sba_iommu.c 58.9 KB
Newer Older
Linus Torvalds's avatar
Linus Torvalds committed
1 2 3
/*
**  IA64 System Bus Adapter (SBA) I/O MMU manager
**
Alex Williamson's avatar
Alex Williamson committed
4
**	(c) Copyright 2002-2005 Alex Williamson
Linus Torvalds's avatar
Linus Torvalds committed
5
**	(c) Copyright 2002-2003 Grant Grundler
Alex Williamson's avatar
Alex Williamson committed
6
**	(c) Copyright 2002-2005 Hewlett-Packard Company
Linus Torvalds's avatar
Linus Torvalds committed
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
**
**	Portions (c) 2000 Grant Grundler (from parisc I/O MMU code)
**	Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
**
**	This program is free software; you can redistribute it and/or modify
**	it under the terms of the GNU General Public License as published by
**      the Free Software Foundation; either version 2 of the License, or
**      (at your option) any later version.
**
**
** This module initializes the IOC (I/O Controller) found on HP
** McKinley machines and their successors.
**
*/

#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/spinlock.h>
#include <linux/slab.h>
#include <linux/init.h>
#include <linux/mm.h>
#include <linux/string.h>
#include <linux/pci.h>
#include <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/acpi.h>
#include <linux/efi.h>
#include <linux/nodemask.h>
#include <linux/bitops.h>         /* hweight64() */
37
#include <linux/crash_dump.h>
38
#include <linux/iommu-helper.h>
39
#include <linux/dma-mapping.h>
40
#include <linux/prefetch.h>
Linus Torvalds's avatar
Linus Torvalds committed
41 42 43 44 45 46 47 48 49

#include <asm/delay.h>		/* ia64_get_itc() */
#include <asm/io.h>
#include <asm/page.h>		/* PAGE_OFFSET */
#include <asm/dma.h>
#include <asm/system.h>		/* wmb() */

#include <asm/acpi-ext.h>

50 51
extern int swiotlb_late_init_with_default_size (size_t size);

Linus Torvalds's avatar
Linus Torvalds committed
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
#define PFX "IOC: "

/*
** Enabling timing search of the pdir resource map.  Output in /proc.
** Disabled by default to optimize performance.
*/
#undef PDIR_SEARCH_TIMING

/*
** This option allows cards capable of 64bit DMA to bypass the IOMMU.  If
** not defined, all DMA will be 32bit and go through the TLB.
** There's potentially a conflict in the bio merge code with us
** advertising an iommu, but then bypassing it.  Since I/O MMU bypassing
** appears to give more performance than bio-level virtual merging, we'll
** do the former for now.  NOTE: BYPASS_SG also needs to be undef'd to
** completely restrict DMA to the IOMMU.
*/
#define ALLOW_IOV_BYPASS

/*
** This option specifically allows/disallows bypassing scatterlists with
** multiple entries.  Coalescing these entries can allow better DMA streaming
** and in some cases shows better performance than entirely bypassing the
** IOMMU.  Performance increase on the order of 1-2% sequential output/input
** using bonnie++ on a RAID0 MD device (sym2 & mpt).
*/
#undef ALLOW_IOV_BYPASS_SG

/*
** If a device prefetches beyond the end of a valid pdir entry, it will cause
** a hard failure, ie. MCA.  Version 3.0 and later of the zx1 LBA should
** disconnect on 4k boundaries and prevent such issues.  If the device is
Matt LaPlante's avatar
Matt LaPlante committed
84
** particularly aggressive, this option will keep the entire pdir valid such
Linus Torvalds's avatar
Linus Torvalds committed
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
** that prefetching will hit a valid address.  This could severely impact
** error containment, and is therefore off by default.  The page that is
** used for spill-over is poisoned, so that should help debugging somewhat.
*/
#undef FULL_VALID_PDIR

#define ENABLE_MARK_CLEAN

/*
** The number of debug flags is a clue - this code is fragile.  NOTE: since
** tightening the use of res_lock the resource bitmap and actual pdir are no
** longer guaranteed to stay in sync.  The sanity checking code isn't going to
** like that.
*/
#undef DEBUG_SBA_INIT
#undef DEBUG_SBA_RUN
#undef DEBUG_SBA_RUN_SG
#undef DEBUG_SBA_RESOURCE
#undef ASSERT_PDIR_SANITY
#undef DEBUG_LARGE_SG_ENTRIES
#undef DEBUG_BYPASS

#if defined(FULL_VALID_PDIR) && defined(ASSERT_PDIR_SANITY)
#error FULL_VALID_PDIR and ASSERT_PDIR_SANITY are mutually exclusive
#endif

#define SBA_INLINE	__inline__
/* #define SBA_INLINE */

#ifdef DEBUG_SBA_INIT
#define DBG_INIT(x...)	printk(x)
#else
#define DBG_INIT(x...)
#endif

#ifdef DEBUG_SBA_RUN
#define DBG_RUN(x...)	printk(x)
#else
#define DBG_RUN(x...)
#endif

#ifdef DEBUG_SBA_RUN_SG
#define DBG_RUN_SG(x...)	printk(x)
#else
#define DBG_RUN_SG(x...)
#endif


#ifdef DEBUG_SBA_RESOURCE
#define DBG_RES(x...)	printk(x)
#else
#define DBG_RES(x...)
#endif

#ifdef DEBUG_BYPASS
#define DBG_BYPASS(x...)	printk(x)
#else
#define DBG_BYPASS(x...)
#endif

#ifdef ASSERT_PDIR_SANITY
#define ASSERT(expr) \
        if(!(expr)) { \
                printk( "\n" __FILE__ ":%d: Assertion " #expr " failed!\n",__LINE__); \
                panic(#expr); \
        }
#else
#define ASSERT(expr)
#endif

/*
** The number of pdir entries to "free" before issuing
** a read to PCOM register to flush out PCOM writes.
** Interacts with allocation granularity (ie 4 or 8 entries
** allocated and free'd/purged at a time might make this
** less interesting).
*/
#define DELAYED_RESOURCE_CNT	64

164 165
#define PCI_DEVICE_ID_HP_SX2000_IOC	0x12ec

Linus Torvalds's avatar
Linus Torvalds committed
166 167 168 169
#define ZX1_IOC_ID	((PCI_DEVICE_ID_HP_ZX1_IOC << 16) | PCI_VENDOR_ID_HP)
#define ZX2_IOC_ID	((PCI_DEVICE_ID_HP_ZX2_IOC << 16) | PCI_VENDOR_ID_HP)
#define REO_IOC_ID	((PCI_DEVICE_ID_HP_REO_IOC << 16) | PCI_VENDOR_ID_HP)
#define SX1000_IOC_ID	((PCI_DEVICE_ID_HP_SX1000_IOC << 16) | PCI_VENDOR_ID_HP)
170
#define SX2000_IOC_ID	((PCI_DEVICE_ID_HP_SX2000_IOC << 16) | PCI_VENDOR_ID_HP)
Linus Torvalds's avatar
Linus Torvalds committed
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251

#define ZX1_IOC_OFFSET	0x1000	/* ACPI reports SBA, we want IOC */

#define IOC_FUNC_ID	0x000
#define IOC_FCLASS	0x008	/* function class, bist, header, rev... */
#define IOC_IBASE	0x300	/* IO TLB */
#define IOC_IMASK	0x308
#define IOC_PCOM	0x310
#define IOC_TCNFG	0x318
#define IOC_PDIR_BASE	0x320

#define IOC_ROPE0_CFG	0x500
#define   IOC_ROPE_AO	  0x10	/* Allow "Relaxed Ordering" */


/* AGP GART driver looks for this */
#define ZX1_SBA_IOMMU_COOKIE	0x0000badbadc0ffeeUL

