iommu.c revision 1.97 1 /* $NetBSD: iommu.c,v 1.97 2010/02/24 06:34:55 skrll 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 1.97 2010/02/24 06:34:55 skrll 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_extern.h>
74
75 #include <machine/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
82 #ifdef DEBUG
83 #define IDB_BUSDMA 0x1
84 #define IDB_IOMMU 0x2
85 #define IDB_INFO 0x4
86 #define IDB_SYNC 0x8
87 int iommudebug = 0x0;
88 #define DPRINTF(l, s) do { if (iommudebug & l) printf s; } while (0)
89 #define IOTTE_DEBUG(n) (n)
90 #else
91 #define DPRINTF(l, s)
92 #define IOTTE_DEBUG(n) 0
93 #endif
94
95 #define iommu_strbuf_flush(i, v) do { \
96 if ((i)->sb_flush) \
97 bus_space_write_8((i)->sb_is->is_bustag, (i)->sb_sb, \
98 STRBUFREG(strbuf_pgflush), (v)); \
99 } while (0)
100
101 static int iommu_strbuf_flush_done(struct strbuf_ctl *);
102 static void _iommu_dvmamap_sync(bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
103 bus_size_t, int);
104
105 /*
106 * initialise the UltraSPARC IOMMU (SBUS or PCI):
107 * - allocate and setup the iotsb.
108 * - enable the IOMMU
109 * - initialise the streaming buffers (if they exist)
110 * - create a private DVMA map.
111 */
112 void
113 iommu_init(char *name, struct iommu_state *is, int tsbsize, uint32_t iovabase)
114 {
115 psize_t size;
116 vaddr_t va;
117 paddr_t pa;
118 struct vm_page *pg;
119 struct pglist pglist;
120
121 /*
122 * Setup the iommu.
123 *
124 * The sun4u iommu is part of the SBUS or PCI controller so we will
125 * deal with it here..
126 *
127 * For sysio and psycho/psycho+ the IOMMU address space always ends at
128 * 0xffffe000, but the starting address depends on the size of the
129 * map. The map size is 1024 * 2 ^ is->is_tsbsize entries, where each
130 * entry is 8 bytes. The start of the map can be calculated by
131 * (0xffffe000 << (8 + is->is_tsbsize)).
132 *
133 * But sabre and hummingbird use a different scheme that seems to
134 * be hard-wired, so we read the start and size from the PROM and
135 * just use those values.
136 */
137 is->is_cr = (tsbsize << 16) | IOMMUCR_EN;
138 is->is_tsbsize = tsbsize;
139 if (iovabase == -1) {
140 is->is_dvmabase = IOTSB_VSTART(is->is_tsbsize);
141 is->is_dvmaend = IOTSB_VEND - 1;
142 } else {
143 is->is_dvmabase = iovabase;
144 is->is_dvmaend = iovabase + IOTSB_VSIZE(tsbsize) - 1;
145 }
146
147 /*
148 * Allocate memory for I/O pagetables. They need to be physically
149 * contiguous.
150 */
151
152 size = PAGE_SIZE << is->is_tsbsize;
153 if (uvm_pglistalloc((psize_t)size, (paddr_t)0, (paddr_t)-1,
154 (paddr_t)PAGE_SIZE, (paddr_t)0, &pglist, 1, 0) != 0)
155 panic("iommu_init: no memory");
156
157 va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY);
158 if (va == 0)
159 panic("iommu_init: no memory");
160 is->is_tsb = (int64_t *)va;
161
162 is->is_ptsb = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pglist));
163
164 /* Map the pages */
165 TAILQ_FOREACH(pg, &pglist, pageq.queue) {
166 pa = VM_PAGE_TO_PHYS(pg);
167 pmap_kenter_pa(va, pa | PMAP_NVC,
168 VM_PROT_READ | VM_PROT_WRITE, 0);
169 va += PAGE_SIZE;
170 }
171 pmap_update(pmap_kernel());
172 memset(is->is_tsb, 0, size);
173
174 #ifdef DEBUG
175 if (iommudebug & IDB_INFO)
176 {
177 /* Probe the iommu */
178
179 printf("iommu regs at: cr=%lx tsb=%lx flush=%lx\n",
180 (u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
181 offsetof (struct iommureg, iommu_cr)),
182 (u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
183 offsetof (struct iommureg, iommu_tsb)),
184 (u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
185 offsetof (struct iommureg, iommu_flush)));
186 printf("iommu cr=%llx tsb=%llx\n",
187 (unsigned long long)bus_space_read_8(is->is_bustag,
188 is->is_iommu,
189 offsetof (struct iommureg, iommu_cr)),
190 (unsigned long long)bus_space_read_8(is->is_bustag,
191 is->is_iommu,
192 offsetof (struct iommureg, iommu_tsb)));
193 printf("TSB base %p phys %llx\n", (void *)is->is_tsb,
194 (unsigned long long)is->is_ptsb);
195 delay(1000000); /* 1 s */
196 }
197 #endif
198
199 /*
200 * now actually start up the IOMMU
201 */
202 iommu_reset(is);
203
204 /*
205 * Now all the hardware's working we need to allocate a dvma map.
