xref: /openbmc/linux/arch/ia64/hp/common/sba_iommu.c (revision c21b37f6)
1 /*
2 **  IA64 System Bus Adapter (SBA) I/O MMU manager
3 **
4 **	(c) Copyright 2002-2005 Alex Williamson
5 **	(c) Copyright 2002-2003 Grant Grundler
6 **	(c) Copyright 2002-2005 Hewlett-Packard Company
7 **
8 **	Portions (c) 2000 Grant Grundler (from parisc I/O MMU code)
9 **	Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
10 **
11 **	This program is free software; you can redistribute it and/or modify
12 **	it under the terms of the GNU General Public License as published by
13 **      the Free Software Foundation; either version 2 of the License, or
14 **      (at your option) any later version.
15 **
16 **
17 ** This module initializes the IOC (I/O Controller) found on HP
18 ** McKinley machines and their successors.
19 **
20 */
21 
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/spinlock.h>
26 #include <linux/slab.h>
27 #include <linux/init.h>
28 #include <linux/mm.h>
29 #include <linux/string.h>
30 #include <linux/pci.h>
31 #include <linux/proc_fs.h>
32 #include <linux/seq_file.h>
33 #include <linux/acpi.h>
34 #include <linux/efi.h>
35 #include <linux/nodemask.h>
36 #include <linux/bitops.h>         /* hweight64() */
37 #include <linux/crash_dump.h>
38 
39 #include <asm/delay.h>		/* ia64_get_itc() */
40 #include <asm/io.h>
41 #include <asm/page.h>		/* PAGE_OFFSET */
42 #include <asm/dma.h>
43 #include <asm/system.h>		/* wmb() */
44 
45 #include <asm/acpi-ext.h>
46 
47 extern int swiotlb_late_init_with_default_size (size_t size);
48 
49 #define PFX "IOC: "
50 
51 /*
52 ** Enabling timing search of the pdir resource map.  Output in /proc.
53 ** Disabled by default to optimize performance.
54 */
55 #undef PDIR_SEARCH_TIMING
56 
57 /*
58 ** This option allows cards capable of 64bit DMA to bypass the IOMMU.  If
59 ** not defined, all DMA will be 32bit and go through the TLB.
60 ** There's potentially a conflict in the bio merge code with us
61 ** advertising an iommu, but then bypassing it.  Since I/O MMU bypassing
62 ** appears to give more performance than bio-level virtual merging, we'll
63 ** do the former for now.  NOTE: BYPASS_SG also needs to be undef'd to
64 ** completely restrict DMA to the IOMMU.
65 */
66 #define ALLOW_IOV_BYPASS
67 
68 /*
69 ** This option specifically allows/disallows bypassing scatterlists with
70 ** multiple entries.  Coalescing these entries can allow better DMA streaming
71 ** and in some cases shows better performance than entirely bypassing the
72 ** IOMMU.  Performance increase on the order of 1-2% sequential output/input
73 ** using bonnie++ on a RAID0 MD device (sym2 & mpt).
74 */
75 #undef ALLOW_IOV_BYPASS_SG
76 
77 /*
78 ** If a device prefetches beyond the end of a valid pdir entry, it will cause
79 ** a hard failure, ie. MCA.  Version 3.0 and later of the zx1 LBA should
80 ** disconnect on 4k boundaries and prevent such issues.  If the device is
81 ** particularly aggressive, this option will keep the entire pdir valid such
82 ** that prefetching will hit a valid address.  This could severely impact
83 ** error containment, and is therefore off by default.  The page that is
84 ** used for spill-over is poisoned, so that should help debugging somewhat.
85 */
86 #undef FULL_VALID_PDIR
87 
88 #define ENABLE_MARK_CLEAN
89 
90 /*
91 ** The number of debug flags is a clue - this code is fragile.  NOTE: since
92 ** tightening the use of res_lock the resource bitmap and actual pdir are no
93 ** longer guaranteed to stay in sync.  The sanity checking code isn't going to
94 ** like that.
95 */
96 #undef DEBUG_SBA_INIT
97 #undef DEBUG_SBA_RUN
98 #undef DEBUG_SBA_RUN_SG
99 #undef DEBUG_SBA_RESOURCE
100 #undef ASSERT_PDIR_SANITY
101 #undef DEBUG_LARGE_SG_ENTRIES
102 #undef DEBUG_BYPASS
103 
104 #if defined(FULL_VALID_PDIR) && defined(ASSERT_PDIR_SANITY)
105 #error FULL_VALID_PDIR and ASSERT_PDIR_SANITY are mutually exclusive
106 #endif
107 
108 #define SBA_INLINE	__inline__
109 /* #define SBA_INLINE */
110 
111 #ifdef DEBUG_SBA_INIT
112 #define DBG_INIT(x...)	printk(x)
113 #else
114 #define DBG_INIT(x...)
115 #endif
116 
117 #ifdef DEBUG_SBA_RUN
118 #define DBG_RUN(x...)	printk(x)
119 #else
120 #define DBG_RUN(x...)
121 #endif
122 
123 #ifdef DEBUG_SBA_RUN_SG
124 #define DBG_RUN_SG(x...)	printk(x)
125 #else
126 #define DBG_RUN_SG(x...)
127 #endif
128 
129 
130 #ifdef DEBUG_SBA_RESOURCE
131 #define DBG_RES(x...)	printk(x)
132 #else
133 #define DBG_RES(x...)
134 #endif
135 
136 #ifdef DEBUG_BYPASS
137 #define DBG_BYPASS(x...)	printk(x)
138 #else
139 #define DBG_BYPASS(x...)
140 #endif
141 
142 #ifdef ASSERT_PDIR_SANITY
143 #define ASSERT(expr) \
144         if(!(expr)) { \
145                 printk( "\n" __FILE__ ":%d: Assertion " #expr " failed!\n",__LINE__); \
146                 panic(#expr); \
147         }
148 #else
149 #define ASSERT(expr)
150 #endif
151 
152 /*
153 ** The number of pdir entries to "free" before issuing
154 ** a read to PCOM register to flush out PCOM writes.
155 ** Interacts with allocation granularity (ie 4 or 8 entries
156 ** allocated and free'd/purged at a time might make this
157 ** less interesting).
158 */
159 #define DELAYED_RESOURCE_CNT	64
160 
161 #define PCI_DEVICE_ID_HP_SX2000_IOC	0x12ec
162 
163 #define ZX1_IOC_ID	((PCI_DEVICE_ID_HP_ZX1_IOC << 16) | PCI_VENDOR_ID_HP)
164 #define ZX2_IOC_ID	((PCI_DEVICE_ID_HP_ZX2_IOC << 16) | PCI_VENDOR_ID_HP)
165 #define REO_IOC_ID	((PCI_DEVICE_ID_HP_REO_IOC << 16) | PCI_VENDOR_ID_HP)
166 #define SX1000_IOC_ID	((PCI_DEVICE_ID_HP_SX1000_IOC << 16) | PCI_VENDOR_ID_HP)
167 #define SX2000_IOC_ID	((PCI_DEVICE_ID_HP_SX2000_IOC << 16) | PCI_VENDOR_ID_HP)
168 
169 #define ZX1_IOC_OFFSET	0x1000	/* ACPI reports SBA, we want IOC */
170 
171 #define IOC_FUNC_ID	0x000
172 #define IOC_FCLASS	0x008	/* function class, bist, header, rev... */
173 #define IOC_IBASE	0x300	/* IO TLB */
174 #define IOC_IMASK	0x308
175 #define IOC_PCOM	0x310
176 #define IOC_TCNFG	0x318
177 #define IOC_PDIR_BASE	0x320
178 
179 #define IOC_ROPE0_CFG	0x500
180 #define   IOC_ROPE_AO	  0x10	/* Allow "Relaxed Ordering" */
181 
182 
183 /* AGP GART driver looks for this */
184 #define ZX1_SBA_IOMMU_COOKIE	0x0000badbadc0ffeeUL
185 
186 /*
187 ** The zx1 IOC supports 4/8/16/64KB page sizes (see TCNFG register)
188 **
189 ** Some IOCs (sx1000) can run at the above pages sizes, but are
190 ** really only supported using the IOC at a 4k page size.
191 **
192 ** iovp_size could only be greater than PAGE_SIZE if we are
193 ** confident the drivers really only touch the next physical
194 ** page iff that driver instance owns it.
195 */
196 static unsigned long iovp_size;
197 static unsigned long iovp_shift;
198 static unsigned long iovp_mask;
199 
200 struct ioc {
201 	void __iomem	*ioc_hpa;	/* I/O MMU base address */
202 	char		*res_map;	/* resource map, bit == pdir entry */
203 	u64		*pdir_base;	/* physical base address */
204 	unsigned long	ibase;		/* pdir IOV Space base */
205 	unsigned long	imask;		/* pdir IOV Space mask */
206 
207 	unsigned long	*res_hint;	/* next avail IOVP - circular search */
208 	unsigned long	dma_mask;
209 	spinlock_t	res_lock;	/* protects the resource bitmap, but must be held when */
210 					/* clearing pdir to prevent races with allocations. */
211 	unsigned int	res_bitshift;	/* from the RIGHT! */
212 	unsigned int	res_size;	/* size of resource map in bytes */
213 #ifdef CONFIG_NUMA
214 	unsigned int	node;		/* node where this IOC lives */
215 #endif
216 #if DELAYED_RESOURCE_CNT > 0
217 	spinlock_t	saved_lock;	/* may want to try to get this on a separate cacheline */
218 					/* than res_lock for bigger systems. */
219 	int		saved_cnt;
220 	struct sba_dma_pair {
221 		dma_addr_t	iova;
222 		size_t		size;
223 	} saved[DELAYED_RESOURCE_CNT];
224 #endif
225 
226 #ifdef PDIR_SEARCH_TIMING
227 #define SBA_SEARCH_SAMPLE	0x100
228 	unsigned long avg_search[SBA_SEARCH_SAMPLE];
229 	unsigned long avg_idx;	/* current index into avg_search */
230 #endif
231 
232 	/* Stuff we don't need in performance path */
233 	struct ioc	*next;		/* list of IOC's in system */
234 	acpi_handle	handle;		/* for multiple IOC's */
235 	const char 	*name;
236 	unsigned int	func_id;
237 	unsigned int	rev;		/* HW revision of chip */
238 	u32		iov_size;
239 	unsigned int	pdir_size;	/* in bytes, determined by IOV Space size */
240 	struct pci_dev	*sac_only_dev;
241 };
242 
243 static struct ioc *ioc_list;
244 static int reserve_sba_gart = 1;
245 
246 static SBA_INLINE void sba_mark_invalid(struct ioc *, dma_addr_t, size_t);
247 static SBA_INLINE void sba_free_range(struct ioc *, dma_addr_t, size_t);
248 
249 #define sba_sg_address(sg)	(page_address((sg)->page) + (sg)->offset)
250 
251 #ifdef FULL_VALID_PDIR
252 static u64 prefetch_spill_page;
253 #endif
254 
255 #ifdef CONFIG_PCI
256 # define GET_IOC(dev)	(((dev)->bus == &pci_bus_type)						\
257 			 ? ((struct ioc *) PCI_CONTROLLER(to_pci_dev(dev))->iommu) : NULL)
258 #else
259 # define GET_IOC(dev)	NULL
260 #endif
261 
262 /*
263 ** DMA_CHUNK_SIZE is used by the SCSI mid-layer to break up
264 ** (or rather not merge) DMAs into manageable chunks.