/*
** The zx1 IOC supports 4/8/16/64KB page sizes (see TCNFG register)
**
** Some IOCs (sx1000) can run at the above pages sizes, but are
** really only supported using the IOC at a 4k page size.
**
** iovp_size could only be greater than PAGE_SIZE if we are
** confident the drivers really only touch the next physical
** page iff that driver instance owns it.
*/
static unsigned long iovp_size;
static unsigned long iovp_shift;
static unsigned long iovp_mask;

struct ioc {
	void __iomem	*ioc_hpa;	/* I/O MMU base address */
	char		*res_map;	/* resource map, bit == pdir entry */
	u64		*pdir_base;	/* physical base address */
	unsigned long	ibase;		/* pdir IOV Space base */
	unsigned long	imask;		/* pdir IOV Space mask */

	unsigned long	*res_hint;	/* next avail IOVP - circular search */
	unsigned long	dma_mask;
	spinlock_t	res_lock;	/* protects the resource bitmap, but must be held when */
					/* clearing pdir to prevent races with allocations. */
	unsigned int	res_bitshift;	/* from the RIGHT! */
	unsigned int	res_size;	/* size of resource map in bytes */
#ifdef CONFIG_NUMA
	unsigned int	node;		/* node where this IOC lives */
#endif
#if DELAYED_RESOURCE_CNT > 0
	spinlock_t	saved_lock;	/* may want to try to get this on a separate cacheline */
					/* than res_lock for bigger systems. */
	int		saved_cnt;
	struct sba_dma_pair {
		dma_addr_t	iova;
		size_t		size;
	} saved[DELAYED_RESOURCE_CNT];
#endif

#ifdef PDIR_SEARCH_TIMING
#define SBA_SEARCH_SAMPLE	0x100
	unsigned long avg_search[SBA_SEARCH_SAMPLE];
	unsigned long avg_idx;	/* current index into avg_search */
#endif

	/* Stuff we don't need in performance path */
	struct ioc	*next;		/* list of IOC's in system */
	acpi_handle	handle;		/* for multiple IOC's */
	const char 	*name;
	unsigned int	func_id;
	unsigned int	rev;		/* HW revision of chip */
	u32		iov_size;
	unsigned int	pdir_size;	/* in bytes, determined by IOV Space size */
	struct pci_dev	*sac_only_dev;
};

static struct ioc *ioc_list;
static int reserve_sba_gart = 1;

static SBA_INLINE void sba_mark_invalid(struct ioc *, dma_addr_t, size_t);
static SBA_INLINE void sba_free_range(struct ioc *, dma_addr_t, size_t);

Jens Axboe's avatar
Jens Axboe committed
252
#define sba_sg_address(sg)	sg_virt((sg))
Linus Torvalds's avatar
Linus Torvalds committed
253 254 255 256 257 258 259 260 261 262 263 264 265 266

#ifdef FULL_VALID_PDIR
static u64 prefetch_spill_page;
#endif

#ifdef CONFIG_PCI
# define GET_IOC(dev)	(((dev)->bus == &pci_bus_type)						\
			 ? ((struct ioc *) PCI_CONTROLLER(to_pci_dev(dev))->iommu) : NULL)
#else
# define GET_IOC(dev)	NULL
#endif

/*
** DMA_CHUNK_SIZE is used by the SCSI mid-layer to break up
Matt LaPlante's avatar
Matt LaPlante committed
267
** (or rather not merge) DMAs into manageable chunks.
Linus Torvalds's avatar
Linus Torvalds committed
268
** On parisc, this is more of the software/tuning constraint
Matt LaPlante's avatar
Matt LaPlante committed
269 270
** rather than the HW. I/O MMU allocation algorithms can be
** faster with smaller sizes (to some degree).
Linus Torvalds's avatar
Linus Torvalds committed
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
*/
#define DMA_CHUNK_SIZE  (BITS_PER_LONG*iovp_size)

#define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))

/************************************
** SBA register read and write support
**
** BE WARNED: register writes are posted.
**  (ie follow writes which must reach HW with a read)
**
*/
#define READ_REG(addr)       __raw_readq(addr)
#define WRITE_REG(val, addr) __raw_writeq(val, addr)

#ifdef DEBUG_SBA_INIT

/**
 * sba_dump_tlb - debugging only - print IOMMU operating parameters
 * @hpa: base address of the IOMMU
 *
 * Print the size/location of the IO MMU PDIR.
 */
static void
sba_dump_tlb(char *hpa)
{
	DBG_INIT("IO TLB at 0x%p\n", (void *)hpa);
	DBG_INIT("IOC_IBASE    : %016lx\n", READ_REG(hpa+IOC_IBASE));
	DBG_INIT("IOC_IMASK    : %016lx\n", READ_REG(hpa+IOC_IMASK));
	DBG_INIT("IOC_TCNFG    : %016lx\n", READ_REG(hpa+IOC_TCNFG));
	DBG_INIT("IOC_PDIR_BASE: %016lx\n", READ_REG(hpa+IOC_PDIR_BASE));
	DBG_INIT("\n");
}
#endif


#ifdef ASSERT_PDIR_SANITY

/**
 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @msg: text to print ont the output line.
 * @pide: pdir index.
 *
 * Print one entry of the IO MMU PDIR in human readable form.
 */
static void
sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
{
	/* start printing from lowest pde in rval */
	u64 *ptr = &ioc->pdir_base[pide  & ~(BITS_PER_LONG - 1)];
	unsigned long *rptr = (unsigned long *) &ioc->res_map[(pide >>3) & -sizeof(unsigned long)];
	uint rcnt;

	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
		 msg, rptr, pide & (BITS_PER_LONG - 1), *rptr);

	rcnt = 0;
	while (rcnt < BITS_PER_LONG) {
		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
		       (rcnt == (pide & (BITS_PER_LONG - 1)))
		       ? "    -->" : "       ",
		       rcnt, ptr, (unsigned long long) *ptr );
		rcnt++;
		ptr++;
	}
	printk(KERN_DEBUG "%s", msg);
}


/**
 * sba_check_pdir - debugging only - consistency checker
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @msg: text to print ont the output line.
 *
 * Verify the resource map and pdir state is consistent
 */
static int
sba_check_pdir(struct ioc *ioc, char *msg)
{
	u64 *rptr_end = (u64 *) &(ioc->res_map[ioc->res_size]);
	u64 *rptr = (u64 *) ioc->res_map;	/* resource map ptr */
	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
	uint pide = 0;

	while (rptr < rptr_end) {
		u64 rval;
		int rcnt; /* number of bits we might check */

		rval = *rptr;
		rcnt = 64;

		while (rcnt) {
			/* Get last byte and highest bit from that */
			u32 pde = ((u32)((*pptr >> (63)) & 0x1));
			if ((rval & 0x1) ^ pde)
			{
				/*
				** BUMMER!  -- res_map != pdir --
				** Dump rval and matching pdir entries
				*/
				sba_dump_pdir_entry(ioc, msg, pide);
				return(1);
			}
			rcnt--;
			rval >>= 1;	/* try the next bit */
			pptr++;
			pide++;
		}
		rptr++;	/* look at next word of res_map */
	}
	/* It'd be nice if we always got here :^) */
	return 0;
}


/**
 * sba_dump_sg - debugging only - print Scatter-Gather list
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @startsg: head of the SG list
 * @nents: number of entries in SG list
 *
 * print the SG list so we can verify it's correct by hand.
 */
static void
sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
{
	while (nents-- > 0) {
		printk(KERN_DEBUG " %d : DMA %08lx/%05x CPU %p\n", nents,
		       startsg->dma_address, startsg->dma_length,
		       sba_sg_address(startsg));
Jens Axboe's avatar
Jens Axboe committed
402
		startsg = sg_next(startsg);
Linus Torvalds's avatar
Linus Torvalds committed
403 404 405 406 407 408 409 410 411 412 413 414
	}
}

static void
sba_check_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
{
	struct scatterlist *the_sg = startsg;
	int the_nents = nents;

	while (the_nents-- > 0) {
		if (sba_sg_address(the_sg) == 0x0UL)
			sba_dump_sg(NULL, startsg, nents);
Jens Axboe's avatar
Jens Axboe committed
415
		the_sg = sg_next(the_sg);
Linus Torvalds's avatar
Linus Torvalds committed
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465
	}
}