206 */
207 printf("DVMA map: %x to %x\n",
208 (unsigned int)is->is_dvmabase,
209 (unsigned int)is->is_dvmaend);
210 printf("IOTSB: %llx to %llx\n",
211 (unsigned long long)is->is_ptsb,
212 (unsigned long long)(is->is_ptsb + size - 1));
213 is->is_dvmamap = extent_create(name,
214 is->is_dvmabase, is->is_dvmaend,
215 M_DEVBUF, 0, 0, EX_NOWAIT);
216 }
217
218 /*
219 * Streaming buffers don't exist on the UltraSPARC IIi; we should have
220 * detected that already and disabled them. If not, we will notice that
221 * they aren't there when the STRBUF_EN bit does not remain.
222 */
223 void
224 iommu_reset(struct iommu_state *is)
225 {
226 int i;
227 struct strbuf_ctl *sb;
228
229 /* Need to do 64-bit stores */
230 bus_space_write_8(is->is_bustag, is->is_iommu, IOMMUREG(iommu_tsb),
231 is->is_ptsb);
232
233 /* Enable IOMMU in diagnostic mode */
234 bus_space_write_8(is->is_bustag, is->is_iommu, IOMMUREG(iommu_cr),
235 is->is_cr|IOMMUCR_DE);
236
237 for (i = 0; i < 2; i++) {
238 if ((sb = is->is_sb[i])) {
239
240 /* Enable diagnostics mode? */
241 bus_space_write_8(is->is_bustag, is->is_sb[i]->sb_sb,
242 STRBUFREG(strbuf_ctl), STRBUF_EN);
243
244 /* No streaming buffers? Disable them */
245 if (bus_space_read_8(is->is_bustag,
246 is->is_sb[i]->sb_sb,
247 STRBUFREG(strbuf_ctl)) == 0) {
248 is->is_sb[i]->sb_flush = NULL;
249 } else {
250
251 /*
252 * locate the pa of the flush buffer.
253 */
254 (void)pmap_extract(pmap_kernel(),
255 (vaddr_t)is->is_sb[i]->sb_flush,
256 &is->is_sb[i]->sb_flushpa);
257 }
258 }
259 }
260 }
261
262 /*
263 * Here are the iommu control routines.
264 */
265 void
266 iommu_enter(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
267 {
268 struct iommu_state *is = sb->sb_is;
269 int strbuf = (flags & BUS_DMA_STREAMING);
270 int64_t tte;
271
272 #ifdef DIAGNOSTIC
273 if (va < is->is_dvmabase || va > is->is_dvmaend)
274 panic("iommu_enter: va %#lx not in DVMA space", va);
275 #endif
276
277 /* Is the streamcache flush really needed? */
278 if (sb->sb_flush)
279 iommu_strbuf_flush(sb, va);
280 else
281 /* If we can't flush the strbuf don't enable it. */
282 strbuf = 0;
283
284 tte = MAKEIOTTE(pa, !(flags & BUS_DMA_NOWRITE),
285 !(flags & BUS_DMA_NOCACHE), (strbuf));
286 #ifdef DEBUG
287 tte |= (flags & 0xff000LL)<<(4*8);
288 #endif
289
290 DPRINTF(IDB_IOMMU, ("Clearing TSB slot %d for va %p\n",
291 (int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va));
292 is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = tte;
293 bus_space_write_8(is->is_bustag, is->is_iommu,
294 IOMMUREG(iommu_flush), va);
295 DPRINTF(IDB_IOMMU, ("iommu_enter: va %lx pa %lx TSB[%lx]@%p=%lx\n",
296 va, (long)pa, (u_long)IOTSBSLOT(va,is->is_tsbsize),
297 (void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
298 (u_long)tte));
299 }
300
301 /*
302 * Find the value of a DVMA address (debug routine).
303 */
304 paddr_t
305 iommu_extract(struct iommu_state *is, vaddr_t dva)
306 {
307 int64_t tte = 0;
308
309 if (dva >= is->is_dvmabase && dva <= is->is_dvmaend)
310 tte = is->is_tsb[IOTSBSLOT(dva, is->is_tsbsize)];
311
312 if ((tte & IOTTE_V) == 0)
313 return ((paddr_t)-1L);
314 return (tte & IOTTE_PAMASK);
315 }
316
317 /*
318 * iommu_remove: removes mappings created by iommu_enter
319 *
320 * Only demap from IOMMU if flag is set.
321 *
322 * XXX: this function needs better internal error checking.
