xref: /netbsd-src/sys/arch/sparc64/dev/iommu.c (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 /*	$NetBSD: iommu.c,v 1.116 2021/04/26 07:18:01 mrg Exp $	*/
2 
3 /*
4  * Copyright (c) 1999, 2000 Matthew R. Green
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
22  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * Copyright (c) 2001, 2002 Eduardo Horvath
31  * All rights reserved.
32  *
33  * Redistribution and use in source and binary forms, with or without
34  * modification, are permitted provided that the following conditions
35  * are met:
36  * 1. Redistributions of source code must retain the above copyright
37  *    notice, this list of conditions and the following disclaimer.
38  * 2. Redistributions in binary form must reproduce the above copyright
39  *    notice, this list of conditions and the following disclaimer in the
40  *    documentation and/or other materials provided with the distribution.
41  * 3. The name of the author may not be used to endorse or promote products
42  *    derived from this software without specific prior written permission.
43  *
44  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
45  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
46  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
47  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
48  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
49  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
50  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
51  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
52  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
53  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
54  * SUCH DAMAGE.
55  */
56 
57 /*
58  * UltraSPARC IOMMU support; used by both the sbus and pci code.
59  */
60 
61 #include <sys/cdefs.h>
62 __KERNEL_RCSID(0, "$NetBSD: iommu.c,v 1.116 2021/04/26 07:18:01 mrg Exp $");
63 
64 #include "opt_ddb.h"
65 
66 #include <sys/param.h>
67 #include <sys/extent.h>
68 #include <sys/malloc.h>
69 #include <sys/systm.h>
70 #include <sys/device.h>
71 #include <sys/proc.h>
72 
73 #include <uvm/uvm.h>
74 
75 #include <sys/bus.h>
76 #include <sparc64/dev/iommureg.h>
77 #include <sparc64/dev/iommuvar.h>
78 
79 #include <machine/autoconf.h>
80 #include <machine/cpu.h>
81 #include <machine/hypervisor.h>
82 
83 #ifdef DEBUG
84 #define IDB_BUSDMA	0x1
85 #define IDB_IOMMU	0x2
86 #define IDB_INFO	0x4
87 #define	IDB_SYNC	0x8
88 int iommudebug = 0x0;
89 #define DPRINTF(l, s)   do { if (iommudebug & l) printf s; } while (0)
90 #define IOTTE_DEBUG(n)	(n)
91 #else
92 #define DPRINTF(l, s)
93 #define IOTTE_DEBUG(n)	0
94 #endif
95 
96 #define iommu_strbuf_flush(i, v) do {					\
97 	if ((i)->sb_flush)						\
98 		bus_space_write_8((i)->sb_is->is_bustag, (i)->sb_sb,	\
99 			STRBUFREG(strbuf_pgflush), (v));		\
100 	} while (0)
101 
102 static	int iommu_strbuf_flush_done(struct strbuf_ctl *);
103 static	void _iommu_dvmamap_sync(bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
104 		bus_size_t, int);
105 static void iommu_enter_sun4u(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags);
106 static void iommu_enter_sun4v(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags);
107 static void iommu_remove_sun4u(struct iommu_state *is, vaddr_t va, size_t len);
108 static void iommu_remove_sun4v(struct iommu_state *is, vaddr_t va, size_t len);
109 
110 /*
111  * initialise the UltraSPARC IOMMU (SBUS or PCI):
112  *	- allocate and setup the iotsb.
113  *	- enable the IOMMU
114  *	- initialise the streaming buffers (if they exist)
115  *	- create a private DVMA map.
116  */
117 void
118 iommu_init(char *name, struct iommu_state *is, int tsbsize, uint32_t iovabase)
119 {
120 	psize_t size;
121 	vaddr_t va;
122 	paddr_t pa;
123 	struct vm_page *pg;
124 	struct pglist pglist;
125 
126 	DPRINTF(IDB_INFO, ("iommu_init: tsbsize %x iovabase %x\n", tsbsize, iovabase));
127 
128 	/*
129 	 * Setup the iommu.
130 	 *
131 	 * The sun4u iommu is part of the SBUS or PCI controller so we will
132 	 * deal with it here..
133 	 *
134 	 * For sysio and psycho/psycho+ the IOMMU address space always ends at
135 	 * 0xffffe000, but the starting address depends on the size of the
136 	 * map.  The map size is 1024 * 2 ^ is->is_tsbsize entries, where each
137 	 * entry is 8 bytes.  The start of the map can be calculated by
138 	 * (0xffffe000 << (8 + is->is_tsbsize)).
139 	 *
140 	 * But sabre and hummingbird use a different scheme that seems to
141 	 * be hard-wired, so we read the start and size from the PROM and
142 	 * just use those values.
143 	 */
144 	if (strncmp(name, "pyro", 4) == 0) {
145 		is->is_cr = IOMMUREG_READ(is, iommu_cr);
146 		is->is_cr &= ~IOMMUCR_FIRE_BE;
147 		is->is_cr |= (IOMMUCR_FIRE_SE | IOMMUCR_FIRE_CM_EN |
148 		    IOMMUCR_FIRE_TE);
149 	} else
150 		is->is_cr = IOMMUCR_EN;
151 	is->is_tsbsize = tsbsize;
152 	if (iovabase == -1) {
153 		is->is_dvmabase = IOTSB_VSTART(is->is_tsbsize);
154 		is->is_dvmaend = IOTSB_VEND - 1;
155 	} else {
156 		is->is_dvmabase = iovabase;
157 		is->is_dvmaend = iovabase + IOTSB_VSIZE(tsbsize) - 1;
158 	}
159 
160 	/*
161 	 * Allocate memory for I/O pagetables.  They need to be physically
162 	 * contiguous.