265 ** On parisc, this is more of the software/tuning constraint
266 ** rather than the HW. I/O MMU allocation algorithms can be
267 ** faster with smaller sizes (to some degree).
268 */
269 #define DMA_CHUNK_SIZE  (BITS_PER_LONG*iovp_size)
270 
271 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
272 
273 /************************************
274 ** SBA register read and write support
275 **
276 ** BE WARNED: register writes are posted.
277 **  (ie follow writes which must reach HW with a read)
278 **
279 */
280 #define READ_REG(addr)       __raw_readq(addr)
281 #define WRITE_REG(val, addr) __raw_writeq(val, addr)
282 
283 #ifdef DEBUG_SBA_INIT
284 
285 /**
286  * sba_dump_tlb - debugging only - print IOMMU operating parameters
287  * @hpa: base address of the IOMMU
288  *
289  * Print the size/location of the IO MMU PDIR.
290  */
291 static void
292 sba_dump_tlb(char *hpa)
293 {
294 	DBG_INIT("IO TLB at 0x%p\n", (void *)hpa);
295 	DBG_INIT("IOC_IBASE    : %016lx\n", READ_REG(hpa+IOC_IBASE));
296 	DBG_INIT("IOC_IMASK    : %016lx\n", READ_REG(hpa+IOC_IMASK));
297 	DBG_INIT("IOC_TCNFG    : %016lx\n", READ_REG(hpa+IOC_TCNFG));
298 	DBG_INIT("IOC_PDIR_BASE: %016lx\n", READ_REG(hpa+IOC_PDIR_BASE));
299 	DBG_INIT("\n");
300 }
301 #endif
302 
303 
304 #ifdef ASSERT_PDIR_SANITY
305 
306 /**
307  * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
308  * @ioc: IO MMU structure which owns the pdir we are interested in.
309  * @msg: text to print ont the output line.
310  * @pide: pdir index.
311  *
312  * Print one entry of the IO MMU PDIR in human readable form.
313  */
314 static void
315 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
316 {
317 	/* start printing from lowest pde in rval */
318 	u64 *ptr = &ioc->pdir_base[pide  & ~(BITS_PER_LONG - 1)];
319 	unsigned long *rptr = (unsigned long *) &ioc->res_map[(pide >>3) & -sizeof(unsigned long)];
320 	uint rcnt;
321 
322 	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
323 		 msg, rptr, pide & (BITS_PER_LONG - 1), *rptr);
324 
325 	rcnt = 0;
326 	while (rcnt < BITS_PER_LONG) {
327 		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
328 		       (rcnt == (pide & (BITS_PER_LONG - 1)))
329 		       ? "    -->" : "       ",
330 		       rcnt, ptr, (unsigned long long) *ptr );
331 		rcnt++;
332 		ptr++;
333 	}
334 	printk(KERN_DEBUG "%s", msg);
335 }
336 
337 
338 /**
339  * sba_check_pdir - debugging only - consistency checker
340  * @ioc: IO MMU structure which owns the pdir we are interested in.
341  * @msg: text to print ont the output line.
342  *
343  * Verify the resource map and pdir state is consistent
344  */
345 static int
346 sba_check_pdir(struct ioc *ioc, char *msg)
347 {
348 	u64 *rptr_end = (u64 *) &(ioc->res_map[ioc->res_size]);
349 	u64 *rptr = (u64 *) ioc->res_map;	/* resource map ptr */
350 	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
351 	uint pide = 0;
352 
353 	while (rptr < rptr_end) {
354 		u64 rval;
355 		int rcnt; /* number of bits we might check */
356 
357 		rval = *rptr;
358 		rcnt = 64;
359 
360 		while (rcnt) {
361 			/* Get last byte and highest bit from that */
362 			u32 pde = ((u32)((*pptr >> (63)) & 0x1));
363 			if ((rval & 0x1) ^ pde)
364 			{
365 				/*
366 				** BUMMER!  -- res_map != pdir --
367 				** Dump rval and matching pdir entries
368 				*/
369 				sba_dump_pdir_entry(ioc, msg, pide);
370 				return(1);
371 			}
372 			rcnt--;
373 			rval >>= 1;	/* try the next bit */
374 			pptr++;
375 			pide++;
376 		}
377 		rptr++;	/* look at next word of res_map */
378 	}
379 	/* It'd be nice if we always got here :^) */
380 	return 0;
381 }
382 
383 
384 /**
385  * sba_dump_sg - debugging only - print Scatter-Gather list
386  * @ioc: IO MMU structure which owns the pdir we are interested in.
387  * @startsg: head of the SG list
388  * @nents: number of entries in SG list
389  *
390  * print the SG list so we can verify it's correct by hand.
391  */
392 static void
393 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
394 {
395 	while (nents-- > 0) {
396 		printk(KERN_DEBUG " %d : DMA %08lx/%05x CPU %p\n", nents,
397 		       startsg->dma_address, startsg->dma_length,
398 		       sba_sg_address(startsg));
399 		startsg++;
400 	}
401 }
402 
403 static void
404 sba_check_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
405 {
406 	struct scatterlist *the_sg = startsg;
407 	int the_nents = nents;
408 
409 	while (the_nents-- > 0) {
410 		if (sba_sg_address(the_sg) == 0x0UL)
411 			sba_dump_sg(NULL, startsg, nents);
412 		the_sg++;
413 	}
414 }
415 
416 #endif /* ASSERT_PDIR_SANITY */
417 
418 
419 
420 
421 /**************************************************************
422 *
423 *   I/O Pdir Resource Management
424 *
425 *   Bits set in the resource map are in use.
426 *   Each bit can represent a number of pages.
427 *   LSbs represent lower addresses (IOVA's).
428 *
429 ***************************************************************/
430 #define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
431 
432 /* Convert from IOVP to IOVA and vice versa. */
433 #define SBA_IOVA(ioc,iovp,offset) ((ioc->ibase) | (iovp) | (offset))
434 #define SBA_IOVP(ioc,iova) ((iova) & ~(ioc->ibase))
435 
436 #define PDIR_ENTRY_SIZE	sizeof(u64)
437 
438 #define PDIR_INDEX(iovp)   ((iovp)>>iovp_shift)
439 
440 #define RESMAP_MASK(n)    ~(~0UL << (n))
441 #define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)
442 
443 
444 /**
445  * For most cases the normal get_order is sufficient, however it limits us
446  * to PAGE_SIZE being the minimum mapping alignment and TC flush granularity.
447  * It only incurs about 1 clock cycle to use this one with the static variable
448  * and makes the code more intuitive.
449  */
450 static SBA_INLINE int
451 get_iovp_order (unsigned long size)
452 {
453 	long double d = size - 1;
454 	long order;
455 
456 	order = ia64_getf_exp(d);
457 	order = order - iovp_shift - 0xffff + 1;
458 	if (order < 0)
459 		order = 0;
460 	return order;
461 }
462 
463 /**
464  * sba_search_bitmap - find free space in IO PDIR resource bitmap
465  * @ioc: IO MMU structure which owns the pdir we are interested in.
466  * @bits_wanted: number of entries we need.
467  * @use_hint: use res_hint to indicate where to start looking
468  *
469  * Find consecutive free bits in resource bitmap.
470  * Each bit represents one entry in the IO Pdir.
471  * Cool perf optimization: search for log2(size) bits at a time.
472  */
473 static SBA_INLINE unsigned long
474 sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted, int use_hint)
475 {
476 	unsigned long *res_ptr;
477 	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
478 	unsigned long flags, pide = ~0UL;
479 
480 	ASSERT(((unsigned long) ioc->res_hint & (sizeof(unsigned long) - 1UL)) == 0);
481 	ASSERT(res_ptr < res_end);
482 
483 	spin_lock_irqsave(&ioc->res_lock, flags);
484 
485 	/* Allow caller to force a search through the entire resource space */
486 	if (likely(use_hint)) {
487 		res_ptr = ioc->res_hint;
488 	} else {
489 		res_ptr = (ulong *)ioc->res_map;
490 		ioc->res_bitshift = 0;
491 	}
492 
493 	/*
494 	 * N.B.  REO/Grande defect AR2305 can cause TLB fetch timeouts
495 	 * if a TLB entry is purged while in use.  sba_mark_invalid()
496 	 * purges IOTLB entries in power-of-two sizes, so we also
497 	 * allocate IOVA space in power-of-two sizes.
498 	 */
499 	bits_wanted = 1UL << get_iovp_order(bits_wanted << iovp_shift);
500 
501 	if (likely(bits_wanted == 1)) {
502 		unsigned int bitshiftcnt;
503 		for(; res_ptr < res_end ; res_ptr++) {
504 			if (likely(*res_ptr != ~0UL)) {
505 				bitshiftcnt = ffz(*res_ptr);
506 				*res_ptr |= (1UL << bitshiftcnt);
507 				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
508 				pide <<= 3;	/* convert to bit address */
509 				pide += bitshiftcnt;
510 				ioc->res_bitshift = bitshiftcnt + bits_wanted;
511 				goto found_it;
512 			}
513 		}
514 		goto not_found;
515 
516 	}
517 
518 	if (likely(bits_wanted <= BITS_PER_LONG/2)) {
519 		/*
520 		** Search the resource bit map on well-aligned values.
521 		** "o" is the alignment.
522 		** We need the alignment to invalidate I/O TLB using
523 		** SBA HW features in the unmap path.
