Home | History | Annotate | Line # | Download | only in dev
iommu.c revision 1.113.18.1
      1 /*	$NetBSD: iommu.c,v 1.113.18.1 2019/06/10 22:06:47 christos 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 1.113.18.1 2019/06/10 22:06:47 christos 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 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 
    393 void
    394 iommu_remove(struct iommu_state *is, vaddr_t va, size_t len)
    395 {
    396 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx len %zu\n", va, len));
    397 	if (!CPU_ISSUN4V)
    398 		iommu_remove_sun4u(is, va, len);
    399 	else
    400 		iommu_remove_sun4v(is, va, len);
    401 }
    402 
    403 void
    404 iommu_remove_sun4u(struct iommu_state *is, vaddr_t va, size_t len)
    405 {
    406 
    407 	int slot;
    408 
    409 #ifdef DIAGNOSTIC
    410 	if (va < is->is_dvmabase || va > is->is_dvmaend)
    411 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
    412 	if ((long)(va + len) < (long)va)
    413 		panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
    414 		      (long) va, (long) len);
    415 	if (len & ~0xfffffff)
    416 		panic("iommu_remove: ridiculous len 0x%lx", (u_long)len);
    417 #endif
    418 
    419 	va = trunc_page(va);
    420 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
    421 		va, (u_long)IOTSBSLOT(va, is->is_tsbsize),
    422 		&is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)]));
    423 	while (len > 0) {
    424 		DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d "
    425 			"for va %p size %lx\n",
    426 			(int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va,
    427 			(u_long)len));
    428 		if (len <= PAGE_SIZE)
    429 			len = 0;
    430 		else
    431 			len -= PAGE_SIZE;
    432 
    433 #if 0
    434 		/*
    435 		 * XXX Zero-ing the entry would not require RMW
    436 		 *
    437 		 * Disabling valid bit while a page is used by a device
    438 		 * causes an uncorrectable DMA error.
    439 		 * Workaround to avoid an uncorrectable DMA error is
    440 		 * eliminating the next line, but the page is mapped
    441 		 * until the next iommu_enter call.
    442 		 */
    443 		is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] &= ~IOTTE_V;
    444 		membar_StoreStore();
    445 #endif
    446 		IOMMUREG_WRITE(is, iommu_flush, va);
    447 
    448 		/* Flush cache if necessary. */
    449 		slot = IOTSBSLOT(trunc_page(va), is->is_tsbsize);
    450 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
    451 		    (len == 0 || (slot % 8) == 7))
    452 			IOMMUREG_WRITE(is, iommu_cache_flush,
    453 			    is->is_ptsb + slot * 8);
    454 
    455 		va += PAGE_SIZE;
    456 	}
    457 }
    458 
    459 void
    460 iommu_remove_sun4v(struct iommu_state *is, vaddr_t va, size_t len)
    461 {
    462 	u_int64_t tsbid = IOTSBSLOT(va, is->is_tsbsize);
    463 	u_int64_t ndemapped;
    464 	int err;
    465 
    466 #ifdef DIAGNOSTIC
    467 	if (va < is->is_dvmabase || (va + PAGE_MASK) > is->is_dvmaend)
    468 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
    469 	if (va != trunc_page(va)) {
    470 		printf("iommu_remove: unaligned va: %lx\n", va);
    471 		va = trunc_page(va);
    472 	}
    473 #endif
    474 
    475 	err = hv_pci_iommu_demap(is->is_devhandle, tsbid, 1, &ndemapped);
    476 	if (err != H_EOK || ndemapped != 1)
    477 		panic("hv_pci_iommu_unmap: err=%d", err);
    478 }
    479 
    480 static int
    481 iommu_strbuf_flush_done(struct strbuf_ctl *sb)
    482 {
    483 	struct iommu_state *is = sb->sb_is;
    484 	struct timeval cur, flushtimeout;
    485 
    486 #define BUMPTIME(t, usec) { \
    487 	register volatile struct timeval *tp = (t); \
    488 	register long us; \
    489  \
    490 	tp->tv_usec = us = tp->tv_usec + (usec); \
    491 	if (us >= 1000000) { \
    492 		tp->tv_usec = us - 1000000; \
    493 		tp->tv_sec++; \
    494 	} \
    495 }
    496 
    497 	if (!sb->sb_flush)
    498 		return (0);
    499 
    500 	/*
    501 	 * Streaming buffer flushes:
    502 	 *
    503 	 *   1 Tell strbuf to flush by storing va to strbuf_pgflush.  If
    504 	 *     we're not on a cache line boundary (64-bits):
    505 	 *   2 Store 0 in flag
    506 	 *   3 Store pointer to flag in flushsync
    507 	 *   4 wait till flushsync becomes 0x1
    508 	 *
    509 	 * If it takes more than .5 sec, something
    510 	 * went wrong.
