlst = kmem_alloc(npages * sizeof (px_iopfn_t),
543 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP))) {
544 		if (waitfp != DDI_DMA_DONTWAIT)
545 			ddi_set_callback(waitfp, dmareq->dmar_arg,
546 				&px_kmem_clid);
547 		return (DDI_DMA_NORESOURCES);
548 	}
549 	/* fill pfn array */
550 	PX_SET_MP_PFN(mp, 0, MP_PFN0(mp) - pfn_adj);	/* pfnlst[0] */
551 	if ((ret = PX_DMA_ISPGPFN(mp) ? px_dma_pgpfn(px_p, mp, npages) :
552 		px_dma_vapfn(px_p, mp, npages)) != DDI_SUCCESS)
553 		goto err;
554 
555 	/* skip pfn0, check mixed mode and adjust peer to peer pfn */
556 	for (i = 1; i < npages; i++) {
557 		px_iopfn_t pfn = PX_GET_MP_PFN1(mp, i);
558 		if (peer ^ TGT_PFN_INBETWEEN(pfn, pfn_base, pfn_last)) {
559 			cmn_err(CE_WARN, "%s%d mixed mode DMA %lx %lx",
560 				NAMEINST(mp->dmai_rdip), MP_PFN0(mp), pfn);
561 			ret = DDI_DMA_NOMAPPING;	/* mixed mode */
562 			goto err;
563 		}
564 		DBG(DBG_DMA_MAP, dip,
565 			"px_dma_pfn: pfnlst[%x]=%x-%x\n", i, pfn, pfn_adj);
566 		if (pfn_adj)
567 			PX_SET_MP_PFN1(mp, i, pfn - pfn_adj);
568 	}
569 	return (DDI_SUCCESS);
570 err:
571 	px_dma_freepfn(mp);
572 	return (ret);
573 }
574 
575 /*
576  * px_dvma_win() - trim requested DVMA size down to window size
577  *	The 1st window starts from offset and ends at page-aligned boundary.
578  *	From the 2nd window on, each window starts and ends at page-aligned
579  *	boundary except the last window ends at wherever requested.
580  *
581  *	accesses the following mp-> members:
582  *	mp->dmai_attr.dma_attr_count_max
583  *	mp->dmai_attr.dma_attr_seg
584  *	mp->dmai_roffset   - start offset of 1st window
585  *	mp->dmai_rflags (redzone)
586  *	mp->dmai_ndvmapages (for 1 page fast path)
587  *
588  *	sets the following mp-> members:
589  *	mp->dmai_size	   - xfer size, != winsize if 1st/last win  (not fixed)
590  *	mp->dmai_winsize   - window size (no redzone), n * page size    (fixed)
591  *	mp->dmai_nwin	   - # of DMA windows of entire object		(fixed)
592  *	mp->dmai_rflags	   - remove partial flag if nwin == 1		(fixed)
593  *	mp->dmai_winlst	   - NULL, window objects not used for DVMA	(fixed)
594  *
595  *	fixed - not changed across different DMA windows
596  */
597 /*ARGSUSED*/
598 int
599 px_dvma_win(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
600 {
601 	uint32_t redzone_sz	= HAS_REDZONE(mp) ? MMU_PAGE_SIZE : 0;
602 	size_t obj_sz		= mp->dmai_object.dmao_size;
603 	size_t xfer_sz;
604 	ulong_t pg_off;
605 
606 	if ((mp->dmai_ndvmapages == 1) && !redzone_sz) {
607 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
608 		mp->dmai_size = obj_sz;
609 		mp->dmai_winsize = MMU_PAGE_SIZE;
610 		mp->dmai_nwin = 1;
611 		goto done;
612 	}
613 
614 	pg_off	= mp->dmai_roffset;
615 	xfer_sz	= obj_sz + redzone_sz;
616 
617 	/* include redzone in nocross check */ {
618 		uint64_t nocross = mp->dmai_attr.dma_attr_seg;
619 		if (xfer_sz + pg_off - 1 > nocross)
620 			xfer_sz = nocross - pg_off + 1;
621 		if (redzone_sz && (xfer_sz <= redzone_sz)) {
622 			DBG(DBG_DMA_MAP, px_p->px_dip,
623 			    "nocross too small: "
624 			    "%lx(%lx)+%lx+%lx < %llx\n",
625 			    xfer_sz, obj_sz, pg_off, redzone_sz, nocross);
626 			return (DDI_DMA_TOOBIG);
627 		}
628 	}
629 	xfer_sz -= redzone_sz;		/* restore transfer size  */
630 	/* check counter max */ {
631 		uint32_t count_max = mp->dmai_attr.dma_attr_count_max;
632 		if (xfer_sz - 1 > count_max)
633 			xfer_sz = count_max + 1;
634 	}
635 	if (xfer_sz >= obj_sz) {
636 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
637 		mp->dmai_size = xfer_sz;
638 		mp->dmai_winsize = P2ROUNDUP(xfer_sz + pg_off, MMU_PAGE_SIZE);
639 		mp->dmai_nwin = 1;
640 		goto done;
641 	}
642 	if (!(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
643 		DBG(DBG_DMA_MAP, px_p->px_dip, "too big: %lx+%lx+%lx > %lx\n",
644 			obj_sz, pg_off, redzone_sz, xfer_sz);
645 		return (DDI_DMA_TOOBIG);
646 	}
647 
648 	xfer_sz = MMU_PTOB(MMU_BTOP(xfer_sz + pg_off)); /* page align */
649 	mp->dmai_size = xfer_sz - pg_off;	/* 1st window xferrable size */
650 	mp->dmai_winsize = xfer_sz;		/* redzone not in winsize */
651 	mp->dmai_nwin = (obj_sz + pg_off + xfer_sz - 1) / xfer_sz;
652 done:
653 	mp->dmai_winlst = NULL;
654 	px_dump_dma_handle(DBG_DMA_MAP, px_p->px_dip, mp);
655 	return (DDI_SUCCESS);
656 }
657 
658 /*
659  * fast track cache entry to mmu context, inserts 3 0 bits between
660  * upper 6-bits and lower 3-bits of the 9-bit cache entry
661  */
662 #define	MMU_FCE_TO_CTX(i)	(((i) << 3) | ((i) & 0x7) | 0x38)
663 
664 /*
665  * px_dvma_map_fast - attempts to map fast trackable DVMA
666  */
667 /*ARGSUSED*/
668 int
669 px_dvma_map_fast(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
670 {
671 	uint_t clustsz = px_dvma_page_cache_clustsz;
672 	uint_t entries = px_dvma_page_cache_entries;
