xref: /openbmc/linux/drivers/parisc/sba_iommu.c (revision 78700c0a)
1 /*
2 **  System Bus Adapter (SBA) I/O MMU manager
3 **
4 **	(c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
5 **	(c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
6 **	(c) Copyright 2000-2004 Hewlett-Packard Company
7 **
8 **	Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
9 **
10 **	This program is free software; you can redistribute it and/or modify
11 **	it under the terms of the GNU General Public License as published by
12 **      the Free Software Foundation; either version 2 of the License, or
13 **      (at your option) any later version.
14 **
15 **
16 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
17 ** J5000/J7000/N-class/L-class machines and their successors.
18 **
19 ** FIXME: add DMA hint support programming in both sba and lba modules.
20 */
21 
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/spinlock.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
27 
28 #include <linux/mm.h>
29 #include <linux/string.h>
30 #include <linux/pci.h>
31 #include <linux/scatterlist.h>
32 #include <linux/iommu-helper.h>
33 
34 #include <asm/byteorder.h>
35 #include <asm/io.h>
36 #include <asm/dma.h>		/* for DMA_CHUNK_SIZE */
37 
38 #include <asm/hardware.h>	/* for register_parisc_driver() stuff */
39 
40 #include <linux/proc_fs.h>
41 #include <linux/seq_file.h>
42 #include <linux/module.h>
43 
44 #include <asm/ropes.h>
45 #include <asm/mckinley.h>	/* for proc_mckinley_root */
46 #include <asm/runway.h>		/* for proc_runway_root */
47 #include <asm/page.h>		/* for PAGE0 */
48 #include <asm/pdc.h>		/* for PDC_MODEL_* */
49 #include <asm/pdcpat.h>		/* for is_pdc_pat() */
50 #include <asm/parisc-device.h>
51 
52 #define MODULE_NAME "SBA"
53 
54 /*
55 ** The number of debug flags is a clue - this code is fragile.
56 ** Don't even think about messing with it unless you have
57 ** plenty of 710's to sacrifice to the computer gods. :^)
58 */
59 #undef DEBUG_SBA_INIT
60 #undef DEBUG_SBA_RUN
61 #undef DEBUG_SBA_RUN_SG
62 #undef DEBUG_SBA_RESOURCE
63 #undef ASSERT_PDIR_SANITY
64 #undef DEBUG_LARGE_SG_ENTRIES
65 #undef DEBUG_DMB_TRAP
66 
67 #ifdef DEBUG_SBA_INIT
68 #define DBG_INIT(x...)	printk(x)
69 #else
70 #define DBG_INIT(x...)
71 #endif
72 
73 #ifdef DEBUG_SBA_RUN
74 #define DBG_RUN(x...)	printk(x)
75 #else
76 #define DBG_RUN(x...)
77 #endif
78 
79 #ifdef DEBUG_SBA_RUN_SG
80 #define DBG_RUN_SG(x...)	printk(x)
81 #else
82 #define DBG_RUN_SG(x...)
83 #endif
84 
85 
86 #ifdef DEBUG_SBA_RESOURCE
87 #define DBG_RES(x...)	printk(x)
88 #else
89 #define DBG_RES(x...)
90 #endif
91 
92 #define SBA_INLINE	__inline__
93 
94 #define DEFAULT_DMA_HINT_REG	0
95 
96 struct sba_device *sba_list;
97 EXPORT_SYMBOL_GPL(sba_list);
98 
99 static unsigned long ioc_needs_fdc = 0;
100 
101 /* global count of IOMMUs in the system */
102 static unsigned int global_ioc_cnt = 0;
103 
104 /* PA8700 (Piranha 2.2) bug workaround */
105 static unsigned long piranha_bad_128k = 0;
106 
107 /* Looks nice and keeps the compiler happy */
108 #define SBA_DEV(d) ((struct sba_device *) (d))
109 
110 #ifdef CONFIG_AGP_PARISC
111 #define SBA_AGP_SUPPORT
112 #endif /*CONFIG_AGP_PARISC*/
113 
114 #ifdef SBA_AGP_SUPPORT
115 static int sba_reserve_agpgart = 1;
116 module_param(sba_reserve_agpgart, int, 0444);
117 MODULE_PARM_DESC(sba_reserve_agpgart, "Reserve half of IO pdir as AGPGART");
118 #endif
119 
120 
121 /************************************
122 ** SBA register read and write support
123 **
124 ** BE WARNED: register writes are posted.