    511 	 */
    512 
    513 	*sb->sb_flush = 0;
    514 	bus_space_write_8(is->is_bustag, sb->sb_sb,
    515 		STRBUFREG(strbuf_flushsync), sb->sb_flushpa);
    516 
    517 	microtime(&flushtimeout);
    518 	cur = flushtimeout;
    519 	BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
    520 
    521 	DPRINTF(IDB_IOMMU, ("%s: flush = %lx at va = %lx pa = %lx now="
    522 		"%"PRIx64":%"PRIx32" until = %"PRIx64":%"PRIx32"\n", __func__,
    523 		(long)*sb->sb_flush, (long)sb->sb_flush, (long)sb->sb_flushpa,
    524 		cur.tv_sec, cur.tv_usec,
    525 		flushtimeout.tv_sec, flushtimeout.tv_usec));
    526 
    527 	/* Bypass non-coherent D$ */
    528 	while ((!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) &&
    529 	       timercmp(&cur, &flushtimeout, <=))
    530 		microtime(&cur);
    531 
    532 #ifdef DIAGNOSTIC
    533 	if (!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) {
    534 		printf("%s: flush timeout %p, at %p\n", __func__,
    535 			(void *)(u_long)*sb->sb_flush,
    536 			(void *)(u_long)sb->sb_flushpa); /* panic? */
    537 #ifdef DDB
    538 		Debugger();
    539 #endif
    540 	}
    541 #endif
    542 	DPRINTF(IDB_IOMMU, ("%s: flushed\n", __func__));
    543 	return (*sb->sb_flush);
    544 }
    545 
    546 /*
    547  * IOMMU DVMA operations, common to SBUS and PCI.
    548  */
    549 int
    550 iommu_dvmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
    551 	bus_size_t buflen, struct proc *p, int flags)
    552 {
    553 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
    554 	struct iommu_state *is = sb->sb_is;
    555 	int err, needsflush;
    556 	bus_size_t sgsize;
    557 	paddr_t curaddr;
    558 	u_long dvmaddr, sgstart, sgend, bmask;
    559 	bus_size_t align, boundary, len;
    560 	vaddr_t vaddr = (vaddr_t)buf;
    561 	int seg;
    562 	struct pmap *pmap;
    563 	int slot;
    564 
    565 	if (map->dm_nsegs) {
    566 		/* Already in use?? */
    567 #ifdef DIAGNOSTIC
    568 		printf("iommu_dvmamap_load: map still in use\n");
    569 #endif
    570 		bus_dmamap_unload(t, map);
    571 	}
    572 
    573 	/*
    574 	 * Make sure that on error condition we return "no valid mappings".
    575 	 */
    576 	map->dm_nsegs = 0;
    577 	KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
    578 
    579 	if (buflen > map->_dm_size) {
    580 		DPRINTF(IDB_BUSDMA,
    581 		    ("iommu_dvmamap_load(): error %d > %d -- "
    582 		     "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
    583 		return (EINVAL);
    584 	}
    585 
    586 	sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
    587 
    588 	/*
    589 	 * A boundary presented to bus_dmamem_alloc() takes precedence
    590 	 * over boundary in the map.
    591 	 */
    592 	if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
    593 		boundary = map->_dm_boundary;
    594 	align = uimax(map->dm_segs[0]._ds_align, PAGE_SIZE);
    595 
    596 	/*
    597 	 * If our segment size is larger than the boundary we need to
    598 	 * split the transfer up int little pieces ourselves.