524 		*/
525 		unsigned long o = 1 << get_iovp_order(bits_wanted << iovp_shift);
526 		uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
527 		unsigned long mask, base_mask;
528 
529 		base_mask = RESMAP_MASK(bits_wanted);
530 		mask = base_mask << bitshiftcnt;
531 
532 		DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
533 		for(; res_ptr < res_end ; res_ptr++)
534 		{
535 			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
536 			ASSERT(0 != mask);
537 			for (; mask ; mask <<= o, bitshiftcnt += o) {
538 				if(0 == ((*res_ptr) & mask)) {
539 					*res_ptr |= mask;     /* mark resources busy! */
540 					pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
541 					pide <<= 3;	/* convert to bit address */
542 					pide += bitshiftcnt;
543 					ioc->res_bitshift = bitshiftcnt + bits_wanted;
544 					goto found_it;
545 				}
546 			}
547 
548 			bitshiftcnt = 0;
549 			mask = base_mask;
550 
551 		}
552 
553 	} else {
554 		int qwords, bits, i;
555 		unsigned long *end;
556 
557 		qwords = bits_wanted >> 6; /* /64 */
558 		bits = bits_wanted - (qwords * BITS_PER_LONG);
559 
560 		end = res_end - qwords;
561 
562 		for (; res_ptr < end; res_ptr++) {
563 			for (i = 0 ; i < qwords ; i++) {
564 				if (res_ptr[i] != 0)
565 					goto next_ptr;
566 			}
567 			if (bits && res_ptr[i] && (__ffs(res_ptr[i]) < bits))
568 				continue;
569 
570 			/* Found it, mark it */
571 			for (i = 0 ; i < qwords ; i++)
572 				res_ptr[i] = ~0UL;
573 			res_ptr[i] |= RESMAP_MASK(bits);
574 
575 			pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
576 			pide <<= 3;	/* convert to bit address */
577 			res_ptr += qwords;
578 			ioc->res_bitshift = bits;
579 			goto found_it;
580 next_ptr:
581 			;
582 		}
583 	}
584 
585 not_found:
586 	prefetch(ioc->res_map);
587 	ioc->res_hint = (unsigned long *) ioc->res_map;
588 	ioc->res_bitshift = 0;
589 	spin_unlock_irqrestore(&ioc->res_lock, flags);
590 	return (pide);
591 
592 found_it:
593 	ioc->res_hint = res_ptr;
594 	spin_unlock_irqrestore(&ioc->res_lock, flags);
595 	return (pide);
596 }
597 
598 
599 /**
600  * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
601  * @ioc: IO MMU structure which owns the pdir we are interested in.
602  * @size: number of bytes to create a mapping for
603  *
604  * Given a size, find consecutive unmarked and then mark those bits in the
605  * resource bit map.
606  */
607 static int
608 sba_alloc_range(struct ioc *ioc, size_t size)
609 {
610 	unsigned int pages_needed = size >> iovp_shift;
611 #ifdef PDIR_SEARCH_TIMING
612 	unsigned long itc_start;
613 #endif
614 	unsigned long pide;
615 
616 	ASSERT(pages_needed);
617 	ASSERT(0 == (size & ~iovp_mask));
618 
619 #ifdef PDIR_SEARCH_TIMING
620 	itc_start = ia64_get_itc();
621 #endif
622 	/*
623 	** "seek and ye shall find"...praying never hurts either...
624 	*/
625 	pide = sba_search_bitmap(ioc, pages_needed, 1);
626 	if (unlikely(pide >= (ioc->res_size << 3))) {
627 		pide = sba_search_bitmap(ioc, pages_needed, 0);
628 		if (unlikely(pide >= (ioc->res_size << 3))) {
629 #if DELAYED_RESOURCE_CNT > 0
630 			unsigned long flags;
631 
632 			/*
633 			** With delayed resource freeing, we can give this one more shot.  We're
634 			** getting close to being in trouble here, so do what we can to make this
635 			** one count.
636 			*/
637 			spin_lock_irqsave(&ioc->saved_lock, flags);
638 			if (ioc->saved_cnt > 0) {
639 				struct sba_dma_pair *d;
640 				int cnt = ioc->saved_cnt;
641 
642 				d = &(ioc->saved[ioc->saved_cnt - 1]);
643 
644 				spin_lock(&ioc->res_lock);
645 				while (cnt--) {
646 					sba_mark_invalid(ioc, d->iova, d->size);
647 					sba_free_range(ioc, d->iova, d->size);
648 					d--;
649 				}
650 				ioc->saved_cnt = 0;
651 				READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
652 				spin_unlock(&ioc->res_lock);
653 			}
654 			spin_unlock_irqrestore(&ioc->saved_lock, flags);
655 
656 			pide = sba_search_bitmap(ioc, pages_needed, 0);
657 			if (unlikely(pide >= (ioc->res_size << 3)))
658 				panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
659 				      ioc->ioc_hpa);
660 #else
661 			panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
662 			      ioc->ioc_hpa);
663 #endif
664 		}
665 	}
666 
667 #ifdef PDIR_SEARCH_TIMING
668 	ioc->avg_search[ioc->avg_idx++] = (ia64_get_itc() - itc_start) / pages_needed;
669 	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
670 #endif
671 
672 	prefetchw(&(ioc->pdir_base[pide]));
673 
674 #ifdef ASSERT_PDIR_SANITY
675 	/* verify the first enable bit is clear */
676 	if(0x00 != ((u8 *) ioc->pdir_base)[pide*PDIR_ENTRY_SIZE + 7]) {
677 		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
678 	}
679 #endif
680 
681 	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
682 		__FUNCTION__, size, pages_needed, pide,
683 		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
684 		ioc->res_bitshift );
685 
686 	return (pide);
687 }
688 
689 
690 /**
691  * sba_free_range - unmark bits in IO PDIR resource bitmap
692  * @ioc: IO MMU structure which owns the pdir we are interested in.
693  * @iova: IO virtual address which was previously allocated.
694  * @size: number of bytes to create a mapping for
695  *
696  * clear bits in the ioc's resource map
697  */
698 static SBA_INLINE void
699 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
700 {
701 	unsigned long iovp = SBA_IOVP(ioc, iova);
702 	unsigned int pide = PDIR_INDEX(iovp);
703 	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
704 	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
705 	int bits_not_wanted = size >> iovp_shift;
706 	unsigned long m;
707 
708 	/* Round up to power-of-two size: see AR2305 note above */
709 	bits_not_wanted = 1UL << get_iovp_order(bits_not_wanted << iovp_shift);
710 	for (; bits_not_wanted > 0 ; res_ptr++) {
711 
712 		if (unlikely(bits_not_wanted > BITS_PER_LONG)) {
713 
714 			/* these mappings start 64bit aligned */
715 			*res_ptr = 0UL;
716 			bits_not_wanted -= BITS_PER_LONG;
717 			pide += BITS_PER_LONG;
718 
719 		} else {
720 
721 			/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
722 			m = RESMAP_MASK(bits_not_wanted) << (pide & (BITS_PER_LONG - 1));
723 			bits_not_wanted = 0;
724 
725 			DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", __FUNCTION__, (uint) iova, size,
726 		        	bits_not_wanted, m, pide, res_ptr, *res_ptr);
727 
728 			ASSERT(m != 0);
729 			ASSERT(bits_not_wanted);
730 			ASSERT((*res_ptr & m) == m); /* verify same bits are set */
731 			*res_ptr &= ~m;
732 		}
733 	}
734 }
735 
736 
737 /**************************************************************
738 *
739 *   "Dynamic DMA Mapping" support (aka "Coherent I/O")
740 *
741 ***************************************************************/
742 
743 /**
744  * sba_io_pdir_entry - fill in one IO PDIR entry
745  * @pdir_ptr:  pointer to IO PDIR entry
746  * @vba: Virtual CPU address of buffer to map
747  *
748  * SBA Mapping Routine
749  *
750  * Given a virtual address (vba, arg1) sba_io_pdir_entry()
751  * loads the I/O PDIR entry pointed to by pdir_ptr (arg0).
752  * Each IO Pdir entry consists of 8 bytes as shown below
753  * (LSB == bit 0):
754  *
755  *  63                    40                                 11    7        0
756  * +-+---------------------+----------------------------------+----+--------+
757  * |V|        U            |            PPN[39:12]            | U  |   FF   |
758  * +-+---------------------+----------------------------------+----+--------+
759  *
760  *  V  == Valid Bit
761  *  U  == Unused
762  * PPN == Physical Page Number
763  *
764  * The physical address fields are filled with the results of virt_to_phys()
765  * on the vba.
766  */
767 
768 #if 1
769 #define sba_io_pdir_entry(pdir_ptr, vba) *pdir_ptr = ((vba & ~0xE000000000000FFFULL)	\
770 						      | 0x8000000000000000ULL)
771 #else
772 void SBA_INLINE
773 sba_io_pdir_entry(u64 *pdir_ptr, unsigned long vba)
774 {
775 	*pdir_ptr = ((vba & ~0xE000000000000FFFULL) | 0x80000000000000FFULL);
776 }
777 #endif
778 
779 #ifdef ENABLE_MARK_CLEAN
780 /**
781  * Since DMA is i-cache coherent, any (complete) pages that were written via
782  * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
783  * flush them when they get mapped into an executable vm-area.
784  */
785 static void
786 mark_clean (void *addr, size_t size)
787 {
788 	unsigned long pg_addr, end;
789 
790 	pg_addr = PAGE_ALIGN((unsigned long) addr);
791 	end = (unsigned long) addr + size;
792 	while (pg_addr + PAGE_SIZE <= end) {
793 		struct page *page = virt_to_page((void *)pg_addr);
794 		set_bit(PG_arch_1, &page->flags);
795 		pg_addr += PAGE_SIZE;
796 	}
797 }
798 #endif
799 
800 /**
801  * sba_mark_invalid - invalidate one or more IO PDIR entries
802  * @ioc: IO MMU structure which owns the pdir we are interested in.
803  * @iova:  IO Virtual Address mapped earlier
804  * @byte_cnt:  number of bytes this mapping covers.
805  *
806  * Marking the IO PDIR entry(ies) as Invalid and invalidate
807  * corresponding IO TLB entry. The PCOM (Purge Command Register)
808  * is to purge stale entries in the IO TLB when unmapping entries.