323 */
324 void
325 iommu_remove(struct iommu_state *is, vaddr_t va, size_t len)
326 {
327
328 #ifdef DIAGNOSTIC
329 if (va < is->is_dvmabase || va > is->is_dvmaend)
330 panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
331 if ((long)(va + len) < (long)va)
332 panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
333 (long) va, (long) len);
334 if (len & ~0xfffffff)
335 panic("iommu_remove: ridiculous len 0x%lx", (u_long)len);
336 #endif
337
338 va = trunc_page(va);
339 DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
340 va, (u_long)IOTSBSLOT(va, is->is_tsbsize),
341 &is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)]));
342 while (len > 0) {
343 DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d "
344 "for va %p size %lx\n",
345 (int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va,
346 (u_long)len));
347 if (len <= PAGE_SIZE)
348 len = 0;
349 else
350 len -= PAGE_SIZE;
351
352 /*
353 * XXX Zero-ing the entry would not require RMW
354 *
355 * Disabling valid bit while a page is used by a device
356 * causes an uncorrectable DMA error.
357 * Workaround to avoid an uncorrectable DMA error is
358 * eliminating the next line, but the page is mapped
359 * until the next iommu_enter call.
360 */
361 is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] &= ~IOTTE_V;
362 membar_storestore();
363 bus_space_write_8(is->is_bustag, is->is_iommu,
364 IOMMUREG(iommu_flush), va);
365 va += PAGE_SIZE;
366 }
367 }
368
369 static int
370 iommu_strbuf_flush_done(struct strbuf_ctl *sb)
371 {
372 struct iommu_state *is = sb->sb_is;
373 struct timeval cur, flushtimeout;
374
375 #define BUMPTIME(t, usec) { \
376 register volatile struct timeval *tp = (t); \
377 register long us; \
378 \
379 tp->tv_usec = us = tp->tv_usec + (usec); \
380 if (us >= 1000000) { \
381 tp->tv_usec = us - 1000000; \
382 tp->tv_sec++; \
383 } \
384 }
385
386 if (!sb->sb_flush)
387 return (0);
388
389 /*
390 * Streaming buffer flushes:
391 *
392 * 1 Tell strbuf to flush by storing va to strbuf_pgflush. If
393 * we're not on a cache line boundary (64-bits):
394 * 2 Store 0 in flag
395 * 3 Store pointer to flag in flushsync
396 * 4 wait till flushsync becomes 0x1
397 *
398 * If it takes more than .5 sec, something
399 * went wrong.
400 */
401
402 *sb->sb_flush = 0;
403 bus_space_write_8(is->is_bustag, sb->sb_sb,
404 STRBUFREG(strbuf_flushsync), sb->sb_flushpa);
405
406 microtime(&flushtimeout);
407 cur = flushtimeout;
408 BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
409
410 DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flush = %lx "
411 "at va = %lx pa = %lx now=%"PRIx64":%"PRIx32" until = %"PRIx64":%"PRIx32"\n",
412 (long)*sb->sb_flush, (long)sb->sb_flush, (long)sb->sb_flushpa,
413 cur.tv_sec, cur.tv_usec,
414 flushtimeout.tv_sec, flushtimeout.tv_usec));
415
416 /* Bypass non-coherent D$ */
417 while ((!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) &&
418 timercmp(&cur, &flushtimeout, <=))
419 microtime(&cur);
420
421 #ifdef DIAGNOSTIC
422 if (!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) {
423 printf("iommu_strbuf_flush_done: flush timeout %p, at %p\n",
424 (void *)(u_long)*sb->sb_flush,
425 (void *)(u_long)sb->sb_flushpa); /* panic? */
426 #ifdef DDB
427 Debugger();
428 #endif
429 }
430 #endif
431 DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flushed\n"));
432 return (*sb->sb_flush);
433 }
434
435 /*
436 * IOMMU DVMA operations, common to SBUS and PCI.
437 */
438 int
439 iommu_dvmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
440 bus_size_t buflen, struct proc *p, int flags)
441 {
442 struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
443 struct iommu_state *is = sb->sb_is;
444 int s;
445 int err, needsflush;
446 bus_size_t sgsize;
447 paddr_t curaddr;
448 u_long dvmaddr, sgstart, sgend, bmask;
449 bus_size_t align, boundary, len;
450 vaddr_t vaddr = (vaddr_t)buf;
451 int seg;
452 struct pmap *pmap;
453
454 if (map->dm_nsegs) {
455 /* Already in use?? */
456 #ifdef DIAGNOSTIC
457 printf("iommu_dvmamap_load: map still in use\n");
458 #endif
459 bus_dmamap_unload(t, map);
460 }
461
462 /*
463 * Make sure that on error condition we return "no valid mappings".