    599 	 */
    600 	KASSERT(is->is_dvmamap);
    601 	mutex_enter(&is->is_lock);
    602 	err = extent_alloc(is->is_dvmamap, sgsize, align,
    603 	    (sgsize > boundary) ? 0 : boundary,
    604 	    EX_NOWAIT|EX_BOUNDZERO, &dvmaddr);
    605 	mutex_exit(&is->is_lock);
    606 
    607 #ifdef DEBUG
    608 	if (err || (dvmaddr == (u_long)-1)) {
    609 		printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
    610 		    (int)sgsize, flags);
    611 #ifdef DDB
    612 		Debugger();
    613 #endif
    614 	}
    615 #endif
    616 	if (err != 0)
    617 		return (err);
    618 
    619 	if (dvmaddr == (u_long)-1)
    620 		return (ENOMEM);
    621 
    622 	/* Set the active DVMA map */
    623 	map->_dm_dvmastart = dvmaddr;
    624 	map->_dm_dvmasize = sgsize;
    625 
    626 	/*
    627 	 * Now split the DVMA range into segments, not crossing
    628 	 * the boundary.
    629 	 */
    630 	seg = 0;
    631 	sgstart = dvmaddr + (vaddr & PGOFSET);
    632 	sgend = sgstart + buflen - 1;
    633 	map->dm_segs[seg].ds_addr = sgstart;
    634 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: boundary %lx boundary - 1 %lx "
    635 	    "~(boundary - 1) %lx\n", (long)boundary, (long)(boundary - 1),
    636 	    (long)~(boundary - 1)));
    637 	bmask = ~(boundary - 1);
    638 	while ((sgstart & bmask) != (sgend & bmask) ||
    639 	       sgend - sgstart + 1 > map->dm_maxsegsz) {
    640 		/* Oops. We crossed a boundary or large seg. Split the xfer. */
    641 		len = map->dm_maxsegsz;
    642 		if ((sgstart & bmask) != (sgend & bmask))
    643 			len = uimin(len, boundary - (sgstart & (boundary - 1)));
    644 		map->dm_segs[seg].ds_len = len;
    645 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
    646 		    "seg %d start %lx size %lx\n", seg,
    647 		    (long)map->dm_segs[seg].ds_addr,
    648 		    (long)map->dm_segs[seg].ds_len));
    649 		if (++seg >= map->_dm_segcnt) {
    650 			/* Too many segments.  Fail the operation. */
    651 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
    652 			    "too many segments %d\n", seg));
    653 			mutex_enter(&is->is_lock);
    654 			err = extent_free(is->is_dvmamap,
    655 			    dvmaddr, sgsize, EX_NOWAIT);
    656 			map->_dm_dvmastart = 0;
    657 			map->_dm_dvmasize = 0;
    658 			mutex_exit(&is->is_lock);
    659 			if (err != 0)
    660 				printf("warning: %s: %" PRId64
    661 				    " of DVMA space lost\n", __func__, sgsize);
    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 	int error;
    746 	bus_size_t sgsize = map->_dm_dvmasize;
    747 
    748 	/* Flush the iommu */
    749 	if (!map->_dm_dvmastart)
    750 		panic("%s: error dvmastart is zero!\n", __func__);
    751 
    752 	if (is->is_flags & IOMMU_SYNC_BEFORE_UNMAP) {
    753 
    754 		/* Flush the caches */
    755 		bus_dmamap_unload(t->_parent, map);
    756 
    757 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
    758 
    759 	} else {
    760 
    761 		iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
    762 
    763 		/* Flush the caches */
    764 		bus_dmamap_unload(t->_parent, map);
    765 	}
    766 
    767 	mutex_enter(&is->is_lock);
    768 	error = extent_free(is->is_dvmamap, map->_dm_dvmastart,
    769 		map->_dm_dvmasize, EX_NOWAIT);
    770 	map->_dm_dvmastart = 0;
    771 	map->_dm_dvmasize = 0;
    772 	mutex_exit(&is->is_lock);
    773 	if (error != 0)
    774 		printf("warning: %s: %" PRId64 " of DVMA space lost\n",
    775 		    __func__, sgsize);
    776 
    777 	/* Clear the map */
    778 }
    779 
    780 
    781 int
    782 iommu_dvmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
    783 	bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
    784 {
    785 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
    786 	struct iommu_state *is = sb->sb_is;
    787 	struct vm_page *pg;
    788 	int i, j;
    789 	int left;
    790 	int err, needsflush;
    791 	bus_size_t sgsize;
    792 	paddr_t pa;
    793 	bus_size_t boundary, align;
    794 	u_long dvmaddr, sgstart, sgend, bmask;
    795 	struct pglist *pglist;
    796 	const int pagesz = PAGE_SIZE;
    797 	int slot;
    798 #ifdef DEBUG
    799 	int npg = 0;
    800 #endif
    801 
    802 	if (map->dm_nsegs) {
    803 		/* Already in use?? */
    804 #ifdef DIAGNOSTIC
    805 		printf("iommu_dvmamap_load_raw: map still in use\n");
    806 #endif
    807 		bus_dmamap_unload(t, map);
    808 	}
    809 
    810 	/*
    811 	 * A boundary presented to bus_dmamem_alloc() takes precedence
    812 	 * over boundary in the map.