125 **  (ie follow writes which must reach HW with a read)
126 **
127 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
128 */
129 #define READ_REG32(addr)	readl(addr)
130 #define READ_REG64(addr)	readq(addr)
131 #define WRITE_REG32(val, addr)	writel((val), (addr))
132 #define WRITE_REG64(val, addr)	writeq((val), (addr))
133 
134 #ifdef CONFIG_64BIT
135 #define READ_REG(addr)		READ_REG64(addr)
136 #define WRITE_REG(value, addr)	WRITE_REG64(value, addr)
137 #else
138 #define READ_REG(addr)		READ_REG32(addr)
139 #define WRITE_REG(value, addr)	WRITE_REG32(value, addr)
140 #endif
141 
142 #ifdef DEBUG_SBA_INIT
143 
144 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
145 
146 /**
147  * sba_dump_ranges - debugging only - print ranges assigned to this IOA
148  * @hpa: base address of the sba
149  *
150  * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
151  * IO Adapter (aka Bus Converter).
152  */
153 static void
154 sba_dump_ranges(void __iomem *hpa)
155 {
156 	DBG_INIT("SBA at 0x%p\n", hpa);
157 	DBG_INIT("IOS_DIST_BASE   : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
158 	DBG_INIT("IOS_DIST_MASK   : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
159 	DBG_INIT("IOS_DIST_ROUTE  : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
160 	DBG_INIT("\n");
161 	DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
162 	DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
163 	DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
164 }
165 
166 /**
167  * sba_dump_tlb - debugging only - print IOMMU operating parameters
168  * @hpa: base address of the IOMMU
169  *
170  * Print the size/location of the IO MMU PDIR.
171  */
172 static void sba_dump_tlb(void __iomem *hpa)
173 {
174 	DBG_INIT("IO TLB at 0x%p\n", hpa);
175 	DBG_INIT("IOC_IBASE    : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
176 	DBG_INIT("IOC_IMASK    : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
177 	DBG_INIT("IOC_TCNFG    : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
178 	DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
179 	DBG_INIT("\n");
180 }
181 #else
182 #define sba_dump_ranges(x)
183 #define sba_dump_tlb(x)
184 #endif	/* DEBUG_SBA_INIT */
185 
186 
187 #ifdef ASSERT_PDIR_SANITY
188 
189 /**
190  * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
191  * @ioc: IO MMU structure which owns the pdir we are interested in.
192  * @msg: text to print ont the output line.
193  * @pide: pdir index.
194  *
195  * Print one entry of the IO MMU PDIR in human readable form.
196  */
197 static void
198 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
199 {
200 	/* start printing from lowest pde in rval */
201 	u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
202 	unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
203 	uint rcnt;
204 
205 	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
206 		 msg,
207 		 rptr, pide & (BITS_PER_LONG - 1), *rptr);
208 
209 	rcnt = 0;
210 	while (rcnt < BITS_PER_LONG) {
211 		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
212 			(rcnt == (pide & (BITS_PER_LONG - 1)))
213 				? "    -->" : "       ",
214 			rcnt, ptr, *ptr );
215 		rcnt++;
216 		ptr++;
217 	}
218 	printk(KERN_DEBUG "%s", msg);
219 }
220 
221 
222 /**
223  * sba_check_pdir - debugging only - consistency checker
224  * @ioc: IO MMU structure which owns the pdir we are interested in.
225  * @msg: text to print ont the output line.
226  *
227  * Verify the resource map and pdir state is consistent
228  */
229 static int
230 sba_check_pdir(struct ioc *ioc, char *msg)
231 {
232 	u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
233 	u32 *rptr = (u32 *) ioc->res_map;	/* resource map ptr */
234 	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
235 	uint pide = 0;
236 
237 	while (rptr < rptr_end) {
238 		u32 rval = *rptr;
239 		int rcnt = 32;	/* number of bits we might check */
240 
241 		while (rcnt) {
242 			/* Get last byte and highest bit from that */
243 			u32 pde = ((u32) (((char *)pptr)[7])) << 24;
244 			if ((rval ^ pde) & 0x80000000)
245 			{
246 				/*
247 				** BUMMER!  -- res_map != pdir --
248 				** Dump rval and matching pdir entries
249 				*/
250 				sba_dump_pdir_entry(ioc, msg, pide);
251 				return(1);
252 			}
253 			rcnt--;
254 			rval <<= 1;	/* try the next bit */
255 			pptr++;
256 			pide++;
257 		}
258 		rptr++;	/* look at next word of res_map */
259 	}
260 	/* It'd be nice if we always got here :^) */
261 	return 0;
262 }
263 
264 
265 /**
266  * sba_dump_sg - debugging only - print Scatter-Gather list
267  * @ioc: IO MMU structure which owns the pdir we are interested in.