163 	 */
164 
165 	size = PAGE_SIZE << is->is_tsbsize;
166 	if (uvm_pglistalloc((psize_t)size, (paddr_t)0, (paddr_t)-1,
167 		(paddr_t)PAGE_SIZE, (paddr_t)0, &pglist, 1, 0) != 0)
168 		panic("iommu_init: no memory");
169 
170 	va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY);
171 	if (va == 0)
172 		panic("iommu_init: no memory");
173 	is->is_tsb = (int64_t *)va;
174 
175 	is->is_ptsb = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pglist));
176 
177 	/* Map the pages */
178 	TAILQ_FOREACH(pg, &pglist, pageq.queue) {
179 		pa = VM_PAGE_TO_PHYS(pg);
180 		pmap_kenter_pa(va, pa | PMAP_NVC,
181 		    VM_PROT_READ | VM_PROT_WRITE, 0);
182 		va += PAGE_SIZE;
183 	}
184 	pmap_update(pmap_kernel());
185 	memset(is->is_tsb, 0, size);
186 
187 #ifdef DEBUG
188 	if (iommudebug & IDB_INFO)
189 	{
190 		/* Probe the iommu */
191 		if (!CPU_ISSUN4V) {
192 			printf("iommu cr=%llx tsb=%llx\n",
193 			    (unsigned long long)bus_space_read_8(is->is_bustag,
194 				is->is_iommu,
195 				offsetof(struct iommureg, iommu_cr)),
196 			    (unsigned long long)bus_space_read_8(is->is_bustag,
197 				is->is_iommu,
198 				offsetof(struct iommureg, iommu_tsb)));
199 			printf("TSB base %p phys %llx\n", (void *)is->is_tsb,
200 			    (unsigned long long)is->is_ptsb);
201 			delay(1000000); /* 1 s */
202 		}
203 	}
204 #endif
205 
206 	/*
207 	 * Now all the hardware's working we need to allocate a dvma map.
208 	 */
209 	aprint_debug("DVMA map: %x to %x\n",
210 		(unsigned int)is->is_dvmabase,
211 		(unsigned int)is->is_dvmaend);
212 	aprint_debug("IOTSB: %llx to %llx\n",
213 		(unsigned long long)is->is_ptsb,
214 		(unsigned long long)(is->is_ptsb + size - 1));
215 	is->is_dvmamap = extent_create(name,
216 	    is->is_dvmabase, is->is_dvmaend,
217 	    0, 0, EX_NOWAIT);
218 	if (!is->is_dvmamap)
219 		panic("iommu_init: extent_create() failed");
220 
221 	mutex_init(&is->is_lock, MUTEX_DEFAULT, IPL_HIGH);
222 
223 	/*
224 	 * Set the TSB size.  The relevant bits were moved to the TSB
225 	 * base register in the PCIe host bridges.
226 	 */
227 	if (is->is_flags & IOMMU_TSBSIZE_IN_PTSB)
228 		is->is_ptsb |= is->is_tsbsize;
229 	else
230 		is->is_cr |= (is->is_tsbsize << 16);
231 
232 	/*
233 	 * now actually start up the IOMMU
234 	 */
235 	iommu_reset(is);
236 }
237 
238 /*
239  * Streaming buffers don't exist on the UltraSPARC IIi; we should have
240  * detected that already and disabled them.  If not, we will notice that
241  * they aren't there when the STRBUF_EN bit does not remain.
242  */
243 void
244 iommu_reset(struct iommu_state *is)
245 {
246 	int i;
247 	struct strbuf_ctl *sb;
248 
249 	if (CPU_ISSUN4V)
250 		return;
251 
252 	IOMMUREG_WRITE(is, iommu_tsb, is->is_ptsb);
253 
254 	/* Enable IOMMU in diagnostic mode */
255 	IOMMUREG_WRITE(is, iommu_cr, is->is_cr|IOMMUCR_DE);
256 
257 	for (i = 0; i < 2; i++) {
258 		if ((sb = is->is_sb[i])) {
259 
260 			/* Enable diagnostics mode? */
261 			bus_space_write_8(is->is_bustag, is->is_sb[i]->sb_sb,
262 				STRBUFREG(strbuf_ctl), STRBUF_EN);
263 
264 			membar_Lookaside();
265 
266 			/* No streaming buffers? Disable them */
267 			if (bus_space_read_8(is->is_bustag,
268 				is->is_sb[i]->sb_sb,
269 				STRBUFREG(strbuf_ctl)) == 0) {
270 				is->is_sb[i]->sb_flush = NULL;
271 			} else {
272 
273 				/*
274 				 * locate the pa of the flush buffer.
275 				 */
276 				if (pmap_extract(pmap_kernel(),
277 				     (vaddr_t)is->is_sb[i]->sb_flush,
278 				     &is->is_sb[i]->sb_flushpa) == FALSE)
279 					is->is_sb[i]->sb_flush = NULL;
280 			}
281 		}
282 	}
283 
284 	if (is->is_flags & IOMMU_FLUSH_CACHE)
285 		IOMMUREG_WRITE(is, iommu_cache_invalidate, -1ULL);
286 }
287 
288 /*
289  * Here are the iommu control routines.