    813 	 */
    814 	if ((boundary = segs[0]._ds_boundary) == 0)
    815 		boundary = map->_dm_boundary;
    816 
    817 	align = uimax(segs[0]._ds_align, pagesz);
    818 
    819 	/*
    820 	 * Make sure that on error condition we return "no valid mappings".
    821 	 */
    822 	map->dm_nsegs = 0;
    823 	/* Count up the total number of pages we need */
    824 	pa = trunc_page(segs[0].ds_addr);
    825 	sgsize = 0;
    826 	left = size;
    827 	for (i = 0; left > 0 && i < nsegs; i++) {
    828 		if (round_page(pa) != round_page(segs[i].ds_addr))
    829 			sgsize = round_page(sgsize) +
    830 			    (segs[i].ds_addr & PGOFSET);
    831 		sgsize += uimin(left, segs[i].ds_len);
    832 		left -= segs[i].ds_len;
    833 		pa = segs[i].ds_addr + segs[i].ds_len;
    834 	}
    835 	sgsize = round_page(sgsize);
    836 
    837 	mutex_enter(&is->is_lock);
    838 	/*
    839 	 * If our segment size is larger than the boundary we need to
    840 	 * split the transfer up into little pieces ourselves.
    841 	 */
    842 	err = extent_alloc(is->is_dvmamap, sgsize, align,
    843 		(sgsize > boundary) ? 0 : boundary,
    844 		((flags & BUS_DMA_NOWAIT) == 0 ? EX_WAITOK : EX_NOWAIT) |
    845 		EX_BOUNDZERO, &dvmaddr);
    846 	mutex_exit(&is->is_lock);
    847 
    848 	if (err != 0)
    849 		return (err);
    850 
    851 #ifdef DEBUG
    852 	if (dvmaddr == (u_long)-1)
    853 	{
    854 		printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
    855 		    (int)sgsize, flags);
    856 #ifdef DDB
    857 		Debugger();
    858 #endif
    859 	}
    860 #endif
    861 	if (dvmaddr == (u_long)-1)
    862 		return (ENOMEM);
    863 
    864 	/* Set the active DVMA map */
    865 	map->_dm_dvmastart = dvmaddr;
    866 	map->_dm_dvmasize = sgsize;
    867 
    868 	bmask = ~(boundary - 1);
    869 	if ((pglist = segs[0]._ds_mlist) == NULL) {
    870 		u_long prev_va = 0UL, last_va = dvmaddr;
    871 		paddr_t prev_pa = 0;
    872 		int end = 0, offset;
    873 		bus_size_t len = size;
    874 
    875 		/*
    876 		 * This segs is made up of individual physical
    877 		 *  segments, probably by _bus_dmamap_load_uio() or
    878 		 * _bus_dmamap_load_mbuf().  Ignore the mlist and
    879 		 * load each one individually.