#endif /* ASSERT_PDIR_SANITY */




/**************************************************************
*
*   I/O Pdir Resource Management
*
*   Bits set in the resource map are in use.
*   Each bit can represent a number of pages.
*   LSbs represent lower addresses (IOVA's).
*
***************************************************************/
#define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */

/* Convert from IOVP to IOVA and vice versa. */
#define SBA_IOVA(ioc,iovp,offset) ((ioc->ibase) | (iovp) | (offset))
#define SBA_IOVP(ioc,iova) ((iova) & ~(ioc->ibase))

#define PDIR_ENTRY_SIZE	sizeof(u64)

#define PDIR_INDEX(iovp)   ((iovp)>>iovp_shift)

#define RESMAP_MASK(n)    ~(~0UL << (n))
#define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)


/**
 * For most cases the normal get_order is sufficient, however it limits us
 * to PAGE_SIZE being the minimum mapping alignment and TC flush granularity.
 * It only incurs about 1 clock cycle to use this one with the static variable
 * and makes the code more intuitive.
 */
static SBA_INLINE int
get_iovp_order (unsigned long size)
{
	long double d = size - 1;
	long order;

	order = ia64_getf_exp(d);
	order = order - iovp_shift - 0xffff + 1;
	if (order < 0)
		order = 0;
	return order;
}

466 467 468 469 470 471 472
static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
				 unsigned int bitshiftcnt)
{
	return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
		+ bitshiftcnt;
}

Linus Torvalds's avatar
Linus Torvalds committed
473 474 475 476
/**
 * sba_search_bitmap - find free space in IO PDIR resource bitmap
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @bits_wanted: number of entries we need.
Alex Williamson's avatar
Alex Williamson committed
477
 * @use_hint: use res_hint to indicate where to start looking
Linus Torvalds's avatar
Linus Torvalds committed
478 479 480 481 482 483
 *
 * Find consecutive free bits in resource bitmap.
 * Each bit represents one entry in the IO Pdir.
 * Cool perf optimization: search for log2(size) bits at a time.
 */
static SBA_INLINE unsigned long
484 485
sba_search_bitmap(struct ioc *ioc, struct device *dev,
		  unsigned long bits_wanted, int use_hint)
Linus Torvalds's avatar
Linus Torvalds committed
486
{
Alex Williamson's avatar
Alex Williamson committed
487
	unsigned long *res_ptr;
Linus Torvalds's avatar
Linus Torvalds committed
488
	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
489 490 491 492
	unsigned long flags, pide = ~0UL, tpide;
	unsigned long boundary_size;
	unsigned long shift;
	int ret;
Linus Torvalds's avatar
Linus Torvalds committed
493 494 495 496

	ASSERT(((unsigned long) ioc->res_hint & (sizeof(unsigned long) - 1UL)) == 0);
	ASSERT(res_ptr < res_end);

497 498 499 500 501 502
	boundary_size = (unsigned long long)dma_get_seg_boundary(dev) + 1;
	boundary_size = ALIGN(boundary_size, 1ULL << iovp_shift) >> iovp_shift;

	BUG_ON(ioc->ibase & ~iovp_mask);
	shift = ioc->ibase >> iovp_shift;

Alex Williamson's avatar
Alex Williamson committed
503 504 505 506 507 508 509 510 511 512
	spin_lock_irqsave(&ioc->res_lock, flags);

	/* Allow caller to force a search through the entire resource space */
	if (likely(use_hint)) {
		res_ptr = ioc->res_hint;
	} else {
		res_ptr = (ulong *)ioc->res_map;
		ioc->res_bitshift = 0;
	}

Linus Torvalds's avatar
Linus Torvalds committed
513 514 515 516 517 518 519 520 521 522 523 524 525 526
	/*
	 * N.B.  REO/Grande defect AR2305 can cause TLB fetch timeouts
	 * if a TLB entry is purged while in use.  sba_mark_invalid()
	 * purges IOTLB entries in power-of-two sizes, so we also
	 * allocate IOVA space in power-of-two sizes.
	 */
	bits_wanted = 1UL << get_iovp_order(bits_wanted << iovp_shift);

	if (likely(bits_wanted == 1)) {
		unsigned int bitshiftcnt;
		for(; res_ptr < res_end ; res_ptr++) {
			if (likely(*res_ptr != ~0UL)) {
				bitshiftcnt = ffz(*res_ptr);
				*res_ptr |= (1UL << bitshiftcnt);
527
				pide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
Linus Torvalds's avatar
Linus Torvalds committed
528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549
				ioc->res_bitshift = bitshiftcnt + bits_wanted;
				goto found_it;
			}
		}
		goto not_found;

	}
	
	if (likely(bits_wanted <= BITS_PER_LONG/2)) {
		/*
		** Search the resource bit map on well-aligned values.
		** "o" is the alignment.
		** We need the alignment to invalidate I/O TLB using
		** SBA HW features in the unmap path.
		*/
		unsigned long o = 1 << get_iovp_order(bits_wanted << iovp_shift);
		uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
		unsigned long mask, base_mask;

		base_mask = RESMAP_MASK(bits_wanted);
		mask = base_mask << bitshiftcnt;

550
		DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
Linus Torvalds's avatar
Linus Torvalds committed
551 552 553 554 555
		for(; res_ptr < res_end ; res_ptr++)
		{ 
			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
			ASSERT(0 != mask);
			for (; mask ; mask <<= o, bitshiftcnt += o) {
556 557 558 559 560
				tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
				ret = iommu_is_span_boundary(tpide, bits_wanted,
							     shift,
							     boundary_size);
				if ((0 == ((*res_ptr) & mask)) && !ret) {
Linus Torvalds's avatar
Linus Torvalds committed
561
					*res_ptr |= mask;     /* mark resources busy! */
562
					pide = tpide;
Linus Torvalds's avatar
Linus Torvalds committed
563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582
					ioc->res_bitshift = bitshiftcnt + bits_wanted;
					goto found_it;
				}
			}

			bitshiftcnt = 0;
			mask = base_mask;

		}

	} else {
		int qwords, bits, i;
		unsigned long *end;

		qwords = bits_wanted >> 6; /* /64 */
		bits = bits_wanted - (qwords * BITS_PER_LONG);

		end = res_end - qwords;

		for (; res_ptr < end; res_ptr++) {
583 584 585 586 587
			tpide = ptr_to_pide(ioc, res_ptr, 0);
			ret = iommu_is_span_boundary(tpide, bits_wanted,
						     shift, boundary_size);
			if (ret)
				goto next_ptr;
Linus Torvalds's avatar
Linus Torvalds committed
588 589 590 591 592 593 594 595 596 597 598 599
			for (i = 0 ; i < qwords ; i++) {
				if (res_ptr[i] != 0)
					goto next_ptr;
			}
			if (bits && res_ptr[i] && (__ffs(res_ptr[i]) < bits))
				continue;

			/* Found it, mark it */
			for (i = 0 ; i < qwords ; i++)
				res_ptr[i] = ~0UL;
			res_ptr[i] |= RESMAP_MASK(bits);

