Home | History | Annotate | Line # | Download | only in dev
      1 /*	$NetBSD: iommu.c,v 1.119 2023/12/20 05:33:58 thorpej 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 1.119 2023/12/20 05:33:58 thorpej Exp $");
     63 
     64 #include "opt_ddb.h"
     65 
     66 #include <sys/param.h>
     67 #include <sys/extent.h>
     68 #include <sys/systm.h>
     69 #include <sys/device.h>
     70 #include <sys/proc.h>
     71 
     72 #include <uvm/uvm.h>
     73 
     74 #include <sys/bus.h>
     75 #include <sparc64/dev/iommureg.h>
     76 #include <sparc64/dev/iommuvar.h>
     77 
     78 #include <machine/autoconf.h>
     79 #include <machine/cpu.h>
     80 #include <machine/hypervisor.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 static void iommu_enter_sun4u(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags);
    105 static void iommu_enter_sun4v(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags);
    106 static void iommu_remove_sun4u(struct iommu_state *is, vaddr_t va, size_t len);
    107 static void iommu_remove_sun4v(struct iommu_state *is, vaddr_t va, size_t len);
    108 
    109 /*
    110  * initialise the UltraSPARC IOMMU (SBUS or PCI):
    111  *	- allocate and setup the iotsb.
    112  *	- enable the IOMMU
    113  *	- initialise the streaming buffers (if they exist)
    114  *	- create a private DVMA map.
    115  */
    116 void
    117 iommu_init(char *name, struct iommu_state *is, int tsbsize, uint32_t iovabase)
    118 {
    119 	psize_t size;
    120 	vaddr_t va;
    121 	paddr_t pa;
    122 	struct vm_page *pg;
    123 	struct pglist pglist;
    124 
    125 	DPRINTF(IDB_INFO, ("iommu_init: tsbsize %x iovabase %x\n", tsbsize, iovabase));
    126 
    127 	/*
    128 	 * Setup the iommu.
    129 	 *
    130 	 * The sun4u iommu is part of the SBUS or PCI controller so we will
    131 	 * deal with it here..
    132 	 *
    133 	 * For sysio and psycho/psycho+ the IOMMU address space always ends at
    134 	 * 0xffffe000, but the starting address depends on the size of the
    135 	 * map.  The map size is 1024 * 2 ^ is->is_tsbsize entries, where each
    136 	 * entry is 8 bytes.  The start of the map can be calculated by
    137 	 * (0xffffe000 << (8 + is->is_tsbsize)).
    138 	 *
    139 	 * But sabre and hummingbird use a different scheme that seems to
    140 	 * be hard-wired, so we read the start and size from the PROM and
    141 	 * just use those values.
    142 	 */
    143 	if (strncmp(name, "pyro", 4) == 0) {
    144 		is->is_cr = IOMMUREG_READ(is, iommu_cr);
    145 		is->is_cr &= ~IOMMUCR_FIRE_BE;
    146 		is->is_cr |= (IOMMUCR_FIRE_SE | IOMMUCR_FIRE_CM_EN |
    147 		    IOMMUCR_FIRE_TE);
    148 	} else
    149 		is->is_cr = IOMMUCR_EN;
    150 	is->is_tsbsize = tsbsize;
    151 	if (iovabase == -1) {
    152 		is->is_dvmabase = IOTSB_VSTART(is->is_tsbsize);
    153 		is->is_dvmaend = IOTSB_VEND - 1;
    154 	} else {
    155 		is->is_dvmabase = iovabase;
    156 		is->is_dvmaend = iovabase + IOTSB_VSIZE(tsbsize) - 1;
    157 	}
    158 
    159 	/*
    160 	 * Allocate memory for I/O pagetables.  They need to be physically
    161 	 * contiguous.
    162 	 */
    163 
    164 	size = PAGE_SIZE << is->is_tsbsize;
    165 	if (uvm_pglistalloc((psize_t)size, (paddr_t)0, (paddr_t)-1,
    166 		(paddr_t)PAGE_SIZE, (paddr_t)0, &pglist, 1, 0) != 0)
    167 		panic("iommu_init: no memory");
    168 
    169 	va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY);
    170 	if (va == 0)
    171 		panic("iommu_init: no memory");
    172 	is->is_tsb = (int64_t *)va;
    173 
    174 	is->is_ptsb = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pglist));
    175 
    176 	/* Map the pages */
    177 	TAILQ_FOREACH(pg, &pglist, pageq.queue) {
    178 		pa = VM_PAGE_TO_PHYS(pg);
    179 		pmap_kenter_pa(va, pa | PMAP_NVC,
    180 		    VM_PROT_READ | VM_PROT_WRITE, 0);
    181 		va += PAGE_SIZE;
    182 	}
    183 	pmap_update(pmap_kernel());
    184 	memset(is->is_tsb, 0, size);
    185 
    186 #ifdef DEBUG
    187 	if (iommudebug & IDB_INFO)
    188 	{
    189 		/* Probe the iommu */
    190 		if (!CPU_ISSUN4V) {
    191 			printf("iommu cr=%llx tsb=%llx\n",
    192 			    (unsigned long long)bus_space_read_8(is->is_bustag,
    193 				is->is_iommu,
    194 				offsetof(struct iommureg, iommu_cr)),
    195 			    (unsigned long long)bus_space_read_8(is->is_bustag,
    196 				is->is_iommu,
    197 				offsetof(struct iommureg, iommu_tsb)));
    198 			printf("TSB base %p phys %llx\n", (void *)is->is_tsb,
    199 			    (unsigned long long)is->is_ptsb);
    200 			delay(1000000); /* 1 s */
    201 		}
    202 	}
    203 #endif
    204 
    205 	/*
    206 	 * Now all the hardware's working we need to allocate a dvma map.