464 */
465 map->dm_nsegs = 0;
466 KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
467
468 if (buflen > map->_dm_size) {
469 DPRINTF(IDB_BUSDMA,
470 ("iommu_dvmamap_load(): error %d > %d -- "
471 "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
472 return (EINVAL);
473 }
474
475 sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
476
477 /*
478 * A boundary presented to bus_dmamem_alloc() takes precedence
479 * over boundary in the map.
480 */
481 if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
482 boundary = map->_dm_boundary;
483 align = max(map->dm_segs[0]._ds_align, PAGE_SIZE);
484
485 /*
486 * If our segment size is larger than the boundary we need to
487 * split the transfer up int little pieces ourselves.
488 */
489 s = splhigh();
490 err = extent_alloc(is->is_dvmamap, sgsize, align,
491 (sgsize > boundary) ? 0 : boundary,
492 EX_NOWAIT|EX_BOUNDZERO, &dvmaddr);
493 splx(s);
494
495 #ifdef DEBUG
496 if (err || (dvmaddr == (u_long)-1)) {
497 printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
498 (int)sgsize, flags);
499 #ifdef DDB
500 Debugger();
501 #endif
502 }
503 #endif
504 if (err != 0)
505 return (err);
506
507 if (dvmaddr == (u_long)-1)
508 return (ENOMEM);
509
510 /* Set the active DVMA map */
511 map->_dm_dvmastart = dvmaddr;
512 map->_dm_dvmasize = sgsize;
513
514 /*
515 * Now split the DVMA range into segments, not crossing
516 * the boundary.
517 */
518 seg = 0;
519 sgstart = dvmaddr + (vaddr & PGOFSET);
520 sgend = sgstart + buflen - 1;
521 map->dm_segs[seg].ds_addr = sgstart;
522 DPRINTF(IDB_INFO, ("iommu_dvmamap_load: boundary %lx boundary - 1 %lx "
523 "~(boundary - 1) %lx\n", (long)boundary, (long)(boundary - 1),
524 (long)~(boundary - 1)));
525 bmask = ~(boundary - 1);
526 while ((sgstart & bmask) != (sgend & bmask) ||
527 sgend - sgstart + 1 > map->dm_maxsegsz) {
528 /* Oops. We crossed a boundary or large seg. Split the xfer. */
529 len = map->dm_maxsegsz;
530 if ((sgstart & bmask) != (sgend & bmask))
531 len = min(len, boundary - (sgstart & (boundary - 1)));
532 map->dm_segs[seg].ds_len = len;
533 DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
534 "seg %d start %lx size %lx\n", seg,
535 (long)map->dm_segs[seg].ds_addr,
536 (long)map->dm_segs[seg].ds_len));
537 if (++seg >= map->_dm_segcnt) {
538 /* Too many segments. Fail the operation. */
539 DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
540 "too many segments %d\n", seg));
541 s = splhigh();
542 err = extent_free(is->is_dvmamap,
543 dvmaddr, sgsize, EX_NOWAIT);
544 map->_dm_dvmastart = 0;
545 map->_dm_dvmasize = 0;
546 splx(s);
547 if (err != 0)
548 printf("warning: %s: %" PRId64
549 " of DVMA space lost\n", __func__, sgsize);
550 return (EFBIG);
551 }
552 sgstart += len;
553 map->dm_segs[seg].ds_addr = sgstart;
554 }
555 map->dm_segs[seg].ds_len = sgend - sgstart + 1;
556 DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
557 "seg %d start %lx size %lx\n", seg,
558 (long)map->dm_segs[seg].ds_addr, (long)map->dm_segs[seg].ds_len));
559 map->dm_nsegs = seg + 1;
560 map->dm_mapsize = buflen;
561
562 if (p != NULL)
563 pmap = p->p_vmspace->vm_map.pmap;
564 else
565 pmap = pmap_kernel();
566
567 needsflush = 0;
568 for (; buflen > 0; ) {
569
570 /*
571 * Get the physical address for this page.
572 */
573 if (pmap_extract(pmap, (vaddr_t)vaddr, &curaddr) == FALSE) {
574 #ifdef DIAGNOSTIC
575 printf("iommu_dvmamap_load: pmap_extract failed %lx\n", vaddr);
576 #endif
577 bus_dmamap_unload(t, map);
578 return (-1);
579 }
580
581 /*
582 * Compute the segment size, and adjust counts.