600
			pide = tpide;
Linus Torvalds's avatar
Linus Torvalds committed
601 602 603 604 605 606 607 608 609 610 611 612
			res_ptr += qwords;
			ioc->res_bitshift = bits;
			goto found_it;
next_ptr:
			;
		}
	}

not_found:
	prefetch(ioc->res_map);
	ioc->res_hint = (unsigned long *) ioc->res_map;
	ioc->res_bitshift = 0;
Alex Williamson's avatar
Alex Williamson committed
613
	spin_unlock_irqrestore(&ioc->res_lock, flags);
Linus Torvalds's avatar
Linus Torvalds committed
614 615 616 617
	return (pide);

found_it:
	ioc->res_hint = res_ptr;
Alex Williamson's avatar
Alex Williamson committed
618
	spin_unlock_irqrestore(&ioc->res_lock, flags);
Linus Torvalds's avatar
Linus Torvalds committed
619 620 621 622 623 624 625 626 627 628 629 630 631
	return (pide);
}


/**
 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @size: number of bytes to create a mapping for
 *
 * Given a size, find consecutive unmarked and then mark those bits in the
 * resource bit map.
 */
static int
632
sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
Linus Torvalds's avatar
Linus Torvalds committed
633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
{
	unsigned int pages_needed = size >> iovp_shift;
#ifdef PDIR_SEARCH_TIMING
	unsigned long itc_start;
#endif
	unsigned long pide;

	ASSERT(pages_needed);
	ASSERT(0 == (size & ~iovp_mask));

#ifdef PDIR_SEARCH_TIMING
	itc_start = ia64_get_itc();
#endif
	/*
	** "seek and ye shall find"...praying never hurts either...
	*/
649
	pide = sba_search_bitmap(ioc, dev, pages_needed, 1);
Linus Torvalds's avatar
Linus Torvalds committed
650
	if (unlikely(pide >= (ioc->res_size << 3))) {
651
		pide = sba_search_bitmap(ioc, dev, pages_needed, 0);
Linus Torvalds's avatar
Linus Torvalds committed
652 653
		if (unlikely(pide >= (ioc->res_size << 3))) {
#if DELAYED_RESOURCE_CNT > 0
Alex Williamson's avatar
Alex Williamson committed
654 655
			unsigned long flags;

Linus Torvalds's avatar
Linus Torvalds committed
656 657 658 659 660
			/*
			** With delayed resource freeing, we can give this one more shot.  We're
			** getting close to being in trouble here, so do what we can to make this
			** one count.
			*/
Alex Williamson's avatar
Alex Williamson committed
661
			spin_lock_irqsave(&ioc->saved_lock, flags);
Linus Torvalds's avatar
Linus Torvalds committed
662 663 664 665
			if (ioc->saved_cnt > 0) {
				struct sba_dma_pair *d;
				int cnt = ioc->saved_cnt;

Alex Williamson's avatar
Alex Williamson committed
666
				d = &(ioc->saved[ioc->saved_cnt - 1]);
Linus Torvalds's avatar
Linus Torvalds committed
667

Alex Williamson's avatar
Alex Williamson committed
668
				spin_lock(&ioc->res_lock);
Linus Torvalds's avatar
Linus Torvalds committed
669 670 671 672 673 674 675
				while (cnt--) {
					sba_mark_invalid(ioc, d->iova, d->size);
					sba_free_range(ioc, d->iova, d->size);
					d--;
				}
				ioc->saved_cnt = 0;
				READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
Alex Williamson's avatar
Alex Williamson committed
676
				spin_unlock(&ioc->res_lock);
Linus Torvalds's avatar
Linus Torvalds committed
677
			}
Alex Williamson's avatar
Alex Williamson committed
678
			spin_unlock_irqrestore(&ioc->saved_lock, flags);
Linus Torvalds's avatar
Linus Torvalds committed
679

680
			pide = sba_search_bitmap(ioc, dev, pages_needed, 0);
681 682 683 684 685 686 687
			if (unlikely(pide >= (ioc->res_size << 3))) {
				printk(KERN_WARNING "%s: I/O MMU @ %p is"
				       "out of mapping resources, %u %u %lx\n",
				       __func__, ioc->ioc_hpa, ioc->res_size,
				       pages_needed, dma_get_seg_boundary(dev));
				return -1;
			}
Linus Torvalds's avatar
Linus Torvalds committed
688
#else
689 690 691 692 693
			printk(KERN_WARNING "%s: I/O MMU @ %p is"
			       "out of mapping resources, %u %u %lx\n",
			       __func__, ioc->ioc_hpa, ioc->res_size,
			       pages_needed, dma_get_seg_boundary(dev));
			return -1;
Linus Torvalds's avatar
Linus Torvalds committed
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
#endif
		}
	}

#ifdef PDIR_SEARCH_TIMING
	ioc->avg_search[ioc->avg_idx++] = (ia64_get_itc() - itc_start) / pages_needed;
	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
#endif

	prefetchw(&(ioc->pdir_base[pide]));

#ifdef ASSERT_PDIR_SANITY
	/* verify the first enable bit is clear */
	if(0x00 != ((u8 *) ioc->pdir_base)[pide*PDIR_ENTRY_SIZE + 7]) {
		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
	}
#endif

	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
713
		__func__, size, pages_needed, pide,
Linus Torvalds's avatar
Linus Torvalds committed
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755
		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
		ioc->res_bitshift );

	return (pide);
}


/**
 * sba_free_range - unmark bits in IO PDIR resource bitmap
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @iova: IO virtual address which was previously allocated.
 * @size: number of bytes to create a mapping for
 *
 * clear bits in the ioc's resource map
 */
static SBA_INLINE void
sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
{
	unsigned long iovp = SBA_IOVP(ioc, iova);
	unsigned int pide = PDIR_INDEX(iovp);
	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
	int bits_not_wanted = size >> iovp_shift;
	unsigned long m;

	/* Round up to power-of-two size: see AR2305 note above */
	bits_not_wanted = 1UL << get_iovp_order(bits_not_wanted << iovp_shift);
	for (; bits_not_wanted > 0 ; res_ptr++) {
		
		if (unlikely(bits_not_wanted > BITS_PER_LONG)) {

			/* these mappings start 64bit aligned */
			*res_ptr = 0UL;
			bits_not_wanted -= BITS_PER_LONG;
			pide += BITS_PER_LONG;

		} else {

			/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
			m = RESMAP_MASK(bits_not_wanted) << (pide & (BITS_PER_LONG - 1));
			bits_not_wanted = 0;

756 757
			DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", __func__, (uint) iova, size,
			        bits_not_wanted, m, pide, res_ptr, *res_ptr);
Linus Torvalds's avatar
Linus Torvalds committed
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910

			ASSERT(m != 0);
			ASSERT(bits_not_wanted);
			ASSERT((*res_ptr & m) == m); /* verify same bits are set */
			*res_ptr &= ~m;
		}
	}
}


/**************************************************************
*
*   "Dynamic DMA Mapping" support (aka "Coherent I/O")
*
***************************************************************/

/**
 * sba_io_pdir_entry - fill in one IO PDIR entry
 * @pdir_ptr:  pointer to IO PDIR entry
 * @vba: Virtual CPU address of buffer to map
 *
 * SBA Mapping Routine
 *
 * Given a virtual address (vba, arg1) sba_io_pdir_entry()
 * loads the I/O PDIR entry pointed to by pdir_ptr (arg0).
 * Each IO Pdir entry consists of 8 bytes as shown below
 * (LSB == bit 0):
 *
 *  63                    40                                 11    7        0
 * +-+---------------------+----------------------------------+----+--------+
 * |V|        U            |            PPN[39:12]            | U  |   FF   |
 * +-+---------------------+----------------------------------+----+--------+
 *
 *  V  == Valid Bit
 *  U  == Unused
 * PPN == Physical Page Number
 *
 * The physical address fields are filled with the results of virt_to_phys()
 * on the vba.
 */

#if 1
#define sba_io_pdir_entry(pdir_ptr, vba) *pdir_ptr = ((vba & ~0xE000000000000FFFULL)	\
						      | 0x8000000000000000ULL)
#else
void SBA_INLINE
sba_io_pdir_entry(u64 *pdir_ptr, unsigned long vba)
{
	*pdir_ptr = ((vba & ~0xE000000000000FFFULL) | 0x80000000000000FFULL);
}
#endif