290  */
291 
292 static void
293 iommu_enter(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
294 {
295 	DPRINTF(IDB_IOMMU, ("iommu_enter: va %lx pa %lx flags %x\n",
296 	    va, (long)pa, flags));
297 	if (!CPU_ISSUN4V)
298 		iommu_enter_sun4u(sb, va, pa, flags);
299 	else
300 		iommu_enter_sun4v(sb, va, pa, flags);
301 }
302 
303 
304 void
305 iommu_enter_sun4u(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
306 {
307 	struct iommu_state *is = sb->sb_is;
308 	int strbuf = (flags & BUS_DMA_STREAMING);
309 	int64_t tte;
310 
311 #ifdef DIAGNOSTIC
312 	if (va < is->is_dvmabase || va > is->is_dvmaend)
313 		panic("iommu_enter: va %#lx not in DVMA space", va);
314 #endif
315 
316 	/* Is the streamcache flush really needed? */
317 	if (sb->sb_flush)
318 		iommu_strbuf_flush(sb, va);
319 	else
320 		/* If we can't flush the strbuf don't enable it. */
321 		strbuf = 0;
322 
323 	tte = MAKEIOTTE(pa, !(flags & BUS_DMA_NOWRITE),
324 		!(flags & BUS_DMA_NOCACHE), (strbuf));
325 #ifdef DEBUG
326 	tte |= (flags & 0xff000LL)<<(4*8);
327 #endif
328 
329 	is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = tte;
330 	bus_space_write_8(is->is_bustag, is->is_iommu,
331 		IOMMUREG(iommu_flush), va);
332 	DPRINTF(IDB_IOMMU, ("iommu_enter: slot %d va %lx pa %lx "
333 		"TSB[%lx]@%p=%lx\n", (int)IOTSBSLOT(va,is->is_tsbsize),
334 		va, (long)pa, (u_long)IOTSBSLOT(va,is->is_tsbsize),
335 		(void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
336 		(u_long)tte));
337 }
338 
339 void
340 iommu_enter_sun4v(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
341 {
342 	struct iommu_state *is = sb->sb_is;
343 	u_int64_t tsbid = IOTSBSLOT(va, is->is_tsbsize);
344 	paddr_t page_list[1], addr;
345 	u_int64_t attr, nmapped;
346 	int err;
347 
348 #ifdef DIAGNOSTIC
349 	if (va < is->is_dvmabase || (va + PAGE_MASK) > is->is_dvmaend)
350 		panic("viommu_enter: va %#lx not in DVMA space", va);
351 #endif
352 
353 	attr = PCI_MAP_ATTR_READ | PCI_MAP_ATTR_WRITE;
354 	if (flags & BUS_DMA_READ)
355 		attr &= ~PCI_MAP_ATTR_READ;
356 	if (flags & BUS_DMA_WRITE)
357 		attr &= ~PCI_MAP_ATTR_WRITE;
358 
359 	page_list[0] = trunc_page(pa);
360 	if (!pmap_extract(pmap_kernel(), (vaddr_t)page_list, &addr))
361 		panic("viommu_enter: pmap_extract failed");
362 	err = hv_pci_iommu_map(is->is_devhandle, tsbid, 1, attr,
363 	    addr, &nmapped);
364 	if (err != H_EOK || nmapped != 1)
365 		panic("hv_pci_iommu_map: err=%d, nmapped=%lu", err, (long unsigned int)nmapped);
366 }
367 
368 /*
369  * Find the value of a DVMA address (debug routine).
370  */
371 paddr_t
372 iommu_extract(struct iommu_state *is, vaddr_t dva)
373 {
374 	int64_t tte = 0;
375 
376 	if (dva >= is->is_dvmabase && dva <= is->is_dvmaend)
377 		tte = is->is_tsb[IOTSBSLOT(dva, is->is_tsbsize)];
378 
379 	if ((tte & IOTTE_V) == 0)
380 		return ((paddr_t)-1L);
381 	return (tte & IOTTE_PAMASK);
382 }
383 
384 /*
385  * iommu_remove: removes mappings created by iommu_enter
386  *
387  * Only demap from IOMMU if flag is set.
388  *
389  * XXX: this function needs better internal error checking.
390  */
391 
392 static void
393 iommu_remove(struct iommu_state *is, vaddr_t va, size_t len)
394 {
395 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx len %zu\n", va, len));
396 	if (!CPU_ISSUN4V)
397 		iommu_remove_sun4u(is, va, len);
398 	else
399 		iommu_remove_sun4v(is, va, len);
400 }
401 
402 void
403 iommu_remove_sun4u(struct iommu_state *is, vaddr_t va, size_t len)
404 {
405 
406 	int slot;
407 
408 #ifdef DIAGNOSTIC
409 	if (va < is->is_dvmabase || va > is->is_dvmaend)
410 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
411 	if ((long)(va + len) < (long)va)
412 		panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
413 		      (long) va, (long) len);
414 	if (len & ~0xfffffff)
415 		panic("iommu_remove: ridiculous len 0x%lx", (u_long)len);
416 #endif
417 
418 	va = trunc_page(va);
419 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
420 		va, (u_long)IOTSBSLOT(va, is->is_tsbsize),
421 		&is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)]));
422 	while (len > 0) {
423 		DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d "
424 			"for va %p size %lx\n",
425 			(int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va,
426 			(u_long)len));
427 		if (len <= PAGE_SIZE)
428 			len = 0;
429 		else
430 			len -= PAGE_SIZE;
431 
432 #if 0
433 		/*
434 		 * XXX Zero-ing the entry would not require RMW
435 		 *
436 		 * Disabling valid bit while a page is used by a device
437 		 * causes an uncorrectable DMA error.
438 		 * Workaround to avoid an uncorrectable DMA error is
439 		 * eliminating the next line, but the page is mapped
440 		 * until the next iommu_enter call.
441 		 */
442 		is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] &= ~IOTTE_V;
443 		membar_StoreStore();
444 #endif
445 		IOMMUREG_WRITE(is, iommu_flush, va);
446 
447 		/* Flush cache if necessary. */
448 		slot = IOTSBSLOT(trunc_page(va), is->is_tsbsize);
449 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
450 		    (len == 0 || (slot % 8) == 7))
451 			IOMMUREG_WRITE(is, iommu_cache_flush,
452 			    is->is_ptsb + slot * 8);
453 
454 		va += PAGE_SIZE;
455 	}
456 }
457 
458 void
459 iommu_remove_sun4v(struct iommu_state *is, vaddr_t va, size_t len)
460 {
461 	u_int64_t tsbid = IOTSBSLOT(va, is->is_tsbsize);
462 	u_int64_t ndemapped;
463 	int err;
464 
465 #ifdef DIAGNOSTIC
466 	if (va < is->is_dvmabase || (va + PAGE_MASK) > is->is_dvmaend)
467 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
468 	if (va != trunc_page(va)) {
469 		printf("iommu_remove: unaligned va: %lx\n", va);
470 		va = trunc_page(va);
471 	}
472 #endif
473 
474 	err = hv_pci_iommu_demap(is->is_devhandle, tsbid, 1, &ndemapped);
475 	if (err != H_EOK || ndemapped != 1)
476 		panic("hv_pci_iommu_unmap: err=%d", err);
477 }
478 
479 static int
480 iommu_strbuf_flush_done(struct strbuf_ctl *sb)
481 {
482 	struct iommu_state *is = sb->sb_is;
483 	struct timeval cur, flushtimeout;
484 
485 #define BUMPTIME(t, usec) { \
486 	register volatile struct timeval *tp = (t); \
487 	register long us; \
488  \
489 	tp->tv_usec = us = tp->tv_usec + (usec); \
490 	if (us >= 1000000) { \
491 		tp->tv_usec = us - 1000000; \
492 		tp->tv_sec++; \
493 	} \
494 }
495 
496 	if (!sb->sb_flush)
497 		return (0);
498 
499 	/*
500 	 * Streaming buffer flushes:
501 	 *
502 	 *   1 Tell strbuf to flush by storing va to strbuf_pgflush.  If
503 	 *     we're not on a cache line boundary (64-bits):
504 	 *   2 Store 0 in flag
505 	 *   3 Store pointer to flag in flushsync
506 	 *   4 wait till flushsync becomes 0x1
507 	 *
508 	 * If it takes more than .5 sec, something
509 	 * went wrong.