809  *
810  * The PCOM register supports purging of multiple pages, with a minium
811  * of 1 page and a maximum of 2GB. Hardware requires the address be
812  * aligned to the size of the range being purged. The size of the range
813  * must be a power of 2. The "Cool perf optimization" in the
814  * allocation routine helps keep that true.
815  */
816 static SBA_INLINE void
817 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
818 {
819 	u32 iovp = (u32) SBA_IOVP(ioc,iova);
820 
821 	int off = PDIR_INDEX(iovp);
822 
823 	/* Must be non-zero and rounded up */
824 	ASSERT(byte_cnt > 0);
825 	ASSERT(0 == (byte_cnt & ~iovp_mask));
826 
827 #ifdef ASSERT_PDIR_SANITY
828 	/* Assert first pdir entry is set */
829 	if (!(ioc->pdir_base[off] >> 60)) {
830 		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
831 	}
832 #endif
833 
834 	if (byte_cnt <= iovp_size)
835 	{
836 		ASSERT(off < ioc->pdir_size);
837 
838 		iovp |= iovp_shift;     /* set "size" field for PCOM */
839 
840 #ifndef FULL_VALID_PDIR
841 		/*
842 		** clear I/O PDIR entry "valid" bit
843 		** Do NOT clear the rest - save it for debugging.
844 		** We should only clear bits that have previously
845 		** been enabled.
846 		*/
847 		ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
848 #else
849 		/*
850   		** If we want to maintain the PDIR as valid, put in
851 		** the spill page so devices prefetching won't
852 		** cause a hard fail.
853 		*/
854 		ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
855 #endif
856 	} else {
857 		u32 t = get_iovp_order(byte_cnt) + iovp_shift;
858 
859 		iovp |= t;
860 		ASSERT(t <= 31);   /* 2GB! Max value of "size" field */
861 
862 		do {
863 			/* verify this pdir entry is enabled */
864 			ASSERT(ioc->pdir_base[off]  >> 63);
865 #ifndef FULL_VALID_PDIR
866 			/* clear I/O Pdir entry "valid" bit first */
867 			ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
868 #else
869 			ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
870 #endif
871 			off++;
872 			byte_cnt -= iovp_size;
873 		} while (byte_cnt > 0);
874 	}
875 
876 	WRITE_REG(iovp | ioc->ibase, ioc->ioc_hpa+IOC_PCOM);
877 }
878 
879 /**
880  * sba_map_single - map one buffer and return IOVA for DMA
881  * @dev: instance of PCI owned by the driver that's asking.
882  * @addr:  driver buffer to map.
883  * @size:  number of bytes to map in driver buffer.
884  * @dir:  R/W or both.
885  *
886  * See Documentation/DMA-mapping.txt
887  */
888 dma_addr_t
889 sba_map_single(struct device *dev, void *addr, size_t size, int dir)
890 {
891 	struct ioc *ioc;
892 	dma_addr_t iovp;
893 	dma_addr_t offset;
894 	u64 *pdir_start;
895 	int pide;
896 #ifdef ASSERT_PDIR_SANITY
897 	unsigned long flags;
898 #endif
899 #ifdef ALLOW_IOV_BYPASS
900 	unsigned long pci_addr = virt_to_phys(addr);
901 #endif
902 
903 #ifdef ALLOW_IOV_BYPASS
904 	ASSERT(to_pci_dev(dev)->dma_mask);
905 	/*
906  	** Check if the PCI device can DMA to ptr... if so, just return ptr
907  	*/
908 	if (likely((pci_addr & ~to_pci_dev(dev)->dma_mask) == 0)) {
909 		/*
910  		** Device is bit capable of DMA'ing to the buffer...
911 		** just return the PCI address of ptr
912  		*/
913 		DBG_BYPASS("sba_map_single() bypass mask/addr: 0x%lx/0x%lx\n",
914 		           to_pci_dev(dev)->dma_mask, pci_addr);
915 		return pci_addr;
916 	}
917 #endif
918 	ioc = GET_IOC(dev);
919 	ASSERT(ioc);
920 
921 	prefetch(ioc->res_hint);
922 
923 	ASSERT(size > 0);
924 	ASSERT(size <= DMA_CHUNK_SIZE);
925 
926 	/* save offset bits */
927 	offset = ((dma_addr_t) (long) addr) & ~iovp_mask;
928 
929 	/* round up to nearest iovp_size */
930 	size = (size + offset + ~iovp_mask) & iovp_mask;
931 
932 #ifdef ASSERT_PDIR_SANITY
933 	spin_lock_irqsave(&ioc->res_lock, flags);
934 	if (sba_check_pdir(ioc,"Check before sba_map_single()"))
935 		panic("Sanity check failed");
936 	spin_unlock_irqrestore(&ioc->res_lock, flags);
937 #endif
938 
939 	pide = sba_alloc_range(ioc, size);
940 
941 	iovp = (dma_addr_t) pide << iovp_shift;
942 
943 	DBG_RUN("%s() 0x%p -> 0x%lx\n",
944 		__FUNCTION__, addr, (long) iovp | offset);
945 
946 	pdir_start = &(ioc->pdir_base[pide]);
947 
948 	while (size > 0) {
949 		ASSERT(((u8 *)pdir_start)[7] == 0); /* verify availability */
950 		sba_io_pdir_entry(pdir_start, (unsigned long) addr);
951 
952 		DBG_RUN("     pdir 0x%p %lx\n", pdir_start, *pdir_start);
953 
954 		addr += iovp_size;
955 		size -= iovp_size;
956 		pdir_start++;
957 	}
958 	/* force pdir update */
959 	wmb();
960 
961 	/* form complete address */
962 #ifdef ASSERT_PDIR_SANITY
963 	spin_lock_irqsave(&ioc->res_lock, flags);
964 	sba_check_pdir(ioc,"Check after sba_map_single()");
965 	spin_unlock_irqrestore(&ioc->res_lock, flags);
966 #endif
967 	return SBA_IOVA(ioc, iovp, offset);
968 }
969 
970 #ifdef ENABLE_MARK_CLEAN
971 static SBA_INLINE void
972 sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size)
973 {
974 	u32	iovp = (u32) SBA_IOVP(ioc,iova);
975 	int	off = PDIR_INDEX(iovp);
976 	void	*addr;
977 
978 	if (size <= iovp_size) {
979 		addr = phys_to_virt(ioc->pdir_base[off] &
980 		                    ~0xE000000000000FFFULL);
981 		mark_clean(addr, size);
982 	} else {
983 		do {
984 			addr = phys_to_virt(ioc->pdir_base[off] &
985 			                    ~0xE000000000000FFFULL);
986 			mark_clean(addr, min(size, iovp_size));
987 			off++;
988 			size -= iovp_size;
989 		} while (size > 0);
990 	}
991 }
992 #endif
993 
994 /**
995  * sba_unmap_single - unmap one IOVA and free resources
996  * @dev: instance of PCI owned by the driver that's asking.
997  * @iova:  IOVA of driver buffer previously mapped.
998  * @size:  number of bytes mapped in driver buffer.
999  * @dir:  R/W or both.
1000  *
1001  * See Documentation/DMA-mapping.txt
1002  */
1003 void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
1004 {
1005 	struct ioc *ioc;
1006 #if DELAYED_RESOURCE_CNT > 0
1007 	struct sba_dma_pair *d;
1008 #endif
1009 	unsigned long flags;
1010 	dma_addr_t offset;
1011 
1012 	ioc = GET_IOC(dev);
1013 	ASSERT(ioc);
1014 
1015 #ifdef ALLOW_IOV_BYPASS
1016 	if (likely((iova & ioc->imask) != ioc->ibase)) {
1017 		/*
1018 		** Address does not fall w/in IOVA, must be bypassing
1019 		*/
1020 		DBG_BYPASS("sba_unmap_single() bypass addr: 0x%lx\n", iova);
1021 
1022 #ifdef ENABLE_MARK_CLEAN
1023 		if (dir == DMA_FROM_DEVICE) {
1024 			mark_clean(phys_to_virt(iova), size);
1025 		}
1026 #endif
1027 		return;
1028 	}
1029 #endif
1030 	offset = iova & ~iovp_mask;
1031 
1032 	DBG_RUN("%s() iovp 0x%lx/%x\n",
1033 		__FUNCTION__, (long) iova, size);
1034 
1035 	iova ^= offset;        /* clear offset bits */
1036 	size += offset;
1037 	size = ROUNDUP(size, iovp_size);
1038 
1039 #ifdef ENABLE_MARK_CLEAN
1040 	if (dir == DMA_FROM_DEVICE)
1041 		sba_mark_clean(ioc, iova, size);
1042 #endif
1043 
1044 #if DELAYED_RESOURCE_CNT > 0
1045 	spin_lock_irqsave(&ioc->saved_lock, flags);
1046 	d = &(ioc->saved[ioc->saved_cnt]);
1047 	d->iova = iova;
1048 	d->size = size;
1049 	if (unlikely(++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT)) {
1050 		int cnt = ioc->saved_cnt;
1051 		spin_lock(&ioc->res_lock);
1052 		while (cnt--) {
1053 			sba_mark_invalid(ioc, d->iova, d->size);
1054 			sba_free_range(ioc, d->iova, d->size);
1055 			d--;
1056 		}
1057 		ioc->saved_cnt = 0;
1058 		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
1059 		spin_unlock(&ioc->res_lock);
1060 	}
1061 	spin_unlock_irqrestore(&ioc->saved_lock, flags);
1062 #else /* DELAYED_RESOURCE_CNT == 0 */
1063 	spin_lock_irqsave(&ioc->res_lock, flags);
1064 	sba_mark_invalid(ioc, iova, size);
1065 	sba_free_range(ioc, iova, size);
1066 	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
1067 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1068 #endif /* DELAYED_RESOURCE_CNT == 0 */
1069 }
1070 
1071 
1072 /**
1073  * sba_alloc_coherent - allocate/map shared mem for DMA
1074  * @dev: instance of PCI owned by the driver that's asking.
1075  * @size:  number of bytes mapped in driver buffer.
1076  * @dma_handle:  IOVA of new buffer.
1077  *
1078  * See Documentation/DMA-mapping.txt
1079  */
1080 void *
1081 sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flags)
1082 {
1083 	struct ioc *ioc;
1084 	void *addr;
1085 
1086 	ioc = GET_IOC(dev);
1087 	ASSERT(ioc);
1088 
1089 #ifdef CONFIG_NUMA
1090 	{
1091 		struct page *page;
1092 		page = alloc_pages_node(ioc->node == MAX_NUMNODES ?