#ifdef ENABLE_MARK_CLEAN
/**
 * Since DMA is i-cache coherent, any (complete) pages that were written via
 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
 * flush them when they get mapped into an executable vm-area.
 */
static void
mark_clean (void *addr, size_t size)
{
	unsigned long pg_addr, end;

	pg_addr = PAGE_ALIGN((unsigned long) addr);
	end = (unsigned long) addr + size;
	while (pg_addr + PAGE_SIZE <= end) {
		struct page *page = virt_to_page((void *)pg_addr);
		set_bit(PG_arch_1, &page->flags);
		pg_addr += PAGE_SIZE;
	}
}
#endif

/**
 * sba_mark_invalid - invalidate one or more IO PDIR entries
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @iova:  IO Virtual Address mapped earlier
 * @byte_cnt:  number of bytes this mapping covers.
 *
 * Marking the IO PDIR entry(ies) as Invalid and invalidate
 * corresponding IO TLB entry. The PCOM (Purge Command Register)
 * is to purge stale entries in the IO TLB when unmapping entries.
 *
 * The PCOM register supports purging of multiple pages, with a minium
 * of 1 page and a maximum of 2GB. Hardware requires the address be
 * aligned to the size of the range being purged. The size of the range
 * must be a power of 2. The "Cool perf optimization" in the
 * allocation routine helps keep that true.
 */
static SBA_INLINE void
sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
{
	u32 iovp = (u32) SBA_IOVP(ioc,iova);

	int off = PDIR_INDEX(iovp);

	/* Must be non-zero and rounded up */
	ASSERT(byte_cnt > 0);
	ASSERT(0 == (byte_cnt & ~iovp_mask));

#ifdef ASSERT_PDIR_SANITY
	/* Assert first pdir entry is set */
	if (!(ioc->pdir_base[off] >> 60)) {
		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
	}
#endif

	if (byte_cnt <= iovp_size)
	{
		ASSERT(off < ioc->pdir_size);

		iovp |= iovp_shift;     /* set "size" field for PCOM */

#ifndef FULL_VALID_PDIR
		/*
		** clear I/O PDIR entry "valid" bit
		** Do NOT clear the rest - save it for debugging.
		** We should only clear bits that have previously
		** been enabled.
		*/
		ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
#else
		/*
  		** If we want to maintain the PDIR as valid, put in
		** the spill page so devices prefetching won't
		** cause a hard fail.
		*/
		ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
#endif
	} else {
		u32 t = get_iovp_order(byte_cnt) + iovp_shift;

		iovp |= t;
		ASSERT(t <= 31);   /* 2GB! Max value of "size" field */

		do {
			/* verify this pdir entry is enabled */
			ASSERT(ioc->pdir_base[off]  >> 63);
#ifndef FULL_VALID_PDIR
			/* clear I/O Pdir entry "valid" bit first */
			ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
#else
			ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
#endif
			off++;
			byte_cnt -= iovp_size;
		} while (byte_cnt > 0);
	}

	WRITE_REG(iovp | ioc->ibase, ioc->ioc_hpa+IOC_PCOM);
}

/**
911
 * sba_map_single_attrs - map one buffer and return IOVA for DMA
Linus Torvalds's avatar
Linus Torvalds committed
912 913 914 915
 * @dev: instance of PCI owned by the driver that's asking.
 * @addr:  driver buffer to map.
 * @size:  number of bytes to map in driver buffer.
 * @dir:  R/W or both.
916
 * @attrs: optional dma attributes
Linus Torvalds's avatar
Linus Torvalds committed
917
 *
918
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
919
 */
920 921 922 923
static dma_addr_t sba_map_page(struct device *dev, struct page *page,
			       unsigned long poff, size_t size,
			       enum dma_data_direction dir,
			       struct dma_attrs *attrs)
Linus Torvalds's avatar
Linus Torvalds committed
924 925
{
	struct ioc *ioc;
926
	void *addr = page_address(page) + poff;
Linus Torvalds's avatar
Linus Torvalds committed
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947
	dma_addr_t iovp;
	dma_addr_t offset;
	u64 *pdir_start;
	int pide;
#ifdef ASSERT_PDIR_SANITY
	unsigned long flags;
#endif
#ifdef ALLOW_IOV_BYPASS
	unsigned long pci_addr = virt_to_phys(addr);
#endif

#ifdef ALLOW_IOV_BYPASS
	ASSERT(to_pci_dev(dev)->dma_mask);
	/*
 	** Check if the PCI device can DMA to ptr... if so, just return ptr
 	*/
	if (likely((pci_addr & ~to_pci_dev(dev)->dma_mask) == 0)) {
		/*
 		** Device is bit capable of DMA'ing to the buffer...
		** just return the PCI address of ptr
 		*/
948 949
		DBG_BYPASS("sba_map_single_attrs() bypass mask/addr: "
			   "0x%lx/0x%lx\n",
Linus Torvalds's avatar
Linus Torvalds committed
950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
		           to_pci_dev(dev)->dma_mask, pci_addr);
		return pci_addr;
	}
#endif
	ioc = GET_IOC(dev);
	ASSERT(ioc);

	prefetch(ioc->res_hint);

	ASSERT(size > 0);
	ASSERT(size <= DMA_CHUNK_SIZE);

	/* save offset bits */
	offset = ((dma_addr_t) (long) addr) & ~iovp_mask;

	/* round up to nearest iovp_size */
	size = (size + offset + ~iovp_mask) & iovp_mask;

#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
970
	if (sba_check_pdir(ioc,"Check before sba_map_single_attrs()"))
Linus Torvalds's avatar
Linus Torvalds committed
971 972 973 974
		panic("Sanity check failed");
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

975
	pide = sba_alloc_range(ioc, dev, size);
976 977
	if (pide < 0)
		return 0;
Linus Torvalds's avatar
Linus Torvalds committed
978 979 980

	iovp = (dma_addr_t) pide << iovp_shift;

981
	DBG_RUN("%s() 0x%p -> 0x%lx\n", __func__, addr, (long) iovp | offset);
Linus Torvalds's avatar
Linus Torvalds committed
982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000

	pdir_start = &(ioc->pdir_base[pide]);

	while (size > 0) {
		ASSERT(((u8 *)pdir_start)[7] == 0); /* verify availability */
		sba_io_pdir_entry(pdir_start, (unsigned long) addr);

		DBG_RUN("     pdir 0x%p %lx\n", pdir_start, *pdir_start);

		addr += iovp_size;
		size -= iovp_size;
		pdir_start++;
	}
	/* force pdir update */
	wmb();

	/* form complete address */
#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
1001
	sba_check_pdir(ioc,"Check after sba_map_single_attrs()");
Linus Torvalds's avatar
Linus Torvalds committed
1002 1003 1004 1005 1006
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif
	return SBA_IOVA(ioc, iovp, offset);
}

1007 1008 1009 1010 1011 1012 1013 1014
static dma_addr_t sba_map_single_attrs(struct device *dev, void *addr,
				       size_t size, enum dma_data_direction dir,
				       struct dma_attrs *attrs)
{
	return sba_map_page(dev, virt_to_page(addr),
			    (unsigned long)addr & ~PAGE_MASK, size, dir, attrs);
}

Alex Williamson's avatar
Alex Williamson committed
1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
#ifdef ENABLE_MARK_CLEAN
static SBA_INLINE void
sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size)
{
	u32	iovp = (u32) SBA_IOVP(ioc,iova);
	int	off = PDIR_INDEX(iovp);
	void	*addr;

	if (size <= iovp_size) {
		addr = phys_to_virt(ioc->pdir_base[off] &
		                    ~0xE000000000000FFFULL);
		mark_clean(addr, size);
	} else {
		do {
			addr = phys_to_virt(ioc->pdir_base[off] &
			                    ~0xE000000000000FFFULL);
			mark_clean(addr, min(size, iovp_size));
			off++;
			size -= iovp_size;
		} while (size > 0);
	}
}
#endif