510 	 */
511 
512 	*sb->sb_flush = 0;
513 	bus_space_write_8(is->is_bustag, sb->sb_sb,
514 		STRBUFREG(strbuf_flushsync), sb->sb_flushpa);
515 
516 	microtime(&flushtimeout);
517 	cur = flushtimeout;
518 	BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
519 
520 	DPRINTF(IDB_IOMMU, ("%s: flush = %lx at va = %lx pa = %lx now="
521 		"%"PRIx64":%"PRIx32" until = %"PRIx64":%"PRIx32"\n", __func__,
522 		(long)*sb->sb_flush, (long)sb->sb_flush, (long)sb->sb_flushpa,
523 		cur.tv_sec, cur.tv_usec,
524 		flushtimeout.tv_sec, flushtimeout.tv_usec));
525 
526 	/* Bypass non-coherent D$ */
527 	while ((!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) &&
528 	       timercmp(&cur, &flushtimeout, <=))
529 		microtime(&cur);
530 
531 #ifdef DIAGNOSTIC
532 	if (!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) {
533 		printf("%s: flush timeout %p, at %p\n", __func__,
534 			(void *)(u_long)*sb->sb_flush,
535 			(void *)(u_long)sb->sb_flushpa); /* panic? */
536 #ifdef DDB
537 		Debugger();
538 #endif
539 	}
540 #endif
541 	DPRINTF(IDB_IOMMU, ("%s: flushed\n", __func__));
542 	return (*sb->sb_flush);
543 }
544 
545 /*
546  * IOMMU DVMA operations, common to SBUS and PCI.
547  */
548 int
549 iommu_dvmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
550 	bus_size_t buflen, struct proc *p, int flags)
551 {
552 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
553 	struct iommu_state *is = sb->sb_is;
554 	int err, needsflush;
555 	bus_size_t sgsize;
556 	paddr_t curaddr;
557 	u_long dvmaddr, sgstart, sgend, bmask;
558 	bus_size_t align, boundary, len;
559 	vaddr_t vaddr = (vaddr_t)buf;
560 	int seg;
561 	struct pmap *pmap;
562 	int slot;
563 
564 	if (map->dm_nsegs) {
565 		/* Already in use?? */
566 #ifdef DIAGNOSTIC
567 		printf("iommu_dvmamap_load: map still in use\n");
568 #endif
569 		bus_dmamap_unload(t, map);
570 	}
571 
572 	/*
573 	 * Make sure that on error condition we return "no valid mappings".
574 	 */
575 	map->dm_nsegs = 0;
576 	KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
577 
578 	if (buflen > map->_dm_size) {
579 		DPRINTF(IDB_BUSDMA,
580 		    ("iommu_dvmamap_load(): error %d > %d -- "
581 		     "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
582 		return (EINVAL);
583 	}
584 
585 	sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
586 
587 	/*
588 	 * A boundary presented to bus_dmamem_alloc() takes precedence
589 	 * over boundary in the map.
590 	 */
591 	if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
592 		boundary = map->_dm_boundary;
593 	align = uimax(map->dm_segs[0]._ds_align, PAGE_SIZE);
594 
595 	/*
596 	 * If our segment size is larger than the boundary we need to
597 	 * split the transfer up int little pieces ourselves.
598 	 */
599 	KASSERT(is->is_dvmamap);
600 	mutex_enter(&is->is_lock);
601 	err = extent_alloc(is->is_dvmamap, sgsize, align,
602 	    (sgsize > boundary) ? 0 : boundary,
603 	    EX_NOWAIT|EX_BOUNDZERO, &dvmaddr);
604 	mutex_exit(&is->is_lock);
605 
606 #ifdef DEBUG
607 	if (err || (dvmaddr == (u_long)-1)) {
608 		printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
609 		    (int)sgsize, flags);
610 #ifdef DDB
611 		Debugger();
612 #endif
613 	}
614 #endif
615 	if (err != 0)
616 		return (err);
617 
618 	if (dvmaddr == (u_long)-1)
619 		return (ENOMEM);
620 
621 	/* Set the active DVMA map */
622 	map->_dm_dvmastart = dvmaddr;
623 	map->_dm_dvmasize = sgsize;
624 
625 	/*
626 	 * Now split the DVMA range into segments, not crossing
627 	 * the boundary.