673 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags);
674 	int i = mmu_p->mmu_dvma_addr_scan_start;
675 	uint8_t *lock_addr = mmu_p->mmu_dvma_cache_locks + i;
676 	px_dvma_addr_t dvma_pg;
677 	size_t npages = MMU_BTOP(mp->dmai_winsize);
678 	dev_info_t *px_dip = mmu_p->mmu_px_p->px_dip;
679 
680 	extern uint8_t ldstub(uint8_t *);
681 	ASSERT(MMU_PTOB(npages) == mp->dmai_winsize);
682 	ASSERT(npages + HAS_REDZONE(mp) <= clustsz);
683 
684 	for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
685 	if (i >= entries) {
686 		lock_addr = mmu_p->mmu_dvma_cache_locks;
687 		i = 0;
688 		for (; i < entries && ldstub(lock_addr); i++, lock_addr++);
689 		if (i >= entries) {
690 #ifdef	PX_DMA_PROF
691 			px_dvmaft_exhaust++;
692 #endif	/* PX_DMA_PROF */
693 			return (DDI_DMA_NORESOURCES);
694 		}
695 	}
696 	mmu_p->mmu_dvma_addr_scan_start = (i + 1) & (entries - 1);
697 
698 	i *= clustsz;
699 	dvma_pg = mmu_p->dvma_base_pg + i;
700 
701 	if (px_lib_iommu_map(px_dip, PCI_TSBID(0, i), npages, attr,
702 	    (void *)mp, 0, MMU_MAP_MP) != DDI_SUCCESS)
703 		return (DDI_FAILURE);
704 
705 #ifdef PX_DMA_PROF
706 	px_dvmaft_success++;
707 #endif
708 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
709 	mp->dmai_offset = 0;
710 	mp->dmai_flags |= DMAI_FLAGS_FASTTRACK;
711 	PX_SAVE_MP_TTE(mp, attr);	/* save TTE template for unmapping */
712 	if (DVMA_DBG_ON(mmu_p))
713 		px_dvma_alloc_debug(mmu_p, (char *)mp->dmai_mapping,
714 			mp->dmai_size, mp);
715 	return (DDI_SUCCESS);
716 }
717 
718 /*
719  * px_dvma_map: map non-fasttrack DMA
720  *		Use quantum cache if single page DMA.
721  */
722 int
723 px_dvma_map(ddi_dma_impl_t *mp, ddi_dma_req_t *dmareq, px_mmu_t *mmu_p)
724 {
725 	uint_t npages = PX_DMA_WINNPGS(mp);
726 	px_dvma_addr_t dvma_pg, dvma_pg_index;
727 	void *dvma_addr;
728 	uint64_t tte = PX_GET_TTE_ATTR(mp->dmai_rflags);
729 	int sleep = dmareq->dmar_fp == DDI_DMA_SLEEP ? VM_SLEEP : VM_NOSLEEP;
730 	dev_info_t *dip = mp->dmai_rdip;
731 	int	ret = DDI_SUCCESS;
732 
733 	/*
734 	 * allocate dvma space resource and map in the first window.
735 	 * (vmem_t *vmp, size_t size,
736 	 *	size_t align, size_t phase, size_t nocross,
737 	 *	void *minaddr, void *maxaddr, int vmflag)
738 	 */
739 	if ((npages == 1) && HAS_NOSYSLIMIT(mp)) {
740 		dvma_addr = vmem_alloc(mmu_p->mmu_dvma_map,
741 			MMU_PAGE_SIZE, sleep);
742 		mp->dmai_flags |= DMAI_FLAGS_VMEMCACHE;
743 #ifdef	PX_DMA_PROF
744 		px_dvma_vmem_alloc++;
745 #endif	/* PX_DMA_PROF */
746 	} else {
747 		dvma_addr = vmem_xalloc(mmu_p->mmu_dvma_map,
748 			MMU_PTOB(npages + HAS_REDZONE(mp)),
749 			MAX(mp->dmai_attr.dma_attr_align, MMU_PAGE_SIZE),
750 			0,
751 			mp->dmai_attr.dma_attr_seg + 1,
752 			(void *)mp->dmai_attr.dma_attr_addr_lo,
753 			(void *)(mp->dmai_attr.dma_attr_addr_hi + 1),
754 			sleep);
755 #ifdef	PX_DMA_PROF
756 		px_dvma_vmem_xalloc++;
757 #endif	/* PX_DMA_PROF */
758 	}
759 	dvma_pg = MMU_BTOP((ulong_t)dvma_addr);
760 	dvma_pg_index = dvma_pg - mmu_p->dvma_base_pg;
761 	DBG(DBG_DMA_MAP, dip, "fallback dvma_pages: dvma_pg=%x index=%x\n",
762 		dvma_pg, dvma_pg_index);
763 	if (dvma_pg == 0)
764 		goto noresource;
765 
766 	mp->dmai_mapping = mp->dmai_roffset | MMU_PTOB(dvma_pg);
767 	mp->dmai_offset = 0;
768 	PX_SAVE_MP_TTE(mp, tte);	/* mp->dmai_tte = tte */
769 
770 	if ((ret = px_mmu_map_pages(mmu_p,
771 	    mp, dvma_pg, npages, 0)) != DDI_SUCCESS) {
772 		if (mp->dmai_flags & DMAI_FLAGS_VMEMCACHE) {
773 			vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
774 			    MMU_PAGE_SIZE);
775 #ifdef PX_DMA_PROF
776 			px_dvma_vmem_free++;
777 #endif /* PX_DMA_PROF */
778 		} else {
779 			vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
780 			    MMU_PTOB(npages + HAS_REDZONE(mp)));
781 #ifdef PX_DMA_PROF
782 			px_dvma_vmem_xfree++;
783 #endif /* PX_DMA_PROF */
784 		}
785 	}
786 
787 	return (ret);
788 noresource:
789 	if (dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
790 		DBG(DBG_DMA_MAP, dip, "dvma_pg 0 - set callback\n");
791 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
792 			&mmu_p->mmu_dvma_clid);
793 	}
794 	DBG(DBG_DMA_MAP, dip, "vmem_xalloc - DDI_DMA_NORESOURCES\n");
795 	return (DDI_DMA_NORESOURCES);
796 }
797 
798 void
799 px_dvma_unmap(px_mmu_t *mmu_p, ddi_dma_impl_t *mp)
800 {
801 	px_dvma_addr_t dvma_addr = (px_dvma_addr_t)mp->dmai_mapping;
802 	px_dvma_addr_t dvma_pg = MMU_BTOP(dvma_addr);
803 	dvma_addr = MMU_PTOB(dvma_pg);
804 
805 	if (mp->dmai_flags & DMAI_FLAGS_FASTTRACK) {
806 		px_iopfn_t index = dvma_pg - mmu_p->dvma_base_pg;
807 		ASSERT(index % px_dvma_page_cache_clustsz == 0);
808 		index /= px_dvma_page_cache_clustsz;
809 		ASSERT(index < px_dvma_page_cache_entries);
810 		mmu_p->mmu_dvma_cache_locks[index] = 0;
811 #ifdef	PX_DMA_PROF
812 		px_dvmaft_free++;
813 #endif	/* PX_DMA_PROF */