Linus Torvalds's avatar
Linus Torvalds committed
1039
/**
1040
 * sba_unmap_single_attrs - unmap one IOVA and free resources
Linus Torvalds's avatar
Linus Torvalds committed
1041 1042 1043 1044
 * @dev: instance of PCI owned by the driver that's asking.
 * @iova:  IOVA of driver buffer previously mapped.
 * @size:  number of bytes mapped in driver buffer.
 * @dir:  R/W or both.
1045
 * @attrs: optional dma attributes
Linus Torvalds's avatar
Linus Torvalds committed
1046
 *
1047
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
1048
 */
1049 1050
static void sba_unmap_page(struct device *dev, dma_addr_t iova, size_t size,
			   enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds's avatar
Linus Torvalds committed
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
{
	struct ioc *ioc;
#if DELAYED_RESOURCE_CNT > 0
	struct sba_dma_pair *d;
#endif
	unsigned long flags;
	dma_addr_t offset;

	ioc = GET_IOC(dev);
	ASSERT(ioc);

#ifdef ALLOW_IOV_BYPASS
	if (likely((iova & ioc->imask) != ioc->ibase)) {
		/*
		** Address does not fall w/in IOVA, must be bypassing
		*/
1067 1068
		DBG_BYPASS("sba_unmap_single_atttrs() bypass addr: 0x%lx\n",
			   iova);
Linus Torvalds's avatar
Linus Torvalds committed
1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079

#ifdef ENABLE_MARK_CLEAN
		if (dir == DMA_FROM_DEVICE) {
			mark_clean(phys_to_virt(iova), size);
		}
#endif
		return;
	}
#endif
	offset = iova & ~iovp_mask;

1080
	DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size);
Linus Torvalds's avatar
Linus Torvalds committed
1081 1082 1083 1084 1085

	iova ^= offset;        /* clear offset bits */
	size += offset;
	size = ROUNDUP(size, iovp_size);

Alex Williamson's avatar
Alex Williamson committed
1086 1087 1088 1089
#ifdef ENABLE_MARK_CLEAN
	if (dir == DMA_FROM_DEVICE)
		sba_mark_clean(ioc, iova, size);
#endif
Linus Torvalds's avatar
Linus Torvalds committed
1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117

#if DELAYED_RESOURCE_CNT > 0
	spin_lock_irqsave(&ioc->saved_lock, flags);
	d = &(ioc->saved[ioc->saved_cnt]);
	d->iova = iova;
	d->size = size;
	if (unlikely(++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT)) {
		int cnt = ioc->saved_cnt;
		spin_lock(&ioc->res_lock);
		while (cnt--) {
			sba_mark_invalid(ioc, d->iova, d->size);
			sba_free_range(ioc, d->iova, d->size);
			d--;
		}
		ioc->saved_cnt = 0;
		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
		spin_unlock(&ioc->res_lock);
	}
	spin_unlock_irqrestore(&ioc->saved_lock, flags);
#else /* DELAYED_RESOURCE_CNT == 0 */
	spin_lock_irqsave(&ioc->res_lock, flags);
	sba_mark_invalid(ioc, iova, size);
	sba_free_range(ioc, iova, size);
	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif /* DELAYED_RESOURCE_CNT == 0 */
}

1118 1119 1120 1121 1122 1123
void sba_unmap_single_attrs(struct device *dev, dma_addr_t iova, size_t size,
			    enum dma_data_direction dir, struct dma_attrs *attrs)
{
	sba_unmap_page(dev, iova, size, dir, attrs);
}

Linus Torvalds's avatar
Linus Torvalds committed
1124 1125 1126 1127 1128 1129
/**
 * sba_alloc_coherent - allocate/map shared mem for DMA
 * @dev: instance of PCI owned by the driver that's asking.
 * @size:  number of bytes mapped in driver buffer.
 * @dma_handle:  IOVA of new buffer.
 *
1130
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
1131
 */
1132
static void *
Al Viro's avatar
Al Viro committed
1133
sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds's avatar
Linus Torvalds committed
1134 1135 1136 1137 1138 1139 1140 1141 1142 1143
{
	struct ioc *ioc;
	void *addr;

	ioc = GET_IOC(dev);
	ASSERT(ioc);

#ifdef CONFIG_NUMA
	{
		struct page *page;
1144
		page = alloc_pages_exact_node(ioc->node == MAX_NUMNODES ?
Linus Torvalds's avatar
Linus Torvalds committed
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178
		                        numa_node_id() : ioc->node, flags,
		                        get_order(size));

		if (unlikely(!page))
			return NULL;

		addr = page_address(page);
	}
#else
	addr = (void *) __get_free_pages(flags, get_order(size));
#endif
	if (unlikely(!addr))
		return NULL;

	memset(addr, 0, size);
	*dma_handle = virt_to_phys(addr);

#ifdef ALLOW_IOV_BYPASS
	ASSERT(dev->coherent_dma_mask);
	/*
 	** Check if the PCI device can DMA to ptr... if so, just return ptr
 	*/
	if (likely((*dma_handle & ~dev->coherent_dma_mask) == 0)) {
		DBG_BYPASS("sba_alloc_coherent() bypass mask/addr: 0x%lx/0x%lx\n",
		           dev->coherent_dma_mask, *dma_handle);

		return addr;
	}
#endif

	/*
	 * If device can't bypass or bypass is disabled, pass the 32bit fake
	 * device to map single to get an iova mapping.
	 */
1179 1180
	*dma_handle = sba_map_single_attrs(&ioc->sac_only_dev->dev, addr,
					   size, 0, NULL);
Linus Torvalds's avatar
Linus Torvalds committed
1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192

	return addr;
}


/**
 * sba_free_coherent - free/unmap shared mem for DMA
 * @dev: instance of PCI owned by the driver that's asking.
 * @size:  number of bytes mapped in driver buffer.
 * @vaddr:  virtual address IOVA of "consistent" buffer.
 * @dma_handler:  IO virtual address of "consistent" buffer.
 *
1193
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
1194
 */
1195 1196
static void sba_free_coherent (struct device *dev, size_t size, void *vaddr,
			       dma_addr_t dma_handle)
Linus Torvalds's avatar
Linus Torvalds committed
1197
{
1198
	sba_unmap_single_attrs(dev, dma_handle, size, 0, NULL);
Linus Torvalds's avatar
Linus Torvalds committed
1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
	free_pages((unsigned long) vaddr, get_order(size));
}


/*
** Since 0 is a valid pdir_base index value, can't use that
** to determine if a value is valid or not. Use a flag to indicate
** the SG list entry contains a valid pdir index.
*/
#define PIDE_FLAG 0x1UL

#ifdef DEBUG_LARGE_SG_ENTRIES
int dump_run_sg = 0;
#endif


/**
 * sba_fill_pdir - write allocated SG entries into IO PDIR
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @startsg:  list of IOVA/size pairs
 * @nents: number of entries in startsg list
 *
 * Take preprocessed SG list and write corresponding entries
 * in the IO PDIR.
 */

static SBA_INLINE int
sba_fill_pdir(
	struct ioc *ioc,
	struct scatterlist *startsg,
	int nents)
{
	struct scatterlist *dma_sg = startsg;	/* pointer to current DMA */
	int n_mappings = 0;
	u64 *pdirp = NULL;
	unsigned long dma_offset = 0;

	while (nents-- > 0) {
		int     cnt = startsg->dma_length;
		startsg->dma_length = 0;