628 	 */
629 	seg = 0;
630 	sgstart = dvmaddr + (vaddr & PGOFSET);
631 	sgend = sgstart + buflen - 1;
632 	map->dm_segs[seg].ds_addr = sgstart;
633 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: boundary %lx boundary - 1 %lx "
634 	    "~(boundary - 1) %lx\n", (long)boundary, (long)(boundary - 1),
635 	    (long)~(boundary - 1)));
636 	bmask = ~(boundary - 1);
637 	while ((sgstart & bmask) != (sgend & bmask) ||
638 	       sgend - sgstart + 1 > map->dm_maxsegsz) {
639 		/* Oops. We crossed a boundary or large seg. Split the xfer. */
640 		len = map->dm_maxsegsz;
641 		if ((sgstart & bmask) != (sgend & bmask))
642 			len = uimin(len, boundary - (sgstart & (boundary - 1)));
643 		map->dm_segs[seg].ds_len = len;
644 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
645 		    "seg %d start %lx size %lx\n", seg,
646 		    (long)map->dm_segs[seg].ds_addr,
647 		    (long)map->dm_segs[seg].ds_len));
648 		if (++seg >= map->_dm_segcnt) {
649 			/* Too many segments.  Fail the operation. */
650 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
651 			    "too many segments %d\n", seg));
652 			mutex_enter(&is->is_lock);
653 			err = extent_free(is->is_dvmamap,
654 			    dvmaddr, sgsize, EX_NOWAIT);
655 			map->_dm_dvmastart = 0;
656 			map->_dm_dvmasize = 0;
657 			mutex_exit(&is->is_lock);
658 			if (err != 0)
659 				printf("warning: %s: %" PRId64
660 				    " of DVMA space lost\n", __func__, sgsize);
661 			return (EFBIG);
662 		}
663 		sgstart += len;
664 		map->dm_segs[seg].ds_addr = sgstart;
665 	}
666 	map->dm_segs[seg].ds_len = sgend - sgstart + 1;
667 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
668 	    "seg %d start %lx size %lx\n", seg,
669 	    (long)map->dm_segs[seg].ds_addr, (long)map->dm_segs[seg].ds_len));
670 	map->dm_nsegs = seg + 1;
671 	map->dm_mapsize = buflen;
672 
673 	if (p != NULL)
674 		pmap = p->p_vmspace->vm_map.pmap;
675 	else
676 		pmap = pmap_kernel();
677 
678 	needsflush = 0;
679 	for (; buflen > 0; ) {
680 
681 		/*
682 		 * Get the physical address for this page.
683 		 */
684 		if (pmap_extract(pmap, (vaddr_t)vaddr, &curaddr) == FALSE) {
685 #ifdef DIAGNOSTIC
686 			printf("iommu_dvmamap_load: pmap_extract failed %lx\n", vaddr);
687 #endif
688 			bus_dmamap_unload(t, map);
689 			return (-1);
690 		}
691 
692 		/*
693 		 * Compute the segment size, and adjust counts.
694 		 */
695 		sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
696 		if (buflen < sgsize)
697 			sgsize = buflen;
698 
699 		DPRINTF(IDB_BUSDMA,
700 		    ("iommu_dvmamap_load: map %p loading va %p "
701 		    "dva %lx at pa %lx\n",
702 		    map, (void *)vaddr, (long)dvmaddr,
703 		    (long)trunc_page(curaddr)));
704 		iommu_enter(sb, trunc_page(dvmaddr), trunc_page(curaddr),
705 		    flags | IOTTE_DEBUG(0x4000));
706 		needsflush = 1;
707 
708 		vaddr += sgsize;
709 		buflen -= sgsize;
710 
711 		/* Flush cache if necessary. */
712 		slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
713 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
714 		    (buflen <= 0 || (slot % 8) == 7))
715 			IOMMUREG_WRITE(is, iommu_cache_flush,
716 			    is->is_ptsb + slot * 8);
717 
718 		dvmaddr += PAGE_SIZE;
719 	}
720 	if (needsflush)
721 		iommu_strbuf_flush_done(sb);
722 #ifdef DIAGNOSTIC
723 	for (seg = 0; seg < map->dm_nsegs; seg++) {
724 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
725 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
726 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
727 			    seg, (long)map->dm_segs[seg].ds_addr,
728 			    is->is_dvmabase, is->is_dvmaend);
729 #ifdef DDB
730 			Debugger();
731 #endif
732 		}
733 	}
734 #endif
735 	return (0);
736 }
737 
738 
739 void
740 iommu_dvmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
741 {
742 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
743 	struct iommu_state *is = sb->sb_is;
744 	int error;
745 	bus_size_t sgsize = map->_dm_dvmasize;
746 
747 	/* Flush the iommu */
748 	if (!map->_dm_dvmastart)
749 		panic("%s: error dvmastart is zero!\n", __func__);
750 
751 	if (is->is_flags & IOMMU_SYNC_BEFORE_UNMAP) {
752 
753 		/* Flush the caches */
754 		bus_dmamap_unload(t->_parent, map);
755 
756 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
757 
758 	} else {
759 
760 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
761 
762 		/* Flush the caches */
763 		bus_dmamap_unload(t->_parent, map);
764 	}
765 
766 	mutex_enter(&is->is_lock);
767 	error = extent_free(is->is_dvmamap, map->_dm_dvmastart,
768 		map->_dm_dvmasize, EX_NOWAIT);
769 	map->_dm_dvmastart = 0;
770 	map->_dm_dvmasize = 0;
771 	mutex_exit(&is->is_lock);
772 	if (error != 0)
773 		printf("warning: %s: %" PRId64 " of DVMA space lost\n",
774 		    __func__, sgsize);
775 
776 	/* Clear the map */
777 }
778 
779 
780 int
781 iommu_dvmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
782 	bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
783 {
784 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
785 	struct iommu_state *is = sb->sb_is;
786 	struct vm_page *pg;
787 	int i, j;
788 	int left;
789 	int err, needsflush;
790 	bus_size_t sgsize;
791 	paddr_t pa;
792 	bus_size_t boundary, align;
793 	u_long dvmaddr, sgstart, sgend, bmask;
794 	struct pglist *pglist;
795 	const int pagesz = PAGE_SIZE;
796 	int slot;
797 #ifdef DEBUG
798 	int npg = 0;
799 #endif
800 
801 	if (map->dm_nsegs) {
802 		/* Already in use?? */
803 #ifdef DIAGNOSTIC
804 		printf("iommu_dvmamap_load_raw: map still in use\n");
805 #endif
806 		bus_dmamap_unload(t, map);
807 	}
808 
809 	/*
810 	 * A boundary presented to bus_dmamem_alloc() takes precedence
811 	 * over boundary in the map.