    880 		 */
    881 		j = 0;
    882 		needsflush = 0;
    883 		for (i = 0; i < nsegs ; i++) {
    884 
    885 			pa = segs[i].ds_addr;
    886 			offset = (pa & PGOFSET);
    887 			pa = trunc_page(pa);
    888 			dvmaddr = trunc_page(dvmaddr);
    889 			left = uimin(len, segs[i].ds_len);
    890 
    891 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: converting "
    892 				"physseg %d start %lx size %lx\n", i,
    893 				(long)segs[i].ds_addr, (long)segs[i].ds_len));
    894 
    895 			if ((pa == prev_pa) &&
    896 				((offset != 0) || (end != offset))) {
    897 				/* We can re-use this mapping */
    898 				dvmaddr = prev_va;
    899 			}
    900 
    901 			sgstart = dvmaddr + offset;
    902 			sgend = sgstart + left - 1;
    903 
    904 			/* Are the segments virtually adjacent? */
    905 			if ((j > 0) && (end == offset) &&
    906 			    ((offset == 0) || (pa == prev_pa)) &&
    907 			    (map->dm_segs[j-1].ds_len + left <=
    908 			     map->dm_maxsegsz)) {
    909 				/* Just append to the previous segment. */
    910 				map->dm_segs[--j].ds_len += left;
    911 				/* Restore sgstart for boundary check */
    912 				sgstart = map->dm_segs[j].ds_addr;
    913 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    914 					"appending seg %d start %lx size %lx\n", j,
    915 					(long)map->dm_segs[j].ds_addr,
    916 					(long)map->dm_segs[j].ds_len));
    917 			} else {
    918 				if (j >= map->_dm_segcnt) {
    919 					iommu_remove(is, map->_dm_dvmastart,
    920 					    last_va - map->_dm_dvmastart);
    921 					goto fail;
    922 				}
    923 				map->dm_segs[j].ds_addr = sgstart;
    924 				map->dm_segs[j].ds_len = left;
    925 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    926 					"seg %d start %lx size %lx\n", j,
    927 					(long)map->dm_segs[j].ds_addr,
    928 					(long)map->dm_segs[j].ds_len));
    929 			}
    930 			end = (offset + left) & PGOFSET;
    931 
    932 			/* Check for boundary issues */
    933 			while ((sgstart & bmask) != (sgend & bmask)) {
    934 				/* Need a new segment. */
    935 				map->dm_segs[j].ds_len =
    936 					boundary - (sgstart & (boundary - 1));
    937 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
    938 					"seg %d start %lx size %lx\n", j,
    939 					(long)map->dm_segs[j].ds_addr,
    940 					(long)map->dm_segs[j].ds_len));
    941 				if (++j >= map->_dm_segcnt) {
    942 					iommu_remove(is, map->_dm_dvmastart,
    943 					    last_va - map->_dm_dvmastart);
    944 					goto fail;
    945 				}
    946 				sgstart += map->dm_segs[j-1].ds_len;
    947 				map->dm_segs[j].ds_addr = sgstart;
    948 				map->dm_segs[j].ds_len = sgend - sgstart + 1;
    949 			}
    950 
    951 			if (sgsize == 0)
    952 				panic("iommu_dmamap_load_raw: size botch");
    953 
    954 			/* Now map a series of pages. */
    955 			while (dvmaddr <= sgend) {
    956 				DPRINTF(IDB_BUSDMA,
    957 					("iommu_dvmamap_load_raw: map %p "
    958 						"loading va %lx at pa %lx\n",
    959 						map, (long)dvmaddr,
    960 						(long)(pa)));
    961 				/* Enter it if we haven't before. */
    962 				if (prev_va != dvmaddr) {
    963 					iommu_enter(sb, prev_va = dvmaddr,
    964 					    prev_pa = pa,
    965 					    flags | IOTTE_DEBUG(++npg << 12));
    966 					needsflush = 1;
    967 
    968 					/* Flush cache if necessary. */
    969 					slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
    970 					if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
    971 					    ((dvmaddr + pagesz) > sgend || (slot % 8) == 7))
    972 						IOMMUREG_WRITE(is, iommu_cache_flush,
    973 						    is->is_ptsb + slot * 8);
    974 				}
    975 
    976 				dvmaddr += pagesz;
    977 				pa += pagesz;
    978 				last_va = dvmaddr;
    979 			}
    980 
    981 			len -= left;
    982 			++j;
    983 		}
    984 		if (needsflush)
    985 			iommu_strbuf_flush_done(sb);
    986 
    987 		map->dm_mapsize = size;
    988 		map->dm_nsegs = j;
    989 #ifdef DIAGNOSTIC
    990 		{ int seg;
    991 	for (seg = 0; seg < map->dm_nsegs; seg++) {
    992 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
    993 		    map->dm_segs[seg].ds_addr > is->is_dvmaend) {
    994 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
    995 				seg, (long)map->dm_segs[seg].ds_addr,
    996 				is->is_dvmabase, is->is_dvmaend);
    997 #ifdef DDB
    998 			Debugger();
    999 #endif
   1000 		}
   1001 	}
   1002 		}
   1003 #endif
   1004 		return (0);
   1005 	}
   1006 
   1007 	/*
   1008 	 * This was allocated with bus_dmamem_alloc.