#ifdef DEBUG_LARGE_SG_ENTRIES
		if (dump_run_sg)
			printk(" %2d : %08lx/%05x %p\n",
				nents, startsg->dma_address, cnt,
				sba_sg_address(startsg));
#else
		DBG_RUN_SG(" %d : %08lx/%05x %p\n",
				nents, startsg->dma_address, cnt,
				sba_sg_address(startsg));
#endif
		/*
		** Look for the start of a new DMA stream
		*/
		if (startsg->dma_address & PIDE_FLAG) {
			u32 pide = startsg->dma_address & ~PIDE_FLAG;
			dma_offset = (unsigned long) pide & ~iovp_mask;
			startsg->dma_address = 0;
1257 1258
			if (n_mappings)
				dma_sg = sg_next(dma_sg);
Linus Torvalds's avatar
Linus Torvalds committed
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284
			dma_sg->dma_address = pide | ioc->ibase;
			pdirp = &(ioc->pdir_base[pide >> iovp_shift]);
			n_mappings++;
		}

		/*
		** Look for a VCONTIG chunk
		*/
		if (cnt) {
			unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
			ASSERT(pdirp);

			/* Since multiple Vcontig blocks could make up
			** one DMA stream, *add* cnt to dma_len.
			*/
			dma_sg->dma_length += cnt;
			cnt += dma_offset;
			dma_offset=0;	/* only want offset on first chunk */
			cnt = ROUNDUP(cnt, iovp_size);
			do {
				sba_io_pdir_entry(pdirp, vaddr);
				vaddr += iovp_size;
				cnt -= iovp_size;
				pdirp++;
			} while (cnt > 0);
		}
Jens Axboe's avatar
Jens Axboe committed
1285
		startsg = sg_next(startsg);
Linus Torvalds's avatar
Linus Torvalds committed
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321
	}
	/* force pdir update */
	wmb();

#ifdef DEBUG_LARGE_SG_ENTRIES
	dump_run_sg = 0;
#endif
	return(n_mappings);
}


/*
** Two address ranges are DMA contiguous *iff* "end of prev" and
** "start of next" are both on an IOV page boundary.
**
** (shift left is a quick trick to mask off upper bits)
*/
#define DMA_CONTIG(__X, __Y) \
	(((((unsigned long) __X) | ((unsigned long) __Y)) << (BITS_PER_LONG - iovp_shift)) == 0UL)


/**
 * sba_coalesce_chunks - preprocess the SG list
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @startsg:  list of IOVA/size pairs
 * @nents: number of entries in startsg list
 *
 * First pass is to walk the SG list and determine where the breaks are
 * in the DMA stream. Allocates PDIR entries but does not fill them.
 * Returns the number of DMA chunks.
 *
 * Doing the fill separate from the coalescing/allocation keeps the
 * code simpler. Future enhancement could make one pass through
 * the sglist do both.
 */
static SBA_INLINE int
1322
sba_coalesce_chunks(struct ioc *ioc, struct device *dev,
Linus Torvalds's avatar
Linus Torvalds committed
1323 1324 1325 1326 1327 1328 1329 1330 1331
	struct scatterlist *startsg,
	int nents)
{
	struct scatterlist *vcontig_sg;    /* VCONTIG chunk head */
	unsigned long vcontig_len;         /* len of VCONTIG chunk */
	unsigned long vcontig_end;
	struct scatterlist *dma_sg;        /* next DMA stream head */
	unsigned long dma_offset, dma_len; /* start/len of DMA stream */
	int n_mappings = 0;
1332
	unsigned int max_seg_size = dma_get_max_seg_size(dev);
1333
	int idx;
Linus Torvalds's avatar
Linus Torvalds committed
1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355

	while (nents > 0) {
		unsigned long vaddr = (unsigned long) sba_sg_address(startsg);

		/*
		** Prepare for first/next DMA stream
		*/
		dma_sg = vcontig_sg = startsg;
		dma_len = vcontig_len = vcontig_end = startsg->length;
		vcontig_end +=  vaddr;
		dma_offset = vaddr & ~iovp_mask;

		/* PARANOID: clear entries */
		startsg->dma_address = startsg->dma_length = 0;

		/*
		** This loop terminates one iteration "early" since
		** it's always looking one "ahead".
		*/
		while (--nents > 0) {
			unsigned long vaddr;	/* tmp */

Jens Axboe's avatar
Jens Axboe committed
1356
			startsg = sg_next(startsg);
Linus Torvalds's avatar
Linus Torvalds committed
1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372

			/* PARANOID */
			startsg->dma_address = startsg->dma_length = 0;

			/* catch brokenness in SCSI layer */
			ASSERT(startsg->length <= DMA_CHUNK_SIZE);

			/*
			** First make sure current dma stream won't
			** exceed DMA_CHUNK_SIZE if we coalesce the
			** next entry.
			*/
			if (((dma_len + dma_offset + startsg->length + ~iovp_mask) & iovp_mask)
			    > DMA_CHUNK_SIZE)
				break;

1373 1374 1375
			if (dma_len + startsg->length > max_seg_size)
				break;

Linus Torvalds's avatar
Linus Torvalds committed
1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394
			/*
			** Then look for virtually contiguous blocks.
			**
			** append the next transaction?
			*/
			vaddr = (unsigned long) sba_sg_address(startsg);
			if  (vcontig_end == vaddr)
			{
				vcontig_len += startsg->length;
				vcontig_end += startsg->length;
				dma_len     += startsg->length;
				continue;
			}

#ifdef DEBUG_LARGE_SG_ENTRIES
			dump_run_sg = (vcontig_len > iovp_size);
#endif

			/*
1395
			** Not virtually contiguous.
Linus Torvalds's avatar
Linus Torvalds committed
1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431
			** Terminate prev chunk.
			** Start a new chunk.
			**
			** Once we start a new VCONTIG chunk, dma_offset
			** can't change. And we need the offset from the first
			** chunk - not the last one. Ergo Successive chunks
			** must start on page boundaries and dove tail
			** with it's predecessor.
			*/
			vcontig_sg->dma_length = vcontig_len;

			vcontig_sg = startsg;
			vcontig_len = startsg->length;

			/*
			** 3) do the entries end/start on page boundaries?
			**    Don't update vcontig_end until we've checked.
			*/
			if (DMA_CONTIG(vcontig_end, vaddr))
			{
				vcontig_end = vcontig_len + vaddr;
				dma_len += vcontig_len;
				continue;
			} else {
				break;
			}
		}

		/*
		** End of DMA Stream
		** Terminate last VCONTIG block.
		** Allocate space for DMA stream.
		*/
		vcontig_sg->dma_length = vcontig_len;
		dma_len = (dma_len + dma_offset + ~iovp_mask) & iovp_mask;
		ASSERT(dma_len <= DMA_CHUNK_SIZE);
1432 1433 1434 1435 1436 1437 1438
		idx = sba_alloc_range(ioc, dev, dma_len);
		if (idx < 0) {
			dma_sg->dma_length = 0;
			return -1;
		}
		dma_sg->dma_address = (dma_addr_t)(PIDE_FLAG | (idx << iovp_shift)
						   | dma_offset);
Linus Torvalds's avatar
Linus Torvalds committed
1439 1440 1441 1442 1443 1444
		n_mappings++;
	}

	return n_mappings;
}

1445 1446 1447
static void sba_unmap_sg_attrs(struct device *dev, struct scatterlist *sglist,
			       int nents, enum dma_data_direction dir,
			       struct dma_attrs *attrs);
Linus Torvalds's avatar
Linus Torvalds committed
1448 1449 1450 1451 1452 1453
/**
 * sba_map_sg - map Scatter/Gather list
 * @dev: instance of PCI owned by the driver that's asking.
 * @sglist:  array of buffer/length pairs
 * @nents:  number of entries in list
 * @dir:  R/W or both.
1454
 * @attrs: optional dma attributes
Linus Torvalds's avatar
Linus Torvalds committed
1455
 *
1456
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
1457
 */
1458
static int sba_map_sg_attrs(struct device *dev, struct scatterlist *sglist,
1459 1460
			    int nents, enum dma_data_direction dir,
			    struct dma_attrs *attrs)
Linus Torvalds's avatar
Linus Torvalds committed
1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
{
	struct ioc *ioc;
	int coalesced, filled = 0;
#ifdef ASSERT_PDIR_SANITY
	unsigned long flags;
#endif
#ifdef ALLOW_IOV_BYPASS_SG
	struct scatterlist *sg;
#endif