812 	 */
813 	if ((boundary = segs[0]._ds_boundary) == 0)
814 		boundary = map->_dm_boundary;
815 
816 	align = uimax(segs[0]._ds_align, pagesz);
817 
818 	/*
819 	 * Make sure that on error condition we return "no valid mappings".
820 	 */
821 	map->dm_nsegs = 0;
822 	/* Count up the total number of pages we need */
823 	pa = trunc_page(segs[0].ds_addr);
824 	sgsize = 0;
825 	left = size;
826 	for (i = 0; left > 0 && i < nsegs; i++) {
827 		if (round_page(pa) != round_page(segs[i].ds_addr))
828 			sgsize = round_page(sgsize) +
829 			    (segs[i].ds_addr & PGOFSET);
830 		sgsize += uimin(left, segs[i].ds_len);
831 		left -= segs[i].ds_len;
832 		pa = segs[i].ds_addr + segs[i].ds_len;
833 	}
834 	sgsize = round_page(sgsize);
835 
836 	mutex_enter(&is->is_lock);
837 	/*
838 	 * If our segment size is larger than the boundary we need to
839 	 * split the transfer up into little pieces ourselves.
840 	 */
841 	err = extent_alloc(is->is_dvmamap, sgsize, align,
842 		(sgsize > boundary) ? 0 : boundary,
843 		((flags & BUS_DMA_NOWAIT) == 0 ? EX_WAITOK : EX_NOWAIT) |
844 		EX_BOUNDZERO, &dvmaddr);
845 	mutex_exit(&is->is_lock);
846 
847 	if (err != 0)
848 		return (err);
849 
850 #ifdef DEBUG
851 	if (dvmaddr == (u_long)-1)
852 	{
853 		printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
854 		    (int)sgsize, flags);
855 #ifdef DDB
856 		Debugger();
857 #endif
858 	}
859 #endif
860 	if (dvmaddr == (u_long)-1)
861 		return (ENOMEM);
862 
863 	/* Set the active DVMA map */
864 	map->_dm_dvmastart = dvmaddr;
865 	map->_dm_dvmasize = sgsize;
866 
867 	bmask = ~(boundary - 1);
868 	if ((pglist = segs[0]._ds_mlist) == NULL) {
869 		u_long prev_va = 0UL, last_va = dvmaddr;
870 		paddr_t prev_pa = 0;
871 		int end = 0, offset;
872 		bus_size_t len = size;
873 
874 		/*
875 		 * This segs is made up of individual physical
876 		 *  segments, probably by _bus_dmamap_load_uio() or
877 		 * _bus_dmamap_load_mbuf().  Ignore the mlist and
878 		 * load each one individually.
879 		 */
880 		j = 0;
881 		needsflush = 0;
882 		for (i = 0; i < nsegs ; i++) {
883 
884 			pa = segs[i].ds_addr;
885 			offset = (pa & PGOFSET);
886 			pa = trunc_page(pa);
887 			dvmaddr = trunc_page(dvmaddr);
888 			left = uimin(len, segs[i].ds_len);
889 
890 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: converting "
891 				"physseg %d start %lx size %lx\n", i,
892 				(long)segs[i].ds_addr, (long)segs[i].ds_len));
893 
894 			if ((pa == prev_pa) &&
895 				((offset != 0) || (end != offset))) {
896 				/* We can re-use this mapping */
897 				dvmaddr = prev_va;
898 			}
899 
900 			sgstart = dvmaddr + offset;
901 			sgend = sgstart + left - 1;
902 
903 			/* Are the segments virtually adjacent? */
904 			if ((j > 0) && (end == offset) &&
905 			    ((offset == 0) || (pa == prev_pa)) &&
906 			    (map->dm_segs[j-1].ds_len + left <=
907 			     map->dm_maxsegsz)) {
908 				/* Just append to the previous segment. */
909 				map->dm_segs[--j].ds_len += left;
910 				/* Restore sgstart for boundary check */
911 				sgstart = map->dm_segs[j].ds_addr;
912 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
913 					"appending seg %d start %lx size %lx\n", j,
914 					(long)map->dm_segs[j].ds_addr,
915 					(long)map->dm_segs[j].ds_len));
916 			} else {
917 				if (j >= map->_dm_segcnt) {
918 					iommu_remove(is, map->_dm_dvmastart,
919 					    last_va - map->_dm_dvmastart);
920 					goto fail;
921 				}
922 				map->dm_segs[j].ds_addr = sgstart;
923 				map->dm_segs[j].ds_len = left;
924 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
925 					"seg %d start %lx size %lx\n", j,
926 					(long)map->dm_segs[j].ds_addr,
927 					(long)map->dm_segs[j].ds_len));
928 			}
929 			end = (offset + left) & PGOFSET;
930 
931 			/* Check for boundary issues */
932 			while ((sgstart & bmask) != (sgend & bmask)) {
933 				/* Need a new segment. */
934 				map->dm_segs[j].ds_len =
935 					boundary - (sgstart & (boundary - 1));
936 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
937 					"seg %d start %lx size %lx\n", j,
938 					(long)map->dm_segs[j].ds_addr,
939 					(long)map->dm_segs[j].ds_len));
940 				if (++j >= map->_dm_segcnt) {
941 					iommu_remove(is, map->_dm_dvmastart,
942 					    last_va - map->_dm_dvmastart);
943 					goto fail;
944 				}
945 				sgstart += map->dm_segs[j-1].ds_len;
946 				map->dm_segs[j].ds_addr = sgstart;
947 				map->dm_segs[j].ds_len = sgend - sgstart + 1;
948 			}
949 
950 			if (sgsize == 0)
951 				panic("iommu_dmamap_load_raw: size botch");
952 
953 			/* Now map a series of pages. */
954 			while (dvmaddr <= sgend) {
955 				DPRINTF(IDB_BUSDMA,
956 					("iommu_dvmamap_load_raw: map %p "
957 						"loading va %lx at pa %lx\n",
958 						map, (long)dvmaddr,
959 						(long)(pa)));
960 				/* Enter it if we haven't before. */
961 				if (prev_va != dvmaddr) {
962 					iommu_enter(sb, prev_va = dvmaddr,
963 					    prev_pa = pa,
964 					    flags | IOTTE_DEBUG(++npg << 12));
965 					needsflush = 1;
966 
967 					/* Flush cache if necessary. */
968 					slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
969 					if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
970 					    ((dvmaddr + pagesz) > sgend || (slot % 8) == 7))
971 						IOMMUREG_WRITE(is, iommu_cache_flush,
972 						    is->is_ptsb + slot * 8);
973 				}
974 
975 				dvmaddr += pagesz;
976 				pa += pagesz;
977 				last_va = dvmaddr;
978 			}
979 
980 			len -= left;
981 			++j;
982 		}
983 		if (needsflush)
984 			iommu_strbuf_flush_done(sb);
985 
986 		map->dm_mapsize = size;
987 		map->dm_nsegs = j;
988 #ifdef DIAGNOSTIC
989 		{ int seg;
990 	for (seg = 0; seg < map->dm_nsegs; seg++) {
991 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
992 		    map->dm_segs[seg].ds_addr > is->is_dvmaend) {
993 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
994 				seg, (long)map->dm_segs[seg].ds_addr,
995 				is->is_dvmabase, is->is_dvmaend);
996 #ifdef DDB
997 			Debugger();
998 #endif
999 		}
1000 	}
1001 		}
1002 #endif
1003 		return (0);
1004 	}
1005 
1006 	/*
1007 	 * This was allocated with bus_dmamem_alloc.
