xref: /titanic_41/usr/src/uts/sun4u/io/pci/pci_dma.c (revision f885d00f4e3c96a769ce0228a732da31ad9d0b78)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * PCI nexus DVMA and DMA core routines:
28  *	dma_map/dma_bind_handle implementation
29  *	bypass and peer-to-peer support
30  *	fast track DVMA space allocation
31  *	runtime DVMA debug
32  */
33 #include <sys/types.h>
34 #include <sys/kmem.h>
35 #include <sys/async.h>
36 #include <sys/sysmacros.h>
37 #include <sys/sunddi.h>
38 #include <sys/machsystm.h>	/* lddphys() */
39 #include <sys/ddi_impldefs.h>
40 #include <vm/hat.h>
41 #include <sys/pci/pci_obj.h>
42 
43 /*LINTLIBRARY*/
44 
45 static void
46 pci_sc_pg_inv(dev_info_t *dip, sc_t *sc_p, ddi_dma_impl_t *mp, off_t off,
47 	size_t len)
48 {
49 	dvma_addr_t dvma_addr, pg_off;
50 	volatile uint64_t *invl_va = sc_p->sc_invl_reg;
51 
52 	if (!len)
53 		len = mp->dmai_size;
54 
55 	pg_off = mp->dmai_offset;			/* start min */
56 	dvma_addr = MAX(off, pg_off);			/* lo */
57 	pg_off += mp->dmai_size;			/* end max */
58 	pg_off = MIN(off + len, pg_off);		/* hi */
59 	if (dvma_addr >= pg_off) {			/* lo >= hi ? */
60 		DEBUG4(DBG_SC, dip, "%x+%x out of window [%x,%x)\n",
61 		    off, len, mp->dmai_offset,
62 		    mp->dmai_offset + mp->dmai_size);
63 		return;
64 	}
65 
66 	len = pg_off - dvma_addr;			/* sz = hi - lo */
67 	dvma_addr += mp->dmai_mapping;			/* start addr */
68 	pg_off = dvma_addr & IOMMU_PAGE_OFFSET;		/* offset in 1st pg */
69 	len = IOMMU_BTOPR(len + pg_off);		/* # of pages */
70 	dvma_addr ^= pg_off;
71 
72 	DEBUG2(DBG_SC, dip, "addr=%x+%x pages: \n", dvma_addr, len);
73 	for (; len; len--, dvma_addr += IOMMU_PAGE_SIZE) {
74 		DEBUG1(DBG_SC|DBG_CONT, dip, " %x", dvma_addr);
75 		*invl_va = (uint64_t)dvma_addr;
76 	}
77 	DEBUG0(DBG_SC|DBG_CONT, dip, "\n");
78 }
79 
80 static void
81 pci_dma_sync_flag_wait(ddi_dma_impl_t *mp, sc_t *sc_p, uint32_t onstack)
82 {
83 	hrtime_t start_time;
84 	uint64_t loops = 0;
85 	uint64_t sync_flag_pa = SYNC_BUF_PA(mp);
86 	uint64_t sync_reg_pa = sc_p->sc_sync_reg_pa;
87 	uint8_t stack_buf[128];
88 
89 	stack_buf[0] = DDI_SUCCESS;
90 
91 	/* check for handle specific sync flag */
92 	if (sync_flag_pa)
93 		goto start;
94 
95 	sync_flag_pa = sc_p->sc_sync_flag_pa;
96 
97 	if (onstack) {
98 		sync_flag_pa = va_to_pa(stack_buf);
99 		sync_flag_pa += PCI_SYNC_FLAG_SIZE;
100 		sync_flag_pa >>= PCI_SYNC_FLAG_SZSHIFT;
101 		sync_flag_pa <<= PCI_SYNC_FLAG_SZSHIFT;
102 		goto start;
103 	}
104 	stack_buf[0] |= PCI_SYNC_FLAG_LOCKED;
105 	mutex_enter(&sc_p->sc_sync_mutex);
106 start:
107 	ASSERT(!(sync_flag_pa & PCI_SYNC_FLAG_SIZE - 1));
108 	stdphys(sync_flag_pa, 0);	/* reset sync flag to 0 */
109 					/* membar  #LoadStore|#StoreStore */
110 	stdphysio(sync_reg_pa, sync_flag_pa);
111 	start_time = gethrtime();
112 
113 	for (; gethrtime() - start_time < pci_sync_buf_timeout; loops++)
114 		if (lddphys(sync_flag_pa))
115 			goto done;
116 
117 	if (!lddphys(sync_flag_pa))
118 		stack_buf[0] |= PCI_SYNC_FLAG_FAILED;
119 done:
120 	DEBUG3(DBG_SC|DBG_CONT, 0, "flag wait loops=%lu ticks=%lu status=%x\n",
121 	    loops, gethrtime() - start_time, stack_buf[0]);
122 
123 	if (stack_buf[0] & PCI_SYNC_FLAG_LOCKED)
124 		mutex_exit(&sc_p->sc_sync_mutex);
125 
126 	if (stack_buf[0] & PCI_SYNC_FLAG_FAILED)
127 		cmn_err(CE_PANIC, "%p pci dma sync %lx %lx timeout!",
128 		    mp, sync_flag_pa, loops);
129 }
130 
131 /*
132  * Cache	RW	Before	During		After
133  *
134  * STREAMING	read	no/no	pg/no		ctx,pg/no
135  * STREAMING	write	no/no	pg/yes		ctx,pg/yes
136  * CONSISTENT	read	no/no	yes,no/no	yes,no/no
137  * CONSISTENT	write	no/no	yes,yes/yes	yes,yes/yes
138  *
139  * STREAMING	read	ctx,pg/no
140  * STREAMING	write	ctx,pg/yes
141  * CONSISTENT	read	yes,no/no
142  * CONSISTENT	write	yes,yes/yes
143  */
144 int
145 pci_dma_sync(dev_info_t *dip, dev_info_t *rdip, ddi_dma_handle_t handle,
146 	off_t off, size_t len, uint32_t sync_flag)
147 {
148 	ddi_dma_impl_t *mp = (ddi_dma_impl_t *)handle;
149 	int ret = ddi_get_instance(dip);
150 	pci_t *pci_p = get_pci_soft_state(ret);
151 	pbm_t *pbm_p = pci_p->pci_pbm_p;
152 	uint32_t dev_flag = mp->dmai_rflags;
153 	sc_t *sc_p;
154 
155 	DEBUG4(DBG_DMA_SYNC, dip, "%s%d flags=%x,%x\n", ddi_driver_name(rdip),
156 	    ddi_get_instance(rdip), dev_flag, sync_flag);
157 	DEBUG4(DBG_SC, dip, "dmai_mapping=%x, dmai_sz=%x off=%x len=%x\n",
158 	    mp->dmai_mapping, mp->dmai_size, off, len);
159 	DEBUG2(DBG_SC, dip, "mp=%p, ctx=%x\n", mp, MP2CTX(mp));
160 
161 	if (!(mp->dmai_flags & DMAI_FLAGS_INUSE)) {
162 		cmn_err(CE_WARN, "Unbound dma handle %p from %s%d", mp,
163 		    ddi_driver_name(rdip), ddi_get_instance(rdip));
164 		return (DDI_FAILURE);
165 	}
166 
167 	if (mp->dmai_flags & DMAI_FLAGS_NOSYNC)
168 		return (DDI_SUCCESS);
169 
170 	if (!(dev_flag & DDI_DMA_CONSISTENT))
171 		goto streaming;
172 
173 	if (sync_flag & PCI_DMA_SYNC_EXT) {
174 		if (sync_flag & (PCI_DMA_SYNC_BEFORE | PCI_DMA_SYNC_POST) ||
175 		    !(sync_flag & PCI_DMA_SYNC_WRITE))
176 			return (DDI_SUCCESS);
177 	} else {
178 		if (!(dev_flag & DDI_DMA_READ) ||
179 		    ((sync_flag & PCI_DMA_SYNC_DDI_FLAGS) ==
180 		    DDI_DMA_SYNC_FORDEV))
181 			return (DDI_SUCCESS);
182 	}
183 
184 	pci_pbm_dma_sync(pbm_p, pbm_p->pbm_sync_ino);
185 	return (DDI_SUCCESS);
186 
187 streaming:
188 	ASSERT(pci_stream_buf_exists && (pci_stream_buf_enable & 1 << ret));
189 	sc_p = pci_p->pci_sc_p;
190 	ret = DDI_FAILURE;
191 
192 	if (sync_flag & PCI_DMA_SYNC_EXT)
193 		goto ext;
194 
195 	if (mp->dmai_flags & DMAI_FLAGS_CONTEXT && pci_sc_use_contexts)
196 		ret = pci_sc_ctx_inv(dip, sc_p, mp);
197 	if (ret)
198 		pci_sc_pg_inv(dip, sc_p, mp, off, len);
199 
200 	if ((dev_flag & DDI_DMA_READ) &&
201 	    ((sync_flag & PCI_DMA_SYNC_DDI_FLAGS) != DDI_DMA_SYNC_FORDEV))
202 		goto wait;
203 
204 	return (DDI_SUCCESS);
205 ext:
206 	if (sync_flag & PCI_DMA_SYNC_BEFORE)
207 		return (DDI_SUCCESS);
208 	if (sync_flag & PCI_DMA_SYNC_BAR)
209 		goto wait_check;
210 	if (sync_flag & PCI_DMA_SYNC_AFTER &&
211 	    mp->dmai_flags & DMAI_FLAGS_CONTEXT && pci_sc_use_contexts)
212 		ret = pci_sc_ctx_inv(dip, sc_p, mp);
213 	if (ret)
214 		pci_sc_pg_inv(dip, sc_p, mp, off, len);
215 wait_check:
216 	if (sync_flag & PCI_DMA_SYNC_POST || !(sync_flag & PCI_DMA_SYNC_WRITE))
217 		return (DDI_SUCCESS);
218 wait:
219 	pci_dma_sync_flag_wait(mp, sc_p, sync_flag & PCI_DMA_SYNC_PRIVATE);
220 	return (DDI_SUCCESS);
221 }
222 
223 int
224 pci_dma_handle_clean(dev_info_t *rdip, ddi_dma_handle_t h)
225 {
226 	ddi_dma_impl_t *mp = (ddi_dma_impl_t *)h;
227 	if ((mp->dmai_flags & DMAI_FLAGS_INUSE) == 0)
228 		return (DDI_FAILURE);
229 	mp->dmai_rflags |= DMP_NOSYNC;
230 	mp->dmai_flags |= DMAI_FLAGS_NOSYNC;
231 	return (DDI_SUCCESS);
232 }
233 
234 /*
235  * pci_dma_allocmp - Allocate a pci dma implementation structure
236  *
237  * An extra ddi_dma_attr structure is bundled with the usual ddi_dma_impl
238  * to hold unmodified device limits. The ddi_dma_attr inside the
239  * ddi_dma_impl structure is augumented with system limits to enhance
240  * DVMA performance at runtime. The unaugumented device limits saved
241  * right after (accessed through the DEV_ATTR macro) is used
242  * strictly for peer-to-peer transfers which do not obey system limits.