268  * @startsg: head of the SG list
269  * @nents: number of entries in SG list
270  *
271  * print the SG list so we can verify it's correct by hand.
272  */
273 static void
274 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
275 {
276 	while (nents-- > 0) {
277 		printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
278 				nents,
279 				(unsigned long) sg_dma_address(startsg),
280 				sg_dma_len(startsg),
281 				sg_virt(startsg), startsg->length);
282 		startsg++;
283 	}
284 }
285 
286 #endif /* ASSERT_PDIR_SANITY */
287 
288 
289 
290 
291 /**************************************************************
292 *
293 *   I/O Pdir Resource Management
294 *
295 *   Bits set in the resource map are in use.
296 *   Each bit can represent a number of pages.
297 *   LSbs represent lower addresses (IOVA's).
298 *
299 ***************************************************************/
300 #define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
301 
302 /* Convert from IOVP to IOVA and vice versa. */
303 
304 #ifdef ZX1_SUPPORT
305 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
306 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
307 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
308 #else
309 /* only support Astro and ancestors. Saves a few cycles in key places */
310 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
311 #define SBA_IOVP(ioc,iova) (iova)
312 #endif
313 
314 #define PDIR_INDEX(iovp)   ((iovp)>>IOVP_SHIFT)
315 
316 #define RESMAP_MASK(n)    (~0UL << (BITS_PER_LONG - (n)))
317 #define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)
318 
319 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
320 				 unsigned int bitshiftcnt)
321 {
322 	return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
323 		+ bitshiftcnt;
324 }
325 
326 /**
327  * sba_search_bitmap - find free space in IO PDIR resource bitmap
328  * @ioc: IO MMU structure which owns the pdir we are interested in.
329  * @bits_wanted: number of entries we need.
330  *
331  * Find consecutive free bits in resource bitmap.
332  * Each bit represents one entry in the IO Pdir.
333  * Cool perf optimization: search for log2(size) bits at a time.
334  */
335 static SBA_INLINE unsigned long
336 sba_search_bitmap(struct ioc *ioc, struct device *dev,
337 		  unsigned long bits_wanted)
338 {
339 	unsigned long *res_ptr = ioc->res_hint;
340 	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
341 	unsigned long pide = ~0UL, tpide;
342 	unsigned long boundary_size;
343 	unsigned long shift;
344 	int ret;
345 
346 	boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
347 			      1ULL << IOVP_SHIFT) >> IOVP_SHIFT;
348 
349 #if defined(ZX1_SUPPORT)
350 	BUG_ON(ioc->ibase & ~IOVP_MASK);
351 	shift = ioc->ibase >> IOVP_SHIFT;
352 #else
353 	shift = 0;
354 #endif
355 
356 	if (bits_wanted > (BITS_PER_LONG/2)) {
357 		/* Search word at a time - no mask needed */
358 		for(; res_ptr < res_end; ++res_ptr) {
359 			tpide = ptr_to_pide(ioc, res_ptr, 0);
360 			ret = iommu_is_span_boundary(tpide, bits_wanted,
361 						     shift,
362 						     boundary_size);
363 			if ((*res_ptr == 0) && !ret) {
364 				*res_ptr = RESMAP_MASK(bits_wanted);
365 				pide = tpide;
366 				break;
367 			}
368 		}
369 		/* point to the next word on next pass */
370 		res_ptr++;
371 		ioc->res_bitshift = 0;
372 	} else {
373 		/*
374 		** Search the resource bit map on well-aligned values.
375 		** "o" is the alignment.
376 		** We need the alignment to invalidate I/O TLB using
377 		** SBA HW features in the unmap path.