814 		return;
815 	}
816 
817 	if (mp->dmai_flags & DMAI_FLAGS_VMEMCACHE) {
818 		vmem_free(mmu_p->mmu_dvma_map, (void *)dvma_addr,
819 			MMU_PAGE_SIZE);
820 #ifdef PX_DMA_PROF
821 		px_dvma_vmem_free++;
822 #endif /* PX_DMA_PROF */
823 	} else {
824 		size_t npages = MMU_BTOP(mp->dmai_winsize) + HAS_REDZONE(mp);
825 		vmem_xfree(mmu_p->mmu_dvma_map, (void *)dvma_addr,
826 			MMU_PTOB(npages));
827 #ifdef PX_DMA_PROF
828 		px_dvma_vmem_xfree++;
829 #endif /* PX_DMA_PROF */
830 	}
831 }
832 
833 /*
834  * DVMA mappings may have multiple windows, but each window always have
835  * one segment.
836  */
837 int
838 px_dvma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
839 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
840 	uint_t cache_flags)
841 {
842 	switch (cmd) {
843 	case DDI_DMA_SYNC:
844 		return (px_lib_dma_sync(dip, rdip, (ddi_dma_handle_t)mp,
845 		    *offp, *lenp, cache_flags));
846 
847 	case DDI_DMA_HTOC: {
848 		int ret;
849 		off_t wo_off, off = *offp;	/* wo_off: wnd's obj offset */
850 		uint_t win_size = mp->dmai_winsize;
851 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)objp;
852 
853 		if (off >= mp->dmai_object.dmao_size) {
854 			cmn_err(CE_WARN, "%s%d invalid dma_htoc offset %lx",
855 				NAMEINST(mp->dmai_rdip), off);
856 			return (DDI_FAILURE);
857 		}
858 		off += mp->dmai_roffset;
859 		ret = px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
860 		    off / win_size, &wo_off, NULL, cp, NULL); /* lenp == NULL */
861 		if (ret)
862 			return (ret);
863 		DBG(DBG_DMA_CTL, dip, "HTOC:cookie=%x+%lx off=%lx,%lx\n",
864 			cp->dmac_address, cp->dmac_size, off, *offp);
865 
866 		/* adjust cookie addr/len if we are not on window boundary */
867 		ASSERT((off % win_size) == (off -
868 			(PX_DMA_CURWIN(mp) ? mp->dmai_roffset : 0) - wo_off));
869 		off = PX_DMA_CURWIN(mp) ? off % win_size : *offp;
870 		ASSERT(cp->dmac_size > off);
871 		cp->dmac_laddress += off;
872 		cp->dmac_size -= off;
873 		DBG(DBG_DMA_CTL, dip, "HTOC:mp=%p cookie=%x+%lx off=%lx,%lx\n",
874 			mp, cp->dmac_address, cp->dmac_size, off, wo_off);
875 		}
876 		return (DDI_SUCCESS);
877 
878 	case DDI_DMA_REPWIN:
879 		*offp = mp->dmai_offset;
880 		*lenp = mp->dmai_size;
881 		return (DDI_SUCCESS);
882 
883 	case DDI_DMA_MOVWIN: {
884 		off_t off = *offp;
885 		if (off >= mp->dmai_object.dmao_size)
886 			return (DDI_FAILURE);
887 		off += mp->dmai_roffset;
888 		return (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
889 		    off / mp->dmai_winsize, offp, lenp,
890 		    (ddi_dma_cookie_t *)objp, NULL));
891 		}
892 
893 	case DDI_DMA_NEXTWIN: {
894 		px_window_t win = PX_DMA_CURWIN(mp);
895 		if (offp) {
896 			if (*(px_window_t *)offp != win) {
897 				/* window not active */
898 				*(px_window_t *)objp = win; /* return cur win */
899 				return (DDI_DMA_STALE);
900 			}
901 			win++;
902 		} else	/* map win 0 */
903 			win = 0;
904 		if (win >= mp->dmai_nwin) {
905 			*(px_window_t *)objp = win - 1;
906 			return (DDI_DMA_DONE);
907 		}
908 		if (px_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
909 		    win, 0, 0, 0, 0)) {
910 			*(px_window_t *)objp = win - 1;
911 			return (DDI_FAILURE);
912 		}
913 		*(px_window_t *)objp = win;
914 		}
915 		return (DDI_SUCCESS);
916 
917 	case DDI_DMA_NEXTSEG:
918 		if (*(px_window_t *)offp != PX_DMA_CURWIN(mp))
919 			return (DDI_DMA_STALE);
920 		if (lenp)				/* only 1 seg allowed */
921 			return (DDI_DMA_DONE);
922 
923 		/* return mp as seg 0 */
924 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
925 		return (DDI_SUCCESS);
926 
927 	case DDI_DMA_SEGTOC:
928 		MAKE_DMA_COOKIE((ddi_dma_cookie_t *)objp, mp->dmai_mapping,
929 			mp->dmai_size);
930 		*offp = mp->dmai_offset;
931 		*lenp = mp->dmai_size;
932 		return (DDI_SUCCESS);
933 
934 	case DDI_DMA_COFF: {
935 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)offp;
936 		if (cp->dmac_address < mp->dmai_mapping ||
937 			(cp->dmac_address + cp->dmac_size) >
938 			(mp->dmai_mapping + mp->dmai_size))
939 			return (DDI_FAILURE);
940 		*objp = (caddr_t)(cp->dmac_address - mp->dmai_mapping +
941 			mp->dmai_offset);
942 		}
943 		return (DDI_SUCCESS);
944 	default:
945 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
946 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
947 		break;
948 	}
949 	return (DDI_FAILURE);
950 }
951 
952 void
953 px_dma_freewin(ddi_dma_impl_t *mp)
954 {
955 	px_dma_win_t *win_p = mp->dmai_winlst, *win2_p;
956 	for (win2_p = win_p; win_p; win2_p = win_p) {
957 		win_p = win2_p->win_next;
958 		kmem_free(win2_p, sizeof (px_dma_win_t) +
959 			sizeof (ddi_dma_cookie_t) * win2_p->win_ncookies);
960 	}
961 	mp->dmai_nwin = 0;
962 	mp->dmai_winlst = NULL;
963 }
964 
965 /*
966  * px_dma_newwin - create a dma window object and cookies
967  *