    207 	 */
    208 	aprint_debug("DVMA map: %x to %x\n",
    209 		(unsigned int)is->is_dvmabase,
    210 		(unsigned int)is->is_dvmaend);
    211 	aprint_debug("IOTSB: %llx to %llx\n",
    212 		(unsigned long long)is->is_ptsb,
    213 		(unsigned long long)(is->is_ptsb + size - 1));
    214 	is->is_dvmamap = vmem_create(name,
    215 				     is->is_dvmabase,
    216 				     (is->is_dvmaend + 1) - is->is_dvmabase,
    217 				     PAGE_SIZE,		/* quantum */
    218 				     NULL,		/* importfn */
    219 				     NULL,		/* releasefn */
    220 				     NULL,		/* source */
    221 				     0,			/* qcache_max */
    222 				     VM_SLEEP,
    223 				     IPL_VM);
    224 	KASSERT(is->is_dvmamap != NULL);
    225 
    226 	/*
    227 	 * Set the TSB size.  The relevant bits were moved to the TSB
    228 	 * base register in the PCIe host bridges.
    229 	 */
    230 	if (is->is_flags & IOMMU_TSBSIZE_IN_PTSB)
    231 		is->is_ptsb |= is->is_tsbsize;
    232 	else
    233 		is->is_cr |= (is->is_tsbsize << 16);
    234 
    235 	/*
    236 	 * now actually start up the IOMMU
    237 	 */
    238 	iommu_reset(is);
    239 }
    240 
    241 /*
    242  * Streaming buffers don't exist on the UltraSPARC IIi; we should have
    243  * detected that already and disabled them.  If not, we will notice that
    244  * they aren't there when the STRBUF_EN bit does not remain.
    245  */
    246 void
    247 iommu_reset(struct iommu_state *is)
    248 {
    249 	int i;
    250 	struct strbuf_ctl *sb;
    251 
    252 	if (CPU_ISSUN4V)
    253 		return;
    254 
    255 	IOMMUREG_WRITE(is, iommu_tsb, is->is_ptsb);
    256 
    257 	/* Enable IOMMU in diagnostic mode */
    258 	IOMMUREG_WRITE(is, iommu_cr, is->is_cr|IOMMUCR_DE);
    259 
    260 	for (i = 0; i < 2; i++) {
    261 		if ((sb = is->is_sb[i])) {
    262 
    263 			/* Enable diagnostics mode? */
    264 			bus_space_write_8(is->is_bustag, is->is_sb[i]->sb_sb,
    265 				STRBUFREG(strbuf_ctl), STRBUF_EN);
    266 
    267 			membar_Lookaside();
    268 
    269 			/* No streaming buffers? Disable them */
    270 			if (bus_space_read_8(is->is_bustag,
    271 				is->is_sb[i]->sb_sb,
    272 				STRBUFREG(strbuf_ctl)) == 0) {
    273 				is->is_sb[i]->sb_flush = NULL;
    274 			} else {
    275 
    276 				/*
    277 				 * locate the pa of the flush buffer.
    278 				 */
    279 				if (pmap_extract(pmap_kernel(),
    280 				     (vaddr_t)is->is_sb[i]->sb_flush,
    281 				     &is->is_sb[i]->sb_flushpa) == FALSE)
    282 					is->is_sb[i]->sb_flush = NULL;
    283 			}
    284 		}
    285 	}
    286 
    287 	if (is->is_flags & IOMMU_FLUSH_CACHE)
    288 		IOMMUREG_WRITE(is, iommu_cache_invalidate, -1ULL);
    289 }
    290 
    291 /*
    292  * Here are the iommu control routines.
    293  */
    294 
    295 static void
    296 iommu_enter(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
    297 {
    298 	DPRINTF(IDB_IOMMU, ("iommu_enter: va %lx pa %lx flags %x\n",
    299 	    va, (long)pa, flags));
    300 	if (!CPU_ISSUN4V)
    301 		iommu_enter_sun4u(sb, va, pa, flags);
    302 	else
    303 		iommu_enter_sun4v(sb, va, pa, flags);
    304 }
    305 
    306 
    307 void
    308 iommu_enter_sun4u(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
    309 {
    310 	struct iommu_state *is = sb->sb_is;
    311 	int strbuf = (flags & BUS_DMA_STREAMING);
    312 	int64_t tte;
    313 
    314 #ifdef DIAGNOSTIC
    315 	if (va < is->is_dvmabase || va > is->is_dvmaend)
    316 		panic("iommu_enter: va %#lx not in DVMA space", va);
    317 #endif
    318 
    319 	/* Is the streamcache flush really needed? */
    320 	if (sb->sb_flush)
    321 		iommu_strbuf_flush(sb, va);
    322 	else
    323 		/* If we can't flush the strbuf don't enable it. */
    324 		strbuf = 0;
    325 
    326 	tte = MAKEIOTTE(pa, !(flags & BUS_DMA_NOWRITE),
    327 		!(flags & BUS_DMA_NOCACHE), (strbuf));
    328 #ifdef DEBUG
    329 	tte |= (flags & 0xff000LL)<<(4*8);
    330 #endif
    331 
    332 	is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = tte;
    333 	bus_space_write_8(is->is_bustag, is->is_iommu,
    334 		IOMMUREG(iommu_flush), va);
    335 	DPRINTF(IDB_IOMMU, ("iommu_enter: slot %d va %lx pa %lx "
    336 		"TSB[%lx]@%p=%lx\n", (int)IOTSBSLOT(va,is->is_tsbsize),
    337 		va, (long)pa, (u_long)IOTSBSLOT(va,is->is_tsbsize),
    338 		(void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
    339 		(u_long)tte));
    340 }
    341 
    342 void
    343 iommu_enter_sun4v(struct strbuf_ctl *sb, vaddr_t va, int64_t pa, int flags)
    344 {
    345 	struct iommu_state *is = sb->sb_is;
    346 	u_int64_t tsbid = IOTSBSLOT(va, is->is_tsbsize);
    347 	paddr_t page_list[1], addr;
    348 	u_int64_t attr, nmapped;
    349 	int err;
    350 
    351 #ifdef DIAGNOSTIC
    352 	if (va < is->is_dvmabase || (va + PAGE_MASK) > is->is_dvmaend)
    353 		panic("viommu_enter: va %#lx not in DVMA space", va);
    354 #endif
    355 
    356 	attr = PCI_MAP_ATTR_READ | PCI_MAP_ATTR_WRITE;
    357 	if (flags & BUS_DMA_READ)
    358 		attr &= ~PCI_MAP_ATTR_READ;
    359 	if (flags & BUS_DMA_WRITE)
    360 		attr &= ~PCI_MAP_ATTR_WRITE;
    361 
    362 	page_list[0] = trunc_page(pa);
    363 	if (!pmap_extract(pmap_kernel(), (vaddr_t)page_list, &addr))
    364 		panic("viommu_enter: pmap_extract failed");
    365 	err = hv_pci_iommu_map(is->is_devhandle, tsbid, 1, attr,
    366 	    addr, &nmapped);
    367 	if (err != H_EOK || nmapped != 1)
    368 		panic("hv_pci_iommu_map: err=%d, nmapped=%lu", err, (long unsigned int)nmapped);
    369 }
    370 
    371 /*
    372  * Find the value of a DVMA address (debug routine).