243  *
244  * return: DDI_SUCCESS DDI_DMA_NORESOURCES
245  */
246 ddi_dma_impl_t *
247 pci_dma_allocmp(dev_info_t *dip, dev_info_t *rdip, int (*waitfp)(caddr_t),
248 	caddr_t arg)
249 {
250 	ddi_dma_impl_t *mp;
251 	int sleep = (waitfp == DDI_DMA_SLEEP) ? KM_SLEEP : KM_NOSLEEP;
252 
253 	/* Caution: we don't use zalloc to enhance performance! */
254 	if ((mp = kmem_alloc(sizeof (pci_dma_hdl_t), sleep)) == 0) {
255 		DEBUG0(DBG_DMA_MAP, dip, "can't alloc dma_handle\n");
256 		if (waitfp != DDI_DMA_DONTWAIT) {
257 			DEBUG0(DBG_DMA_MAP, dip, "alloc_mp kmem cb\n");
258 			ddi_set_callback(waitfp, arg, &pci_kmem_clid);
259 		}
260 		return (mp);
261 	}
262 
263 	mp->dmai_rdip = rdip;
264 	mp->dmai_flags = 0;
265 	mp->dmai_pfnlst = NULL;
266 	mp->dmai_winlst = NULL;
267 
268 	/*
269 	 * kmem_alloc debug: the following fields are not zero-ed
270 	 * mp->dmai_mapping = 0;
271 	 * mp->dmai_size = 0;
272 	 * mp->dmai_offset = 0;
273 	 * mp->dmai_minxfer = 0;
274 	 * mp->dmai_burstsizes = 0;
275 	 * mp->dmai_ndvmapages = 0;
276 	 * mp->dmai_pool/roffset = 0;
277 	 * mp->dmai_rflags = 0;
278 	 * mp->dmai_inuse/flags
279 	 * mp->dmai_nwin = 0;
280 	 * mp->dmai_winsize = 0;
281 	 * mp->dmai_nexus_private/tte = 0;
282 	 * mp->dmai_iopte/pfnlst
283 	 * mp->dmai_sbi/pfn0 = 0;
284 	 * mp->dmai_minfo/winlst/fdvma
285 	 * mp->dmai_rdip
286 	 * bzero(&mp->dmai_object, sizeof (ddi_dma_obj_t));
287 	 * mp->dmai_cookie = 0;
288 	 */
289 
290 	mp->dmai_attr.dma_attr_version = (uint_t)DMA_ATTR_VERSION;
291 	mp->dmai_attr.dma_attr_flags = (uint_t)0;
292 	mp->dmai_fault = 0;
293 	mp->dmai_fault_check = NULL;
294 	mp->dmai_fault_notify = NULL;
295 
296 	mp->dmai_error.err_ena = 0;
297 	mp->dmai_error.err_status = DDI_FM_OK;
298 	mp->dmai_error.err_expected = DDI_FM_ERR_UNEXPECTED;
299 	mp->dmai_error.err_ontrap = NULL;
300 	mp->dmai_error.err_fep = NULL;
301 	mp->dmai_error.err_cf = NULL;
302 	ndi_fmc_insert(rdip, DMA_HANDLE, mp, NULL);
303 
304 	SYNC_BUF_PA(mp) = 0ull;
305 	return (mp);
306 }
307 
308 void
309 pci_dma_freemp(ddi_dma_impl_t *mp)
310 {
311 	ndi_fmc_remove(mp->dmai_rdip, DMA_HANDLE, mp);
312 	if (mp->dmai_ndvmapages > 1)
313 		pci_dma_freepfn(mp);
314 	if (mp->dmai_winlst)
315 		pci_dma_freewin(mp);
316 	kmem_free(mp, sizeof (pci_dma_hdl_t));
317 }
318 
319 void
320 pci_dma_freepfn(ddi_dma_impl_t *mp)
321 {
322 	void *addr = mp->dmai_pfnlst;
323 	ASSERT(!PCI_DMA_CANRELOC(mp));
324 	if (addr) {
325 		size_t npages = mp->dmai_ndvmapages;
326 		if (npages > 1)
327 			kmem_free(addr, npages * sizeof (iopfn_t));
328 		mp->dmai_pfnlst = NULL;
329 	}
330 	mp->dmai_ndvmapages = 0;
331 }
332 
333 /*
334  * pci_dma_lmts2hdl - alloate a ddi_dma_impl_t, validate practical limits
335  *			and convert dmareq->dmar_limits to mp->dmai_attr
336  *
337  * ddi_dma_impl_t member modified     input
338  * ------------------------------------------------------------------------
339  * mp->dmai_minxfer		    - dev
340  * mp->dmai_burstsizes		    - dev
341  * mp->dmai_flags		    - no limit? peer-to-peer only?
342  *
343  * ddi_dma_attr member modified       input
344  * ------------------------------------------------------------------------
345  * mp->dmai_attr.dma_attr_addr_lo   - dev lo, sys lo
346  * mp->dmai_attr.dma_attr_addr_hi   - dev hi, sys hi
347  * mp->dmai_attr.dma_attr_count_max - dev count max, dev/sys lo/hi delta
348  * mp->dmai_attr.dma_attr_seg       - 0         (no nocross   restriction)
349  * mp->dmai_attr.dma_attr_align     - 1		(no alignment restriction)
350  *
351  * The dlim_dmaspeed member of dmareq->dmar_limits is ignored.
352  */
353 ddi_dma_impl_t *
354 pci_dma_lmts2hdl(dev_info_t *dip, dev_info_t *rdip, iommu_t *iommu_p,
355 	ddi_dma_req_t *dmareq)
356 {
357 	ddi_dma_impl_t *mp;
358 	ddi_dma_attr_t *attr_p;
359 	uint64_t syslo		= iommu_p->iommu_dvma_base;
360 	uint64_t syshi		= iommu_p->iommu_dvma_end;
361 	uint64_t fasthi		= iommu_p->iommu_dvma_fast_end;
362 	ddi_dma_lim_t *lim_p	= dmareq->dmar_limits;
363 	uint32_t count_max	= lim_p->dlim_cntr_max;
364 	uint64_t lo		= lim_p->dlim_addr_lo;
365 	uint64_t hi		= lim_p->dlim_addr_hi;
366 	if (hi <= lo) {
367 		DEBUG0(DBG_DMA_MAP, dip, "Bad limits\n");
368 		return ((ddi_dma_impl_t *)DDI_DMA_NOMAPPING);
369 	}
370 	if (!count_max)
371 		count_max--;
372 
373 	if (!(mp = pci_dma_allocmp(dip, rdip, dmareq->dmar_fp,
374 	    dmareq->dmar_arg)))
375 		return (NULL);
376 
377 	/* store original dev input at the 2nd ddi_dma_attr */
378 	attr_p = DEV_ATTR(mp);
379 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
380 	SET_DMAALIGN(attr_p, 1);
381 
382 	lo = MAX(lo, syslo);
383 	hi = MIN(hi, syshi);
384 	if (hi <= lo)
385 		mp->dmai_flags |= DMAI_FLAGS_PEER_ONLY;
386 	count_max = MIN(count_max, hi - lo);
387 
388 	if (DEV_NOSYSLIMIT(lo, hi, syslo, fasthi, 1))
389 		mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT |
390 		    DMAI_FLAGS_NOSYSLIMIT;
391 	else {
392 		if (DEV_NOFASTLIMIT(lo, hi, syslo, syshi, 1))
393 			mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT;
394 	}
395 	if (PCI_DMA_NOCTX(rdip))
396 		mp->dmai_flags |= DMAI_FLAGS_NOCTX;
397 
398 	/* store augumented dev input to mp->dmai_attr */
399 	mp->dmai_minxfer	= lim_p->dlim_minxfer;
400 	mp->dmai_burstsizes	= lim_p->dlim_burstsizes;
401 	attr_p = &mp->dmai_attr;
402 	SET_DMAATTR(attr_p, lo, hi, -1, count_max);
403 	SET_DMAALIGN(attr_p, 1);
404 	return (mp);
405 }
406 
407 /*
408  * pci_dma_attr2hdl
409  *
410  * This routine is called from the alloc handle entry point to sanity check the
411  * dma attribute structure.
412  *
413  * use by: pci_dma_allochdl()
414  *
415  * return value:
416  *
417  *	DDI_SUCCESS		- on success
418  *	DDI_DMA_BADATTR		- attribute has invalid version number
419  *				  or address limits exclude dvma space
420  */
421 int
422 pci_dma_attr2hdl(pci_t *pci_p, ddi_dma_impl_t *mp)
423 {
424 	iommu_t *iommu_p = pci_p->pci_iommu_p;
425 	uint64_t syslo, syshi;
426 	ddi_dma_attr_t *attrp		= DEV_ATTR(mp);
427 	uint64_t hi		= attrp->dma_attr_addr_hi;
428 	uint64_t lo		= attrp->dma_attr_addr_lo;
429 	uint64_t align		= attrp->dma_attr_align;
430 	uint64_t nocross	= attrp->dma_attr_seg;
431 	uint64_t count_max	= attrp->dma_attr_count_max;
432 
433 	DEBUG3(DBG_DMA_ALLOCH, pci_p->pci_dip, "attrp=%p cntr_max=%x.%08x\n",
434 	    attrp, HI32(count_max), LO32(count_max));
435 	DEBUG4(DBG_DMA_ALLOCH, pci_p->pci_dip, "hi=%x.%08x lo=%x.%08x\n",
436 	    HI32(hi), LO32(hi), HI32(lo), LO32(lo));
437 	DEBUG4(DBG_DMA_ALLOCH, pci_p->pci_dip, "seg=%x.%08x align=%x.%08x\n",
438 	    HI32(nocross), LO32(nocross), HI32(align), LO32(align));
439 
440 	if (!nocross)
441 		nocross--;
442 	if (attrp->dma_attr_flags & DDI_DMA_FORCE_PHYSICAL) { /* BYPASS */
443 
444 		DEBUG0(DBG_DMA_ALLOCH, pci_p->pci_dip, "bypass mode\n");
445 		/* if tomatillo ver <= 2.3 don't allow bypass */
446 		if (tomatillo_disallow_bypass)
447 			return (DDI_DMA_BADATTR);
448 
449 		mp->dmai_flags |= DMAI_FLAGS_BYPASSREQ;
450 		if (nocross != UINT64_MAX)
451 			return (DDI_DMA_BADATTR);
452 		if (align && (align > IOMMU_PAGE_SIZE))
453 			return (DDI_DMA_BADATTR);
454 		align = 1; /* align on 1 page boundary */
455 		syslo = iommu_p->iommu_dma_bypass_base;
456 		syshi = iommu_p->iommu_dma_bypass_end;
457 
458 	} else { /* IOMMU_XLATE or PEER_TO_PEER */
459 		align = MAX(align, IOMMU_PAGE_SIZE) - 1;
460 		if ((align & nocross) != align) {
461 			dev_info_t *rdip = mp->dmai_rdip;
462 			cmn_err(CE_WARN, "%s%d dma_attr_seg not aligned",
463 			    NAMEINST(rdip));
464 			return (DDI_DMA_BADATTR);
465 		}
466 		align = IOMMU_BTOP(align + 1);
467 		syslo = iommu_p->iommu_dvma_base;
468 		syshi = iommu_p->iommu_dvma_end;
469 	}
470 	if (hi <= lo) {
471 		dev_info_t *rdip = mp->dmai_rdip;
472 		cmn_err(CE_WARN, "%s%d limits out of range", NAMEINST(rdip));
473 		return (DDI_DMA_BADATTR);
474 	}
475 	lo = MAX(lo, syslo);
476 	hi = MIN(hi, syshi);
477 	if (!count_max)
478 		count_max--;
479 
480 	DEBUG4(DBG_DMA_ALLOCH, pci_p->pci_dip, "hi=%x.%08x, lo=%x.%08x\n",
481 	    HI32(hi), LO32(hi), HI32(lo), LO32(lo));
482 	if (hi <= lo) { /* peer transfers cannot have alignment & nocross */
483 		dev_info_t *rdip = mp->dmai_rdip;
484 		cmn_err(CE_WARN, "%s%d peer only dev %p", NAMEINST(rdip), mp);
485 		if ((nocross < UINT32_MAX) || (align > 1)) {
486 			cmn_err(CE_WARN, "%s%d peer only device bad attr",
487 			    NAMEINST(rdip));
488 			return (DDI_DMA_BADATTR);
489 		}
490 		mp->dmai_flags |= DMAI_FLAGS_PEER_ONLY;
491 	} else /* set practical counter_max value */
492 		count_max = MIN(count_max, hi - lo);
493 
494 	if (DEV_NOSYSLIMIT(lo, hi, syslo, syshi, align))
495 		mp->dmai_flags |= DMAI_FLAGS_NOSYSLIMIT |
496 		    DMAI_FLAGS_NOFASTLIMIT;
497 	else {
498 		syshi = iommu_p->iommu_dvma_fast_end;
499 		if (DEV_NOFASTLIMIT(lo, hi, syslo, syshi, align))
500 			mp->dmai_flags |= DMAI_FLAGS_NOFASTLIMIT;
501 	}
502 	if (PCI_DMA_NOCTX(mp->dmai_rdip))
503 		mp->dmai_flags |= DMAI_FLAGS_NOCTX;
504 
505 	mp->dmai_minxfer	= attrp->dma_attr_minxfer;
506 	mp->dmai_burstsizes	= attrp->dma_attr_burstsizes;
507 	attrp = &mp->dmai_attr;
508 	SET_DMAATTR(attrp, lo, hi, nocross, count_max);
509 	return (DDI_SUCCESS);
510 }
511 
512 /*
513  * set up consistent dma flags according to hardware capability
514  */
515 uint32_t
516 pci_dma_consist_check(uint32_t req_flags, pbm_t *pbm_p)
517 {
518 	if (!pci_stream_buf_enable || !pci_stream_buf_exists)
519 		req_flags |= DDI_DMA_CONSISTENT;
520 	if (req_flags & DDI_DMA_CONSISTENT && !pbm_p->pbm_sync_reg_pa)
521 		req_flags |= DMP_NOSYNC;
522 	return (req_flags);
523 }
524 
525 #define	TGT_PFN_INBETWEEN(pfn, bgn, end) ((pfn >= bgn) && (pfn <= end))
526 
527 /*
528  * pci_dma_type - determine which of the three types DMA (peer-to-peer,
529  *		iommu bypass, or iommu translate) we are asked to do.