1471
	DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
Linus Torvalds's avatar
Linus Torvalds committed
1472 1473 1474 1475 1476 1477
	ioc = GET_IOC(dev);
	ASSERT(ioc);

#ifdef ALLOW_IOV_BYPASS_SG
	ASSERT(to_pci_dev(dev)->dma_mask);
	if (likely((ioc->dma_mask & ~to_pci_dev(dev)->dma_mask) == 0)) {
Jens Axboe's avatar
Jens Axboe committed
1478
		for_each_sg(sglist, sg, nents, filled) {
Linus Torvalds's avatar
Linus Torvalds committed
1479 1480 1481 1482 1483 1484 1485 1486 1487
			sg->dma_length = sg->length;
			sg->dma_address = virt_to_phys(sba_sg_address(sg));
		}
		return filled;
	}
#endif
	/* Fast path single entry scatterlists. */
	if (nents == 1) {
		sglist->dma_length = sglist->length;
1488
		sglist->dma_address = sba_map_single_attrs(dev, sba_sg_address(sglist), sglist->length, dir, attrs);
Linus Torvalds's avatar
Linus Torvalds committed
1489 1490 1491 1492 1493
		return 1;
	}

#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
1494
	if (sba_check_pdir(ioc,"Check before sba_map_sg_attrs()"))
Linus Torvalds's avatar
Linus Torvalds committed
1495 1496
	{
		sba_dump_sg(ioc, sglist, nents);
1497
		panic("Check before sba_map_sg_attrs()");
Linus Torvalds's avatar
Linus Torvalds committed
1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511
	}
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

	prefetch(ioc->res_hint);

	/*
	** First coalesce the chunks and allocate I/O pdir space
	**
	** If this is one DMA stream, we can properly map using the
	** correct virtual address associated with each DMA page.
	** w/o this association, we wouldn't have coherent DMA!
	** Access to the virtual address is what forces a two pass algorithm.
	*/
1512
	coalesced = sba_coalesce_chunks(ioc, dev, sglist, nents);
1513 1514 1515 1516
	if (coalesced < 0) {
		sba_unmap_sg_attrs(dev, sglist, nents, dir, attrs);
		return 0;
	}
Linus Torvalds's avatar
Linus Torvalds committed
1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529

	/*
	** Program the I/O Pdir
	**
	** map the virtual addresses to the I/O Pdir
	** o dma_address will contain the pdir index
	** o dma_len will contain the number of bytes to map
	** o address contains the virtual address.
	*/
	filled = sba_fill_pdir(ioc, sglist, nents);

#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
1530
	if (sba_check_pdir(ioc,"Check after sba_map_sg_attrs()"))
Linus Torvalds's avatar
Linus Torvalds committed
1531 1532
	{
		sba_dump_sg(ioc, sglist, nents);
1533
		panic("Check after sba_map_sg_attrs()\n");
Linus Torvalds's avatar
Linus Torvalds committed
1534 1535 1536 1537 1538
	}
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

	ASSERT(coalesced == filled);
1539
	DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
Linus Torvalds's avatar
Linus Torvalds committed
1540 1541 1542 1543 1544

	return filled;
}

/**
1545
 * sba_unmap_sg_attrs - unmap Scatter/Gather list
Linus Torvalds's avatar
Linus Torvalds committed
1546 1547 1548 1549
 * @dev: instance of PCI owned by the driver that's asking.
 * @sglist:  array of buffer/length pairs
 * @nents:  number of entries in list
 * @dir:  R/W or both.
1550
 * @attrs: optional dma attributes
Linus Torvalds's avatar
Linus Torvalds committed
1551
 *
1552
 * See Documentation/PCI/PCI-DMA-mapping.txt
Linus Torvalds's avatar
Linus Torvalds committed
1553
 */
1554
static void sba_unmap_sg_attrs(struct device *dev, struct scatterlist *sglist,
1555 1556
			       int nents, enum dma_data_direction dir,
			       struct dma_attrs *attrs)
Linus Torvalds's avatar
Linus Torvalds committed
1557 1558 1559 1560 1561 1562 1563
{
#ifdef ASSERT_PDIR_SANITY
	struct ioc *ioc;
	unsigned long flags;
#endif

	DBG_RUN_SG("%s() START %d entries,  %p,%x\n",
1564
		   __func__, nents, sba_sg_address(sglist), sglist->length);
Linus Torvalds's avatar
Linus Torvalds committed
1565 1566 1567 1568 1569 1570

#ifdef ASSERT_PDIR_SANITY
	ioc = GET_IOC(dev);
	ASSERT(ioc);

	spin_lock_irqsave(&ioc->res_lock, flags);
1571
	sba_check_pdir(ioc,"Check before sba_unmap_sg_attrs()");
Linus Torvalds's avatar
Linus Torvalds committed
1572 1573 1574 1575 1576
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

	while (nents && sglist->dma_length) {

1577 1578
		sba_unmap_single_attrs(dev, sglist->dma_address,
				       sglist->dma_length, dir, attrs);
Jens Axboe's avatar
Jens Axboe committed
1579
		sglist = sg_next(sglist);
Linus Torvalds's avatar
Linus Torvalds committed
1580 1581 1582
		nents--;
	}

1583
	DBG_RUN_SG("%s() DONE (nents %d)\n", __func__,  nents);
Linus Torvalds's avatar
Linus Torvalds committed
1584 1585 1586

#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
1587
	sba_check_pdir(ioc,"Check after sba_unmap_sg_attrs()");
Linus Torvalds's avatar
Linus Torvalds committed
1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

}

/**************************************************************
*
*   Initialization and claim
*
***************************************************************/

static void __init
ioc_iova_init(struct ioc *ioc)
{
	int tcnfg;
	int agp_found = 0;
	struct pci_dev *device = NULL;
#ifdef FULL_VALID_PDIR
	unsigned long index;
#endif

	/*
	** Firmware programs the base and size of a "safe IOVA space"
	** (one that doesn't overlap memory or LMMIO space) in the
	** IBASE and IMASK registers.
	*/
	ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1UL;
	ioc->imask = READ_REG(ioc->ioc_hpa + IOC_IMASK) | 0xFFFFFFFF00000000UL;

	ioc->iov_size = ~ioc->imask + 1;

	DBG_INIT("%s() hpa %p IOV base 0x%lx mask 0x%lx (%dMB)\n",
1620
		__func__, ioc->ioc_hpa, ioc->ibase, ioc->imask,
Linus Torvalds's avatar
Linus Torvalds committed
1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
		ioc->iov_size >> 20);

	switch (iovp_size) {
		case  4*1024: tcnfg = 0; break;
		case  8*1024: tcnfg = 1; break;
		case 16*1024: tcnfg = 2; break;
		case 64*1024: tcnfg = 3; break;
		default:
			panic(PFX "Unsupported IOTLB page size %ldK",
				iovp_size >> 10);
			break;
	}
	WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);

	ioc->pdir_size = (ioc->iov_size / iovp_size) * PDIR_ENTRY_SIZE;
	ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
						   get_order(ioc->pdir_size));
	if (!ioc->pdir_base)
		panic(PFX "Couldn't allocate I/O Page Table\n");

	memset(ioc->pdir_base, 0, ioc->pdir_size);

1643
	DBG_INIT("%s() IOV page size %ldK pdir %p size %x\n", __func__,
Linus Torvalds's avatar
Linus Torvalds committed
1644 1645 1646 1647 1648 1649 1650