    373  */
    374 paddr_t
    375 iommu_extract(struct iommu_state *is, vaddr_t dva)
    376 {
    377 	int64_t tte = 0;
    378 
    379 	if (dva >= is->is_dvmabase && dva <= is->is_dvmaend)
    380 		tte = is->is_tsb[IOTSBSLOT(dva, is->is_tsbsize)];
    381 
    382 	if ((tte & IOTTE_V) == 0)
    383 		return ((paddr_t)-1L);
    384 	return (tte & IOTTE_PAMASK);
    385 }
    386 
    387 /*
    388  * iommu_remove: removes mappings created by iommu_enter
    389  *
    390  * Only demap from IOMMU if flag is set.
    391  *
    392  * XXX: this function needs better internal error checking.
    393  */
    394 
    395 static void
    396 iommu_remove(struct iommu_state *is, vaddr_t va, size_t len)
    397 {
    398 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx len %zu\n", va, len));
    399 	if (!CPU_ISSUN4V)
    400 		iommu_remove_sun4u(is, va, len);
    401 	else
    402 		iommu_remove_sun4v(is, va, len);
    403 }
    404 
    405 void
    406 iommu_remove_sun4u(struct iommu_state *is, vaddr_t va, size_t len)
    407 {
    408 
    409 	int slot;
    410 
    411 #ifdef DIAGNOSTIC
    412 	if (va < is->is_dvmabase || va > is->is_dvmaend)
    413 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
    414 	if ((long)(va + len) < (long)va)
    415 		panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
    416 		      (long) va, (long) len);
    417 	if (len & ~0xfffffff)
    418 		panic("iommu_remove: ridiculous len 0x%lx", (u_long)len);
    419 #endif
    420 
    421 	va = trunc_page(va);
    422 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
    423 		va, (u_long)IOTSBSLOT(va, is->is_tsbsize),
    424 		&is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)]));
    425 	while (len > 0) {
    426 		DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d "
    427 			"for va %p size %lx\n",
    428 			(int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va,
    429 			(u_long)len));
    430 		if (len <= PAGE_SIZE)
    431 			len = 0;
    432 		else
    433 			len -= PAGE_SIZE;
    434 
    435 #if 0
    436 		/*
    437 		 * XXX Zero-ing the entry would not require RMW
    438 		 *
    439 		 * Disabling valid bit while a page is used by a device
    440 		 * causes an uncorrectable DMA error.
    441 		 * Workaround to avoid an uncorrectable DMA error is
    442 		 * eliminating the next line, but the page is mapped
    443 		 * until the next iommu_enter call.
    444 		 */
    445 		is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] &= ~IOTTE_V;
    446 		membar_StoreStore();
    447 #endif
    448 		IOMMUREG_WRITE(is, iommu_flush, va);
    449 
    450 		/* Flush cache if necessary. */
    451 		slot = IOTSBSLOT(trunc_page(va), is->is_tsbsize);
    452 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
    453 		    (len == 0 || (slot % 8) == 7))
    454 			IOMMUREG_WRITE(is, iommu_cache_flush,
    455 			    is->is_ptsb + slot * 8);
    456 
    457 		va += PAGE_SIZE;
    458 	}
    459 }
    460 
    461 void
    462 iommu_remove_sun4v(struct iommu_state *is, vaddr_t va, size_t len)
    463 {
    464 	u_int64_t tsbid = IOTSBSLOT(va, is->is_tsbsize);
    465 	u_int64_t ndemapped;
    466 	int err;
    467 
    468 #ifdef DIAGNOSTIC
    469 	if (va < is->is_dvmabase || (va + PAGE_MASK) > is->is_dvmaend)
    470 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
    471 	if (va != trunc_page(va)) {
    472 		printf("iommu_remove: unaligned va: %lx\n", va);
    473 		va = trunc_page(va);
    474 	}
    475 #endif
    476 
    477 	err = hv_pci_iommu_demap(is->is_devhandle, tsbid, 1, &ndemapped);
    478 	if (err != H_EOK || ndemapped != 1)
    479 		panic("hv_pci_iommu_unmap: err=%d", err);
    480 }
    481 
    482 static int
    483 iommu_strbuf_flush_done(struct strbuf_ctl *sb)
    484 {
    485 	struct iommu_state *is = sb->sb_is;
    486 	struct timeval cur, flushtimeout;
    487 
    488 #define BUMPTIME(t, usec) { \
    489 	register volatile struct timeval *tp = (t); \
    490 	register long us; \
    491  \
    492 	tp->tv_usec = us = tp->tv_usec + (usec); \
    493 	if (us >= 1000000) { \
    494 		tp->tv_usec = us - 1000000; \
    495 		tp->tv_sec++; \
    496 	} \
    497 }
    498 
    499 	if (!sb->sb_flush)
    500 		return (0);
    501 
    502 	/*
    503 	 * Streaming buffer flushes:
    504 	 *
    505 	 *   1 Tell strbuf to flush by storing va to strbuf_pgflush.  If
    506 	 *     we're not on a cache line boundary (64-bits):
    507 	 *   2 Store 0 in flag
    508 	 *   3 Store pointer to flag in flushsync
    509 	 *   4 wait till flushsync becomes 0x1
    510 	 *
    511 	 * If it takes more than .5 sec, something
    512 	 * went wrong.