583 */
584 sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
585 if (buflen < sgsize)
586 sgsize = buflen;
587
588 DPRINTF(IDB_BUSDMA,
589 ("iommu_dvmamap_load: map %p loading va %p "
590 "dva %lx at pa %lx\n",
591 map, (void *)vaddr, (long)dvmaddr,
592 (long)trunc_page(curaddr)));
593 iommu_enter(sb, trunc_page(dvmaddr), trunc_page(curaddr),
594 flags | IOTTE_DEBUG(0x4000));
595 needsflush = 1;
596
597 dvmaddr += PAGE_SIZE;
598 vaddr += sgsize;
599 buflen -= sgsize;
600 }
601 if (needsflush)
602 iommu_strbuf_flush_done(sb);
603 #ifdef DIAGNOSTIC
604 for (seg = 0; seg < map->dm_nsegs; seg++) {
605 if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
606 map->dm_segs[seg].ds_addr > is->is_dvmaend) {
607 printf("seg %d dvmaddr %lx out of range %x - %x\n",
608 seg, (long)map->dm_segs[seg].ds_addr,
609 is->is_dvmabase, is->is_dvmaend);
610 #ifdef DDB
611 Debugger();
612 #endif
613 }
614 }
615 #endif
616 return (0);
617 }
618
619
620 void
621 iommu_dvmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
622 {
623 struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
624 struct iommu_state *is = sb->sb_is;
625 int error, s;
626 bus_size_t sgsize = map->_dm_dvmasize;
627
628 /* Flush the iommu */
629 #ifdef DEBUG
630 if (!map->_dm_dvmastart) {
631 printf("iommu_dvmamap_unload: No dvmastart is zero\n");
632 #ifdef DDB
633 Debugger();
634 #endif
635 }
636 #endif
637 iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
638
639 /* Flush the caches */
640 bus_dmamap_unload(t->_parent, map);
641
642 /* Mark the mappings as invalid. */
643 map->dm_mapsize = 0;
644 map->dm_nsegs = 0;
645
646 s = splhigh();
647 error = extent_free(is->is_dvmamap, map->_dm_dvmastart,
648 map->_dm_dvmasize, EX_NOWAIT);
649 map->_dm_dvmastart = 0;
650 map->_dm_dvmasize = 0;
651 splx(s);
652 if (error != 0)
653 printf("warning: %s: %" PRId64 " of DVMA space lost\n",
654 __func__, sgsize);
655
656 /* Clear the map */
657 }
658
659
660 int
661 iommu_dvmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
662 bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
663 {
664 struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
665 struct iommu_state *is = sb->sb_is;
666 struct vm_page *pg;
667 int i, j, s;
668 int left;
669 int err, needsflush;
670 bus_size_t sgsize;
671 paddr_t pa;
672 bus_size_t boundary, align;
673 u_long dvmaddr, sgstart, sgend, bmask;
674 struct pglist *pglist;
675 const int pagesz = PAGE_SIZE;
676 #ifdef DEBUG
677 int npg = 0;
678 #endif
679
680 if (map->dm_nsegs) {
681 /* Already in use?? */
682 #ifdef DIAGNOSTIC
683 printf("iommu_dvmamap_load_raw: map still in use\n");
684 #endif
685 bus_dmamap_unload(t, map);
686 }
687
688 /*
689 * A boundary presented to bus_dmamem_alloc() takes precedence
690 * over boundary in the map.
691 */
692 if ((boundary = segs[0]._ds_boundary) == 0)
693 boundary = map->_dm_boundary;
694
695 align = max(segs[0]._ds_align, pagesz);
696
697 /*
698 * Make sure that on error condition we return "no valid mappings".
699 */
700 map->dm_nsegs = 0;
701 /* Count up the total number of pages we need */
702 pa = trunc_page(segs[0].ds_addr);
703 sgsize = 0;
704 left = size;
705 for (i = 0; left > 0 && i < nsegs; i++) {
706 if (round_page(pa) != round_page(segs[i].ds_addr))
707 sgsize = round_page(sgsize) +
708 (segs[i].ds_addr & PGOFSET);
709 sgsize += min(left, segs[i].ds_len);
710 left -= segs[i].ds_len;
711 pa = segs[i].ds_addr + segs[i].ds_len;
712 }
713 sgsize = round_page(sgsize);
714
715 s = splhigh();
716 /*
717 * If our segment size is larger than the boundary we need to
718 * split the transfer up into little pieces ourselves.
719 */
720 err = extent_alloc(is->is_dvmamap, sgsize, align,
721 (sgsize > boundary) ? 0 : boundary,
722 ((flags & BUS_DMA_NOWAIT) == 0 ? EX_WAITOK : EX_NOWAIT) |
723 EX_BOUNDZERO, &dvmaddr);
724 splx(s);
725
726 if (err != 0)
727 return (err);
728
729 #ifdef DEBUG
730 if (dvmaddr == (u_long)-1)
731 {
732 printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
733 (int)sgsize, flags);
734 #ifdef DDB
735 Debugger();
736 #endif
737 }
738 #endif
739 if (dvmaddr == (u_long)-1)
740 return (ENOMEM);
741
742 /* Set the active DVMA map */
743 map->_dm_dvmastart = dvmaddr;
744 map->_dm_dvmasize = sgsize;
745
746 bmask = ~(boundary - 1);
747 if ((pglist = segs[0]._ds_mlist) == NULL) {
748 u_long prev_va = 0UL, last_va = dvmaddr;
749 paddr_t prev_pa = 0;
750 int end = 0, offset;
751 bus_size_t len = size;
752
753 /*
754 * This segs is made up of individual physical
755 * segments, probably by _bus_dmamap_load_uio() or
756 * _bus_dmamap_load_mbuf(). Ignore the mlist and
757 * load each one individually.