530  *		Also checks pfn0 and rejects any non-peer-to-peer
531  *		requests for peer-only devices.
532  *
533  *	return values:
534  *		DDI_DMA_NOMAPPING - can't get valid pfn0, or bad dma type
535  *		DDI_SUCCESS
536  *
537  *	dma handle members affected (set on exit):
538  *	mp->dmai_object		- dmareq->dmar_object
539  *	mp->dmai_rflags		- consistent?, nosync?, dmareq->dmar_flags
540  *	mp->dmai_flags   	- DMA type
541  *	mp->dmai_pfn0   	- 1st page pfn (if va/size pair and not shadow)
542  *	mp->dmai_roffset 	- initialized to starting IOMMU page offset
543  *	mp->dmai_ndvmapages	- # of total IOMMU pages of entire object
544  *	mp->pdh_sync_buf_pa	- dma sync buffer PA is DMA flow is supported
545  */
546 int
547 pci_dma_type(pci_t *pci_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
548 {
549 	dev_info_t *dip = pci_p->pci_dip;
550 	ddi_dma_obj_t *dobj_p = &dmareq->dmar_object;
551 	pbm_t *pbm_p = pci_p->pci_pbm_p;
552 	page_t **pplist;
553 	struct as *as_p;
554 	uint32_t offset;
555 	caddr_t vaddr;
556 	pfn_t pfn0;
557 
558 	mp->dmai_rflags = pci_dma_consist_check(dmareq->dmar_flags, pbm_p);
559 	mp->dmai_flags |= mp->dmai_rflags & DMP_NOSYNC ? DMAI_FLAGS_NOSYNC : 0;
560 
561 	switch (dobj_p->dmao_type) {
562 	case DMA_OTYP_BUFVADDR:
563 	case DMA_OTYP_VADDR: {
564 		vaddr = dobj_p->dmao_obj.virt_obj.v_addr;
565 		pplist = dobj_p->dmao_obj.virt_obj.v_priv;
566 		as_p = dobj_p->dmao_obj.virt_obj.v_as;
567 		if (as_p == NULL)
568 			as_p = &kas;
569 
570 		DEBUG2(DBG_DMA_MAP, dip, "vaddr=%p pplist=%p\n", vaddr, pplist);
571 		offset = (ulong_t)vaddr & IOMMU_PAGE_OFFSET;
572 
573 		if (pplist) {				/* shadow list */
574 			mp->dmai_flags |= DMAI_FLAGS_PGPFN;
575 			ASSERT(PAGE_LOCKED(*pplist));
576 			pfn0 = page_pptonum(*pplist);
577 		} else if (pci_dvma_remap_enabled && as_p == &kas &&
578 		    dobj_p->dmao_type != DMA_OTYP_BUFVADDR) {
579 			int (*waitfp)(caddr_t) = dmareq->dmar_fp;
580 			uint_t flags = ((waitfp == DDI_DMA_SLEEP)?
581 			    HAC_SLEEP : HAC_NOSLEEP) | HAC_PAGELOCK;
582 			int ret;
583 
584 			ret = hat_add_callback(pci_dvma_cbid, vaddr,
585 			    IOMMU_PAGE_SIZE - offset, flags, mp, &pfn0,
586 			    MP_HAT_CB_COOKIE_PTR(mp, 0));
587 
588 			if (pfn0 == PFN_INVALID && ret == ENOMEM) {
589 				ASSERT(waitfp != DDI_DMA_SLEEP);
590 				if (waitfp != DDI_DMA_DONTWAIT) {
591 					ddi_set_callback(waitfp,
592 					    dmareq->dmar_arg,
593 					    &pci_kmem_clid);
594 					return (DDI_DMA_NORESOURCES);
595 					}
596 			}
597 			mp->dmai_flags |= DMAI_FLAGS_RELOC;
598 		} else
599 			pfn0 = hat_getpfnum(as_p->a_hat, vaddr);
600 		}
601 		break;
602 
603 	case DMA_OTYP_PAGES:
604 		offset = dobj_p->dmao_obj.pp_obj.pp_offset;
605 		mp->dmai_flags |= DMAI_FLAGS_PGPFN;
606 		pfn0 = page_pptonum(dobj_p->dmao_obj.pp_obj.pp_pp);
607 		ASSERT(PAGE_LOCKED(dobj_p->dmao_obj.pp_obj.pp_pp));
608 		break;
609 
610 	case DMA_OTYP_PADDR:
611 	default:
612 		cmn_err(CE_WARN, "%s%d requested unsupported dma type %x",
613 		    NAMEINST(mp->dmai_rdip), dobj_p->dmao_type);
614 		return (DDI_DMA_NOMAPPING);
615 	}
616 	if (pfn0 == PFN_INVALID) {
617 		cmn_err(CE_WARN, "%s%d: invalid pfn0 for DMA object %p",
618 		    NAMEINST(dip), dobj_p);
619 		return (DDI_DMA_NOMAPPING);
620 	}
621 	if (TGT_PFN_INBETWEEN(pfn0, pbm_p->pbm_base_pfn, pbm_p->pbm_last_pfn)) {
622 		mp->dmai_flags |= DMAI_FLAGS_PEER_TO_PEER;
623 		goto done;	/* leave bypass and dvma flag as 0 */
624 	}
625 	if (PCI_DMA_ISPEERONLY(mp)) {
626 		dev_info_t *rdip = mp->dmai_rdip;
627 		cmn_err(CE_WARN, "Bad peer-to-peer req %s%d", NAMEINST(rdip));
628 		return (DDI_DMA_NOMAPPING);
629 	}
630 	mp->dmai_flags |= (mp->dmai_flags & DMAI_FLAGS_BYPASSREQ) ?
631 	    DMAI_FLAGS_BYPASS : DMAI_FLAGS_DVMA;
632 done:
633 	mp->dmai_object	 = *dobj_p;			/* whole object    */
634 	mp->dmai_pfn0	 = (void *)pfn0;		/* cache pfn0	   */
635 	mp->dmai_roffset = offset;			/* win0 pg0 offset */
636 	mp->dmai_ndvmapages = IOMMU_BTOPR(offset + mp->dmai_object.dmao_size);
637 
638 	return (DDI_SUCCESS);
639 }
640 
641 /*
642  * pci_dma_pgpfn - set up pfnlst array according to pages
643  *	VA/size pair: <shadow IO, bypass, peer-to-peer>, or OTYP_PAGES
644  */
645 /*ARGSUSED*/
646 static int
647 pci_dma_pgpfn(pci_t *pci_p, ddi_dma_impl_t *mp, uint_t npages)
648 {
649 	int i;
650 #ifdef DEBUG
651 	dev_info_t *dip = pci_p->pci_dip;
652 #endif
653 	switch (mp->dmai_object.dmao_type) {
654 	case DMA_OTYP_BUFVADDR:
655 	case DMA_OTYP_VADDR: {
656 		page_t **pplist = mp->dmai_object.dmao_obj.virt_obj.v_priv;
657 		DEBUG2(DBG_DMA_MAP, dip, "shadow pplist=%p, %x pages, pfns=",
658 		    pplist, npages);
659 		for (i = 1; i < npages; i++) {
660 			iopfn_t pfn = page_pptonum(pplist[i]);
661 			ASSERT(PAGE_LOCKED(pplist[i]));
662 			PCI_SET_MP_PFN1(mp, i, pfn);
663 			DEBUG1(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
664 		}
665 		DEBUG0(DBG_DMA_MAP|DBG_CONT, dip, "\n");
666 		}
667 		break;
668 
669 	case DMA_OTYP_PAGES: {
670 		page_t *pp = mp->dmai_object.dmao_obj.pp_obj.pp_pp->p_next;
671 		DEBUG1(DBG_DMA_MAP, dip, "pp=%p pfns=", pp);
672 		for (i = 1; i < npages; i++, pp = pp->p_next) {
673 			iopfn_t pfn = page_pptonum(pp);
674 			ASSERT(PAGE_LOCKED(pp));
675 			PCI_SET_MP_PFN1(mp, i, pfn);
676 			DEBUG1(DBG_DMA_MAP|DBG_CONT, dip, "%x ", pfn);
677 		}
678 		DEBUG0(DBG_DMA_MAP|DBG_CONT, dip, "\n");
679 		}
680 		break;
681 
682 	default:	/* check is already done by pci_dma_type */
683 		ASSERT(0);
684 		break;
685 	}
686 	return (DDI_SUCCESS);
687 }
688 
689 /*
690  * pci_dma_vapfn - set up pfnlst array according to VA
691  *	VA/size pair: <normal, bypass, peer-to-peer>
692  *	pfn0 is skipped as it is already done.
693  *	In this case, the cached pfn0 is used to fill pfnlst[0]
694  */
695 static int
696 pci_dma_vapfn(pci_t *pci_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp,
697 	uint_t npages)
698 {
699 	dev_info_t *dip = pci_p->pci_dip;
700 	int i;
701 	caddr_t vaddr = (caddr_t)mp->dmai_object.dmao_obj.virt_obj.v_as;
702 	struct hat *hat_p = vaddr ? ((struct as *)vaddr)->a_hat : kas.a_hat;
703 	caddr_t sva;
704 	int needcb = 0;
705 
706 	sva = (caddr_t)(((uintptr_t)mp->dmai_object.dmao_obj.virt_obj.v_addr +
707 	    IOMMU_PAGE_SIZE) & IOMMU_PAGE_MASK);
708 
709 	if (pci_dvma_remap_enabled && hat_p == kas.a_hat &&
710 	    mp->dmai_object.dmao_type != DMA_OTYP_BUFVADDR)
711 		needcb = 1;
712 
713 	for (vaddr = sva, i = 1; i < npages; i++, vaddr += IOMMU_PAGE_SIZE) {
714 		pfn_t pfn;
715 
716 		if (needcb) {
717 			int (*waitfp)(caddr_t) = dmareq->dmar_fp;
718 			uint_t flags = ((waitfp == DDI_DMA_SLEEP)?