    513 	 */
    514 
    515 	*sb->sb_flush = 0;
    516 	bus_space_write_8(is->is_bustag, sb->sb_sb,
    517 		STRBUFREG(strbuf_flushsync), sb->sb_flushpa);
    518 
    519 	microtime(&flushtimeout);
    520 	cur = flushtimeout;
    521 	BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
    522 
    523 	DPRINTF(IDB_IOMMU, ("%s: flush = %lx at va = %lx pa = %lx now="
    524 		"%"PRIx64":%"PRIx32" until = %"PRIx64":%"PRIx32"\n", __func__,
    525 		(long)*sb->sb_flush, (long)sb->sb_flush, (long)sb->sb_flushpa,
    526 		cur.tv_sec, cur.tv_usec,
    527 		flushtimeout.tv_sec, flushtimeout.tv_usec));
    528 
    529 	/* Bypass non-coherent D$ */
    530 	while ((!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) &&
    531 	       timercmp(&cur, &flushtimeout, <=))
    532 		microtime(&cur);
    533 
    534 #ifdef DIAGNOSTIC
    535 	if (!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) {
    536 		printf("%s: flush timeout %p, at %p\n", __func__,
    537 			(void *)(u_long)*sb->sb_flush,
    538 			(void *)(u_long)sb->sb_flushpa); /* panic? */
    539 #ifdef DDB
    540 		Debugger();
    541 #endif
    542 	}
    543 #endif
    544 	DPRINTF(IDB_IOMMU, ("%s: flushed\n", __func__));
    545 	return (*sb->sb_flush);
    546 }
    547 
    548 /*
    549  * IOMMU DVMA operations, common to SBUS and PCI.
    550  */
    551 int
    552 iommu_dvmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
    553 	bus_size_t buflen, struct proc *p, int flags)
    554 {
    555 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
    556 	struct iommu_state *is = sb->sb_is;
    557 	int err, needsflush;
    558 	bus_size_t sgsize;
    559 	paddr_t curaddr;
    560 	u_long sgstart, sgend, bmask;
    561 	vmem_addr_t dvmaddr;
    562 	bus_size_t align, boundary, len;
    563 	vaddr_t vaddr = (vaddr_t)buf;
    564 	int seg;
    565 	struct pmap *pmap;
    566 	int slot;
    567 
    568 	if (map->dm_nsegs) {
    569 		/* Already in use?? */
    570 #ifdef DIAGNOSTIC
    571 		printf("iommu_dvmamap_load: map still in use\n");
    572 #endif
    573 		bus_dmamap_unload(t, map);
    574 	}
    575 
    576 	/*
    577 	 * Make sure that on error condition we return "no valid mappings".
    578 	 */
    579 	map->dm_nsegs = 0;
    580 	KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
    581 
    582 	if (buflen > map->_dm_size) {
    583 		DPRINTF(IDB_BUSDMA,
    584 		    ("iommu_dvmamap_load(): error %d > %d -- "
    585 		     "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
    586 		return (EINVAL);
    587 	}
    588 
    589 	sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
    590 
    591 	/*
    592 	 * A boundary presented to bus_dmamem_alloc() takes precedence
    593 	 * over boundary in the map.
    594 	 */
    595 	if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
    596 		boundary = map->_dm_boundary;
    597 	align = uimax(map->dm_segs[0]._ds_align, PAGE_SIZE);
    598 
    599 	/*
    600 	 * If our segment size is larger than the boundary we need to
    601 	 * split the transfer up int little pieces ourselves.
    602 	 */
    603 	KASSERT(is->is_dvmamap != NULL);
    604 	err = vmem_xalloc(is->is_dvmamap, sgsize,
    605 			  align,		/* alignment */
    606 			  0,			/* phase */
    607 			  (sgsize > boundary) ? 0 : boundary,
    608 			  VMEM_ADDR_MIN,	/* minaddr */
    609 			  VMEM_ADDR_MAX,	/* maxaddr */
    610 			  VM_NOSLEEP | VM_BESTFIT,
    611 			  &dvmaddr);
    612 
    613 #ifdef DEBUG
    614 	if (err || (dvmaddr == (u_long)-1)) {
    615 		printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
    616 		    (int)sgsize, flags);
    617 #ifdef DDB
    618 		Debugger();
    619 #endif
    620 	}
    621 #endif
    622 	if (err != 0)
    623 		return (err);
    624 
    625 	if (dvmaddr == (u_long)-1)
    626 		return (ENOMEM);
    627 
    628 	/* Set the active DVMA map */
    629 	map->_dm_dvmastart = dvmaddr;
    630 	map->_dm_dvmasize = sgsize;
    631 
    632 	/*
    633 	 * Now split the DVMA range into segments, not crossing
    634 	 * the boundary.
    635 	 */
    636 	seg = 0;
    637 	sgstart = dvmaddr + (vaddr & PGOFSET);
    638 	sgend = sgstart + buflen - 1;
    639 	map->dm_segs[seg].ds_addr = sgstart;
    640 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: boundary %lx boundary - 1 %lx "
    641 	    "~(boundary - 1) %lx\n", (long)boundary, (long)(boundary - 1),
    642 	    (long)~(boundary - 1)));
    643 	bmask = ~(boundary - 1);
    644 	while ((sgstart & bmask) != (sgend & bmask) ||
    645 	       sgend - sgstart + 1 > map->dm_maxsegsz) {
    646 		/* Oops. We crossed a boundary or large seg. Split the xfer. */
    647 		len = map->dm_maxsegsz;
    648 		if ((sgstart & bmask) != (sgend & bmask))
    649 			len = uimin(len, boundary - (sgstart & (boundary - 1)));
    650 		map->dm_segs[seg].ds_len = len;
    651 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
    652 		    "seg %d start %lx size %lx\n", seg,
    653 		    (long)map->dm_segs[seg].ds_addr,
    654 		    (long)map->dm_segs[seg].ds_len));
    655 		if (++seg >= map->_dm_segcnt) {
    656 			/* Too many segments.  Fail the operation. */
    657 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
    658 			    "too many segments %d\n", seg));
    659 			vmem_xfree(is->is_dvmamap, dvmaddr, sgsize);
    660 			map->_dm_dvmastart = 0;
    661 			map->_dm_dvmasize = 0;
    662 			return (EFBIG);
    663 		}
    664 		sgstart += len;
    665 		map->dm_segs[seg].ds_addr = sgstart;
    666 	}
    667 	map->dm_segs[seg].ds_len = sgend - sgstart + 1;
    668 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
    669 	    "seg %d start %lx size %lx\n", seg,
    670 	    (long)map->dm_segs[seg].ds_addr, (long)map->dm_segs[seg].ds_len));
    671 	map->dm_nsegs = seg + 1;
    672 	map->dm_mapsize = buflen;
    673 
    674 	if (p != NULL)
    675 		pmap = p->p_vmspace->vm_map.pmap;
    676 	else
    677 		pmap = pmap_kernel();
    678 
    679 	needsflush = 0;
    680 	for (; buflen > 0; ) {
    681 
    682 		/*
    683 		 * Get the physical address for this page.