758 */
759 j = 0;
760 needsflush = 0;
761 for (i = 0; i < nsegs ; i++) {
762
763 pa = segs[i].ds_addr;
764 offset = (pa & PGOFSET);
765 pa = trunc_page(pa);
766 dvmaddr = trunc_page(dvmaddr);
767 left = min(len, segs[i].ds_len);
768
769 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: converting "
770 "physseg %d start %lx size %lx\n", i,
771 (long)segs[i].ds_addr, (long)segs[i].ds_len));
772
773 if ((pa == prev_pa) &&
774 ((offset != 0) || (end != offset))) {
775 /* We can re-use this mapping */
776 dvmaddr = prev_va;
777 }
778
779 sgstart = dvmaddr + offset;
780 sgend = sgstart + left - 1;
781
782 /* Are the segments virtually adjacent? */
783 if ((j > 0) && (end == offset) &&
784 ((offset == 0) || (pa == prev_pa)) &&
785 (map->dm_segs[j-1].ds_len + left <=
786 map->dm_maxsegsz)) {
787 /* Just append to the previous segment. */
788 map->dm_segs[--j].ds_len += left;
789 /* Restore sgstart for boundary check */
790 sgstart = map->dm_segs[j].ds_addr;
791 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
792 "appending seg %d start %lx size %lx\n", j,
793 (long)map->dm_segs[j].ds_addr,
794 (long)map->dm_segs[j].ds_len));
795 } else {
796 if (j >= map->_dm_segcnt) {
797 iommu_remove(is, map->_dm_dvmastart,
798 last_va - map->_dm_dvmastart);
799 goto fail;
800 }
801 map->dm_segs[j].ds_addr = sgstart;
802 map->dm_segs[j].ds_len = left;
803 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
804 "seg %d start %lx size %lx\n", j,
805 (long)map->dm_segs[j].ds_addr,
806 (long)map->dm_segs[j].ds_len));
807 }
808 end = (offset + left) & PGOFSET;
809
810 /* Check for boundary issues */
811 while ((sgstart & bmask) != (sgend & bmask)) {
812 /* Need a new segment. */
813 map->dm_segs[j].ds_len =
814 boundary - (sgstart & (boundary - 1));
815 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
816 "seg %d start %lx size %lx\n", j,
817 (long)map->dm_segs[j].ds_addr,
818 (long)map->dm_segs[j].ds_len));
819 if (++j >= map->_dm_segcnt) {
820 iommu_remove(is, map->_dm_dvmastart,
821 last_va - map->_dm_dvmastart);
822 goto fail;
823 }
824 sgstart += map->dm_segs[j-1].ds_len;
825 map->dm_segs[j].ds_addr = sgstart;
826 map->dm_segs[j].ds_len = sgend - sgstart + 1;
827 }
828
829 if (sgsize == 0)
830 panic("iommu_dmamap_load_raw: size botch");
831
832 /* Now map a series of pages. */
833 while (dvmaddr <= sgend) {
834 DPRINTF(IDB_BUSDMA,
835 ("iommu_dvmamap_load_raw: map %p "
836 "loading va %lx at pa %lx\n",
837 map, (long)dvmaddr,
838 (long)(pa)));
839 /* Enter it if we haven't before. */
840 if (prev_va != dvmaddr) {
841 iommu_enter(sb, prev_va = dvmaddr,
842 prev_pa = pa,
843 flags | IOTTE_DEBUG(++npg << 12));
844 needsflush = 1;
845 }
846 dvmaddr += pagesz;
847 pa += pagesz;
848 last_va = dvmaddr;
849 }
850
851 len -= left;
852 ++j;
853 }
854 if (needsflush)
855 iommu_strbuf_flush_done(sb);
856
857 map->dm_mapsize = size;
858 map->dm_nsegs = j;
859 #ifdef DIAGNOSTIC
860 { int seg;
861 for (seg = 0; seg < map->dm_nsegs; seg++) {
862 if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
863 map->dm_segs[seg].ds_addr > is->is_dvmaend) {
864 printf("seg %d dvmaddr %lx out of range %x - %x\n",
865 seg, (long)map->dm_segs[seg].ds_addr,
866 is->is_dvmabase, is->is_dvmaend);
867 #ifdef DDB
868 Debugger();
869 #endif
870 }
871 }
872 }
873 #endif
874 return (0);
875 }
876
877 /*
878 * This was allocated with bus_dmamem_alloc.