719 			    HAC_SLEEP : HAC_NOSLEEP) | HAC_PAGELOCK;
720 			int ret;
721 
722 			ret = hat_add_callback(pci_dvma_cbid, vaddr,
723 			    IOMMU_PAGE_SIZE, flags, mp, &pfn,
724 			    MP_HAT_CB_COOKIE_PTR(mp, i));
725 
726 			if (pfn == PFN_INVALID && ret == ENOMEM) {
727 				ASSERT(waitfp != DDI_DMA_SLEEP);
728 				if (waitfp != DDI_DMA_DONTWAIT)
729 					ddi_set_callback(waitfp,
730 					    dmareq->dmar_arg, &pci_kmem_clid);
731 				return (DDI_DMA_NORESOURCES);
732 			}
733 		} else
734 			pfn = hat_getpfnum(hat_p, vaddr);
735 		if (pfn == PFN_INVALID)
736 			goto err_badpfn;
737 		PCI_SET_MP_PFN1(mp, i, (iopfn_t)pfn);
738 		DEBUG3(DBG_DMA_MAP, dip, "pci_dma_vapfn: mp=%p pfnlst[%x]=%x\n",
739 		    mp, i, (iopfn_t)pfn);
740 	}
741 	return (DDI_SUCCESS);
742 err_badpfn:
743 	cmn_err(CE_WARN, "%s%d: bad page frame vaddr=%p", NAMEINST(dip), vaddr);
744 	return (DDI_DMA_NOMAPPING);
745 }
746 
747 /*
748  * pci_dma_pfn - Fills pfn list for all pages being DMA-ed.
749  *
750  * dependencies:
751  *	mp->dmai_ndvmapages	- set to total # of dma pages
752  *
753  * return value:
754  *	DDI_SUCCESS
755  *	DDI_DMA_NOMAPPING
756  */
757 int
758 pci_dma_pfn(pci_t *pci_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
759 {
760 	uint32_t npages = mp->dmai_ndvmapages;
761 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
762 	int i, ret, peer = PCI_DMA_ISPTP(mp);
763 
764 	pbm_t *pbm_p = pci_p->pci_pbm_p;
765 	iopfn_t pfn_base = pbm_p->pbm_base_pfn;
766 	iopfn_t pfn_last = pbm_p->pbm_last_pfn;
767 	iopfn_t pfn_adj = peer ? pfn_base : 0;
768 
769 	DEBUG2(DBG_DMA_MAP, pci_p->pci_dip, "pci_dma_pfn: mp=%p pfn0=%x\n",
770 	    mp, MP_PFN0(mp) - pfn_adj);
771 	/* 1 page: no array alloc/fill, no mixed mode check */
772 	if (npages == 1) {
773 		PCI_SET_MP_PFN(mp, 0, MP_PFN0(mp) - pfn_adj);
774 		return (DDI_SUCCESS);
775 	}
776 	/* allocate pfn array */
777 	if (!(mp->dmai_pfnlst = kmem_alloc(npages * sizeof (iopfn_t),
778 	    waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP))) {
779 		if (waitfp != DDI_DMA_DONTWAIT)
780 			ddi_set_callback(waitfp, dmareq->dmar_arg,
781 			    &pci_kmem_clid);
782 		return (DDI_DMA_NORESOURCES);
783 	}
784 	/* fill pfn array */
785 	PCI_SET_MP_PFN(mp, 0, MP_PFN0(mp) - pfn_adj);	/* pfnlst[0] */
786 	if ((ret = PCI_DMA_ISPGPFN(mp) ? pci_dma_pgpfn(pci_p, mp, npages) :
787 	    pci_dma_vapfn(pci_p, dmareq, mp, npages)) != DDI_SUCCESS)
788 		goto err;
789 
790 	/* skip pfn0, check mixed mode and adjust peer to peer pfn */
791 	for (i = 1; i < npages; i++) {
792 		iopfn_t pfn = PCI_GET_MP_PFN1(mp, i);
793 		if (peer ^ TGT_PFN_INBETWEEN(pfn, pfn_base, pfn_last)) {
794 			cmn_err(CE_WARN, "%s%d mixed mode DMA %lx %lx",
795 			    NAMEINST(mp->dmai_rdip), MP_PFN0(mp), pfn);
796 			ret = DDI_DMA_NOMAPPING;	/* mixed mode */
797 			goto err;
798 		}
799 		DEBUG3(DBG_DMA_MAP, pci_p->pci_dip,
800 		    "pci_dma_pfn: pfnlst[%x]=%x-%x\n", i, pfn, pfn_adj);
801 		if (pfn_adj)
802 			PCI_SET_MP_PFN1(mp, i, pfn - pfn_adj);
803 	}
804 	return (DDI_SUCCESS);
805 err:
806 	pci_dvma_unregister_callbacks(pci_p, mp);
807 	pci_dma_freepfn(mp);
808 	return (ret);
809 }
810 
811 /*
812  * pci_dvma_win() - trim requested DVMA size down to window size
813  *	The 1st window starts from offset and ends at page-aligned boundary.
814  *	From the 2nd window on, each window starts and ends at page-aligned
815  *	boundary except the last window ends at wherever requested.
816  *
817  *	accesses the following mp-> members:
818  *	mp->dmai_attr.dma_attr_count_max
819  *	mp->dmai_attr.dma_attr_seg
820  *	mp->dmai_roffset   - start offset of 1st window
821  *	mp->dmai_rflags (redzone)
822  *	mp->dmai_ndvmapages (for 1 page fast path)
823  *
824  *	sets the following mp-> members:
825  *	mp->dmai_size	   - xfer size, != winsize if 1st/last win  (not fixed)
826  *	mp->dmai_winsize   - window size (no redzone), n * page size    (fixed)
827  *	mp->dmai_nwin	   - # of DMA windows of entire object		(fixed)
828  *	mp->dmai_rflags	   - remove partial flag if nwin == 1		(fixed)
829  *	mp->dmai_winlst	   - NULL, window objects not used for DVMA	(fixed)
830  *
831  *	fixed - not changed across different DMA windows
832  */
833 /*ARGSUSED*/
834 int
835 pci_dvma_win(pci_t *pci_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
836 {
837 	uint32_t redzone_sz	= HAS_REDZONE(mp) ? IOMMU_PAGE_SIZE : 0;
838 	size_t obj_sz	= mp->dmai_object.dmao_size;
839 	size_t xfer_sz;
840 	ulong_t pg_off;
841 
842 	if ((mp->dmai_ndvmapages == 1) && !redzone_sz) {
843 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
844 		mp->dmai_size = obj_sz;
845 		mp->dmai_winsize = IOMMU_PAGE_SIZE;
846 		mp->dmai_nwin = 1;
847 		goto done;
848 	}
849 
850 	pg_off	= mp->dmai_roffset;
851 	xfer_sz	= obj_sz + redzone_sz;
852 
853 	/* include redzone in nocross check */
854 	{
855 		uint64_t nocross = mp->dmai_attr.dma_attr_seg;
856 		if (xfer_sz + pg_off - 1 > nocross)
857 			xfer_sz = nocross - pg_off + 1;
858 		if (redzone_sz && (xfer_sz <= redzone_sz)) {
859 			DEBUG5(DBG_DMA_MAP, pci_p->pci_dip,
860 			    "nocross too small %lx(%lx)+%lx+%x < %" PRIx64 "\n",
861 			    xfer_sz, obj_sz, pg_off, redzone_sz, nocross);
862 			return (DDI_DMA_TOOBIG);
863 		}
864 	}
865 	xfer_sz -= redzone_sz;	/* restore transfer size  */
866 	/* check counter max */
867 	{
868 		uint32_t count_max = mp->dmai_attr.dma_attr_count_max;
869 		if (xfer_sz - 1 > count_max)
870 			xfer_sz = count_max + 1;
871 	}
872 	if (xfer_sz >= obj_sz) {
873 		mp->dmai_rflags &= ~DDI_DMA_PARTIAL;
874 		mp->dmai_size = xfer_sz;
875 		mp->dmai_winsize = P2ROUNDUP(xfer_sz + pg_off, IOMMU_PAGE_SIZE);
876 		mp->dmai_nwin = 1;
877 		goto done;
878 	}
879 	if (!(dmareq->dmar_flags & DDI_DMA_PARTIAL)) {
880 		DEBUG4(DBG_DMA_MAP, pci_p->pci_dip,
881 		    "too big: %lx+%lx+%x > %lx\n",
882 		    obj_sz, pg_off, redzone_sz, xfer_sz);
883 		return (DDI_DMA_TOOBIG);
884 	}
885 
886 	xfer_sz = IOMMU_PTOB(IOMMU_BTOP(xfer_sz + pg_off)); /* page align */
887 	mp->dmai_size = xfer_sz - pg_off;	/* 1st window xferrable size */
888 	mp->dmai_winsize = xfer_sz;		/* redzone not in winsize */
889 	mp->dmai_nwin = (obj_sz + pg_off + xfer_sz - 1) / xfer_sz;
890 done:
891 	mp->dmai_winlst = NULL;
892 	dump_dma_handle(DBG_DMA_MAP, pci_p->pci_dip, mp);
893 	return (DDI_SUCCESS);
894 }
895 
896 /*
897  * fast track cache entry to iommu context, inserts 3 0 bits between
898  * upper 6-bits and lower 3-bits of the 9-bit cache entry
899  */
900 #define	IOMMU_FCE_TO_CTX(i)	(((i) << 3) | ((i) & 0x7) | 0x38)
901 
902 /*
903  * pci_dvma_map_fast - attempts to map fast trackable DVMA
904  */
905 int
906 pci_dvma_map_fast(iommu_t *iommu_p, ddi_dma_impl_t *mp)
907 {
908 	uint_t clustsz = pci_dvma_page_cache_clustsz;
909 	uint_t entries = pci_dvma_page_cache_entries;
910 	uint64_t *tte_addr;
911 	uint64_t tte = GET_TTE_TEMPLATE(mp);
912 	int i = iommu_p->iommu_dvma_addr_scan_start;
913 	uint8_t *lock_addr = iommu_p->iommu_dvma_cache_locks + i;
914 	iopfn_t *pfn_addr;
915 	dvma_addr_t dvma_pg;
916 	size_t npages = IOMMU_BTOP(mp->dmai_winsize);
917 #ifdef DEBUG
918 	dev_info_t *dip = mp->dmai_rdip;
919 #endif
920 	extern uint8_t ldstub(uint8_t *);
921 	ASSERT(IOMMU_PTOB(npages) == mp->dmai_winsize);
922 	ASSERT(npages + HAS_REDZONE(mp) <= clustsz);
923 
924 	for (; i < entries && ldstub(lock_addr); i++, lock_addr++)
925 		;
926 	if (i >= entries) {
927 		lock_addr = iommu_p->iommu_dvma_cache_locks;
928 		i = 0;
929 		for (; i < entries && ldstub(lock_addr); i++, lock_addr++)
930 			;
931 		if (i >= entries) {
932 #ifdef PCI_DMA_PROF
933 			pci_dvmaft_exhaust++;
934 #endif
935 			return (DDI_DMA_NORESOURCES);
936 		}
937 	}
938 	iommu_p->iommu_dvma_addr_scan_start = (i + 1) & (entries - 1);
939 	if (PCI_DMA_USECTX(mp)) {
940 		dvma_context_t ctx = IOMMU_FCE_TO_CTX(i);
941 		tte |= IOMMU_CTX2TTE(ctx);
942 		mp->dmai_flags |= DMAI_FLAGS_CONTEXT;
943 		DEBUG1(DBG_DMA_MAP, dip, "fast: ctx=0x%x\n", ctx);
944 	}
945 	i *= clustsz;
946 	tte_addr = iommu_p->iommu_tsb_vaddr + i;
947 	dvma_pg = iommu_p->dvma_base_pg + i;
948 #ifdef DEBUG
949 	for (i = 0; i < clustsz; i++)
950 		ASSERT(TTE_IS_INVALID(tte_addr[i]));
951 #endif
952 	*tte_addr = tte | IOMMU_PTOB(MP_PFN0(mp)); /* map page 0 */
953 	DEBUG5(DBG_DMA_MAP, dip, "fast %p:dvma_pg=%x tte0(%p)=%08x.%08x\n", mp,
954 	    dvma_pg, tte_addr, HI32(*tte_addr), LO32(*tte_addr));
955 	if (npages == 1)
956 		goto tte_done;
957 	pfn_addr = PCI_GET_MP_PFN1_ADDR(mp); /* short iommu_map_pages() */
958 	for (tte_addr++, i = 1; i < npages; i++, tte_addr++, pfn_addr++) {
959 		*tte_addr = tte | IOMMU_PTOB(*pfn_addr);
960 		DEBUG5(DBG_DMA_MAP, dip, "fast %p:tte(%p, %p)=%08x.%08x\n", mp,
961 		    tte_addr, pfn_addr, HI32(*tte_addr), LO32(*tte_addr));
962 	}
963 tte_done:
964 #ifdef PCI_DMA_PROF
965 	pci_dvmaft_success++;
966 #endif
967 	mp->dmai_mapping = mp->dmai_roffset | IOMMU_PTOB(dvma_pg);
968 	mp->dmai_offset = 0;
969 	mp->dmai_flags |= DMAI_FLAGS_FASTTRACK;
970 	PCI_SAVE_MP_TTE(mp, tte);	/* save TTE template for unmapping */
971 	if (DVMA_DBG_ON(iommu_p))
972 		pci_dvma_alloc_debug(iommu_p, (char *)mp->dmai_mapping,
973 		    mp->dmai_size, mp);
974 	return (DDI_SUCCESS);
975 }
976 
977 /*
978  * pci_dvma_map: map non-fasttrack DMA
979  *		Use quantum cache if single page DMA.