1093 		                        numa_node_id() : ioc->node, flags,
1094 		                        get_order(size));
1095 
1096 		if (unlikely(!page))
1097 			return NULL;
1098 
1099 		addr = page_address(page);
1100 	}
1101 #else
1102 	addr = (void *) __get_free_pages(flags, get_order(size));
1103 #endif
1104 	if (unlikely(!addr))
1105 		return NULL;
1106 
1107 	memset(addr, 0, size);
1108 	*dma_handle = virt_to_phys(addr);
1109 
1110 #ifdef ALLOW_IOV_BYPASS
1111 	ASSERT(dev->coherent_dma_mask);
1112 	/*
1113  	** Check if the PCI device can DMA to ptr... if so, just return ptr
1114  	*/
1115 	if (likely((*dma_handle & ~dev->coherent_dma_mask) == 0)) {
1116 		DBG_BYPASS("sba_alloc_coherent() bypass mask/addr: 0x%lx/0x%lx\n",
1117 		           dev->coherent_dma_mask, *dma_handle);
1118 
1119 		return addr;
1120 	}
1121 #endif
1122 
1123 	/*
1124 	 * If device can't bypass or bypass is disabled, pass the 32bit fake
1125 	 * device to map single to get an iova mapping.
1126 	 */
1127 	*dma_handle = sba_map_single(&ioc->sac_only_dev->dev, addr, size, 0);
1128 
1129 	return addr;
1130 }
1131 
1132 
1133 /**
1134  * sba_free_coherent - free/unmap shared mem for DMA
1135  * @dev: instance of PCI owned by the driver that's asking.
1136  * @size:  number of bytes mapped in driver buffer.
1137  * @vaddr:  virtual address IOVA of "consistent" buffer.
1138  * @dma_handler:  IO virtual address of "consistent" buffer.
1139  *
1140  * See Documentation/DMA-mapping.txt
1141  */
1142 void sba_free_coherent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle)
1143 {
1144 	sba_unmap_single(dev, dma_handle, size, 0);
1145 	free_pages((unsigned long) vaddr, get_order(size));
1146 }
1147 
1148 
1149 /*
1150 ** Since 0 is a valid pdir_base index value, can't use that
1151 ** to determine if a value is valid or not. Use a flag to indicate
1152 ** the SG list entry contains a valid pdir index.
1153 */
1154 #define PIDE_FLAG 0x1UL
1155 
1156 #ifdef DEBUG_LARGE_SG_ENTRIES
1157 int dump_run_sg = 0;
1158 #endif
1159 
1160 
1161 /**
1162  * sba_fill_pdir - write allocated SG entries into IO PDIR
1163  * @ioc: IO MMU structure which owns the pdir we are interested in.
1164  * @startsg:  list of IOVA/size pairs
1165  * @nents: number of entries in startsg list
1166  *
1167  * Take preprocessed SG list and write corresponding entries
1168  * in the IO PDIR.
1169  */
1170 
1171 static SBA_INLINE int
1172 sba_fill_pdir(
1173 	struct ioc *ioc,
1174 	struct scatterlist *startsg,
1175 	int nents)
1176 {
1177 	struct scatterlist *dma_sg = startsg;	/* pointer to current DMA */
1178 	int n_mappings = 0;
1179 	u64 *pdirp = NULL;
1180 	unsigned long dma_offset = 0;
1181 
1182 	dma_sg--;
1183 	while (nents-- > 0) {
1184 		int     cnt = startsg->dma_length;
1185 		startsg->dma_length = 0;
1186 
1187 #ifdef DEBUG_LARGE_SG_ENTRIES
1188 		if (dump_run_sg)
1189 			printk(" %2d : %08lx/%05x %p\n",
1190 				nents, startsg->dma_address, cnt,
1191 				sba_sg_address(startsg));
1192 #else
1193 		DBG_RUN_SG(" %d : %08lx/%05x %p\n",
1194 				nents, startsg->dma_address, cnt,
1195 				sba_sg_address(startsg));
1196 #endif
1197 		/*
1198 		** Look for the start of a new DMA stream
1199 		*/
1200 		if (startsg->dma_address & PIDE_FLAG) {
1201 			u32 pide = startsg->dma_address & ~PIDE_FLAG;
1202 			dma_offset = (unsigned long) pide & ~iovp_mask;
1203 			startsg->dma_address = 0;
1204 			dma_sg++;
1205 			dma_sg->dma_address = pide | ioc->ibase;
1206 			pdirp = &(ioc->pdir_base[pide >> iovp_shift]);
1207 			n_mappings++;
1208 		}
1209 
1210 		/*
1211 		** Look for a VCONTIG chunk
1212 		*/
1213 		if (cnt) {
1214 			unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
1215 			ASSERT(pdirp);
1216 
1217 			/* Since multiple Vcontig blocks could make up
1218 			** one DMA stream, *add* cnt to dma_len.
1219 			*/
1220 			dma_sg->dma_length += cnt;
1221 			cnt += dma_offset;
1222 			dma_offset=0;	/* only want offset on first chunk */
1223 			cnt = ROUNDUP(cnt, iovp_size);
1224 			do {
1225 				sba_io_pdir_entry(pdirp, vaddr);
1226 				vaddr += iovp_size;
1227 				cnt -= iovp_size;
1228 				pdirp++;
1229 			} while (cnt > 0);
1230 		}
1231 		startsg++;
1232 	}
1233 	/* force pdir update */
1234 	wmb();
1235 
1236 #ifdef DEBUG_LARGE_SG_ENTRIES
1237 	dump_run_sg = 0;
1238 #endif
1239 	return(n_mappings);
1240 }
1241 
1242 
1243 /*
1244 ** Two address ranges are DMA contiguous *iff* "end of prev" and
1245 ** "start of next" are both on an IOV page boundary.
1246 **
1247 ** (shift left is a quick trick to mask off upper bits)
1248 */
1249 #define DMA_CONTIG(__X, __Y) \
1250 	(((((unsigned long) __X) | ((unsigned long) __Y)) << (BITS_PER_LONG - iovp_shift)) == 0UL)
1251 
1252 
1253 /**
1254  * sba_coalesce_chunks - preprocess the SG list
1255  * @ioc: IO MMU structure which owns the pdir we are interested in.
1256  * @startsg:  list of IOVA/size pairs
1257  * @nents: number of entries in startsg list
1258  *
1259  * First pass is to walk the SG list and determine where the breaks are
1260  * in the DMA stream. Allocates PDIR entries but does not fill them.
1261  * Returns the number of DMA chunks.
1262  *
1263  * Doing the fill separate from the coalescing/allocation keeps the
1264  * code simpler. Future enhancement could make one pass through
1265  * the sglist do both.
1266  */
1267 static SBA_INLINE int
1268 sba_coalesce_chunks( struct ioc *ioc,
1269 	struct scatterlist *startsg,
1270 	int nents)
1271 {
1272 	struct scatterlist *vcontig_sg;    /* VCONTIG chunk head */
1273 	unsigned long vcontig_len;         /* len of VCONTIG chunk */
1274 	unsigned long vcontig_end;
1275 	struct scatterlist *dma_sg;        /* next DMA stream head */
1276 	unsigned long dma_offset, dma_len; /* start/len of DMA stream */
1277 	int n_mappings = 0;
1278 
1279 	while (nents > 0) {
1280 		unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
1281 
1282 		/*
1283 		** Prepare for first/next DMA stream
1284 		*/
1285 		dma_sg = vcontig_sg = startsg;
1286 		dma_len = vcontig_len = vcontig_end = startsg->length;
1287 		vcontig_end +=  vaddr;
1288 		dma_offset = vaddr & ~iovp_mask;
1289 
1290 		/* PARANOID: clear entries */
1291 		startsg->dma_address = startsg->dma_length = 0;
1292 
1293 		/*
1294 		** This loop terminates one iteration "early" since
1295 		** it's always looking one "ahead".
1296 		*/
1297 		while (--nents > 0) {
1298 			unsigned long vaddr;	/* tmp */
1299 
1300 			startsg++;
1301 
1302 			/* PARANOID */
1303 			startsg->dma_address = startsg->dma_length = 0;
1304 
1305 			/* catch brokenness in SCSI layer */
1306 			ASSERT(startsg->length <= DMA_CHUNK_SIZE);
1307 
1308 			/*
1309 			** First make sure current dma stream won't
1310 			** exceed DMA_CHUNK_SIZE if we coalesce the
1311 			** next entry.
1312 			*/
1313 			if (((dma_len + dma_offset + startsg->length + ~iovp_mask) & iovp_mask)
1314 			    > DMA_CHUNK_SIZE)
1315 				break;
1316 
1317 			/*
1318 			** Then look for virtually contiguous blocks.
1319 			**
1320 			** append the next transaction?
1321 			*/
1322 			vaddr = (unsigned long) sba_sg_address(startsg);
1323 			if  (vcontig_end == vaddr)
1324 			{
1325 				vcontig_len += startsg->length;
1326 				vcontig_end += startsg->length;
1327 				dma_len     += startsg->length;
1328 				continue;
1329 			}
1330 
1331 #ifdef DEBUG_LARGE_SG_ENTRIES
1332 			dump_run_sg = (vcontig_len > iovp_size);
1333 #endif
1334 
1335 			/*
1336 			** Not virtually contigous.
1337 			** Terminate prev chunk.
1338 			** Start a new chunk.
1339 			**
1340 			** Once we start a new VCONTIG chunk, dma_offset
1341 			** can't change. And we need the offset from the first
1342 			** chunk - not the last one. Ergo Successive chunks
1343 			** must start on page boundaries and dove tail
1344 			** with it's predecessor.
1345 			*/
1346 			vcontig_sg->dma_length = vcontig_len;
1347 
1348 			vcontig_sg = startsg;
1349 			vcontig_len = startsg->length;
1350 
1351 			/*
1352 			** 3) do the entries end/start on page boundaries?
1353 			**    Don't update vcontig_end until we've checked.
1354 			*/
1355 			if (DMA_CONTIG(vcontig_end, vaddr))
1356 			{
1357 				vcontig_end = vcontig_len + vaddr;
1358 				dma_len += vcontig_len;
1359 				continue;
1360 			} else {
1361 				break;
1362 			}
1363 		}
1364 
1365 		/*
1366 		** End of DMA Stream
1367 		** Terminate last VCONTIG block.
1368 		** Allocate space for DMA stream.