    684 		 */
    685 		if (pmap_extract(pmap, (vaddr_t)vaddr, &curaddr) == FALSE) {
    686 #ifdef DIAGNOSTIC
    687 			printf("iommu_dvmamap_load: pmap_extract failed %lx\n", vaddr);
    688 #endif
    689 			bus_dmamap_unload(t, map);
    690 			return (-1);
    691 		}
    692 
    693 		/*
    694 		 * Compute the segment size, and adjust counts.
    695 		 */
    696 		sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
    697 		if (buflen < sgsize)
    698 			sgsize = buflen;
    699 
    700 		DPRINTF(IDB_BUSDMA,
    701 		    ("iommu_dvmamap_load: map %p loading va %p "
    702 		    "dva %lx at pa %lx\n",
    703 		    map, (void *)vaddr, (long)dvmaddr,
    704 		    (long)trunc_page(curaddr)));
    705 		iommu_enter(sb, trunc_page(dvmaddr), trunc_page(curaddr),
    706 		    flags | IOTTE_DEBUG(0x4000));
    707 		needsflush = 1;
    708 
    709 		vaddr += sgsize;
    710 		buflen -= sgsize;
    711 
    712 		/* Flush cache if necessary. */
    713 		slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
    714 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
    715 		    (buflen <= 0 || (slot % 8) == 7))
    716 			IOMMUREG_WRITE(is, iommu_cache_flush,
    717 			    is->is_ptsb + slot * 8);
    718 
    719 		dvmaddr += PAGE_SIZE;
    720 	}
    721 	if (needsflush)
    722 		iommu_strbuf_flush_done(sb);
    723 #ifdef DIAGNOSTIC
    724 	for (seg = 0; seg < map->dm_nsegs; seg++) {
    725 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
    726 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
    727 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
    728 			    seg, (long)map->dm_segs[seg].ds_addr,
    729 			    is->is_dvmabase, is->is_dvmaend);
    730 #ifdef DDB
    731 			Debugger();
    732 #endif
    733 		}
    734 	}
    735 #endif
    736 	return (0);
    737 }
    738 
    739 
    740 void
    741 iommu_dvmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
    742 {
    743 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
    744 	struct iommu_state *is = sb->sb_is;
    745 
    746 	/* Flush the iommu */
    747 	if (!map->_dm_dvmastart)
    748 		panic("%s: error dvmastart is zero!\n", __func__);
    749 
    750 	if (is->is_flags & IOMMU_SYNC_BEFORE_UNMAP) {
    751 
    752 		/* Flush the caches */
    753 		bus_dmamap_unload(t->_parent, map);
    754 
    755 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
    756 
    757 	} else {
    758 
    759 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
    760 
    761 		/* Flush the caches */
    762 		bus_dmamap_unload(t->_parent, map);
    763 	}
    764 
    765 	vmem_xfree(is->is_dvmamap, map->_dm_dvmastart, map->_dm_dvmasize);
    766 	map->_dm_dvmastart = 0;
    767 	map->_dm_dvmasize = 0;
    768 
    769 	/* Clear the map */
    770 }
    771 
    772 
    773 int
    774 iommu_dvmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
    775 	bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
    776 {
    777 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
    778 	struct iommu_state *is = sb->sb_is;
    779 	struct vm_page *pg;
    780 	int i, j;
    781 	int left;
    782 	int err, needsflush;
    783 	bus_size_t sgsize;
    784 	paddr_t pa;
    785 	bus_size_t boundary, align;
    786 	u_long dvmaddr, sgstart, sgend, bmask;
    787 	struct pglist *pglist;
    788 	const int pagesz = PAGE_SIZE;
    789 	int slot;
    790 #ifdef DEBUG
    791 	int npg = 0;
    792 #endif
    793 
    794 	if (map->dm_nsegs) {
    795 		/* Already in use?? */
    796 #ifdef DIAGNOSTIC
    797 		printf("iommu_dvmamap_load_raw: map still in use\n");
    798 #endif
    799 		bus_dmamap_unload(t, map);
    800 	}
    801 
    802 	/*
    803 	 * A boundary presented to bus_dmamem_alloc() takes precedence
    804 	 * over boundary in the map.
    805 	 */
    806 	if ((boundary = segs[0]._ds_boundary) == 0)
    807 		boundary = map->_dm_boundary;
    808 
    809 	align = uimax(segs[0]._ds_align, pagesz);
    810 
    811 	/*
    812 	 * Make sure that on error condition we return "no valid mappings".
    813 	 */
    814 	map->dm_nsegs = 0;
    815 	/* Count up the total number of pages we need */
    816 	pa = trunc_page(segs[0].ds_addr);
    817 	sgsize = 0;
    818 	left = size;
    819 	for (i = 0; left > 0 && i < nsegs; i++) {
    820 		if (round_page(pa) != round_page(segs[i].ds_addr))
    821 			sgsize = round_page(sgsize) +
    822 			    (segs[i].ds_addr & PGOFSET);
    823 		sgsize += uimin(left, segs[i].ds_len);
    824 		left -= segs[i].ds_len;
    825 		pa = segs[i].ds_addr + segs[i].ds_len;
    826 	}
    827 	sgsize = round_page(sgsize);
    828 
    829 	/*
    830 	 * If our segment size is larger than the boundary we need to
    831 	 * split the transfer up into little pieces ourselves.