980  */
981 int
982 pci_dvma_map(ddi_dma_impl_t *mp, ddi_dma_req_t *dmareq, iommu_t *iommu_p)
983 {
984 	uint_t npages = PCI_DMA_WINNPGS(mp);
985 	dvma_addr_t dvma_pg, dvma_pg_index;
986 	void *dvma_addr;
987 	uint64_t tte = GET_TTE_TEMPLATE(mp);
988 	int sleep = dmareq->dmar_fp == DDI_DMA_SLEEP ? VM_SLEEP : VM_NOSLEEP;
989 #ifdef DEBUG
990 	dev_info_t *dip = mp->dmai_rdip;
991 #endif
992 	/*
993 	 * allocate dvma space resource and map in the first window.
994 	 * (vmem_t *vmp, size_t size,
995 	 *	size_t align, size_t phase, size_t nocross,
996 	 *	void *minaddr, void *maxaddr, int vmflag)
997 	 */
998 	if ((npages == 1) && !HAS_REDZONE(mp) && HAS_NOSYSLIMIT(mp)) {
999 		dvma_addr = vmem_alloc(iommu_p->iommu_dvma_map,
1000 		    IOMMU_PAGE_SIZE, sleep);
1001 		mp->dmai_flags |= DMAI_FLAGS_VMEMCACHE;
1002 #ifdef PCI_DMA_PROF
1003 		pci_dvma_vmem_alloc++;
1004 #endif
1005 	} else {
1006 		dvma_addr = vmem_xalloc(iommu_p->iommu_dvma_map,
1007 		    IOMMU_PTOB(npages + HAS_REDZONE(mp)),
1008 		    MAX(mp->dmai_attr.dma_attr_align, IOMMU_PAGE_SIZE),
1009 		    0,
1010 		    mp->dmai_attr.dma_attr_seg + 1,
1011 		    (void *)mp->dmai_attr.dma_attr_addr_lo,
1012 		    (void *)(mp->dmai_attr.dma_attr_addr_hi + 1),
1013 		    sleep);
1014 #ifdef PCI_DMA_PROF
1015 		pci_dvma_vmem_xalloc++;
1016 #endif
1017 	}
1018 	dvma_pg = IOMMU_BTOP((ulong_t)dvma_addr);
1019 	dvma_pg_index = dvma_pg - iommu_p->dvma_base_pg;
1020 	DEBUG2(DBG_DMA_MAP, dip, "fallback dvma_pages: dvma_pg=%x index=%x\n",
1021 	    dvma_pg, dvma_pg_index);
1022 	if (dvma_pg == 0)
1023 		goto noresource;
1024 
1025 	/* allocate DVMA context */
1026 	if ((npages >= pci_context_minpages) && PCI_DMA_USECTX(mp)) {
1027 		dvma_context_t ctx;
1028 		if (ctx = pci_iommu_get_dvma_context(iommu_p, dvma_pg_index)) {
1029 			tte |= IOMMU_CTX2TTE(ctx);
1030 			mp->dmai_flags |= DMAI_FLAGS_CONTEXT;
1031 		}
1032 	}
1033 	mp->dmai_mapping = mp->dmai_roffset | IOMMU_PTOB(dvma_pg);
1034 	mp->dmai_offset = 0;
1035 	PCI_SAVE_MP_TTE(mp, tte);	/* mp->dmai_tte = tte */
1036 	iommu_map_pages(iommu_p, mp, dvma_pg, npages, 0);
1037 	return (DDI_SUCCESS);
1038 noresource:
1039 	if (dmareq->dmar_fp != DDI_DMA_DONTWAIT) {
1040 		DEBUG0(DBG_DMA_MAP, dip, "dvma_pg 0 - set callback\n");
1041 		ddi_set_callback(dmareq->dmar_fp, dmareq->dmar_arg,
1042 		    &iommu_p->iommu_dvma_clid);
1043 	}
1044 	DEBUG0(DBG_DMA_MAP, dip, "vmem_xalloc - DDI_DMA_NORESOURCES\n");
1045 	return (DDI_DMA_NORESOURCES);
1046 }
1047 
1048 void
1049 pci_dvma_unmap(iommu_t *iommu_p, ddi_dma_impl_t *mp)
1050 {
1051 	size_t npages;
1052 	dvma_addr_t dvma_addr = (dvma_addr_t)mp->dmai_mapping;
1053 	dvma_addr_t dvma_pg = IOMMU_BTOP(dvma_addr);
1054 	dvma_addr = IOMMU_PTOB(dvma_pg);
1055 
1056 	if (mp->dmai_flags & DMAI_FLAGS_FASTTRACK) {
1057 		iopfn_t index = dvma_pg - iommu_p->dvma_base_pg;
1058 		ASSERT(index % pci_dvma_page_cache_clustsz == 0);
1059 		index /= pci_dvma_page_cache_clustsz;
1060 		ASSERT(index < pci_dvma_page_cache_entries);
1061 		iommu_p->iommu_dvma_cache_locks[index] = 0;
1062 #ifdef PCI_DMA_PROF
1063 		pci_dvmaft_free++;
1064 #endif
1065 		return;
1066 	}
1067 	npages = IOMMU_BTOP(mp->dmai_winsize) + HAS_REDZONE(mp);
1068 	pci_vmem_free(iommu_p, mp, (void *)dvma_addr, npages);
1069 
1070 	if (mp->dmai_flags & DMAI_FLAGS_CONTEXT)
1071 		pci_iommu_free_dvma_context(iommu_p, MP2CTX(mp));
1072 }
1073 
1074 void
1075 pci_dma_sync_unmap(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp)
1076 {
1077 	pci_t *pci_p = get_pci_soft_state(ddi_get_instance(dip));
1078 	iommu_t *iommu_p = pci_p->pci_iommu_p;
1079 	uint64_t sync_buf_save = SYNC_BUF_PA(mp);
1080 	uint32_t fast_track = mp->dmai_flags & DMAI_FLAGS_FASTTRACK;
1081 
1082 	if (fast_track) {
1083 		dvma_addr_t dvma_pg = IOMMU_BTOP(mp->dmai_mapping);
1084 
1085 		SYNC_BUF_PA(mp) = IOMMU_PAGE_TTEPA(iommu_p, dvma_pg);
1086 		ASSERT(!(SYNC_BUF_PA(mp) & PCI_SYNC_FLAG_SIZE - 1));
1087 	}
1088 
1089 	if (pci_dvma_sync_before_unmap) {
1090 		pci_dma_sync(dip, rdip, (ddi_dma_handle_t)mp, 0, 0,
1091 		    DDI_DMA_SYNC_FORCPU);
1092 		iommu_unmap_window(iommu_p, mp);
1093 	} else {
1094 		iommu_unmap_window(iommu_p, mp);
1095 		pci_dma_sync(dip, rdip, (ddi_dma_handle_t)mp, 0, 0,
1096 		    DDI_DMA_SYNC_FORCPU);
1097 	}
1098 
1099 	if (fast_track)
1100 		SYNC_BUF_PA(mp) = sync_buf_save;
1101 }
1102 
1103 /*
1104  * DVMA mappings may have multiple windows, but each window always have
1105  * one segment.