968  *	After the initial scan in px_dma_physwin(), which identifies
969  *	a portion of the pfn array that belongs to a dma window,
970  *	we are called to allocate and initialize representing memory
971  *	resources. We know from the 1st scan the number of cookies
972  *	or dma segment in this window so we can allocate a contiguous
973  *	memory array for the dma cookies (The implementation of
974  *	ddi_dma_nextcookie(9f) dictates dma cookies be contiguous).
975  *
976  *	A second round scan is done on the pfn array to identify
977  *	each dma segment and initialize its corresponding dma cookie.
978  *	We don't need to do all the safety checking and we know they
979  *	all belong to the same dma window.
980  *
981  *	Input:	cookie_no - # of cookies identified by the 1st scan
982  *		start_idx - subscript of the pfn array for the starting pfn
983  *		end_idx   - subscript of the last pfn in dma window
984  *		win_pp    - pointer to win_next member of previous window
985  *	Return:	DDI_SUCCESS - with **win_pp as newly created window object
986  *		DDI_DMA_NORESROUCE - caller frees all previous window objs
987  *	Note:	Each cookie and window size are all initialized on page
988  *		boundary. This is not true for the 1st cookie of the 1st
989  *		window and the last cookie of the last window.
990  *		We fix that later in upper layer which has access to size
991  *		and offset info.
992  *
993  */
994 /*ARGSUSED*/
995 static int
996 px_dma_newwin(dev_info_t *dip, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp,
997 	uint32_t cookie_no, uint32_t start_idx, uint32_t end_idx,
998 	px_dma_win_t **win_pp, uint64_t count_max, uint64_t bypass)
999 {
1000 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
1001 	ddi_dma_cookie_t *cookie_p;
1002 	uint32_t pfn_no = 1;
1003 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, start_idx);
1004 	px_iopfn_t prev_pfn = pfn;
1005 	uint64_t baddr, seg_pfn0 = pfn;
1006 	size_t sz = cookie_no * sizeof (ddi_dma_cookie_t);
1007 	px_dma_win_t *win_p = kmem_zalloc(sizeof (px_dma_win_t) + sz,
1008 		waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP);
1009 	io_attributes_t	attr = PX_GET_TTE_ATTR(mp->dmai_rflags);
1010 
1011 	if (!win_p)
1012 		goto noresource;
1013 
1014 	win_p->win_next = NULL;
1015 	win_p->win_ncookies = cookie_no;
1016 	win_p->win_curseg = 0;	/* start from segment 0 */
1017 	win_p->win_size = MMU_PTOB(end_idx - start_idx + 1);
1018 	/* win_p->win_offset is left uninitialized */
1019 
1020 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1021 	start_idx++;
1022 	for (; start_idx <= end_idx; start_idx++, prev_pfn = pfn, pfn_no++) {
1023 		pfn = PX_GET_MP_PFN1(mp, start_idx);
1024 		if ((pfn == prev_pfn + 1) &&
1025 			(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1026 			continue;
1027 
1028 		/* close up the cookie up to (including) prev_pfn */
1029 		baddr = MMU_PTOB(seg_pfn0);
1030 		if (bypass && (px_lib_iommu_getbypass(dip,
1031 				baddr, attr, &baddr) != DDI_SUCCESS))
1032 			return (DDI_FAILURE);
1033 
1034 		MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1035 		DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages)\n",
1036 			MMU_PTOB(seg_pfn0), pfn_no);
1037 
1038 		cookie_p++;	/* advance to next available cookie cell */
1039 		pfn_no = 0;
1040 		seg_pfn0 = pfn;	/* start a new segment from current pfn */
1041 	}
1042 
1043 	baddr = MMU_PTOB(seg_pfn0);
1044 	if (bypass && (px_lib_iommu_getbypass(dip,
1045 			baddr, attr, &baddr) != DDI_SUCCESS))
1046 		return (DDI_FAILURE);
1047 
1048 	MAKE_DMA_COOKIE(cookie_p, baddr, MMU_PTOB(pfn_no));
1049 	DBG(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages) of total %x\n",
1050 		MMU_PTOB(seg_pfn0), pfn_no, cookie_no);
1051 #ifdef	DEBUG
1052 	cookie_p++;
1053 	ASSERT((cookie_p - (ddi_dma_cookie_t *)(win_p + 1)) == cookie_no);
1054 #endif	/* DEBUG */
1055 	*win_pp = win_p;
1056 	return (DDI_SUCCESS);
1057 noresource:
1058 	if (waitfp != DDI_DMA_DONTWAIT)
1059 		ddi_set_callback(waitfp, dmareq->dmar_arg, &px_kmem_clid);
1060 	return (DDI_DMA_NORESOURCES);
1061 }
1062 
1063 /*
1064  * px_dma_adjust - adjust 1st and last cookie and window sizes
1065  *	remove initial dma page offset from 1st cookie and window size
1066  *	remove last dma page remainder from last cookie and window size
1067  *	fill win_offset of each dma window according to just fixed up
1068  *		each window sizes
1069  *	px_dma_win_t members modified:
1070  *	win_p->win_offset - this window's offset within entire DMA object
1071  *	win_p->win_size	  - xferrable size (in bytes) for this window
1072  *
1073  *	ddi_dma_impl_t members modified:
1074  *	mp->dmai_size	  - 1st window xferrable size
1075  *	mp->dmai_offset   - 0, which is the dma offset of the 1st window
1076  *
1077  *	ddi_dma_cookie_t members modified:
1078  *	cookie_p->dmac_size - 1st and last cookie remove offset or remainder
1079  *	cookie_p->dmac_laddress - 1st cookie add page offset
1080  */
1081 static void
1082 px_dma_adjust(ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp, px_dma_win_t *win_p)
1083 {
1084 	ddi_dma_cookie_t *cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1085 	size_t pg_offset = mp->dmai_roffset;
1086 	size_t win_offset = 0;
1087 
1088 	cookie_p->dmac_size -= pg_offset;
1089 	cookie_p->dmac_laddress |= pg_offset;
1090 	win_p->win_size -= pg_offset;
1091 	DBG(DBG_BYPASS, mp->dmai_rdip, "pg0 adjust %lx\n", pg_offset);
1092 
1093 	mp->dmai_size = win_p->win_size;
1094 	mp->dmai_offset = 0;
1095 
1096 	pg_offset += mp->dmai_object.dmao_size;
1097 	pg_offset &= MMU_PAGE_OFFSET;
1098 	if (pg_offset)
1099 		pg_offset = MMU_PAGE_SIZE - pg_offset;
1100 	DBG(DBG_BYPASS, mp->dmai_rdip, "last pg adjust %lx\n", pg_offset);
1101 
1102 	for (; win_p->win_next; win_p = win_p->win_next) {
1103 		DBG(DBG_BYPASS, mp->dmai_rdip, "win off %p\n", win_offset);
1104 		win_p->win_offset = win_offset;
1105 		win_offset += win_p->win_size;
1106 	}
1107 	/* last window */
1108 	win_p->win_offset = win_offset;
1109 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1110 	cookie_p[win_p->win_ncookies - 1].dmac_size -= pg_offset;
1111 	win_p->win_size -= pg_offset;
1112 	ASSERT((win_offset + win_p->win_size) == mp->dmai_object.dmao_size);
1113 }
1114 
1115 /*
1116  * px_dma_physwin() - carve up dma windows using physical addresses.
1117  *	Called to handle mmu bypass and pci peer-to-peer transfers.
1118  *	Calls px_dma_newwin() to allocate window objects.
1119  *
1120  * Dependency: mp->dmai_pfnlst points to an array of pfns
1121  *
1122  * 1. Each dma window is represented by a px_dma_win_t object.
1123  *	The object will be casted to ddi_dma_win_t and returned
1124  *	to leaf driver through the DDI interface.
1125  * 2. Each dma window can have several dma segments with each
1126  *	segment representing a physically contiguous either memory
1127  *	space (if we are doing an mmu bypass transfer) or pci address
1128  *	space (if we are doing a peer-to-peer transfer).
1129  * 3. Each segment has a DMA cookie to program the DMA engine.
1130  *	The cookies within each DMA window must be located in a
1131  *	contiguous array per ddi_dma_nextcookie(9f).
1132  * 4. The number of DMA segments within each DMA window cannot exceed
1133  *	mp->dmai_attr.dma_attr_sgllen. If the transfer size is
1134  *	too large to fit in the sgllen, the rest needs to be
1135  *	relocated to the next dma window.
1136  * 5. Peer-to-peer DMA segment follows device hi, lo, count_max,
1137  *	and nocross restrictions while bypass DMA follows the set of
1138  *	restrictions with system limits factored in.
1139  *
1140  * Return:
1141  *	mp->dmai_winlst	 - points to a link list of px_dma_win_t objects.
1142  *		Each px_dma_win_t object on the link list contains
1143  *		infomation such as its window size (# of pages),
1144  *		starting offset (also see Restriction), an array of
1145  *		DMA cookies, and # of cookies in the array.
1146  *	mp->dmai_pfnlst	 - NULL, the pfn list is freed to conserve memory.
1147  *	mp->dmai_nwin	 - # of total DMA windows on mp->dmai_winlst.
1148  *	mp->dmai_mapping - starting cookie address
1149  *	mp->dmai_rflags	 - consistent, nosync, no redzone
1150  *	mp->dmai_cookie	 - start of cookie table of the 1st DMA window
1151  *
1152  * Restriction:
1153  *	Each px_dma_win_t object can theoratically start from any offset
1154  *	since the mmu is not involved. However, this implementation
1155  *	always make windows start from page aligned offset (except
1156  *	the 1st window, which follows the requested offset) due to the
1157  *	fact that we are handed a pfn list. This does require device's
1158  *	count_max and attr_seg to be at least MMU_PAGE_SIZE aligned.
1159  */
1160 int
1161 px_dma_physwin(px_t *px_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
1162 {
1163 	uint_t npages = mp->dmai_ndvmapages;
1164 	int ret, sgllen = mp->dmai_attr.dma_attr_sgllen;
1165 	px_iopfn_t pfn_lo, pfn_hi, prev_pfn;
1166 	px_iopfn_t pfn = PX_GET_MP_PFN(mp, 0);
1167 	uint32_t i, win_no = 0, pfn_no = 1, win_pfn0_index = 0, cookie_no = 0;
1168 	uint64_t count_max, bypass_addr = 0;
1169 	px_dma_win_t **win_pp = (px_dma_win_t **)&mp->dmai_winlst;
1170 	ddi_dma_cookie_t *cookie0_p;
1171 	io_attributes_t attr = PX_GET_TTE_ATTR(mp->dmai_rflags);
1172 	dev_info_t *dip = px_p->px_dip;
1173 
1174 	ASSERT(PX_DMA_ISPTP(mp) || PX_DMA_ISBYPASS(mp));
1175 	if (PX_DMA_ISPTP(mp)) { /* ignore sys limits for peer-to-peer */
1176 		ddi_dma_attr_t *dev_attr_p = DEV_ATTR(mp);
1177 		uint64_t nocross = dev_attr_p->dma_attr_seg;
1178 		px_pec_t *pec_p = px_p->px_pec_p;
1179 		px_iopfn_t pfn_last = PX_DMA_ISPTP32(mp) ?