879 * The pages are on a `pglist'.
880 */
881 i = 0;
882 sgstart = dvmaddr;
883 sgend = sgstart + size - 1;
884 map->dm_segs[i].ds_addr = sgstart;
885 while ((sgstart & bmask) != (sgend & bmask)) {
886 /* Oops. We crossed a boundary. Split the xfer. */
887 map->dm_segs[i].ds_len = boundary - (sgstart & (boundary - 1));
888 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
889 "seg %d start %lx size %lx\n", i,
890 (long)map->dm_segs[i].ds_addr,
891 (long)map->dm_segs[i].ds_len));
892 if (++i >= map->_dm_segcnt) {
893 /* Too many segments. Fail the operation. */
894 goto fail;
895 }
896 sgstart += map->dm_segs[i-1].ds_len;
897 map->dm_segs[i].ds_addr = sgstart;
898 }
899 DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
900 "seg %d start %lx size %lx\n", i,
901 (long)map->dm_segs[i].ds_addr, (long)map->dm_segs[i].ds_len));
902 map->dm_segs[i].ds_len = sgend - sgstart + 1;
903
904 needsflush = 0;
905 TAILQ_FOREACH(pg, pglist, pageq.queue) {
906 if (sgsize == 0)
907 panic("iommu_dmamap_load_raw: size botch");
908 pa = VM_PAGE_TO_PHYS(pg);
909
910 DPRINTF(IDB_BUSDMA,
911 ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
912 map, (long)dvmaddr, (long)(pa)));
913 iommu_enter(sb, dvmaddr, pa, flags | IOTTE_DEBUG(0x8000));
914 needsflush = 1;
915
916 dvmaddr += pagesz;
917 sgsize -= pagesz;
918 }
919 if (needsflush)
920 iommu_strbuf_flush_done(sb);
921 map->dm_mapsize = size;
922 map->dm_nsegs = i+1;
923 #ifdef DIAGNOSTIC
924 { int seg;
925 for (seg = 0; seg < map->dm_nsegs; seg++) {
926 if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
927 map->dm_segs[seg].ds_addr > is->is_dvmaend) {
928 printf("seg %d dvmaddr %lx out of range %x - %x\n",
929 seg, (long)map->dm_segs[seg].ds_addr,
930 is->is_dvmabase, is->is_dvmaend);
931 #ifdef DDB
932 Debugger();
933 #endif
934 }
935 }
936 }
937 #endif
938 return (0);
939
940 fail:
941 s = splhigh();
942 err = extent_free(is->is_dvmamap, map->_dm_dvmastart, sgsize,
943 EX_NOWAIT);
944 map->_dm_dvmastart = 0;
945 map->_dm_dvmasize = 0;
946 splx(s);
947 if (err != 0)
948 printf("warning: %s: %" PRId64 " of DVMA space lost\n",
949 __func__, sgsize);
950 return (EFBIG);
951 }
952
953
954 /*
955 * Flush an individual dma segment, returns non-zero if the streaming buffers
956 * need flushing afterwards.
957 */
958 static int
959 iommu_dvmamap_sync_range(struct strbuf_ctl *sb, vaddr_t va, bus_size_t len)
960 {
961 vaddr_t vaend;
962 struct iommu_state *is = sb->sb_is;
963
964 #ifdef DIAGNOSTIC
965 if (va < is->is_dvmabase || va > is->is_dvmaend)
966 panic("invalid va: %llx", (long long)va);
967 #endif
968
969 if ((is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)] & IOTTE_STREAM) == 0) {
970 DPRINTF(IDB_BUSDMA,
971 ("iommu_dvmamap_sync_range: attempting to flush "
972 "non-streaming entry\n"));
973 return (0);
974 }
975
976 vaend = round_page(va + len) - 1;
977 va = trunc_page(va);
978
979 #ifdef DIAGNOSTIC
980 if (va < is->is_dvmabase || vaend > is->is_dvmaend)
981 panic("invalid va range: %llx to %llx (%x to %x)",
982 (long long)va, (long long)vaend,
983 is->is_dvmabase,
984 is->is_dvmaend);
985 #endif
986
987 for ( ; va <= vaend; va += PAGE_SIZE) {
988 DPRINTF(IDB_BUSDMA,
989 ("iommu_dvmamap_sync_range: flushing va %p\n",
990 (void *)(u_long)va));
991 iommu_strbuf_flush(sb, va);
992 }
993
994 return (1);
995 }
996
997 static void
998 _iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
999 bus_size_t len, int ops)
1000 {
1001 struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
1002 bus_size_t count;
1003 int i, needsflush = 0;
1004
1005 if (!sb->sb_flush)
1006 return;
1007
1008 for (i = 0; i < map->dm_nsegs; i++) {
1009 if (offset < map->dm_segs[i].ds_len)
1010 break;
1011 offset -= map->dm_segs[i].ds_len;
1012 }
1013
1014 if (i == map->dm_nsegs)
1015 panic("iommu_dvmamap_sync: segment too short %llu",
1016 (unsigned long long)offset);
1017
1018 if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_POSTWRITE)) {
1019 /* Nothing to do */;
1020 }
1021
1022 if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_PREWRITE)) {