   1009 	 * The pages are on a `pglist'.
   1010 	 */
   1011 	i = 0;
   1012 	sgstart = dvmaddr;
   1013 	sgend = sgstart + size - 1;
   1014 	map->dm_segs[i].ds_addr = sgstart;
   1015 	while ((sgstart & bmask) != (sgend & bmask)) {
   1016 		/* Oops.  We crossed a boundary.  Split the xfer. */
   1017 		map->dm_segs[i].ds_len = boundary - (sgstart & (boundary - 1));
   1018 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
   1019 			"seg %d start %lx size %lx\n", i,
   1020 			(long)map->dm_segs[i].ds_addr,
   1021 			(long)map->dm_segs[i].ds_len));
   1022 		if (++i >= map->_dm_segcnt) {
   1023 			/* Too many segments.  Fail the operation. */
   1024 			goto fail;
   1025 		}
   1026 		sgstart += map->dm_segs[i-1].ds_len;
   1027 		map->dm_segs[i].ds_addr = sgstart;
   1028 	}
   1029 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
   1030 			"seg %d start %lx size %lx\n", i,
   1031 			(long)map->dm_segs[i].ds_addr, (long)map->dm_segs[i].ds_len));
   1032 	map->dm_segs[i].ds_len = sgend - sgstart + 1;
   1033 
   1034 	needsflush = 0;
   1035 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
   1036 		if (sgsize == 0)
   1037 			panic("iommu_dmamap_load_raw: size botch");
   1038 		pa = VM_PAGE_TO_PHYS(pg);
   1039 
   1040 		DPRINTF(IDB_BUSDMA,
   1041 		    ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
   1042 		    map, (long)dvmaddr, (long)(pa)));
   1043 		iommu_enter(sb, dvmaddr, pa, flags | IOTTE_DEBUG(0x8000));
   1044 		needsflush = 1;
   1045 
   1046 		sgsize -= pagesz;
   1047 
   1048 		/* Flush cache if necessary. */
   1049 		slot = IOTSBSLOT(trunc_page(dvmaddr), is->is_tsbsize);
   1050 		if ((is->is_flags & IOMMU_FLUSH_CACHE) &&
   1051 		    (sgsize == 0 || (slot % 8) == 7))
   1052 			IOMMUREG_WRITE(is, iommu_cache_flush,
   1053 			    is->is_ptsb + slot * 8);
   1054 
   1055 		dvmaddr += pagesz;
   1056 	}
   1057 	if (needsflush)
   1058 		iommu_strbuf_flush_done(sb);
   1059 	map->dm_mapsize = size;
   1060 	map->dm_nsegs = i+1;
   1061 #ifdef DIAGNOSTIC
   1062 	{ int seg;
   1063 	for (seg = 0; seg < map->dm_nsegs; seg++) {
   1064 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
   1065 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
   1066 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
   1067 				seg, (long)map->dm_segs[seg].ds_addr,
   1068 				is->is_dvmabase, is->is_dvmaend);
   1069 #ifdef DDB
   1070 			Debugger();
   1071 #endif
   1072 		}
   1073 	}
   1074 	}
   1075 #endif
   1076 	return (0);
   1077 
   1078 fail:
   1079 	mutex_enter(&is->is_lock);
   1080 	err = extent_free(is->is_dvmamap, map->_dm_dvmastart, sgsize,
   1081 	    EX_NOWAIT);
   1082 	map->_dm_dvmastart = 0;
   1083 	map->_dm_dvmasize = 0;
   1084 	mutex_exit(&is->is_lock);
   1085 	if (err != 0)
   1086 		printf("warning: %s: %" PRId64 " of DVMA space lost\n",
   1087 		    __func__, sgsize);
   1088 	return (EFBIG);
   1089 }
   1090 
   1091 
   1092 /*
   1093  * Flush an individual dma segment, returns non-zero if the streaming buffers
   1094  * need flushing afterwards.