1369 		*/
1370 		vcontig_sg->dma_length = vcontig_len;
1371 		dma_len = (dma_len + dma_offset + ~iovp_mask) & iovp_mask;
1372 		ASSERT(dma_len <= DMA_CHUNK_SIZE);
1373 		dma_sg->dma_address = (dma_addr_t) (PIDE_FLAG
1374 			| (sba_alloc_range(ioc, dma_len) << iovp_shift)
1375 			| dma_offset);
1376 		n_mappings++;
1377 	}
1378 
1379 	return n_mappings;
1380 }
1381 
1382 
1383 /**
1384  * sba_map_sg - map Scatter/Gather list
1385  * @dev: instance of PCI owned by the driver that's asking.
1386  * @sglist:  array of buffer/length pairs
1387  * @nents:  number of entries in list
1388  * @dir:  R/W or both.
1389  *
1390  * See Documentation/DMA-mapping.txt
1391  */
1392 int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int dir)
1393 {
1394 	struct ioc *ioc;
1395 	int coalesced, filled = 0;
1396 #ifdef ASSERT_PDIR_SANITY
1397 	unsigned long flags;
1398 #endif
1399 #ifdef ALLOW_IOV_BYPASS_SG
1400 	struct scatterlist *sg;
1401 #endif
1402 
1403 	DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents);
1404 	ioc = GET_IOC(dev);
1405 	ASSERT(ioc);
1406 
1407 #ifdef ALLOW_IOV_BYPASS_SG
1408 	ASSERT(to_pci_dev(dev)->dma_mask);
1409 	if (likely((ioc->dma_mask & ~to_pci_dev(dev)->dma_mask) == 0)) {
1410 		for (sg = sglist ; filled < nents ; filled++, sg++){
1411 			sg->dma_length = sg->length;
1412 			sg->dma_address = virt_to_phys(sba_sg_address(sg));
1413 		}
1414 		return filled;
1415 	}
1416 #endif
1417 	/* Fast path single entry scatterlists. */
1418 	if (nents == 1) {
1419 		sglist->dma_length = sglist->length;
1420 		sglist->dma_address = sba_map_single(dev, sba_sg_address(sglist), sglist->length, dir);
1421 		return 1;
1422 	}
1423 
1424 #ifdef ASSERT_PDIR_SANITY
1425 	spin_lock_irqsave(&ioc->res_lock, flags);
1426 	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
1427 	{
1428 		sba_dump_sg(ioc, sglist, nents);
1429 		panic("Check before sba_map_sg()");
1430 	}
1431 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1432 #endif
1433 
1434 	prefetch(ioc->res_hint);
1435 
1436 	/*
1437 	** First coalesce the chunks and allocate I/O pdir space
1438 	**
1439 	** If this is one DMA stream, we can properly map using the
1440 	** correct virtual address associated with each DMA page.
1441 	** w/o this association, we wouldn't have coherent DMA!
1442 	** Access to the virtual address is what forces a two pass algorithm.
1443 	*/
1444 	coalesced = sba_coalesce_chunks(ioc, sglist, nents);
1445 
1446 	/*
1447 	** Program the I/O Pdir
1448 	**
1449 	** map the virtual addresses to the I/O Pdir
1450 	** o dma_address will contain the pdir index
1451 	** o dma_len will contain the number of bytes to map
1452 	** o address contains the virtual address.
1453 	*/
1454 	filled = sba_fill_pdir(ioc, sglist, nents);
1455 
1456 #ifdef ASSERT_PDIR_SANITY
1457 	spin_lock_irqsave(&ioc->res_lock, flags);
1458 	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1459 	{
1460 		sba_dump_sg(ioc, sglist, nents);
1461 		panic("Check after sba_map_sg()\n");
1462 	}
1463 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1464 #endif
1465 
1466 	ASSERT(coalesced == filled);
1467 	DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled);
1468 
1469 	return filled;
1470 }
1471 
1472 
1473 /**
1474  * sba_unmap_sg - unmap Scatter/Gather list
1475  * @dev: instance of PCI owned by the driver that's asking.
1476  * @sglist:  array of buffer/length pairs
1477  * @nents:  number of entries in list
1478  * @dir:  R/W or both.
1479  *
1480  * See Documentation/DMA-mapping.txt
1481  */
1482 void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir)
1483 {
1484 #ifdef ASSERT_PDIR_SANITY
1485 	struct ioc *ioc;
1486 	unsigned long flags;
1487 #endif
1488 
1489 	DBG_RUN_SG("%s() START %d entries,  %p,%x\n",
1490 		__FUNCTION__, nents, sba_sg_address(sglist), sglist->length);
1491 
1492 #ifdef ASSERT_PDIR_SANITY
1493 	ioc = GET_IOC(dev);
1494 	ASSERT(ioc);
1495 
1496 	spin_lock_irqsave(&ioc->res_lock, flags);
1497 	sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1498 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1499 #endif
1500 
1501 	while (nents && sglist->dma_length) {
1502 
1503 		sba_unmap_single(dev, sglist->dma_address, sglist->dma_length, dir);
1504 		sglist++;
1505 		nents--;
1506 	}
1507 
1508 	DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__,  nents);
1509 
1510 #ifdef ASSERT_PDIR_SANITY
1511 	spin_lock_irqsave(&ioc->res_lock, flags);
1512 	sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1513 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1514 #endif
1515 
1516 }
1517 
1518 /**************************************************************
1519 *
1520 *   Initialization and claim
1521 *
1522 ***************************************************************/
1523 
1524 static void __init
1525 ioc_iova_init(struct ioc *ioc)
1526 {
1527 	int tcnfg;
1528 	int agp_found = 0;
1529 	struct pci_dev *device = NULL;
1530 #ifdef FULL_VALID_PDIR
1531 	unsigned long index;
1532 #endif
1533 
1534 	/*
1535 	** Firmware programs the base and size of a "safe IOVA space"
1536 	** (one that doesn't overlap memory or LMMIO space) in the
1537 	** IBASE and IMASK registers.
1538 	*/
1539 	ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1UL;
1540 	ioc->imask = READ_REG(ioc->ioc_hpa + IOC_IMASK) | 0xFFFFFFFF00000000UL;
1541 
1542 	ioc->iov_size = ~ioc->imask + 1;
1543 
1544 	DBG_INIT("%s() hpa %p IOV base 0x%lx mask 0x%lx (%dMB)\n",
1545 		__FUNCTION__, ioc->ioc_hpa, ioc->ibase, ioc->imask,
1546 		ioc->iov_size >> 20);
1547 
1548 	switch (iovp_size) {
1549 		case  4*1024: tcnfg = 0; break;
1550 		case  8*1024: tcnfg = 1; break;
1551 		case 16*1024: tcnfg = 2; break;
1552 		case 64*1024: tcnfg = 3; break;
1553 		default:
1554 			panic(PFX "Unsupported IOTLB page size %ldK",
1555 				iovp_size >> 10);
1556 			break;
1557 	}
1558 	WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1559 
1560 	ioc->pdir_size = (ioc->iov_size / iovp_size) * PDIR_ENTRY_SIZE;
1561 	ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1562 						   get_order(ioc->pdir_size));
1563 	if (!ioc->pdir_base)
1564 		panic(PFX "Couldn't allocate I/O Page Table\n");
1565 
1566 	memset(ioc->pdir_base, 0, ioc->pdir_size);
1567 
1568 	DBG_INIT("%s() IOV page size %ldK pdir %p size %x\n", __FUNCTION__,
1569 		iovp_size >> 10, ioc->pdir_base, ioc->pdir_size);
1570 
1571 	ASSERT(ALIGN((unsigned long) ioc->pdir_base, 4*1024) == (unsigned long) ioc->pdir_base);
1572 	WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1573 
1574 	/*
1575 	** If an AGP device is present, only use half of the IOV space
1576 	** for PCI DMA.  Unfortunately we can't know ahead of time
1577 	** whether GART support will actually be used, for now we
1578 	** can just key on an AGP device found in the system.
1579 	** We program the next pdir index after we stop w/ a key for
1580 	** the GART code to handshake on.
1581 	*/
1582 	for_each_pci_dev(device)
1583 		agp_found |= pci_find_capability(device, PCI_CAP_ID_AGP);
1584 
1585 	if (agp_found && reserve_sba_gart) {
1586 		printk(KERN_INFO PFX "reserving %dMb of IOVA space at 0x%lx for agpgart\n",
1587 		      ioc->iov_size/2 >> 20, ioc->ibase + ioc->iov_size/2);
1588 		ioc->pdir_size /= 2;
1589 		((u64 *)ioc->pdir_base)[PDIR_INDEX(ioc->iov_size/2)] = ZX1_SBA_IOMMU_COOKIE;
1590 	}
1591 #ifdef FULL_VALID_PDIR
1592 	/*
1593   	** Check to see if the spill page has been allocated, we don't need more than
1594 	** one across multiple SBAs.