    832 	 */
    833 	const vm_flag_t vmflags = VM_BESTFIT |
    834 	    ((flags & BUS_DMA_NOWAIT) ? VM_NOSLEEP : VM_SLEEP);
    835 
    836 	err = vmem_xalloc(is->is_dvmamap, sgsize,
    837 			  align,		/* alignment */
    838 			  0,			/* phase */
    839 			  (sgsize > boundary) ? 0 : boundary,
    840 			  VMEM_ADDR_MIN,	/* minaddr */
    841 			  VMEM_ADDR_MAX,	/* maxaddr */
    842 			  vmflags,
    843 			  &dvmaddr);
    844 	if (err != 0)
    845 		return (err);
    846 
    847 #ifdef DEBUG
    848 	if (dvmaddr == (u_long)-1)
    849 	{
    850 		printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
    851 		    (int)sgsize, flags);
    852 #ifdef DDB
    853 		Debugger();
    854 #endif
    855 	}
    856 #endif
    857 	if (dvmaddr == (u_long)-1)
    858 		return (ENOMEM);
    859 
    860 	/* Set the active DVMA map */
    861 	map->_dm_dvmastart = dvmaddr;
    862 	map->_dm_dvmasize = sgsize;
    863 
    864 	bmask = ~(boundary - 1);
    865 	if ((pglist = segs[0]._ds_mlist) == NULL) {
    866 		u_long prev_va = 0UL, last_va = dvmaddr;
    867 		paddr_t prev_pa = 0;
    868 		int end = 0, offset;
    869 		bus_size_t len = size;
    870 
    871 		/*
    872 		 * This segs is made up of individual physical
    873 		 *  segments, probably by _bus_dmamap_load_uio() or
    874 		 * _bus_dmamap_load_mbuf().  Ignore the mlist and
    875 		 * load each one individually.
    876 		 */
    877 		j = 0;
    878 		needsflush = 0;
    879 		for (i = 0; i < nsegs ; i++) {
    880 
    881 			pa = segs[i].ds_addr;
    882 			offset = (pa & PGOFSET);
    883 			pa = trunc_page(pa);
    884 			dvmaddr = trunc_page(dvmaddr);
    885 			left = uimin(len, segs[i].ds_len);
    886 
    887 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: converting "
    888 				"physseg %d start %lx size %lx\n", i,
    889 				(long)segs[i].ds_addr, (long)segs[i].ds_len));
    890 
    891 			if ((pa == prev_pa) &&
    892 				((offset != 0) || (end != offset))) {
    893 				/* We can re-use this mapping */
    894 				dvmaddr = prev_va;
    895 			}
    896 
    897 			sgstart = dvmaddr + offset;
    898 			sgend = sgstart + left - 1;
    899 
    900 			/* Are the segments virtually adjacent? */
    901 			if ((j > 0) && (end == offset) &&
    902 			    ((offset == 0) || (pa == prev_pa)) &&
    903 			    (map->dm_segs[j-1].ds_len + left <=
    904 			     map->dm_maxsegsz)) {
    905 				/* Just append to the previous segment. */
    906 				map->dm_segs[--j].ds_len += left;
    907 				/* Restore sgstart for boundary check */
    908 				sgstart = map->dm_segs[j].ds_addr;
    909 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    910 					"appending seg %d start %lx size %lx\n", j,
    911 					(long)map->dm_segs[j].ds_addr,
    912 					(long)map->dm_segs[j].ds_len));
    913 			} else {
    914 				if (j >= map->_dm_segcnt) {
    915 					iommu_remove(is, map->_dm_dvmastart,
    916 					    last_va - map->_dm_dvmastart);
    917 					goto fail;
    918 				}
    919 				map->dm_segs[j].ds_addr = sgstart;
    920 				map->dm_segs[j].ds_len = left;
    921 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    922 					"seg %d start %lx size %lx\n", j,
    923 					(long)map->dm_segs[j].ds_addr,
    924 					(long)map->dm_segs[j].ds_len));
    925 			}
    926 			end = (offset + left) & PGOFSET;
    927 
    928 			/* Check for boundary issues */
    929 			while ((sgstart & bmask) != (sgend & bmask)) {
    930 				/* Need a new segment. */
    931 				map->dm_segs[j].ds_len =
    932 					boundary - (sgstart & (boundary - 1));
    933 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    934 					"seg %d start %lx size %lx\n", j,
    935 					(long)map->dm_segs[j].ds_addr,
    936 					(long)map->dm_segs[j].ds_len));
    937 				if (++j >= map->_dm_segcnt) {
    938 					iommu_remove(is, map->_dm_dvmastart,
    939 					    last_va - map->_dm_dvmastart);
    940 					goto fail;
    941 				}
    942 				sgstart += map->dm_segs[j-1].ds_len;
    943 				map->dm_segs[j].ds_addr = sgstart;
    944 				map->dm_segs[j].ds_len = sgend - sgstart + 1;
    945 			}
    946 
    947 			if (sgsize == 0)
    948 				panic("iommu_dmamap_load_raw: size botch");
    949 
    950 			/* Now map a series of pages. */
    951 			while (dvmaddr <= sgend) {
    952 				DPRINTF(IDB_BUSDMA,
    953 					("iommu_dvmamap_load_raw: map %p "
    954 						"loading va %lx at pa %lx\n",
    955 						map, (long)dvmaddr,
    956 						(long)(pa)));
    957 				/* Enter it if we haven't before. */
    958 				if (prev_va != dvmaddr) {
    959 					iommu_enter(sb, prev_va = dvmaddr,
    960 					    prev_pa = pa,
    961 					    flags | IOTTE_DEBUG(++npg << 12));
    962 					needsflush = 1;
    963 
    964 					/* Flush cache if necessary. */
    965 					slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
    966 					if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
    967 					    ((dvmaddr + pagesz) > sgend || (slot % 8) == 7))
    968 						IOMMUREG_WRITE(is, iommu_cache_flush,
    969 						    is->is_ptsb + slot * 8);
    970 				}
    971 
    972 				dvmaddr += pagesz;
    973 				pa += pagesz;
    974 				last_va = dvmaddr;
    975 			}
    976 
    977 			len -= left;
    978 			++j;
    979 		}
    980 		if (needsflush)
    981 			iommu_strbuf_flush_done(sb);
    982 
    983 		map->dm_mapsize = size;
    984 		map->dm_nsegs = j;
    985 #ifdef DIAGNOSTIC
    986 		{ int seg;
    987 	for (seg = 0; seg < map->dm_nsegs; seg++) {
    988 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
    989 		    map->dm_segs[seg].ds_addr > is->is_dvmaend) {
    990 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
    991 				seg, (long)map->dm_segs[seg].ds_addr,
    992 				is->is_dvmabase, is->is_dvmaend);
    993 #ifdef DDB
    994 			Debugger();
    995 #endif
    996 		}
    997 	}
    998 		}
    999 #endif
   1000 		return (0);
   1001 	}
   1002 
   1003 	/*
   1004 	 * This was allocated with bus_dmamem_alloc.