1106  */
1107 int
1108 pci_dvma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
1109 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
1110 	uint_t cache_flags)
1111 {
1112 	switch (cmd) {
1113 	case DDI_DMA_SYNC:
1114 		return (pci_dma_sync(dip, rdip, (ddi_dma_handle_t)mp,
1115 		    *offp, *lenp, cache_flags));
1116 
1117 	case DDI_DMA_HTOC: {
1118 		int ret;
1119 		off_t wo_off, off = *offp;	/* wo_off: wnd's obj offset */
1120 		uint_t win_size = mp->dmai_winsize;
1121 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)objp;
1122 
1123 		if (off >= mp->dmai_object.dmao_size) {
1124 			cmn_err(CE_WARN, "%s%d invalid dma_htoc offset %lx",
1125 			    NAMEINST(mp->dmai_rdip), off);
1126 			return (DDI_FAILURE);
1127 		}
1128 		off += mp->dmai_roffset;
1129 		ret = pci_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
1130 		    off / win_size, &wo_off, NULL, cp, NULL); /* lenp == NULL */
1131 		if (ret)
1132 			return (ret);
1133 		DEBUG4(DBG_DMA_CTL, dip, "HTOC:cookie=%x+%lx off=%lx,%lx\n",
1134 		    cp->dmac_address, cp->dmac_size, off, *offp);
1135 
1136 		/* adjust cookie addr/len if we are not on window boundary */
1137 		ASSERT((off % win_size) == (off -
1138 		    (PCI_DMA_CURWIN(mp) ? mp->dmai_roffset : 0) - wo_off));
1139 		off = PCI_DMA_CURWIN(mp) ? off % win_size : *offp;
1140 		ASSERT(cp->dmac_size > off);
1141 		cp->dmac_laddress += off;
1142 		cp->dmac_size -= off;
1143 		DEBUG5(DBG_DMA_CTL, dip,
1144 		    "HTOC:mp=%p cookie=%x+%lx off=%lx,%lx\n",
1145 		    mp, cp->dmac_address, cp->dmac_size, off, wo_off);
1146 		}
1147 		return (DDI_SUCCESS);
1148 
1149 	case DDI_DMA_REPWIN:
1150 		*offp = mp->dmai_offset;
1151 		*lenp = mp->dmai_size;
1152 		return (DDI_SUCCESS);
1153 
1154 	case DDI_DMA_MOVWIN: {
1155 		off_t off = *offp;
1156 		if (off >= mp->dmai_object.dmao_size)
1157 			return (DDI_FAILURE);
1158 		off += mp->dmai_roffset;
1159 		return (pci_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
1160 		    off / mp->dmai_winsize, offp, lenp,
1161 		    (ddi_dma_cookie_t *)objp, NULL));
1162 		}
1163 
1164 	case DDI_DMA_NEXTWIN: {
1165 		window_t win = PCI_DMA_CURWIN(mp);
1166 		if (offp) {
1167 			if (*(window_t *)offp != win) {  /* window not active */
1168 				*(window_t *)objp = win; /* return cur win */
1169 				return (DDI_DMA_STALE);
1170 			}
1171 			win++;
1172 		} else	/* map win 0 */
1173 			win = 0;
1174 		if (win >= mp->dmai_nwin) {
1175 			*(window_t *)objp = win - 1;
1176 			return (DDI_DMA_DONE);
1177 		}
1178 		if (pci_dma_win(dip, rdip, (ddi_dma_handle_t)mp,
1179 		    win, 0, 0, 0, 0)) {
1180 			*(window_t *)objp = win - 1;
1181 			return (DDI_FAILURE);
1182 		}
1183 		*(window_t *)objp = win;
1184 		}
1185 		return (DDI_SUCCESS);
1186 
1187 	case DDI_DMA_NEXTSEG:
1188 		if (*(window_t *)offp != PCI_DMA_CURWIN(mp))
1189 			return (DDI_DMA_STALE);
1190 		if (lenp)				/* only 1 seg allowed */
1191 			return (DDI_DMA_DONE);
1192 							/* return mp as seg 0 */
1193 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
1194 		return (DDI_SUCCESS);
1195 
1196 	case DDI_DMA_SEGTOC:
1197 		MAKE_DMA_COOKIE((ddi_dma_cookie_t *)objp, mp->dmai_mapping,
1198 		    mp->dmai_size);
1199 		*offp = mp->dmai_offset;
1200 		*lenp = mp->dmai_size;
1201 		return (DDI_SUCCESS);
1202 
1203 	case DDI_DMA_COFF: {
1204 		ddi_dma_cookie_t *cp = (ddi_dma_cookie_t *)offp;
1205 		if (cp->dmac_address < mp->dmai_mapping ||
1206 		    (cp->dmac_address + cp->dmac_size) >
1207 		    (mp->dmai_mapping + mp->dmai_size))
1208 			return (DDI_FAILURE);
1209 		*objp = (caddr_t)(cp->dmac_address - mp->dmai_mapping +
1210 		    mp->dmai_offset);
1211 		}
1212 		return (DDI_SUCCESS);
1213 
1214 	case DDI_DMA_REMAP:
1215 		if (pci_dvma_remap_enabled)
1216 			return (pci_dvma_remap(dip, rdip, mp, *offp, *lenp));
1217 		return (DDI_FAILURE);
1218 
1219 	default:
1220 		DEBUG3(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
1221 		    cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
1222 		break;
1223 	}
1224 	return (DDI_FAILURE);
1225 }
1226 
1227 void
1228 pci_dma_freewin(ddi_dma_impl_t *mp)
1229 {
1230 	pci_dma_win_t *win_p = mp->dmai_winlst, *win2_p;
1231 	for (win2_p = win_p; win_p; win2_p = win_p) {
1232 		win_p = win2_p->win_next;
1233 		kmem_free(win2_p, sizeof (pci_dma_win_t) +
1234 		    sizeof (ddi_dma_cookie_t) * win2_p->win_ncookies);
1235 	}
1236 	mp->dmai_nwin = 0;
1237 	mp->dmai_winlst = NULL;
1238 }
1239 
1240 /*
1241  * pci_dma_newwin - create a dma window object and cookies
1242  *
1243  *	After the initial scan in pci_dma_physwin(), which identifies
1244  *	a portion of the pfn array that belongs to a dma window,
1245  *	we are called to allocate and initialize representing memory
1246  *	resources. We know from the 1st scan the number of cookies
1247  *	or dma segment in this window so we can allocate a contiguous
1248  *	memory array for the dma cookies (The implementation of
1249  *	ddi_dma_nextcookie(9f) dictates dma cookies be contiguous).
1250  *
1251  *	A second round scan is done on the pfn array to identify
1252  *	each dma segment and initialize its corresponding dma cookie.
1253  *	We don't need to do all the safety checking and we know they
1254  *	all belong to the same dma window.
1255  *
1256  *	Input:	cookie_no - # of cookies identified by the 1st scan
1257  *		start_idx - subscript of the pfn array for the starting pfn
1258  *		end_idx   - subscript of the last pfn in dma window
1259  *		win_pp    - pointer to win_next member of previous window
1260  *	Return:	DDI_SUCCESS - with **win_pp as newly created window object
1261  *		DDI_DMA_NORESROUCE - caller frees all previous window objs
1262  *	Note:	Each cookie and window size are all initialized on page
1263  *		boundary. This is not true for the 1st cookie of the 1st
1264  *		window and the last cookie of the last window.
1265  *		We fix that later in upper layer which has access to size
1266  *		and offset info.
1267  *
1268  */
1269 static int
1270 pci_dma_newwin(ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp, uint32_t cookie_no,
1271 	uint32_t start_idx, uint32_t end_idx, pci_dma_win_t **win_pp,
1272 	uint64_t count_max, uint64_t bypass_prefix)
1273 {
1274 	int (*waitfp)(caddr_t) = dmareq->dmar_fp;
1275 	ddi_dma_cookie_t *cookie_p;
1276 	uint32_t pfn_no = 1;
1277 	iopfn_t pfn = PCI_GET_MP_PFN(mp, start_idx);
1278 	iopfn_t prev_pfn = pfn;
1279 	uint64_t seg_pfn0 = pfn;
1280 	size_t sz = cookie_no * sizeof (ddi_dma_cookie_t);
1281 	pci_dma_win_t *win_p = kmem_alloc(sizeof (pci_dma_win_t) + sz,
1282 	    waitfp == DDI_DMA_SLEEP ? KM_SLEEP : KM_NOSLEEP);
1283 	if (!win_p)
1284 		goto noresource;
1285 
1286 	win_p->win_next = NULL;
1287 	win_p->win_ncookies = cookie_no;
1288 	win_p->win_curseg = 0;	/* start from segment 0 */
1289 	win_p->win_size = IOMMU_PTOB(end_idx - start_idx + 1);
1290 	/* win_p->win_offset is left uninitialized */
1291 
1292 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1293 	start_idx++;
1294 	for (; start_idx <= end_idx; start_idx++, prev_pfn = pfn, pfn_no++) {
1295 		pfn = PCI_GET_MP_PFN1(mp, start_idx);
1296 		if ((pfn == prev_pfn + 1) &&
1297 		    (IOMMU_PTOB(pfn_no + 1) - 1 <= count_max))
1298 			continue;
1299 
1300 		/* close up the cookie up to (including) prev_pfn */
1301 		MAKE_DMA_COOKIE(cookie_p, IOMMU_PTOB(seg_pfn0) | bypass_prefix,
1302 		    IOMMU_PTOB(pfn_no));
1303 		DEBUG2(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages)\n",
1304 		    IOMMU_PTOB(seg_pfn0) | bypass_prefix, pfn_no);
1305 
1306 		cookie_p++;	/* advance to next available cookie cell */
1307 		pfn_no = 0;
1308 		seg_pfn0 = pfn;	/* start a new segment from current pfn */
1309 	}
1310 	MAKE_DMA_COOKIE(cookie_p, IOMMU_PTOB(seg_pfn0) | bypass_prefix,
1311 	    IOMMU_PTOB(pfn_no));
1312 	DEBUG3(DBG_BYPASS, mp->dmai_rdip, "cookie %p (%x pages) of total %x\n",
1313 	    IOMMU_PTOB(seg_pfn0) | bypass_prefix, pfn_no, cookie_no);
1314 #ifdef DEBUG
1315 	cookie_p++;
1316 	ASSERT((cookie_p - (ddi_dma_cookie_t *)(win_p + 1)) == cookie_no);
1317 #endif
1318 	*win_pp = win_p;
1319 	return (DDI_SUCCESS);
1320 noresource:
1321 	if (waitfp != DDI_DMA_DONTWAIT)
1322 		ddi_set_callback(waitfp, dmareq->dmar_arg, &pci_kmem_clid);
1323 	return (DDI_DMA_NORESOURCES);
1324 }
1325 
1326 /*
1327  * pci_dma_adjust - adjust 1st and last cookie and window sizes
1328  *	remove initial dma page offset from 1st cookie and window size
1329  *	remove last dma page remainder from last cookie and window size
1330  *	fill win_offset of each dma window according to just fixed up
1331  *		each window sizes
1332  *	pci_dma_win_t members modified:
1333  *	win_p->win_offset - this window's offset within entire DMA object
1334  *	win_p->win_size	  - xferrable size (in bytes) for this window
1335  *
1336  *	ddi_dma_impl_t members modified:
1337  *	mp->dmai_size	  - 1st window xferrable size
1338  *	mp->dmai_offset   - 0, which is the dma offset of the 1st window
1339  *
1340  *	ddi_dma_cookie_t members modified:
1341  *	cookie_p->dmac_size - 1st and last cookie remove offset or remainder
1342  *	cookie_p->dmac_laddress - 1st cookie add page offset
1343  */
1344 static void
1345 pci_dma_adjust(ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp, pci_dma_win_t *win_p)
1346 {
1347 	ddi_dma_cookie_t *cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1348 	size_t pg_offset = mp->dmai_roffset;
1349 	size_t win_offset = 0;
1350 
1351 	cookie_p->dmac_size -= pg_offset;
1352 	cookie_p->dmac_laddress |= pg_offset;
1353 	win_p->win_size -= pg_offset;
1354 	DEBUG1(DBG_BYPASS, mp->dmai_rdip, "pg0 adjust %lx\n", pg_offset);
1355 
1356 	mp->dmai_size = win_p->win_size;
1357 	mp->dmai_offset = 0;
1358 
1359 	pg_offset += mp->dmai_object.dmao_size;
1360 	pg_offset &= IOMMU_PAGE_OFFSET;
1361 	if (pg_offset)
1362 		pg_offset = IOMMU_PAGE_SIZE - pg_offset;
1363 	DEBUG1(DBG_BYPASS, mp->dmai_rdip, "last pg adjust %lx\n", pg_offset);
1364 
1365 	for (; win_p->win_next; win_p = win_p->win_next) {
1366 		DEBUG1(DBG_BYPASS, mp->dmai_rdip, "win off %p\n", win_offset);
1367 		win_p->win_offset = win_offset;
1368 		win_offset += win_p->win_size;
1369 	}
1370 	/* last window */
1371 	win_p->win_offset = win_offset;
1372 	cookie_p = (ddi_dma_cookie_t *)(win_p + 1);
1373 	cookie_p[win_p->win_ncookies - 1].dmac_size -= pg_offset;
1374 	win_p->win_size -= pg_offset;
1375 	ASSERT((win_offset + win_p->win_size) == mp->dmai_object.dmao_size);
1376 }
1377 
1378 /*
1379  * pci_dma_physwin() - carve up dma windows using physical addresses.
1380  *	Called to handle iommu bypass and pci peer-to-peer transfers.