1595 	*/
1596 	if (!prefetch_spill_page) {
1597 		char *spill_poison = "SBAIOMMU POISON";
1598 		int poison_size = 16;
1599 		void *poison_addr, *addr;
1600 
1601 		addr = (void *)__get_free_pages(GFP_KERNEL, get_order(iovp_size));
1602 		if (!addr)
1603 			panic(PFX "Couldn't allocate PDIR spill page\n");
1604 
1605 		poison_addr = addr;
1606 		for ( ; (u64) poison_addr < addr + iovp_size; poison_addr += poison_size)
1607 			memcpy(poison_addr, spill_poison, poison_size);
1608 
1609 		prefetch_spill_page = virt_to_phys(addr);
1610 
1611 		DBG_INIT("%s() prefetch spill addr: 0x%lx\n", __FUNCTION__, prefetch_spill_page);
1612 	}
1613 	/*
1614   	** Set all the PDIR entries valid w/ the spill page as the target
1615 	*/
1616 	for (index = 0 ; index < (ioc->pdir_size / PDIR_ENTRY_SIZE) ; index++)
1617 		((u64 *)ioc->pdir_base)[index] = (0x80000000000000FF | prefetch_spill_page);
1618 #endif
1619 
1620 	/* Clear I/O TLB of any possible entries */
1621 	WRITE_REG(ioc->ibase | (get_iovp_order(ioc->iov_size) + iovp_shift), ioc->ioc_hpa + IOC_PCOM);
1622 	READ_REG(ioc->ioc_hpa + IOC_PCOM);
1623 
1624 	/* Enable IOVA translation */
1625 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1626 	READ_REG(ioc->ioc_hpa + IOC_IBASE);
1627 }
1628 
1629 static void __init
1630 ioc_resource_init(struct ioc *ioc)
1631 {
1632 	spin_lock_init(&ioc->res_lock);
1633 #if DELAYED_RESOURCE_CNT > 0
1634 	spin_lock_init(&ioc->saved_lock);
1635 #endif
1636 
1637 	/* resource map size dictated by pdir_size */
1638 	ioc->res_size = ioc->pdir_size / PDIR_ENTRY_SIZE; /* entries */
1639 	ioc->res_size >>= 3;  /* convert bit count to byte count */
1640 	DBG_INIT("%s() res_size 0x%x\n", __FUNCTION__, ioc->res_size);
1641 
1642 	ioc->res_map = (char *) __get_free_pages(GFP_KERNEL,
1643 						 get_order(ioc->res_size));
1644 	if (!ioc->res_map)
1645 		panic(PFX "Couldn't allocate resource map\n");
1646 
1647 	memset(ioc->res_map, 0, ioc->res_size);
1648 	/* next available IOVP - circular search */
1649 	ioc->res_hint = (unsigned long *) ioc->res_map;
1650 
1651 #ifdef ASSERT_PDIR_SANITY
1652 	/* Mark first bit busy - ie no IOVA 0 */
1653 	ioc->res_map[0] = 0x1;
1654 	ioc->pdir_base[0] = 0x8000000000000000ULL | ZX1_SBA_IOMMU_COOKIE;
1655 #endif
1656 #ifdef FULL_VALID_PDIR
1657 	/* Mark the last resource used so we don't prefetch beyond IOVA space */
1658 	ioc->res_map[ioc->res_size - 1] |= 0x80UL; /* res_map is chars */
1659 	ioc->pdir_base[(ioc->pdir_size / PDIR_ENTRY_SIZE) - 1] = (0x80000000000000FF
1660 							      | prefetch_spill_page);
1661 #endif
1662 
1663 	DBG_INIT("%s() res_map %x %p\n", __FUNCTION__,
1664 		 ioc->res_size, (void *) ioc->res_map);
1665 }
1666 
1667 static void __init
1668 ioc_sac_init(struct ioc *ioc)
1669 {
1670 	struct pci_dev *sac = NULL;
1671 	struct pci_controller *controller = NULL;
1672 
1673 	/*
1674 	 * pci_alloc_coherent() must return a DMA address which is
1675 	 * SAC (single address cycle) addressable, so allocate a
1676 	 * pseudo-device to enforce that.
1677 	 */
1678 	sac = kzalloc(sizeof(*sac), GFP_KERNEL);
1679 	if (!sac)
1680 		panic(PFX "Couldn't allocate struct pci_dev");
1681 
1682 	controller = kzalloc(sizeof(*controller), GFP_KERNEL);
1683 	if (!controller)
1684 		panic(PFX "Couldn't allocate struct pci_controller");
1685 
1686 	controller->iommu = ioc;
1687 	sac->sysdata = controller;
1688 	sac->dma_mask = 0xFFFFFFFFUL;
1689 #ifdef CONFIG_PCI
1690 	sac->dev.bus = &pci_bus_type;
1691 #endif
1692 	ioc->sac_only_dev = sac;
1693 }
1694 
1695 static void __init
1696 ioc_zx1_init(struct ioc *ioc)
1697 {
1698 	unsigned long rope_config;
1699 	unsigned int i;
1700 
1701 	if (ioc->rev < 0x20)
1702 		panic(PFX "IOC 2.0 or later required for IOMMU support\n");
1703 
1704 	/* 38 bit memory controller + extra bit for range displaced by MMIO */
1705 	ioc->dma_mask = (0x1UL << 39) - 1;
1706 
1707 	/*
1708 	** Clear ROPE(N)_CONFIG AO bit.
1709 	** Disables "NT Ordering" (~= !"Relaxed Ordering")
1710 	** Overrides bit 1 in DMA Hint Sets.
1711 	** Improves netperf UDP_STREAM by ~10% for tg3 on bcm5701.
1712 	*/
1713 	for (i=0; i<(8*8); i+=8) {
1714 		rope_config = READ_REG(ioc->ioc_hpa + IOC_ROPE0_CFG + i);
1715 		rope_config &= ~IOC_ROPE_AO;
1716 		WRITE_REG(rope_config, ioc->ioc_hpa + IOC_ROPE0_CFG + i);
1717 	}
1718 }
1719 
1720 typedef void (initfunc)(struct ioc *);
1721 
1722 struct ioc_iommu {
1723 	u32 func_id;
1724 	char *name;
1725 	initfunc *init;
1726 };
1727 
1728 static struct ioc_iommu ioc_iommu_info[] __initdata = {
1729 	{ ZX1_IOC_ID, "zx1", ioc_zx1_init },
1730 	{ ZX2_IOC_ID, "zx2", NULL },
1731 	{ SX1000_IOC_ID, "sx1000", NULL },
1732 	{ SX2000_IOC_ID, "sx2000", NULL },
1733 };
1734 
1735 static struct ioc * __init
1736 ioc_init(u64 hpa, void *handle)
1737 {
1738 	struct ioc *ioc;
1739 	struct ioc_iommu *info;
1740 
1741 	ioc = kzalloc(sizeof(*ioc), GFP_KERNEL);
1742 	if (!ioc)
1743 		return NULL;
1744 
1745 	ioc->next = ioc_list;
1746 	ioc_list = ioc;
1747 
1748 	ioc->handle = handle;
1749 	ioc->ioc_hpa = ioremap(hpa, 0x1000);
1750 
1751 	ioc->func_id = READ_REG(ioc->ioc_hpa + IOC_FUNC_ID);
1752 	ioc->rev = READ_REG(ioc->ioc_hpa + IOC_FCLASS) & 0xFFUL;
1753 	ioc->dma_mask = 0xFFFFFFFFFFFFFFFFUL;	/* conservative */
1754 
1755 	for (info = ioc_iommu_info; info < ioc_iommu_info + ARRAY_SIZE(ioc_iommu_info); info++) {
1756 		if (ioc->func_id == info->func_id) {
1757 			ioc->name = info->name;
1758 			if (info->init)
1759 				(info->init)(ioc);
1760 		}
1761 	}
1762 
1763 	iovp_size = (1 << iovp_shift);
1764 	iovp_mask = ~(iovp_size - 1);
1765 
1766 	DBG_INIT("%s: PAGE_SIZE %ldK, iovp_size %ldK\n", __FUNCTION__,
1767 		PAGE_SIZE >> 10, iovp_size >> 10);
1768 
1769 	if (!ioc->name) {
1770 		ioc->name = kmalloc(24, GFP_KERNEL);
1771 		if (ioc->name)
1772 			sprintf((char *) ioc->name, "Unknown (%04x:%04x)",
1773 				ioc->func_id & 0xFFFF, (ioc->func_id >> 16) & 0xFFFF);
1774 		else
1775 			ioc->name = "Unknown";
1776 	}
1777 
1778 	ioc_iova_init(ioc);
1779 	ioc_resource_init(ioc);
1780 	ioc_sac_init(ioc);
1781 
1782 	if ((long) ~iovp_mask > (long) ia64_max_iommu_merge_mask)
1783 		ia64_max_iommu_merge_mask = ~iovp_mask;
1784 
1785 	printk(KERN_INFO PFX
1786 		"%s %d.%d HPA 0x%lx IOVA space %dMb at 0x%lx\n",
1787 		ioc->name, (ioc->rev >> 4) & 0xF, ioc->rev & 0xF,
1788 		hpa, ioc->iov_size >> 20, ioc->ibase);
1789 
1790 	return ioc;
1791 }
1792 
1793 
1794 
1795 /**************************************************************************
1796 **
1797 **   SBA initialization code (HW and SW)
1798 **
1799 **   o identify SBA chip itself
1800 **   o FIXME: initialize DMA hints for reasonable defaults
1801 **
1802 **************************************************************************/
1803 
1804 #ifdef CONFIG_PROC_FS
1805 static void *
1806 ioc_start(struct seq_file *s, loff_t *pos)
1807 {
1808 	struct ioc *ioc;
1809 	loff_t n = *pos;
1810 
1811 	for (ioc = ioc_list; ioc; ioc = ioc->next)
1812 		if (!n--)
1813 			return ioc;
1814 
1815 	return NULL;
1816 }
1817 
1818 static void *
1819 ioc_next(struct seq_file *s, void *v, loff_t *pos)
1820 {
1821 	struct ioc *ioc = v;
1822 
1823 	++*pos;
1824 	return ioc->next;
1825 }
1826 
1827 static void
1828 ioc_stop(struct seq_file *s, void *v)
1829 {
1830 }
1831 
1832 static int
1833 ioc_show(struct seq_file *s, void *v)
1834 {
1835 	struct ioc *ioc = v;
1836 	unsigned long *res_ptr = (unsigned long *)ioc->res_map;
1837 	int i, used = 0;
1838 
1839 	seq_printf(s, "Hewlett Packard %s IOC rev %d.%d\n",
1840 		ioc->name, ((ioc->rev >> 4) & 0xF), (ioc->rev & 0xF));
1841 #ifdef CONFIG_NUMA
1842 	if (ioc->node != MAX_NUMNODES)
1843 		seq_printf(s, "NUMA node       : %d\n", ioc->node);
1844 #endif
1845 	seq_printf(s, "IOVA size       : %ld MB\n", ((ioc->pdir_size >> 3) * iovp_size)/(1024*1024));
1846 	seq_printf(s, "IOVA page size  : %ld kb\n", iovp_size/1024);
1847 
1848 	for (i = 0; i < (ioc->res_size / sizeof(unsigned long)); ++i, ++res_ptr)
1849 		used += hweight64(*res_ptr);
1850 
1851 	seq_printf(s, "PDIR size       : %d entries\n", ioc->pdir_size >> 3);
1852 	seq_printf(s, "PDIR used       : %d entries\n", used);
1853 
1854 #ifdef PDIR_SEARCH_TIMING
1855 	{
1856 		unsigned long i = 0, avg = 0, min, max;
1857 		min = max = ioc->avg_search[0];
1858 		for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1859 			avg += ioc->avg_search[i];
1860 			if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1861 			if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1862 		}
1863 		avg /= SBA_SEARCH_SAMPLE;
1864 		seq_printf(s, "Bitmap search   : %ld/%ld/%ld (min/avg/max CPU Cycles/IOVA page)\n",
1865 		           min, avg, max);
1866 	}
1867 #endif
1868 #ifndef ALLOW_IOV_BYPASS
1869 	 seq_printf(s, "IOVA bypass disabled\n");
1870 #endif
1871 	return 0;
1872 }
1873 
1874 static struct seq_operations ioc_seq_ops = {
1875 	.start = ioc_start,
1876 	.next  = ioc_next,
1877 	.stop  = ioc_stop,
1878 	.show  = ioc_show
1879 };
1880 
1881 static int
1882 ioc_open(struct inode *inode, struct file *file)
1883 {
1884 	return seq_open(file, &ioc_seq_ops);
1885 }
1886 
1887 static const struct file_operations ioc_fops = {
1888 	.open    = ioc_open,
1889 	.read    = seq_read,
1890 	.llseek  = seq_lseek,
1891 	.release = seq_release
1892 };
1893 
1894 static void __init
1895 ioc_proc_init(void)
1896 {
1897 	struct proc_dir_entry *dir, *entry;
1898 
1899 	dir = proc_mkdir("bus/mckinley", NULL);
1900 	if (!dir)
1901 		return;
1902 
1903 	entry = create_proc_entry(ioc_list->name, 0, dir);
1904 	if (entry)
1905 		entry->proc_fops = &ioc_fops;
1906 }
1907 #endif
1908 
1909 static void
1910 sba_connect_bus(struct pci_bus *bus)
1911 {
1912 	acpi_handle handle, parent;
1913 	acpi_status status;
1914 	struct ioc *ioc;
1915 
1916 	if (!PCI_CONTROLLER(bus))
1917 		panic(PFX "no sysdata on bus %d!\n", bus->number);
1918 
1919 	if (PCI_CONTROLLER(bus)->iommu)
1920 		return;
1921 
1922 	handle = PCI_CONTROLLER(bus)->acpi_handle;
1923 	if (!handle)
1924 		return;
1925 
1926 	/*
1927 	 * The IOC scope encloses PCI root bridges in the ACPI
1928 	 * namespace, so work our way out until we find an IOC we
1929 	 * claimed previously.