   1005 	 * The pages are on a `pglist'.
   1006 	 */
   1007 	i = 0;
   1008 	sgstart = dvmaddr;
   1009 	sgend = sgstart + size - 1;
   1010 	map->dm_segs[i].ds_addr = sgstart;
   1011 	while ((sgstart & bmask) != (sgend & bmask)) {
   1012 		/* Oops.  We crossed a boundary.  Split the xfer. */
   1013 		map->dm_segs[i].ds_len = boundary - (sgstart & (boundary - 1));
   1014 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
   1015 			"seg %d start %lx size %lx\n", i,
   1016 			(long)map->dm_segs[i].ds_addr,
   1017 			(long)map->dm_segs[i].ds_len));
   1018 		if (++i >= map->_dm_segcnt) {
   1019 			/* Too many segments.  Fail the operation. */
   1020 			goto fail;
   1021 		}
   1022 		sgstart += map->dm_segs[i-1].ds_len;
   1023 		map->dm_segs[i].ds_addr = sgstart;
   1024 	}
   1025 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
   1026 			"seg %d start %lx size %lx\n", i,
   1027 			(long)map->dm_segs[i].ds_addr, (long)map->dm_segs[i].ds_len));
   1028 	map->dm_segs[i].ds_len = sgend - sgstart + 1;
   1029 
   1030 	needsflush = 0;
   1031 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
   1032 		if (sgsize == 0)
   1033 			panic("iommu_dmamap_load_raw: size botch");
   1034 		pa = VM_PAGE_TO_PHYS(pg);
   1035 
   1036 		DPRINTF(IDB_BUSDMA,
   1037 		    ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
   1038 		    map, (long)dvmaddr, (long)(pa)));
   1039 		iommu_enter(sb, dvmaddr, pa, flags | IOTTE_DEBUG(0x8000));
   1040 		needsflush = 1;
   1041 
   1042 		sgsize -= pagesz;
   1043 
   1044 		/* Flush cache if necessary. */
   1045 		slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
   1046 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
   1047 		    (sgsize == 0 || (slot % 8) == 7))
   1048 			IOMMUREG_WRITE(is, iommu_cache_flush,
   1049 			    is->is_ptsb + slot * 8);
   1050 
   1051 		dvmaddr += pagesz;
   1052 	}
   1053 	if (needsflush)
   1054 		iommu_strbuf_flush_done(sb);
   1055 	map->dm_mapsize = size;
   1056 	map->dm_nsegs = i+1;
   1057 #ifdef DIAGNOSTIC
   1058 	{ int seg;
   1059 	for (seg = 0; seg < map->dm_nsegs; seg++) {
   1060 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
   1061 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
   1062 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
   1063 				seg, (long)map->dm_segs[seg].ds_addr,
   1064 				is->is_dvmabase, is->is_dvmaend);
   1065 #ifdef DDB
   1066 			Debugger();
   1067 #endif
   1068 		}
   1069 	}
   1070 	}
   1071 #endif
   1072 	return (0);
   1073 
   1074 fail:
   1075 	vmem_xfree(is->is_dvmamap, map->_dm_dvmastart, sgsize);
   1076 	map->_dm_dvmastart = 0;
   1077 	map->_dm_dvmasize = 0;
   1078 	return (EFBIG);
   1079 }
   1080 
   1081 
   1082 /*
   1083  * Flush an individual dma segment, returns non-zero if the streaming buffers
   1084  * need flushing afterwards.