   1095  */
   1096 static int
   1097 iommu_dvmamap_sync_range(struct strbuf_ctl *sb, vaddr_t va, bus_size_t len)
   1098 {
   1099 	vaddr_t vaend;
   1100 	struct iommu_state *is = sb->sb_is;
   1101 
   1102 #ifdef DIAGNOSTIC
   1103 	if (va < is->is_dvmabase || va > is->is_dvmaend)
   1104 		panic("invalid va: %llx", (long long)va);
   1105 #endif
   1106 
   1107 	if ((is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)] & IOTTE_STREAM) == 0) {
   1108 		DPRINTF(IDB_SYNC,
   1109 			("iommu_dvmamap_sync_range: attempting to flush "
   1110 			 "non-streaming entry\n"));
   1111 		return (0);
   1112 	}
   1113 
   1114 	vaend = round_page(va + len) - 1;
   1115 	va = trunc_page(va);
   1116 
   1117 #ifdef DIAGNOSTIC
   1118 	if (va < is->is_dvmabase || vaend > is->is_dvmaend)
   1119 		panic("invalid va range: %llx to %llx (%x to %x)",
   1120 		    (long long)va, (long long)vaend,
   1121 		    is->is_dvmabase,
   1122 		    is->is_dvmaend);
   1123 #endif
   1124 
   1125 	for ( ; va <= vaend; va += PAGE_SIZE) {
   1126 		DPRINTF(IDB_SYNC,
   1127 		    ("iommu_dvmamap_sync_range: flushing va %p\n",
   1128 		    (void *)(u_long)va));
   1129 		iommu_strbuf_flush(sb, va);
   1130 	}
   1131 
   1132 	return (1);
   1133 }
   1134 
   1135 static void
   1136 _iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
   1137 	bus_size_t len, int ops)
   1138 {
   1139 	struct strbuf_ctl *sb = (struct strbuf_ctl *)map->_dm_cookie;
   1140 	bus_size_t count;
   1141 	int i, needsflush = 0;
   1142 
   1143 	if (!sb->sb_flush)
   1144 		return;
   1145 
   1146 	for (i = 0; i < map->dm_nsegs; i++) {
   1147 		if (offset < map->dm_segs[i].ds_len)
   1148 			break;
   1149 		offset -= map->dm_segs[i].ds_len;
   1150 	}
   1151 
   1152 	if (i == map->dm_nsegs)
   1153 		panic("%s: segment too short %llu", __func__,
   1154 		    (unsigned long long)offset);
   1155 
   1156 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_POSTWRITE)) {
   1157 		/* Nothing to do */;
   1158 	}
   1159 
   1160 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_PREWRITE)) {
   1161 
   1162 		for (; len > 0 && i < map->dm_nsegs; i++) {
   1163 			count = MIN(map->dm_segs[i].ds_len - offset, len);
   1164 			if (count > 0 &&
   1165 			    iommu_dvmamap_sync_range(sb,
   1166 				map->dm_segs[i].ds_addr + offset, count))
   1167 				needsflush = 1;
   1168 			offset = 0;
   1169 			len -= count;
   1170 		}
   1171 #ifdef DIAGNOSTIC
   1172 		if (i == map->dm_nsegs && len > 0)
   1173 			panic("%s: leftover %llu", __func__,
   1174 			    (unsigned long long)len);
   1175 #endif
   1176 
   1177 		if (needsflush)
   1178 			iommu_strbuf_flush_done(sb);
   1179 	}
   1180 }
   1181 
   1182 void
   1183 iommu_dvmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
   1184 	bus_size_t len, int ops)
   1185 {
   1186 
   1187 	/* If len is 0, then there is nothing to do */
   1188 	if (len == 0)
   1189 		return;
   1190 
   1191 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) {
   1192 		/* Flush the CPU then the IOMMU */
   1193 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
   1194 		_iommu_dvmamap_sync(t, map, offset, len, ops);
   1195 	}
   1196 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE)) {
   1197 		/* Flush the IOMMU then the CPU */
   1198 		_iommu_dvmamap_sync(t, map, offset, len, ops);