378 		*/
379 		unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
380 		uint bitshiftcnt = ALIGN(ioc->res_bitshift, o);
381 		unsigned long mask;
382 
383 		if (bitshiftcnt >= BITS_PER_LONG) {
384 			bitshiftcnt = 0;
385 			res_ptr++;
386 		}
387 		mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
388 
389 		DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
390 		while(res_ptr < res_end)
391 		{
392 			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
393 			WARN_ON(mask == 0);
394 			tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
395 			ret = iommu_is_span_boundary(tpide, bits_wanted,
396 						     shift,
397 						     boundary_size);
398 			if ((((*res_ptr) & mask) == 0) && !ret) {
399 				*res_ptr |= mask;     /* mark resources busy! */
400 				pide = tpide;
401 				break;
402 			}
403 			mask >>= o;
404 			bitshiftcnt += o;
405 			if (mask == 0) {
406 				mask = RESMAP_MASK(bits_wanted);
407 				bitshiftcnt=0;
408 				res_ptr++;
409 			}
410 		}
411 		/* look in the same word on the next pass */
412 		ioc->res_bitshift = bitshiftcnt + bits_wanted;
413 	}
414 
415 	/* wrapped ? */
416 	if (res_end <= res_ptr) {
417 		ioc->res_hint = (unsigned long *) ioc->res_map;
418 		ioc->res_bitshift = 0;
419 	} else {
420 		ioc->res_hint = res_ptr;
421 	}
422 	return (pide);
423 }
424 
425 
426 /**
427  * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
428  * @ioc: IO MMU structure which owns the pdir we are interested in.
429  * @size: number of bytes to create a mapping for
430  *
431  * Given a size, find consecutive unmarked and then mark those bits in the
432  * resource bit map.
433  */
434 static int
435 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
436 {
437 	unsigned int pages_needed = size >> IOVP_SHIFT;
438 #ifdef SBA_COLLECT_STATS
439 	unsigned long cr_start = mfctl(16);
440 #endif
441 	unsigned long pide;
442 
443 	pide = sba_search_bitmap(ioc, dev, pages_needed);
444 	if (pide >= (ioc->res_size << 3)) {
445 		pide = sba_search_bitmap(ioc, dev, pages_needed);
446 		if (pide >= (ioc->res_size << 3))
447 			panic("%s: I/O MMU @ %p is out of mapping resources\n",
448 			      __FILE__, ioc->ioc_hpa);
449 	}
450 
451 #ifdef ASSERT_PDIR_SANITY
452 	/* verify the first enable bit is clear */
453 	if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
454 		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
455 	}
456 #endif
457 
458 	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
459 		__func__, size, pages_needed, pide,
460 		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
461 		ioc->res_bitshift );
462 
463 #ifdef SBA_COLLECT_STATS
464 	{
465 		unsigned long cr_end = mfctl(16);
466 		unsigned long tmp = cr_end - cr_start;
467 		/* check for roll over */
468 		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
469 	}
470 	ioc->avg_search[ioc->avg_idx++] = cr_start;
471 	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
472 
473 	ioc->used_pages += pages_needed;
474 #endif
475 
476 	return (pide);
477 }
478 
479 
480 /**
481  * sba_free_range - unmark bits in IO PDIR resource bitmap
482  * @ioc: IO MMU structure which owns the pdir we are interested in.
483  * @iova: IO virtual address which was previously allocated.
484  * @size: number of bytes to create a mapping for
485  *
486  * clear bits in the ioc's resource map
487  */
488 static SBA_INLINE void
489 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
490 {
491 	unsigned long iovp = SBA_IOVP(ioc, iova);
492 	unsigned int pide = PDIR_INDEX(iovp);
493 	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
494 	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
495 
496 	int bits_not_wanted = size >> IOVP_SHIFT;
497 
498 	/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
499 	unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
500 
501 	DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
502 		__func__, (uint) iova, size,
503 		bits_not_wanted, m, pide, res_ptr, *res_ptr);
504 
505 #ifdef SBA_COLLECT_STATS
506 	ioc->used_pages -= bits_not_wanted;
507 #endif
508 
509 	*res_ptr &= ~m;
510 }
511 
512 
513 /**************************************************************
514 *
515 *   "Dynamic DMA Mapping" support (aka "Coherent I/O")
516 *
517 ***************************************************************/
518 
519 #ifdef SBA_HINT_SUPPORT
520 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
521 #endif
522 
523 typedef unsigned long space_t;
524 #define KERNEL_SPACE 0
525 
526 /**
527  * sba_io_pdir_entry - fill in one IO PDIR entry
528  * @pdir_ptr:  pointer to IO PDIR entry
529  * @sid: process Space ID - currently only support KERNEL_SPACE
530  * @vba: Virtual CPU address of buffer to map
531  * @hint: DMA hint set to use for this mapping
532  *
533  * SBA Mapping Routine
534  *
535  * Given a virtual address (vba, arg2) and space id, (sid, arg1)
536  * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
537  * pdir_ptr (arg0).