   1085  */
   1086 static int
   1087 iommu_dvmamap_sync_range(struct strbuf_ctl *sb, vaddr_t va, bus_size_t len)
   1088 {
   1089 	vaddr_t vaend;
   1090 	struct iommu_state *is = sb->sb_is;
   1091 
   1092 #ifdef DIAGNOSTIC
   1093 	if (va < is->is_dvmabase || va > is->is_dvmaend)
   1094 		panic("invalid va: %llx", (long long)va);
   1095 #endif
   1096 
   1097 	if ((is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)] & IOTTE_STREAM) == 0) {
   1098 		DPRINTF(IDB_SYNC,
   1099 			("iommu_dvmamap_sync_range: attempting to flush "
   1100 			 "non-streaming entry\n"));
   1101 		return (0);
   1102 	}
   1103 
   1104 	vaend = round_page(va + len) - 1;
   1105 	va = trunc_page(va);
   1106 
   1107 #ifdef DIAGNOSTIC
   1108 	if (va < is->is_dvmabase || vaend > is->is_dvmaend)
   1109 		panic("invalid va range: %llx to %llx (%x to %x)",
   1110 		    (long long)va, (long long)vaend,
   1111 		    is->is_dvmabase,
   1112 		    is->is_dvmaend);
   1113 #endif
   1114 
   1115 	for ( ; va <= vaend; va += PAGE_SIZE) {
   1116 		DPRINTF(IDB_SYNC,
   1117 		    ("iommu_dvmamap_sync_range: flushing va %p\n",
   1118 		    (void *)(u_long)va));
   1119 		iommu_strbuf_flush(sb, va);
   1120 	}
   1121 
   1122 	return (1);
   1123 }
   1124 
   1125 static void
   1126 _iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
   1127 	bus_size_t len, int ops)
   1128 {
   1129 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
   1130 	bus_size_t count;
   1131 	int i, needsflush = 0;
   1132 
   1133 	if (!sb->sb_flush)
   1134 		return;
   1135 
   1136 	for (i = 0; i < map->dm_nsegs; i++) {
   1137 		if (offset < map->dm_segs[i].ds_len)
   1138 			break;
   1139 		offset -= map->dm_segs[i].ds_len;
   1140 	}
   1141 
   1142 	if (i == map->dm_nsegs)
   1143 		panic("%s: segment too short %llu", __func__,
   1144 		    (unsigned long long)offset);
   1145 
   1146 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_POSTWRITE)) {
   1147 		/* Nothing to do */;
   1148 	}
   1149 
   1150 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_PREWRITE)) {
   1151 
   1152 		for (; len > 0 && i < map->dm_nsegs; i++) {
   1153 			count = MIN(map->dm_segs[i].ds_len - offset, len);
   1154 			if (count > 0 &&
   1155 			    iommu_dvmamap_sync_range(sb,
   1156 				map->dm_segs[i].ds_addr + offset, count))
   1157 				needsflush = 1;
   1158 			offset = 0;
   1159 			len -= count;
   1160 		}
   1161 #ifdef DIAGNOSTIC
   1162 		if (i == map->dm_nsegs && len > 0)
   1163 			panic("%s: leftover %llu", __func__,
   1164 			    (unsigned long long)len);
   1165 #endif
   1166 
   1167 		if (needsflush)
   1168 			iommu_strbuf_flush_done(sb);
   1169 	}
   1170 }
   1171 
   1172 void
   1173 iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
   1174 	bus_size_t len, int ops)
   1175 {
   1176 
   1177 	/* If len is 0, then there is nothing to do */
   1178 	if (len == 0)
   1179 		return;
   1180 
   1181 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) {
   1182 		/* Flush the CPU then the IOMMU */
   1183 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
   1184 		_iommu_dvmamap_sync(t, map, offset, len, ops);
   1185 	}
   1186 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE)) {
   1187 		/* Flush the IOMMU then the CPU */
   1188 		_iommu_dvmamap_sync(t, map, offset, len, ops);
   1189 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
   1190 	}
   1191 }
   1192 
   1193 int
   1194 iommu_dvmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
   1195 	bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
   1196 	int flags)
   1197 {
   1198 
   1199 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
   1200 	   "segp %p flags %d\n", (unsigned long long)size,
   1201 	   (unsigned long long)alignment, (unsigned long long)boundary,
   1202 	   segs, flags));
   1203 	return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
   1204 	    segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
   1205 }
   1206 
   1207 void
   1208 iommu_dvmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
   1209 {
   1210 
   1211 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
   1212 	    segs, nsegs));
   1213 	bus_dmamem_free(t->_parent, segs, nsegs);
   1214 }
   1215 
   1216 /*
   1217  * Map the DVMA mappings into the kernel pmap.
   1218  * Check the flags to see whether we're streaming or coherent.
   1219  */
   1220 int
   1221 iommu_dvmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
   1222 	size_t size, void **kvap, int flags)
   1223 {
   1224 	struct vm_page *pg;
   1225 	vaddr_t va;
   1226 	bus_addr_t addr;
   1227 	struct pglist *pglist;
   1228 	int cbit;
   1229 	const uvm_flag_t kmflags =
   1230 	    (flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0;
   1231 
   1232 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
   1233 	    segs, nsegs, size));
   1234 
   1235 	/*
   1236 	 * Allocate some space in the kernel map, and then map these pages
   1237 	 * into this space.
   1238 	 */
   1239 	size = round_page(size);
   1240 	va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY | kmflags);
   1241 	if (va == 0)
   1242 		return (ENOMEM);
   1243 
   1244 	*kvap = (void *)va;
   1245 
   1246 	/*
   1247 	 * digest flags:
   1248 	 */
   1249 	cbit = 0;
   1250 	if (flags & BUS_DMA_COHERENT)	/* Disable vcache */
   1251 		cbit |= PMAP_NVC;
   1252 	if (flags & BUS_DMA_NOCACHE)	/* side effects */
   1253 		cbit |= PMAP_NC;
   1254 
   1255 	/*
   1256 	 * Now take this and map it into the CPU.
   1257 	 */
   1258 	pglist = segs[0]._ds_mlist;
   1259 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
   1260 #ifdef DIAGNOSTIC
   1261 		if (size == 0)
   1262 			panic("iommu_dvmamem_map: size botch");
   1263 #endif
   1264 		addr = VM_PAGE_TO_PHYS(pg);
   1265 		DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
   1266 		    "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
   1267 		pmap_kenter_pa(va, addr | cbit,
   1268 		    VM_PROT_READ | VM_PROT_WRITE, 0);
   1269 		va += PAGE_SIZE;
   1270 		size -= PAGE_SIZE;
   1271 	}
   1272 	pmap_update(pmap_kernel());
   1273 	return (0);
   1274 }
   1275 
   1276 /*
   1277  * Unmap DVMA mappings from kernel
   1278  */
   1279 void
   1280 iommu_dvmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
   1281 {
   1282 
   1283 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
   1284 	    kva, size));
   1285 
   1286 #ifdef DIAGNOSTIC
   1287 	if ((u_long)kva & PGOFSET)
   1288 		panic("iommu_dvmamem_unmap");
   1289 #endif
   1290 
   1291 	size = round_page(size);
   1292 	pmap_kremove((vaddr_t)kva, size);
   1293 	pmap_update(pmap_kernel());
   1294 	uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
   1295 }
   1296