   1199 		bus_dmamap_sync(t->_parent, map, offset, len, ops);
   1200 	}
   1201 }
   1202 
   1203 int
   1204 iommu_dvmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
   1205 	bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
   1206 	int flags)
   1207 {
   1208 
   1209 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
   1210 	   "segp %p flags %d\n", (unsigned long long)size,
   1211 	   (unsigned long long)alignment, (unsigned long long)boundary,
   1212 	   segs, flags));
   1213 	return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
   1214 	    segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
   1215 }
   1216 
   1217 void
   1218 iommu_dvmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
   1219 {
   1220 
   1221 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
   1222 	    segs, nsegs));
   1223 	bus_dmamem_free(t->_parent, segs, nsegs);
   1224 }
   1225 
   1226 /*
   1227  * Map the DVMA mappings into the kernel pmap.
   1228  * Check the flags to see whether we're streaming or coherent.
   1229  */
   1230 int
   1231 iommu_dvmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
   1232 	size_t size, void **kvap, int flags)
   1233 {
   1234 	struct vm_page *pg;
   1235 	vaddr_t va;
   1236 	bus_addr_t addr;
   1237 	struct pglist *pglist;
   1238 	int cbit;
   1239 	const uvm_flag_t kmflags =
   1240 	    (flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0;
   1241 
   1242 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
   1243 	    segs, nsegs, size));
   1244 
   1245 	/*
   1246 	 * Allocate some space in the kernel map, and then map these pages
   1247 	 * into this space.
   1248 	 */
   1249 	size = round_page(size);
   1250 	va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY | kmflags);
   1251 	if (va == 0)
   1252 		return (ENOMEM);
   1253 
   1254 	*kvap = (void *)va;
   1255 
   1256 	/*
   1257 	 * digest flags:
   1258 	 */
   1259 	cbit = 0;
   1260 	if (flags & BUS_DMA_COHERENT)	/* Disable vcache */
   1261 		cbit |= PMAP_NVC;
   1262 	if (flags & BUS_DMA_NOCACHE)	/* side effects */
   1263 		cbit |= PMAP_NC;
   1264 
   1265 	/*
   1266 	 * Now take this and map it into the CPU.
   1267 	 */
   1268 	pglist = segs[0]._ds_mlist;
   1269 	TAILQ_FOREACH(pg, pglist, pageq.queue) {
   1270 #ifdef DIAGNOSTIC
   1271 		if (size == 0)
   1272 			panic("iommu_dvmamem_map: size botch");
   1273 #endif
   1274 		addr = VM_PAGE_TO_PHYS(pg);
   1275 		DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
   1276 		    "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
   1277 		pmap_kenter_pa(va, addr | cbit,
   1278 		    VM_PROT_READ | VM_PROT_WRITE, 0);
   1279 		va += PAGE_SIZE;
   1280 		size -= PAGE_SIZE;
   1281 	}
   1282 	pmap_update(pmap_kernel());
   1283 	return (0);
   1284 }
   1285 
   1286 /*
   1287  * Unmap DVMA mappings from kernel
   1288  */
   1289 void
   1290 iommu_dvmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
   1291 {
   1292 
   1293 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
   1294 	    kva, size));
   1295 
   1296 #ifdef DIAGNOSTIC
   1297 	if ((u_long)kva & PGOFSET)
   1298 		panic("iommu_dvmamem_unmap");
   1299 #endif
   1300 
   1301 	size = round_page(size);
   1302 	pmap_kremove((vaddr_t)kva, size);
   1303 	pmap_update(pmap_kernel());
   1304 	uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
   1305 }
   1306