1381  *	Calls pci_dma_newwin() to allocate window objects.
1382  *
1383  * Dependency: mp->dmai_pfnlst points to an array of pfns
1384  *
1385  * 1. Each dma window is represented by a pci_dma_win_t object.
1386  *	The object will be casted to ddi_dma_win_t and returned
1387  *	to leaf driver through the DDI interface.
1388  * 2. Each dma window can have several dma segments with each
1389  *	segment representing a physically contiguous either memory
1390  *	space (if we are doing an iommu bypass transfer) or pci address
1391  *	space (if we are doing a peer-to-peer transfer).
1392  * 3. Each segment has a DMA cookie to program the DMA engine.
1393  *	The cookies within each DMA window must be located in a
1394  *	contiguous array per ddi_dma_nextcookie(9f).
1395  * 4. The number of DMA segments within each DMA window cannot exceed
1396  *	mp->dmai_attr.dma_attr_sgllen. If the transfer size is
1397  *	too large to fit in the sgllen, the rest needs to be
1398  *	relocated to the next dma window.
1399  * 5. Peer-to-peer DMA segment follows device hi, lo, count_max,
1400  *	and nocross restrictions while bypass DMA follows the set of
1401  *	restrictions with system limits factored in.
1402  *
1403  * Return:
1404  *	mp->dmai_winlst	 - points to a link list of pci_dma_win_t objects.
1405  *		Each pci_dma_win_t object on the link list contains
1406  *		infomation such as its window size (# of pages),
1407  *		starting offset (also see Restriction), an array of
1408  *		DMA cookies, and # of cookies in the array.
1409  *	mp->dmai_pfnlst	 - NULL, the pfn list is freed to conserve memory.
1410  *	mp->dmai_nwin	 - # of total DMA windows on mp->dmai_winlst.
1411  *	mp->dmai_mapping - starting cookie address
1412  *	mp->dmai_rflags	 - consistent, nosync, no redzone
1413  *	mp->dmai_cookie	 - start of cookie table of the 1st DMA window
1414  *
1415  * Restriction:
1416  *	Each pci_dma_win_t object can theoratically start from any offset
1417  *	since the iommu is not involved. However, this implementation
1418  *	always make windows start from page aligned offset (except
1419  *	the 1st window, which follows the requested offset) due to the
1420  *	fact that we are handed a pfn list. This does require device's
1421  *	count_max and attr_seg to be at least IOMMU_PAGE_SIZE aligned.
1422  */
1423 int
1424 pci_dma_physwin(pci_t *pci_p, ddi_dma_req_t *dmareq, ddi_dma_impl_t *mp)
1425 {
1426 	uint_t npages = mp->dmai_ndvmapages;
1427 	int ret, sgllen = mp->dmai_attr.dma_attr_sgllen;
1428 	iopfn_t pfn_lo, pfn_hi, prev_pfn, bypass_pfn;
1429 	iopfn_t pfn = PCI_GET_MP_PFN(mp, 0);
1430 	uint32_t i, win_no = 0, pfn_no = 1, win_pfn0_index = 0, cookie_no = 0;
1431 	uint64_t count_max, bypass = PCI_DMA_BYPASS_PREFIX(mp, pfn);
1432 	pci_dma_win_t **win_pp = (pci_dma_win_t **)&mp->dmai_winlst;
1433 	ddi_dma_cookie_t *cookie0_p;
1434 
1435 	if (PCI_DMA_ISPTP(mp)) { /* ignore sys limits for peer-to-peer */
1436 		ddi_dma_attr_t *dev_attr_p = DEV_ATTR(mp);
1437 		iopfn_t pfn_base = pci_p->pci_pbm_p->pbm_base_pfn;
1438 		iopfn_t pfn_last = pci_p->pci_pbm_p->pbm_last_pfn - pfn_base;
1439 		uint64_t nocross = dev_attr_p->dma_attr_seg;
1440 		if (nocross && (nocross < UINT32_MAX))
1441 			return (DDI_DMA_NOMAPPING);
1442 		if (dev_attr_p->dma_attr_align > IOMMU_PAGE_SIZE)
1443 			return (DDI_DMA_NOMAPPING);
1444 		pfn_lo = IOMMU_BTOP(dev_attr_p->dma_attr_addr_lo);
1445 		pfn_hi = IOMMU_BTOP(dev_attr_p->dma_attr_addr_hi);
1446 		pfn_hi = MIN(pfn_hi, pfn_last);
1447 		if ((pfn_lo > pfn_hi) || (pfn < pfn_lo))
1448 			return (DDI_DMA_NOMAPPING);
1449 		count_max = dev_attr_p->dma_attr_count_max;
1450 		count_max = MIN(count_max, nocross);
1451 		/*
1452 		 * the following count_max trim is not done because we are
1453 		 * making sure pfn_lo <= pfn <= pfn_hi inside the loop
1454 		 * count_max=MIN(count_max, IOMMU_PTOB(pfn_hi - pfn_lo + 1)-1);
1455 		 */
1456 	} else { /* bypass hi/lo/count_max have been processed by attr2hdl() */
1457 		count_max = mp->dmai_attr.dma_attr_count_max;
1458 		pfn_lo = IOMMU_BTOP(mp->dmai_attr.dma_attr_addr_lo);
1459 		pfn_hi = IOMMU_BTOP(mp->dmai_attr.dma_attr_addr_hi);
1460 	}
1461 
1462 	bypass_pfn = IOMMU_BTOP(bypass);
1463 
1464 	for (prev_pfn = (bypass_pfn | pfn), i = 1; i < npages;
1465 	    i++, prev_pfn = pfn, pfn_no++) {
1466 		pfn = bypass_pfn | PCI_GET_MP_PFN1(mp, i);
1467 		if ((pfn == prev_pfn + 1) &&
1468 		    (IOMMU_PTOB(pfn_no + 1) - 1 <= count_max))
1469 			continue;
1470 		if ((pfn < pfn_lo) || (prev_pfn > pfn_hi)) {
1471 			ret = DDI_DMA_NOMAPPING;
1472 			goto err;
1473 		}
1474 		cookie_no++;
1475 		pfn_no = 0;
1476 		if (cookie_no < sgllen)
1477 			continue;
1478 
1479 		DEBUG3(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1480 		    win_pfn0_index, i - 1, cookie_no);
1481 		if (ret = pci_dma_newwin(dmareq, mp, cookie_no,
1482 		    win_pfn0_index, i - 1, win_pp, count_max, bypass))
1483 			goto err;
1484 
1485 		win_pp = &(*win_pp)->win_next;	/* win_pp = *(win_pp) */
1486 		win_no++;
1487 		win_pfn0_index = i;
1488 		cookie_no = 0;
1489 	}
1490 	if (pfn > pfn_hi) {
1491 		ret = DDI_DMA_NOMAPPING;
1492 		goto err;
1493 	}
1494 	cookie_no++;
1495 	DEBUG3(DBG_BYPASS, mp->dmai_rdip, "newwin pfn[%x-%x] %x cks\n",
1496 	    win_pfn0_index, i - 1, cookie_no);
1497 	if (ret = pci_dma_newwin(dmareq, mp, cookie_no, win_pfn0_index,
1498 	    i - 1, win_pp, count_max, bypass))
1499 		goto err;
1500 	win_no++;
1501 	pci_dma_adjust(dmareq, mp, mp->dmai_winlst);
1502 	mp->dmai_nwin = win_no;
1503 	mp->dmai_rflags |= DDI_DMA_CONSISTENT;
1504 	if (!pci_p->pci_pbm_p->pbm_sync_reg_pa) {
1505 		mp->dmai_rflags |= DMP_NOSYNC;
1506 		mp->dmai_flags |= DMAI_FLAGS_NOSYNC;
1507 	}
1508 	mp->dmai_rflags &= ~DDI_DMA_REDZONE;
1509 	cookie0_p = (ddi_dma_cookie_t *)(WINLST(mp) + 1);
1510 	mp->dmai_cookie = WINLST(mp)->win_ncookies > 1 ? cookie0_p + 1 : 0;
1511 	mp->dmai_mapping = cookie0_p->dmac_laddress;
1512 
1513 	pci_dma_freepfn(mp);
1514 	return (DDI_DMA_MAPPED);
1515 err:
1516 	pci_dma_freewin(mp);
1517 	return (ret);
1518 }
1519 
1520 /*ARGSUSED*/
1521 int
1522 pci_dma_ctl(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
1523 	enum ddi_dma_ctlops cmd, off_t *offp, size_t *lenp, caddr_t *objp,
1524 	uint_t cache_flags)
1525 {
1526 	switch (cmd) {
1527 	case DDI_DMA_SYNC: /* XXX */
1528 		return (DDI_SUCCESS);
1529 
1530 	case DDI_DMA_HTOC: {
1531 		off_t off = *offp;
1532 		ddi_dma_cookie_t *loop_cp, *cp;
1533 		pci_dma_win_t *win_p = mp->dmai_winlst;
1534 
1535 		if (off >= mp->dmai_object.dmao_size)
1536 			return (DDI_FAILURE);
1537 
1538 		/* locate window */
1539 		while (win_p->win_offset + win_p->win_size <= off)
1540 			win_p = win_p->win_next;
1541 
1542 		loop_cp = cp = (ddi_dma_cookie_t *)(win_p + 1);
1543 		mp->dmai_offset = win_p->win_offset;
1544 		mp->dmai_size   = win_p->win_size;
1545 		mp->dmai_mapping = cp->dmac_laddress; /* cookie0 start addr */
1546 
1547 		/* adjust cookie addr/len if we are not on cookie boundary */
1548 		off -= win_p->win_offset;	   /* offset within window */
1549 		for (; off >= loop_cp->dmac_size; loop_cp++)
1550 			off -= loop_cp->dmac_size; /* offset within cookie */
1551 
1552 		mp->dmai_cookie = loop_cp + 1;
1553 		win_p->win_curseg = loop_cp - cp;
1554 		cp = (ddi_dma_cookie_t *)objp;
1555 		MAKE_DMA_COOKIE(cp, loop_cp->dmac_laddress + off,
1556 		    loop_cp->dmac_size - off);
1557 
1558 		DEBUG2(DBG_DMA_CTL, dip,
1559 		    "HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1560 		    cp->dmac_laddress, cp->dmac_size);
1561 		}
1562 		return (DDI_SUCCESS);
1563 
1564 	case DDI_DMA_REPWIN:
1565 		*offp = mp->dmai_offset;
1566 		*lenp = mp->dmai_size;
1567 		return (DDI_SUCCESS);
1568 
1569 	case DDI_DMA_MOVWIN: {
1570 		off_t off = *offp;
1571 		ddi_dma_cookie_t *cp;
1572 		pci_dma_win_t *win_p = mp->dmai_winlst;
1573 
1574 		if (off >= mp->dmai_object.dmao_size)
1575 			return (DDI_FAILURE);
1576 
1577 		/* locate window */
1578 		while (win_p->win_offset + win_p->win_size <= off)
1579 			win_p = win_p->win_next;
1580 
1581 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1582 		mp->dmai_offset = win_p->win_offset;