1180 				pec_p->pec_last32_pfn - pec_p->pec_base32_pfn :
1181 				pec_p->pec_last64_pfn - pec_p->pec_base64_pfn;
1182 
1183 		if (nocross && (nocross < UINT32_MAX))
1184 			return (DDI_DMA_NOMAPPING);
1185 		if (dev_attr_p->dma_attr_align > MMU_PAGE_SIZE)
1186 			return (DDI_DMA_NOMAPPING);
1187 		pfn_lo = MMU_BTOP(dev_attr_p->dma_attr_addr_lo);
1188 		pfn_hi = MMU_BTOP(dev_attr_p->dma_attr_addr_hi);
1189 		pfn_hi = MIN(pfn_hi, pfn_last);
1190 		if ((pfn_lo > pfn_hi) || (pfn < pfn_lo))
1191 			return (DDI_DMA_NOMAPPING);
1192 
1193 		count_max = dev_attr_p->dma_attr_count_max;
1194 		count_max = MIN(count_max, nocross);
1195 		/*
1196 		 * the following count_max trim is not done because we are
1197 		 * making sure pfn_lo <= pfn <= pfn_hi inside the loop
1198 		 * count_max=MIN(count_max, MMU_PTOB(pfn_hi - pfn_lo + 1)-1);
1199 		 */
1200 	} else { /* bypass hi/lo/count_max have been processed by attr2hdl() */
1201 		count_max = mp->dmai_attr.dma_attr_count_max;
1202 		pfn_lo = MMU_BTOP(mp->dmai_attr.dma_attr_addr_lo);
1203 		pfn_hi = MMU_BTOP(mp->dmai_attr.dma_attr_addr_hi);
1204 
1205 		if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn),
1206 				attr, &bypass_addr) != DDI_SUCCESS) {
1207 			cmn_err(CE_WARN, "bypass cookie failure %lx\n", pfn);
1208 			return (DDI_DMA_NOMAPPING);
1209 		}
1210 		pfn = MMU_BTOP(bypass_addr);
1211 	}
1212 
1213 	/* pfn: absolute (bypass mode) or relative (p2p mode) */
1214 	for (prev_pfn = pfn, i = 1; i < npages;
1215 	    i++, prev_pfn = pfn, pfn_no++) {
1216 		pfn = PX_GET_MP_PFN1(mp, i);
1217 		if (bypass_addr) {
1218 			if (px_lib_iommu_getbypass(dip, MMU_PTOB(pfn), attr,
1219 					&bypass_addr) != DDI_SUCCESS) {
1220 				ret = DDI_DMA_NOMAPPING;
1221 				goto err;
1222 			}
1223 			pfn = MMU_BTOP(bypass_addr);
1224 		}
1225 		if ((pfn == prev_pfn + 1) &&
1226 				(MMU_PTOB(pfn_no + 1) - 1 <= count_max))
1227 			continue;
1228 		if ((pfn < pfn_lo) || (prev_pfn > pfn_hi)) {
1229 			ret = DDI_DMA_NOMAPPING;
1230 			goto err;
1231 		}
1232 		cookie_no++;
1233 		pfn_no = 0;
1234 		if (cookie_no < sgllen)
1235 			continue;
1236 
1237 		DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1238 			win_pfn0_index, i - 1, cookie_no);
1239 		if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no,
1240 			win_pfn0_index, i - 1, win_pp, count_max, bypass_addr))
1241 			goto err;
1242 
1243 		win_pp = &(*win_pp)->win_next;	/* win_pp = *(win_pp) */
1244 		win_no++;
1245 		win_pfn0_index = i;
1246 		cookie_no = 0;
1247 	}
1248 	if (pfn > pfn_hi) {
1249 		ret = DDI_DMA_NOMAPPING;
1250 		goto err;
1251 	}
1252 	cookie_no++;
1253 	DBG(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1254 		win_pfn0_index, i - 1, cookie_no);
1255 	if (ret = px_dma_newwin(dip, dmareq, mp, cookie_no, win_pfn0_index,
1256 		i - 1, win_pp, count_max, bypass_addr))
1257 		goto err;
1258 	win_no++;
1259 	px_dma_adjust(dmareq, mp, mp->dmai_winlst);
1260 	mp->dmai_nwin = win_no;
1261 	mp->dmai_rflags |= DDI_DMA_CONSISTENT | DMP_NOSYNC;
1262 	mp->dmai_rflags &= ~DDI_DMA_REDZONE;
1263 	mp->dmai_flags |= DMAI_FLAGS_NOSYNC;
1264 	cookie0_p = (ddi_dma_cookie_t *)(WINLST(mp) + 1);
1265 	mp->dmai_cookie = WINLST(mp)->win_ncookies > 1 ? cookie0_p + 1 : 0;
1266 	mp->dmai_mapping = cookie0_p->dmac_laddress;
1267 
1268 	px_dma_freepfn(mp);
1269 	return (DDI_DMA_MAPPED);
1270 err:
1271 	px_dma_freewin(mp);
1272 	return (ret);
1273 }
1274 
1275 int
1276 px_dma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
1277 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
1278 	uint_t cache_flags)
1279 {
1280 	switch (cmd) {
1281 	case DDI_DMA_SYNC:
1282 		return (DDI_SUCCESS);
1283 
1284 	case DDI_DMA_HTOC: {
1285 		off_t off = *offp;
1286 		ddi_dma_cookie_t *loop_cp, *cp;
1287 		px_dma_win_t *win_p = mp->dmai_winlst;
1288 
1289 		if (off >= mp->dmai_object.dmao_size)
1290 			return (DDI_FAILURE);
1291 
1292 		/* locate window */
1293 		while (win_p->win_offset + win_p->win_size <= off)
1294 			win_p = win_p->win_next;
1295 
1296 		loop_cp = cp = (ddi_dma_cookie_t *)(win_p + 1);
1297 		mp->dmai_offset = win_p->win_offset;
1298 		mp->dmai_size   = win_p->win_size;
1299 		mp->dmai_mapping = cp->dmac_laddress; /* cookie0 start addr */
1300 