1930 	 */
1931 	do {
1932 		for (ioc = ioc_list; ioc; ioc = ioc->next)
1933 			if (ioc->handle == handle) {
1934 				PCI_CONTROLLER(bus)->iommu = ioc;
1935 				return;
1936 			}
1937 
1938 		status = acpi_get_parent(handle, &parent);
1939 		handle = parent;
1940 	} while (ACPI_SUCCESS(status));
1941 
1942 	printk(KERN_WARNING "No IOC for PCI Bus %04x:%02x in ACPI\n", pci_domain_nr(bus), bus->number);
1943 }
1944 
1945 #ifdef CONFIG_NUMA
1946 static void __init
1947 sba_map_ioc_to_node(struct ioc *ioc, acpi_handle handle)
1948 {
1949 	unsigned int node;
1950 	int pxm;
1951 
1952 	ioc->node = MAX_NUMNODES;
1953 
1954 	pxm = acpi_get_pxm(handle);
1955 
1956 	if (pxm < 0)
1957 		return;
1958 
1959 	node = pxm_to_node(pxm);
1960 
1961 	if (node >= MAX_NUMNODES || !node_online(node))
1962 		return;
1963 
1964 	ioc->node = node;
1965 	return;
1966 }
1967 #else
1968 #define sba_map_ioc_to_node(ioc, handle)
1969 #endif
1970 
1971 static int __init
1972 acpi_sba_ioc_add(struct acpi_device *device)
1973 {
1974 	struct ioc *ioc;
1975 	acpi_status status;
1976 	u64 hpa, length;
1977 	struct acpi_buffer buffer;
1978 	struct acpi_device_info *dev_info;
1979 
1980 	status = hp_acpi_csr_space(device->handle, &hpa, &length);
1981 	if (ACPI_FAILURE(status))
1982 		return 1;
1983 
1984 	buffer.length = ACPI_ALLOCATE_LOCAL_BUFFER;
1985 	status = acpi_get_object_info(device->handle, &buffer);
1986 	if (ACPI_FAILURE(status))
1987 		return 1;
1988 	dev_info = buffer.pointer;
1989 
1990 	/*
1991 	 * For HWP0001, only SBA appears in ACPI namespace.  It encloses the PCI
1992 	 * root bridges, and its CSR space includes the IOC function.
1993 	 */
1994 	if (strncmp("HWP0001", dev_info->hardware_id.value, 7) == 0) {
1995 		hpa += ZX1_IOC_OFFSET;
1996 		/* zx1 based systems default to kernel page size iommu pages */
1997 		if (!iovp_shift)
1998 			iovp_shift = min(PAGE_SHIFT, 16);
1999 	}
2000 	kfree(dev_info);
2001 
2002 	/*
2003 	 * default anything not caught above or specified on cmdline to 4k
2004 	 * iommu page size
2005 	 */
2006 	if (!iovp_shift)
2007 		iovp_shift = 12;
2008 
2009 	ioc = ioc_init(hpa, device->handle);
2010 	if (!ioc)
2011 		return 1;
2012 
2013 	/* setup NUMA node association */
2014 	sba_map_ioc_to_node(ioc, device->handle);
2015 	return 0;
2016 }
2017 
2018 static const struct acpi_device_id hp_ioc_iommu_device_ids[] = {
2019 	{"HWP0001", 0},
2020 	{"HWP0004", 0},
2021 	{"", 0},
2022 };
2023 static struct acpi_driver acpi_sba_ioc_driver = {
2024 	.name		= "IOC IOMMU Driver",
2025 	.ids		= hp_ioc_iommu_device_ids,
2026 	.ops		= {
2027 		.add	= acpi_sba_ioc_add,
2028 	},
2029 };
2030 
2031 static int __init
2032 sba_init(void)
2033 {
2034 	if (!ia64_platform_is("hpzx1") && !ia64_platform_is("hpzx1_swiotlb"))
2035 		return 0;
2036 
2037 #if defined(CONFIG_IA64_GENERIC) && defined(CONFIG_CRASH_DUMP)
2038 	/* If we are booting a kdump kernel, the sba_iommu will
2039 	 * cause devices that were not shutdown properly to MCA
2040 	 * as soon as they are turned back on.  Our only option for
2041 	 * a successful kdump kernel boot is to use the swiotlb.
2042 	 */
2043 	if (elfcorehdr_addr < ELFCORE_ADDR_MAX) {
2044 		if (swiotlb_late_init_with_default_size(64 * (1<<20)) != 0)
2045 			panic("Unable to initialize software I/O TLB:"
2046 				  " Try machvec=dig boot option");
2047 		machvec_init("dig");
2048 		return 0;
2049 	}
2050 #endif
2051 
2052 	acpi_bus_register_driver(&acpi_sba_ioc_driver);
2053 	if (!ioc_list) {
2054 #ifdef CONFIG_IA64_GENERIC
2055 		/*
2056 		 * If we didn't find something sba_iommu can claim, we
2057 		 * need to setup the swiotlb and switch to the dig machvec.
2058 		 */
2059 		if (swiotlb_late_init_with_default_size(64 * (1<<20)) != 0)
2060 			panic("Unable to find SBA IOMMU or initialize "
2061 			      "software I/O TLB: Try machvec=dig boot option");
2062 		machvec_init("dig");
2063 #else
2064 		panic("Unable to find SBA IOMMU: Try a generic or DIG kernel");
2065 #endif
2066 		return 0;
2067 	}
2068 
2069 #if defined(CONFIG_IA64_GENERIC) || defined(CONFIG_IA64_HP_ZX1_SWIOTLB)
2070 	/*
2071 	 * hpzx1_swiotlb needs to have a fairly small swiotlb bounce
2072 	 * buffer setup to support devices with smaller DMA masks than
2073 	 * sba_iommu can handle.
2074 	 */
2075 	if (ia64_platform_is("hpzx1_swiotlb")) {
2076 		extern void hwsw_init(void);
2077 
2078 		hwsw_init();
2079 	}
2080 #endif
2081 
2082 #ifdef CONFIG_PCI
2083 	{
2084 		struct pci_bus *b = NULL;
2085 		while ((b = pci_find_next_bus(b)) != NULL)
2086 			sba_connect_bus(b);
2087 	}
2088 #endif
2089 
2090 #ifdef CONFIG_PROC_FS
2091 	ioc_proc_init();
2092 #endif
2093 	return 0;
2094 }
2095 
2096 subsys_initcall(sba_init); /* must be initialized after ACPI etc., but before any drivers... */
2097 
2098 static int __init
2099 nosbagart(char *str)
2100 {
2101 	reserve_sba_gart = 0;
2102 	return 1;
2103 }
2104 
2105 int
2106 sba_dma_supported (struct device *dev, u64 mask)
2107 {
2108 	/* make sure it's at least 32bit capable */
2109 	return ((mask & 0xFFFFFFFFUL) == 0xFFFFFFFFUL);
2110 }
2111 
2112 int
2113 sba_dma_mapping_error (dma_addr_t dma_addr)
2114 {
2115 	return 0;
2116 }
2117 
2118 __setup("nosbagart", nosbagart);
2119 
2120 static int __init
2121 sba_page_override(char *str)
2122 {
2123 	unsigned long page_size;
2124 
2125 	page_size = memparse(str, &str);
2126 	switch (page_size) {
2127 		case 4096:
2128 		case 8192:
2129 		case 16384:
2130 		case 65536:
2131 			iovp_shift = ffs(page_size) - 1;
2132 			break;
2133 		default:
2134 			printk("%s: unknown/unsupported iommu page size %ld\n",
2135 			       __FUNCTION__, page_size);
2136 	}
2137 
2138 	return 1;
2139 }
2140 
2141 __setup("sbapagesize=",sba_page_override);
2142 
2143 EXPORT_SYMBOL(sba_dma_mapping_error);
2144 EXPORT_SYMBOL(sba_map_single);
2145 EXPORT_SYMBOL(sba_unmap_single);
2146 EXPORT_SYMBOL(sba_map_sg);
2147 EXPORT_SYMBOL(sba_unmap_sg);
2148 EXPORT_SYMBOL(sba_dma_supported);
2149 EXPORT_SYMBOL(sba_alloc_coherent);
2150 EXPORT_SYMBOL(sba_free_coherent);
2151