1008 	 * The pages are on a `pglist'.
1009 	 */
1010 	i = 0;
1011 	sgstart = dvmaddr;
1012 	sgend = sgstart + size - 1;
1013 	map->dm_segs[i].ds_addr = sgstart;
1014 	while ((sgstart & bmask) != (sgend & bmask)) {
1015 		/* Oops.  We crossed a boundary.  Split the xfer. */
1016 		map->dm_segs[i].ds_len = boundary - (sgstart & (boundary - 1));
1017 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
1018 			"seg %d start %lx size %lx\n", i,
1019 			(long)map->dm_segs[i].ds_addr,
1020 			(long)map->dm_segs[i].ds_len));
1021 		if (++i >= map->_dm_segcnt) {
1022 			/* Too many segments.  Fail the operation. */
1023 			goto fail;
1024 		}
1025 		sgstart += map->dm_segs[i-1].ds_len;
1026 		map->dm_segs[i].ds_addr = sgstart;
1027 	}
1028 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
1029 			"seg %d start %lx size %lx\n", i,
1030 			(long)map->dm_segs[i].ds_addr, (long)map->dm_segs[i].ds_len));
1031 	map->dm_segs[i].ds_len = sgend - sgstart + 1;
1032 
1033 	needsflush = 0;
1034 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
1035 		if (sgsize == 0)
1036 			panic("iommu_dmamap_load_raw: size botch");
1037 		pa = VM_PAGE_TO_PHYS(pg);
1038 
1039 		DPRINTF(IDB_BUSDMA,
1040 		    ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
1041 		    map, (long)dvmaddr, (long)(pa)));
1042 		iommu_enter(sb, dvmaddr, pa, flags | IOTTE_DEBUG(0x8000));
1043 		needsflush = 1;
1044 
1045 		sgsize -= pagesz;
1046 
1047 		/* Flush cache if necessary. */
1048 		slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
1049 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
1050 		    (sgsize == 0 || (slot % 8) == 7))
1051 			IOMMUREG_WRITE(is, iommu_cache_flush,
1052 			    is->is_ptsb + slot * 8);
1053 
1054 		dvmaddr += pagesz;
1055 	}
1056 	if (needsflush)
1057 		iommu_strbuf_flush_done(sb);
1058 	map->dm_mapsize = size;
1059 	map->dm_nsegs = i+1;
1060 #ifdef DIAGNOSTIC
1061 	{ int seg;
1062 	for (seg = 0; seg < map->dm_nsegs; seg++) {
1063 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
1064 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
1065 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
1066 				seg, (long)map->dm_segs[seg].ds_addr,
1067 				is->is_dvmabase, is->is_dvmaend);
1068 #ifdef DDB
1069 			Debugger();
1070 #endif
1071 		}
1072 	}
1073 	}
1074 #endif
1075 	return (0);
1076 
1077 fail:
1078 	mutex_enter(&is->is_lock);
1079 	err = extent_free(is->is_dvmamap, map->_dm_dvmastart, sgsize,
1080 	    EX_NOWAIT);
1081 	map->_dm_dvmastart = 0;
1082 	map->_dm_dvmasize = 0;
1083 	mutex_exit(&is->is_lock);
1084 	if (err != 0)
1085 		printf("warning: %s: %" PRId64 " of DVMA space lost\n",
1086 		    __func__, sgsize);
1087 	return (EFBIG);
1088 }
1089 
1090 
1091 /*
1092  * Flush an individual dma segment, returns non-zero if the streaming buffers
1093  * need flushing afterwards.