1301 		/* adjust cookie addr/len if we are not on cookie boundary */
1302 		off -= win_p->win_offset;	   /* offset within window */
1303 		for (; off >= loop_cp->dmac_size; loop_cp++)
1304 			off -= loop_cp->dmac_size; /* offset within cookie */
1305 
1306 		mp->dmai_cookie = loop_cp + 1;
1307 		win_p->win_curseg = loop_cp - cp;
1308 		cp = (ddi_dma_cookie_t *)objp;
1309 		MAKE_DMA_COOKIE(cp, loop_cp->dmac_laddress + off,
1310 			loop_cp->dmac_size - off);
1311 
1312 		DBG(DBG_DMA_CTL, dip,
1313 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1314 			cp->dmac_laddress, cp->dmac_size);
1315 		}
1316 		return (DDI_SUCCESS);
1317 
1318 	case DDI_DMA_REPWIN:
1319 		*offp = mp->dmai_offset;
1320 		*lenp = mp->dmai_size;
1321 		return (DDI_SUCCESS);
1322 
1323 	case DDI_DMA_MOVWIN: {
1324 		off_t off = *offp;
1325 		ddi_dma_cookie_t *cp;
1326 		px_dma_win_t *win_p = mp->dmai_winlst;
1327 
1328 		if (off >= mp->dmai_object.dmao_size)
1329 			return (DDI_FAILURE);
1330 
1331 		/* locate window */
1332 		while (win_p->win_offset + win_p->win_size <= off)
1333 			win_p = win_p->win_next;
1334 
1335 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1336 		mp->dmai_offset = win_p->win_offset;
1337 		mp->dmai_size   = win_p->win_size;
1338 		mp->dmai_mapping = cp->dmac_laddress;	/* cookie0 star addr */
1339 		mp->dmai_cookie = cp + 1;
1340 		win_p->win_curseg = 0;
1341 
1342 		*(ddi_dma_cookie_t *)objp = *cp;
1343 		*offp = win_p->win_offset;
1344 		*lenp = win_p->win_size;
1345 		DBG(DBG_DMA_CTL, dip,
1346 			"HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1347 			cp->dmac_laddress, cp->dmac_size);
1348 		}
1349 		return (DDI_SUCCESS);
1350 
1351 	case DDI_DMA_NEXTWIN: {
1352 		px_dma_win_t *win_p = *(px_dma_win_t **)offp;
1353 		px_dma_win_t **nw_pp = (px_dma_win_t **)objp;
1354 		ddi_dma_cookie_t *cp;
1355 		if (!win_p) {
1356 			*nw_pp = mp->dmai_winlst;
1357 			return (DDI_SUCCESS);
1358 		}
1359 
1360 		if (win_p->win_offset != mp->dmai_offset)
1361 			return (DDI_DMA_STALE);
1362 		if (!win_p->win_next)
1363 			return (DDI_DMA_DONE);
1364 		win_p = win_p->win_next;
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 		*nw_pp = win_p;
1372 		}
1373 		return (DDI_SUCCESS);
1374 
1375 	case DDI_DMA_NEXTSEG: {
1376 		px_dma_win_t *w_p = *(px_dma_win_t **)offp;
1377 		if (w_p->win_offset != mp->dmai_offset)
1378 			return (DDI_DMA_STALE);
1379 		if (w_p->win_curseg + 1 >= w_p->win_ncookies)
1380 			return (DDI_DMA_DONE);
1381 		w_p->win_curseg++;
1382 		}
1383 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
1384 		return (DDI_SUCCESS);
1385 
1386 	case DDI_DMA_SEGTOC: {
1387 		px_dma_win_t *win_p = mp->dmai_winlst;
1388 		off_t off = mp->dmai_offset;
1389 		ddi_dma_cookie_t *cp;
1390 		int i;
1391 
1392 		/* locate active window */
1393 		for (; win_p->win_offset != off; win_p = win_p->win_next);
1394 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1395 		for (i = 0; i < win_p->win_curseg; i++, cp++)
1396 			off += cp->dmac_size;
1397 		*offp = off;
1398 		*lenp = cp->dmac_size;
1399 		*(ddi_dma_cookie_t *)objp = *cp;	/* copy cookie */
1400 		}
1401 		return (DDI_SUCCESS);
1402 
1403 	case DDI_DMA_COFF: {
1404 		px_dma_win_t *win_p;
1405 		ddi_dma_cookie_t *cp;
1406 		uint64_t addr, key = ((ddi_dma_cookie_t *)offp)->dmac_laddress;
1407 		size_t win_off;
1408 
1409 		for (win_p = mp->dmai_winlst; win_p; win_p = win_p->win_next) {
1410 			int i;
1411 			win_off = 0;
1412 			cp = (ddi_dma_cookie_t *)(win_p + 1);
1413 			for (i = 0; i < win_p->win_ncookies; i++, cp++) {
1414 				size_t sz = cp->dmac_size;
1415 
1416 				addr = cp->dmac_laddress;
1417 				if ((addr <= key) && (addr + sz >= key))
1418 					goto found;
1419 				win_off += sz;
1420 			}
1421 		}
1422 		return (DDI_FAILURE);
1423 found:
1424 		*objp = (caddr_t)(win_p->win_offset + win_off + (key - addr));
1425 		return (DDI_SUCCESS);
1426 		}
1427 	default:
1428 		DBG(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
1429 			cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
1430 		break;
1431 	}
1432 	return (DDI_FAILURE);
1433 }
1434 
1435 static void
1436 px_dvma_debug_init(px_mmu_t *mmu_p)
1437 {
1438 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1439 	ASSERT(MUTEX_HELD(&mmu_p->dvma_debug_lock));