1583 		mp->dmai_size   = win_p->win_size;
1584 		mp->dmai_mapping = cp->dmac_laddress;	/* cookie0 star addr */
1585 		mp->dmai_cookie = cp + 1;
1586 		win_p->win_curseg = 0;
1587 
1588 		*(ddi_dma_cookie_t *)objp = *cp;
1589 		*offp = win_p->win_offset;
1590 		*lenp = win_p->win_size;
1591 		DEBUG2(DBG_DMA_CTL, dip,
1592 		    "HTOC: cookie - dmac_laddress=%p dmac_size=%x\n",
1593 		    cp->dmac_laddress, cp->dmac_size);
1594 		}
1595 		return (DDI_SUCCESS);
1596 
1597 	case DDI_DMA_NEXTWIN: {
1598 		pci_dma_win_t *win_p = *(pci_dma_win_t **)offp;
1599 		pci_dma_win_t **nw_pp = (pci_dma_win_t **)objp;
1600 		ddi_dma_cookie_t *cp;
1601 		if (!win_p) {
1602 			*nw_pp = mp->dmai_winlst;
1603 			return (DDI_SUCCESS);
1604 		}
1605 
1606 		if (win_p->win_offset != mp->dmai_offset)
1607 			return (DDI_DMA_STALE);
1608 		if (!win_p->win_next)
1609 			return (DDI_DMA_DONE);
1610 		win_p = win_p->win_next;
1611 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1612 		mp->dmai_offset = win_p->win_offset;
1613 		mp->dmai_size   = win_p->win_size;
1614 		mp->dmai_mapping = cp->dmac_laddress;   /* cookie0 star addr */
1615 		mp->dmai_cookie = cp + 1;
1616 		win_p->win_curseg = 0;
1617 		*nw_pp = win_p;
1618 		}
1619 		return (DDI_SUCCESS);
1620 
1621 	case DDI_DMA_NEXTSEG: {
1622 		pci_dma_win_t *w_p = *(pci_dma_win_t **)offp;
1623 		if (w_p->win_offset != mp->dmai_offset)
1624 			return (DDI_DMA_STALE);
1625 		if (w_p->win_curseg + 1 >= w_p->win_ncookies)
1626 			return (DDI_DMA_DONE);
1627 		w_p->win_curseg++;
1628 		}
1629 		*(ddi_dma_seg_t *)objp = (ddi_dma_seg_t)mp;
1630 		return (DDI_SUCCESS);
1631 
1632 	case DDI_DMA_SEGTOC: {
1633 		pci_dma_win_t *win_p = mp->dmai_winlst;
1634 		off_t off = mp->dmai_offset;
1635 		ddi_dma_cookie_t *cp;
1636 		int i;
1637 
1638 		/* locate active window */
1639 		for (; win_p->win_offset != off; win_p = win_p->win_next)
1640 			;
1641 		cp = (ddi_dma_cookie_t *)(win_p + 1);
1642 		for (i = 0; i < win_p->win_curseg; i++, cp++)
1643 			off += cp->dmac_size;
1644 		*offp = off;
1645 		*lenp = cp->dmac_size;
1646 		*(ddi_dma_cookie_t *)objp = *cp;	/* copy cookie */
1647 		}
1648 		return (DDI_SUCCESS);
1649 
1650 	case DDI_DMA_COFF: {
1651 		pci_dma_win_t *win_p;
1652 		ddi_dma_cookie_t *cp;
1653 		uint64_t addr, key = ((ddi_dma_cookie_t *)offp)->dmac_laddress;
1654 		size_t win_off;
1655 
1656 		for (win_p = mp->dmai_winlst; win_p; win_p = win_p->win_next) {
1657 			int i;
1658 			win_off = 0;
1659 			cp = (ddi_dma_cookie_t *)(win_p + 1);
1660 			for (i = 0; i < win_p->win_ncookies; i++, cp++) {
1661 				size_t sz = cp->dmac_size;
1662 
1663 				addr = cp->dmac_laddress;
1664 				if ((addr <= key) && (addr + sz >= key))
1665 					goto found;
1666 				win_off += sz;
1667 			}
1668 		}
1669 		return (DDI_FAILURE);
1670 found:
1671 		*objp = (caddr_t)(win_p->win_offset + win_off + (key - addr));
1672 		return (DDI_SUCCESS);
1673 		}
1674 
1675 	case DDI_DMA_REMAP:
1676 		return (DDI_FAILURE);
1677 
1678 	default:
1679 		DEBUG3(DBG_DMA_CTL, dip, "unknown command (%x): rdip=%s%d\n",
1680 		    cmd, ddi_driver_name(rdip), ddi_get_instance(rdip));
1681 		break;
1682 	}
1683 	return (DDI_FAILURE);
1684 }
1685 
1686 static void
1687 pci_dvma_debug_init(iommu_t *iommu_p)
1688 {
1689 	size_t sz = sizeof (struct dvma_rec) * pci_dvma_debug_rec;
1690 	ASSERT(MUTEX_HELD(&iommu_p->dvma_debug_lock));
1691 	cmn_err(CE_NOTE, "PCI DVMA %p stat ON", iommu_p);
1692 
1693 	iommu_p->dvma_alloc_rec = kmem_zalloc(sz, KM_SLEEP);
1694 	iommu_p->dvma_free_rec = kmem_zalloc(sz, KM_SLEEP);
1695 
1696 	iommu_p->dvma_active_list = NULL;
1697 	iommu_p->dvma_alloc_rec_index = 0;
1698 	iommu_p->dvma_free_rec_index = 0;
1699 	iommu_p->dvma_active_count = 0;
1700 }
1701 
1702 void
1703 pci_dvma_debug_fini(iommu_t *iommu_p)
1704 {
1705 	struct dvma_rec *prev, *ptr;
1706 	size_t sz = sizeof (struct dvma_rec) * pci_dvma_debug_rec;
1707 	uint64_t mask = ~(1ull << iommu_p->iommu_inst);
1708 	cmn_err(CE_NOTE, "PCI DVMA %p stat OFF", iommu_p);
1709 
1710 	kmem_free(iommu_p->dvma_alloc_rec, sz);
1711 	kmem_free(iommu_p->dvma_free_rec, sz);
1712 	iommu_p->dvma_alloc_rec = iommu_p->dvma_free_rec = NULL;
1713 
1714 	prev = iommu_p->dvma_active_list;
1715 	if (!prev)
1716 		return;
1717 	for (ptr = prev->next; ptr; prev = ptr, ptr = ptr->next)
1718 		kmem_free(prev, sizeof (struct dvma_rec));
1719 	kmem_free(prev, sizeof (struct dvma_rec));
1720 
1721 	iommu_p->dvma_active_list = NULL;
1722 	iommu_p->dvma_alloc_rec_index = 0;
1723 	iommu_p->dvma_free_rec_index = 0;
1724 	iommu_p->dvma_active_count = 0;
1725 
1726 	pci_dvma_debug_on  &= mask;
1727 	pci_dvma_debug_off &= mask;
1728 }
1729 
1730 void
1731 pci_dvma_alloc_debug(iommu_t *iommu_p, char *address, uint_t len,
1732 	ddi_dma_impl_t *mp)
1733 {
1734 	struct dvma_rec *ptr;
1735 	mutex_enter(&iommu_p->dvma_debug_lock);
1736 
1737 	if (!iommu_p->dvma_alloc_rec)
1738 		pci_dvma_debug_init(iommu_p);
1739 	if (DVMA_DBG_OFF(iommu_p)) {
1740 		pci_dvma_debug_fini(iommu_p);
1741 		goto done;
1742 	}
1743 
1744 	ptr = &iommu_p->dvma_alloc_rec[iommu_p->dvma_alloc_rec_index];
1745 	ptr->dvma_addr = address;
1746 	ptr->len = len;
1747 	ptr->mp = mp;
1748 	if (++iommu_p->dvma_alloc_rec_index == pci_dvma_debug_rec)
1749 		iommu_p->dvma_alloc_rec_index = 0;
1750 
1751 	ptr = kmem_alloc(sizeof (struct dvma_rec), KM_SLEEP);
1752 	ptr->dvma_addr = address;
1753 	ptr->len = len;
1754 	ptr->mp = mp;
1755 
1756 	ptr->next = iommu_p->dvma_active_list;
1757 	iommu_p->dvma_active_list = ptr;
1758 	iommu_p->dvma_active_count++;
1759 done:
1760 	mutex_exit(&iommu_p->dvma_debug_lock);
1761 }
1762 
1763 void
1764 pci_dvma_free_debug(iommu_t *iommu_p, char *address, uint_t len,
1765 	ddi_dma_impl_t *mp)
1766 {
1767 	struct dvma_rec *ptr, *ptr_save;
1768 	mutex_enter(&iommu_p->dvma_debug_lock);
1769 
1770 	if (!iommu_p->dvma_alloc_rec)
1771 		pci_dvma_debug_init(iommu_p);
1772 	if (DVMA_DBG_OFF(iommu_p)) {
1773 		pci_dvma_debug_fini(iommu_p);
1774 		goto done;
1775 	}
1776 
1777 	ptr = &iommu_p->dvma_free_rec[iommu_p->dvma_free_rec_index];
1778 	ptr->dvma_addr = address;
1779 	ptr->len = len;
1780 	ptr->mp = mp;
1781 	if (++iommu_p->dvma_free_rec_index == pci_dvma_debug_rec)
1782 		iommu_p->dvma_free_rec_index = 0;
1783 
1784 	ptr_save = iommu_p->dvma_active_list;
1785 	for (ptr = ptr_save; ptr; ptr = ptr->next) {
1786 		if ((ptr->dvma_addr == address) && (ptr->len = len))
1787 			break;
1788 		ptr_save = ptr;
1789 	}
1790 	if (!ptr) {
1791 		cmn_err(CE_WARN, "bad dvma free addr=%lx len=%x",
1792 		    (long)address, len);
1793 		goto done;
1794 	}
1795 	if (ptr == iommu_p->dvma_active_list)
1796 		iommu_p->dvma_active_list = ptr->next;
1797 	else
1798 		ptr_save->next = ptr->next;
1799 	kmem_free(ptr, sizeof (struct dvma_rec));
1800 	iommu_p->dvma_active_count--;
1801 done:
1802 	mutex_exit(&iommu_p->dvma_debug_lock);
1803 }
1804 
1805 #ifdef DEBUG
1806 void
1807 dump_dma_handle(uint64_t flag, dev_info_t *dip, ddi_dma_impl_t *hp)
1808 {
1809 	DEBUG4(flag, dip, "mp(%p): flags=%x mapping=%lx xfer_size=%x\n",
1810 	    hp, hp->dmai_inuse, hp->dmai_mapping, hp->dmai_size);
1811 	DEBUG4(flag|DBG_CONT, dip, "\tnpages=%x roffset=%x rflags=%x nwin=%x\n",
1812 	    hp->dmai_ndvmapages, hp->dmai_roffset, hp->dmai_rflags,
1813 	    hp->dmai_nwin);
1814 	DEBUG4(flag|DBG_CONT, dip, "\twinsize=%x tte=%p pfnlst=%p pfn0=%p\n",
1815 	    hp->dmai_winsize, hp->dmai_tte, hp->dmai_pfnlst, hp->dmai_pfn0);
1816 	DEBUG4(flag|DBG_CONT, dip, "\twinlst=%x obj=%p attr=%p ckp=%p\n",
1817 	    hp->dmai_winlst, &hp->dmai_object, &hp->dmai_attr,
1818 	    hp->dmai_cookie);
1819 }
1820 #endif
1821 
1822 void
1823 pci_vmem_do_free(iommu_t *iommu_p, void *base_addr, size_t npages,
1824     int vmemcache)
1825 {
1826 	vmem_t *map_p = iommu_p->iommu_dvma_map;
1827 
1828 	if (vmemcache) {
1829 		vmem_free(map_p, base_addr, IOMMU_PAGE_SIZE);
1830 #ifdef PCI_DMA_PROF
1831 		pci_dvma_vmem_free++;
1832 #endif
1833 		return;
1834 	}
1835 
1836 	vmem_xfree(map_p, base_addr, IOMMU_PTOB(npages));
1837 #ifdef PCI_DMA_PROF
1838 		pci_dvma_vmem_xfree++;
1839 #endif
1840 }
1841