1023
1024 for (; len > 0 && i < map->dm_nsegs; i++) {
1025 count = MIN(map->dm_segs[i].ds_len - offset, len);
1026 if (count > 0 &&
1027 iommu_dvmamap_sync_range(sb,
1028 map->dm_segs[i].ds_addr + offset, count))
1029 needsflush = 1;
1030 offset = 0;
1031 len -= count;
1032 }
1033 #ifdef DIAGNOSTIC
1034 if (i == map->dm_nsegs && len > 0)
1035 panic("iommu_dvmamap_sync: leftover %llu",
1036 (unsigned long long)len);
1037 #endif
1038
1039 if (needsflush)
1040 iommu_strbuf_flush_done(sb);
1041 }
1042 }
1043
1044 void
1045 iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1046 bus_size_t len, int ops)
1047 {
1048
1049 /* If len is 0, then there is nothing to do */
1050 if (len == 0)
1051 return;
1052
1053 if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) {
1054 /* Flush the CPU then the IOMMU */
1055 bus_dmamap_sync(t->_parent, map, offset, len, ops);
1056 _iommu_dvmamap_sync(t, map, offset, len, ops);
1057 }
1058 if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE)) {
1059 /* Flush the IOMMU then the CPU */
1060 _iommu_dvmamap_sync(t, map, offset, len, ops);
1061 bus_dmamap_sync(t->_parent, map, offset, len, ops);
1062 }
1063 }
1064
1065 int
1066 iommu_dvmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1067 bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1068 int flags)
1069 {
1070
1071 DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
1072 "segp %p flags %d\n", (unsigned long long)size,
1073 (unsigned long long)alignment, (unsigned long long)boundary,
1074 segs, flags));
1075 return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
1076 segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
1077 }
1078
1079 void
1080 iommu_dvmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
1081 {
1082
1083 DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
1084 segs, nsegs));
1085 bus_dmamem_free(t->_parent, segs, nsegs);
1086 }
1087
1088 /*
1089 * Map the DVMA mappings into the kernel pmap.
1090 * Check the flags to see whether we're streaming or coherent.
1091 */
1092 int
1093 iommu_dvmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1094 size_t size, void **kvap, int flags)
1095 {
1096 struct vm_page *pg;
1097 vaddr_t va;
1098 bus_addr_t addr;
1099 struct pglist *pglist;
1100 int cbit;
1101 const uvm_flag_t kmflags =
1102 (flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0;
1103
1104 DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
1105 segs, nsegs, size));
1106
1107 /*
1108 * Allocate some space in the kernel map, and then map these pages
1109 * into this space.
1110 */
1111 size = round_page(size);
1112 va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY | kmflags);
1113 if (va == 0)
1114 return (ENOMEM);
1115
1116 *kvap = (void *)va;
1117
1118 /*
1119 * digest flags:
1120 */
1121 cbit = 0;
1122 if (flags & BUS_DMA_COHERENT) /* Disable vcache */
1123 cbit |= PMAP_NVC;
1124 if (flags & BUS_DMA_NOCACHE) /* side effects */
1125 cbit |= PMAP_NC;
1126
1127 /*
1128 * Now take this and map it into the CPU.
1129 */
1130 pglist = segs[0]._ds_mlist;
1131 TAILQ_FOREACH(pg, pglist, pageq.queue) {
1132 #ifdef DIAGNOSTIC
1133 if (size == 0)
1134 panic("iommu_dvmamem_map: size botch");
1135 #endif
1136 addr = VM_PAGE_TO_PHYS(pg);
1137 DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
1138 "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
1139 pmap_kenter_pa(va, addr | cbit,
1140 VM_PROT_READ | VM_PROT_WRITE, 0);
1141 va += PAGE_SIZE;
1142 size -= PAGE_SIZE;
1143 }
1144 pmap_update(pmap_kernel());
1145 return (0);
1146 }
1147
1148 /*
1149 * Unmap DVMA mappings from kernel
1150 */
1151 void
1152 iommu_dvmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
1153 {
1154
1155 DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
1156 kva, size));
1157
1158 #ifdef DIAGNOSTIC
1159 if ((u_long)kva & PGOFSET)
1160 panic("iommu_dvmamem_unmap");
1161 #endif
1162
1163 size = round_page(size);
1164 pmap_kremove((vaddr_t)kva, size);
1165 pmap_update(pmap_kernel());
1166 uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
1167 }
1168