1440 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat ON", mmu_p);
1441 
1442 	mmu_p->dvma_alloc_rec = kmem_alloc(sz, KM_SLEEP);
1443 	mmu_p->dvma_free_rec = kmem_alloc(sz, KM_SLEEP);
1444 
1445 	mmu_p->dvma_active_list = NULL;
1446 	mmu_p->dvma_alloc_rec_index = 0;
1447 	mmu_p->dvma_free_rec_index = 0;
1448 	mmu_p->dvma_active_count = 0;
1449 }
1450 
1451 void
1452 px_dvma_debug_fini(px_mmu_t *mmu_p)
1453 {
1454 	struct px_dvma_rec *prev, *ptr;
1455 	size_t sz = sizeof (struct px_dvma_rec) * px_dvma_debug_rec;
1456 	uint64_t mask = ~(1ull << mmu_p->mmu_inst);
1457 	cmn_err(CE_NOTE, "PCI Express DVMA %p stat OFF", mmu_p);
1458 
1459 	kmem_free(mmu_p->dvma_alloc_rec, sz);
1460 	kmem_free(mmu_p->dvma_free_rec, sz);
1461 	mmu_p->dvma_alloc_rec = mmu_p->dvma_free_rec = NULL;
1462 
1463 	prev = mmu_p->dvma_active_list;
1464 	if (!prev)
1465 		return;
1466 	for (ptr = prev->next; ptr; prev = ptr, ptr = ptr->next)
1467 		kmem_free(prev, sizeof (struct px_dvma_rec));
1468 	kmem_free(prev, sizeof (struct px_dvma_rec));
1469 
1470 	mmu_p->dvma_active_list = NULL;
1471 	mmu_p->dvma_alloc_rec_index = 0;
1472 	mmu_p->dvma_free_rec_index = 0;
1473 	mmu_p->dvma_active_count = 0;
1474 
1475 	px_dvma_debug_off &= mask;
1476 	px_dvma_debug_on &= mask;
1477 }
1478 
1479 void
1480 px_dvma_alloc_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1481 	ddi_dma_impl_t *mp)
1482 {
1483 	struct px_dvma_rec *ptr;
1484 	mutex_enter(&mmu_p->dvma_debug_lock);
1485 
1486 	if (!mmu_p->dvma_alloc_rec)
1487 		px_dvma_debug_init(mmu_p);
1488 	if (DVMA_DBG_OFF(mmu_p)) {
1489 		px_dvma_debug_fini(mmu_p);
1490 		goto done;
1491 	}
1492 
1493 	ptr = &mmu_p->dvma_alloc_rec[mmu_p->dvma_alloc_rec_index];
1494 	ptr->dvma_addr = address;
1495 	ptr->len = len;
1496 	ptr->mp = mp;
1497 	if (++mmu_p->dvma_alloc_rec_index == px_dvma_debug_rec)
1498 		mmu_p->dvma_alloc_rec_index = 0;
1499 
1500 	ptr = kmem_alloc(sizeof (struct px_dvma_rec), KM_SLEEP);
1501 	ptr->dvma_addr = address;
1502 	ptr->len = len;
1503 	ptr->mp = mp;
1504 
1505 	ptr->next = mmu_p->dvma_active_list;
1506 	mmu_p->dvma_active_list = ptr;
1507 	mmu_p->dvma_active_count++;
1508 done:
1509 	mutex_exit(&mmu_p->dvma_debug_lock);
1510 }
1511 
1512 void
1513 px_dvma_free_debug(px_mmu_t *mmu_p, char *address, uint_t len,
1514     ddi_dma_impl_t *mp)
1515 {
1516 	struct px_dvma_rec *ptr, *ptr_save;
1517 	mutex_enter(&mmu_p->dvma_debug_lock);
1518 
1519 	if (!mmu_p->dvma_alloc_rec)
1520 		px_dvma_debug_init(mmu_p);
1521 	if (DVMA_DBG_OFF(mmu_p)) {
1522 		px_dvma_debug_fini(mmu_p);
1523 		goto done;
1524 	}
1525 
1526 	ptr = &mmu_p->dvma_free_rec[mmu_p->dvma_free_rec_index];
1527 	ptr->dvma_addr = address;
1528 	ptr->len = len;
1529 	ptr->mp = mp;
1530 	if (++mmu_p->dvma_free_rec_index == px_dvma_debug_rec)
1531 		mmu_p->dvma_free_rec_index = 0;
1532 
1533 	ptr_save = mmu_p->dvma_active_list;
1534 	for (ptr = ptr_save; ptr; ptr = ptr->next) {
1535 		if ((ptr->dvma_addr == address) && (ptr->len = len))
1536 			break;
1537 		ptr_save = ptr;
1538 	}
1539 	if (!ptr) {
1540 		cmn_err(CE_WARN, "bad dvma free addr=%lx len=%x",
1541 			(long)address, len);
1542 		goto done;
1543 	}
1544 	if (ptr == mmu_p->dvma_active_list)
1545 		mmu_p->dvma_active_list = ptr->next;
1546 	else
1547 		ptr_save->next = ptr->next;
1548 	kmem_free(ptr, sizeof (struct px_dvma_rec));
1549 	mmu_p->dvma_active_count--;
1550 done:
1551 	mutex_exit(&mmu_p->dvma_debug_lock);
1552 }
1553 
1554 #ifdef	DEBUG
1555 void
1556 px_dump_dma_handle(uint64_t flag, dev_info_t *dip, ddi_dma_impl_t *hp)
1557 {
1558 	DBG(flag, dip, "mp(%p): flags=%x mapping=%lx xfer_size=%x\n",
1559 		hp, hp->dmai_inuse, hp->dmai_mapping, hp->dmai_size);
1560 	DBG(flag|DBG_CONT, dip, "\tnpages=%x roffset=%x rflags=%x nwin=%x\n",
1561 		hp->dmai_ndvmapages, hp->dmai_roffset, hp->dmai_rflags,
1562 		hp->dmai_nwin);
1563 	DBG(flag|DBG_CONT, dip, "\twinsize=%x tte=%p pfnlst=%p pfn0=%p\n",
1564 		hp->dmai_winsize, hp->dmai_tte, hp->dmai_pfnlst, hp->dmai_pfn0);
1565 	DBG(flag|DBG_CONT, dip, "\twinlst=%x obj=%p attr=%p ckp=%p\n",
1566 		hp->dmai_winlst, &hp->dmai_object, &hp->dmai_attr,
1567 		hp->dmai_cookie);
1568 }
1569 #endif	/* DEBUG */
1570