1094  */
1095 static int
1096 iommu_dvmamap_sync_range(struct strbuf_ctl *sb, vaddr_t va, bus_size_t len)
1097 {
1098 	vaddr_t vaend;
1099 	struct iommu_state *is = sb->sb_is;
1100 
1101 #ifdef DIAGNOSTIC
1102 	if (va < is->is_dvmabase || va > is->is_dvmaend)
1103 		panic("invalid va: %llx", (long long)va);
1104 #endif
1105 
1106 	if ((is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)] & IOTTE_STREAM) == 0) {
1107 		DPRINTF(IDB_SYNC,
1108 			("iommu_dvmamap_sync_range: attempting to flush "
1109 			 "non-streaming entry\n"));
1110 		return (0);
1111 	}
1112 
1113 	vaend = round_page(va + len) - 1;
1114 	va = trunc_page(va);
1115 
1116 #ifdef DIAGNOSTIC
1117 	if (va < is->is_dvmabase || vaend > is->is_dvmaend)
1118 		panic("invalid va range: %llx to %llx (%x to %x)",
1119 		    (long long)va, (long long)vaend,
1120 		    is->is_dvmabase,
1121 		    is->is_dvmaend);
1122 #endif
1123 
1124 	for ( ; va <= vaend; va += PAGE_SIZE) {
1125 		DPRINTF(IDB_SYNC,
1126 		    ("iommu_dvmamap_sync_range: flushing va %p\n",
1127 		    (void *)(u_long)va));
1128 		iommu_strbuf_flush(sb, va);
1129 	}
1130 
1131 	return (1);
1132 }
1133 
1134 static void
1135 _iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1136 	bus_size_t len, int ops)
1137 {
1138 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
1139 	bus_size_t count;
1140 	int i, needsflush = 0;
1141 
1142 	if (!sb->sb_flush)
1143 		return;
1144 
1145 	for (i = 0; i < map->dm_nsegs; i++) {
1146 		if (offset < map->dm_segs[i].ds_len)
1147 			break;
1148 		offset -= map->dm_segs[i].ds_len;
1149 	}
1150 
1151 	if (i == map->dm_nsegs)
1152 		panic("%s: segment too short %llu", __func__,
1153 		    (unsigned long long)offset);
1154 
1155 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_POSTWRITE)) {
1156 		/* Nothing to do */;
1157 	}
1158 
1159 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_PREWRITE)) {
1160 
1161 		for (; len > 0 && i < map->dm_nsegs; i++) {
1162 			count = MIN(map->dm_segs[i].ds_len - offset, len);
1163 			if (count > 0 &&
1164 			    iommu_dvmamap_sync_range(sb,
1165 				map->dm_segs[i].ds_addr + offset, count))
1166 				needsflush = 1;
1167 			offset = 0;
1168 			len -= count;
1169 		}
1170 #ifdef DIAGNOSTIC
1171 		if (i == map->dm_nsegs && len > 0)
1172 			panic("%s: leftover %llu", __func__,
1173 			    (unsigned long long)len);
1174 #endif
1175 
1176 		if (needsflush)
1177 			iommu_strbuf_flush_done(sb);
1178 	}
1179 }
1180 
1181 void
1182 iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1183 	bus_size_t len, int ops)
1184 {
1185 
1186 	/* If len is 0, then there is nothing to do */
1187 	if (len == 0)
1188 		return;
1189 
1190 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) {
1191 		/* Flush the CPU then the IOMMU */
1192 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
1193 		_iommu_dvmamap_sync(t, map, offset, len, ops);
1194 	}
1195 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE)) {
1196 		/* Flush the IOMMU then the CPU */
1197 		_iommu_dvmamap_sync(t, map, offset, len, ops);
1198 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
1199 	}
1200 }
1201 
1202 int
1203 iommu_dvmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1204 	bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1205 	int flags)
1206 {
1207 
1208 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
1209 	   "segp %p flags %d\n", (unsigned long long)size,
1210 	   (unsigned long long)alignment, (unsigned long long)boundary,
1211 	   segs, flags));
1212 	return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
1213 	    segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
1214 }
1215 
1216 void
1217 iommu_dvmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
1218 {
1219 
1220 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
1221 	    segs, nsegs));
1222 	bus_dmamem_free(t->_parent, segs, nsegs);
1223 }
1224 
1225 /*
1226  * Map the DVMA mappings into the kernel pmap.
1227  * Check the flags to see whether we're streaming or coherent.
1228  */
1229 int
1230 iommu_dvmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1231 	size_t size, void **kvap, int flags)
1232 {
1233 	struct vm_page *pg;
1234 	vaddr_t va;
1235 	bus_addr_t addr;
1236 	struct pglist *pglist;
1237 	int cbit;
1238 	const uvm_flag_t kmflags =
1239 	    (flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0;
1240 
1241 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
1242 	    segs, nsegs, size));
1243 
1244 	/*
1245 	 * Allocate some space in the kernel map, and then map these pages
1246 	 * into this space.
1247 	 */
1248 	size = round_page(size);
1249 	va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY | kmflags);
1250 	if (va == 0)
1251 		return (ENOMEM);
1252 
1253 	*kvap = (void *)va;
1254 
1255 	/*
1256 	 * digest flags:
1257 	 */
1258 	cbit = 0;
1259 	if (flags & BUS_DMA_COHERENT)	/* Disable vcache */
1260 		cbit |= PMAP_NVC;
1261 	if (flags & BUS_DMA_NOCACHE)	/* side effects */
1262 		cbit |= PMAP_NC;
1263 
1264 	/*
1265 	 * Now take this and map it into the CPU.
1266 	 */
1267 	pglist = segs[0]._ds_mlist;
1268 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
1269 #ifdef DIAGNOSTIC
1270 		if (size == 0)
1271 			panic("iommu_dvmamem_map: size botch");
1272 #endif
1273 		addr = VM_PAGE_TO_PHYS(pg);
1274 		DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
1275 		    "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
1276 		pmap_kenter_pa(va, addr | cbit,
1277 		    VM_PROT_READ | VM_PROT_WRITE, 0);
1278 		va += PAGE_SIZE;
1279 		size -= PAGE_SIZE;
1280 	}
1281 	pmap_update(pmap_kernel());
1282 	return (0);
1283 }
1284 
1285 /*
1286  * Unmap DVMA mappings from kernel
1287  */
1288 void
1289 iommu_dvmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
1290 {
1291 
1292 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
1293 	    kva, size));
1294 
1295 #ifdef DIAGNOSTIC
1296 	if ((u_long)kva & PGOFSET)
1297 		panic("iommu_dvmamem_unmap");
1298 #endif
1299 
1300 	size = round_page(size);
1301 	pmap_kremove((vaddr_t)kva, size);
1302 	pmap_update(pmap_kernel());
1303 	uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
1304 }
1305