538  * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
539  * for Astro/Ike looks like:
540  *
541  *
542  *  0                    19                                 51   55       63
543  * +-+---------------------+----------------------------------+----+--------+
544  * |V|        U            |            PPN[43:12]            | U  |   VI   |
545  * +-+---------------------+----------------------------------+----+--------+
546  *
547  * Pluto is basically identical, supports fewer physical address bits:
548  *
549  *  0                       23                              51   55       63
550  * +-+------------------------+-------------------------------+----+--------+
551  * |V|        U               |         PPN[39:12]            | U  |   VI   |
552  * +-+------------------------+-------------------------------+----+--------+
553  *
554  *  V  == Valid Bit  (Most Significant Bit is bit 0)
555  *  U  == Unused
556  * PPN == Physical Page Number
557  * VI  == Virtual Index (aka Coherent Index)
558  *
559  * LPA instruction output is put into PPN field.
560  * LCI (Load Coherence Index) instruction provides the "VI" bits.
561  *
562  * We pre-swap the bytes since PCX-W is Big Endian and the
563  * IOMMU uses little endian for the pdir.
564  */
565 
566 static void SBA_INLINE
567 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
568 		  unsigned long hint)
569 {
570 	u64 pa; /* physical address */
571 	register unsigned ci; /* coherent index */
572 
573 	pa = virt_to_phys(vba);
574 	pa &= IOVP_MASK;
575 
576 	mtsp(sid,1);
577 	asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
578 	pa |= (ci >> PAGE_SHIFT) & 0xff;  /* move CI (8 bits) into lowest byte */
579 
580 	pa |= SBA_PDIR_VALID_BIT;	/* set "valid" bit */
581 	*pdir_ptr = cpu_to_le64(pa);	/* swap and store into I/O Pdir */
582 
583 	/*
584 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
585 	 * (bit #61, big endian), we have to flush and sync every time
586 	 * IO-PDIR is changed in Ike/Astro.
587 	 */
588 	if (ioc_needs_fdc)
589 		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
590 }
591 
592 
593 /**
594  * sba_mark_invalid - invalidate one or more IO PDIR entries
595  * @ioc: IO MMU structure which owns the pdir we are interested in.
596  * @iova:  IO Virtual Address mapped earlier
597  * @byte_cnt:  number of bytes this mapping covers.
598  *
599  * Marking the IO PDIR entry(ies) as Invalid and invalidate
600  * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
601  * is to purge stale entries in the IO TLB when unmapping entries.
602  *
603  * The PCOM register supports purging of multiple pages, with a minium
604  * of 1 page and a maximum of 2GB. Hardware requires the address be
605  * aligned to the size of the range being purged. The size of the range
606  * must be a power of 2. The "Cool perf optimization" in the
607  * allocation routine helps keep that true.
608  */
609 static SBA_INLINE void
610 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
611 {
612 	u32 iovp = (u32) SBA_IOVP(ioc,iova);
613 	u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
614 
615 #ifdef ASSERT_PDIR_SANITY
616 	/* Assert first pdir entry is set.
617 	**
618 	** Even though this is a big-endian machine, the entries
619 	** in the iopdir are little endian. That's why we look at
620 	** the byte at +7 instead of at +0.
621 	*/
622 	if (0x80 != (((u8 *) pdir_ptr)[7])) {
623 		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
624 	}
625 #endif
626 
627 	if (byte_cnt > IOVP_SIZE)
628 	{
629 #if 0
630 		unsigned long entries_per_cacheline = ioc_needs_fdc ?
631 				L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
632 					- (unsigned long) pdir_ptr;
633 				: 262144;
634 #endif
635 
636 		/* set "size" field for PCOM */
637 		iovp |= get_order(byte_cnt) + PAGE_SHIFT;
638 
639 		do {
640 			/* clear I/O Pdir entry "valid" bit first */
641 			((u8 *) pdir_ptr)[7] = 0;
642 			if (ioc_needs_fdc) {
643 				asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
644 #if 0
645 				entries_per_cacheline = L1_CACHE_SHIFT - 3;
646 #endif
647 			}
648 			pdir_ptr++;
649 			byte_cnt -= IOVP_SIZE;
650 		} while (byte_cnt > IOVP_SIZE);
651 	} else
652 		iovp |= IOVP_SHIFT;     /* set "size" field for PCOM */
653 
654 	/*
655 	** clear I/O PDIR entry "valid" bit.
656 	** We have to R/M/W the cacheline regardless how much of the
657 	** pdir entry that we clobber.
658 	** The rest of the entry would be useful for debugging if we
659 	** could dump core on HPMC.
660 	*/
661 	((u8 *) pdir_ptr)[7] = 0;
662 	if (ioc_needs_fdc)
663 		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
664 
665 	WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
666 }
667 
668 /**
669  * sba_dma_supported - PCI driver can query DMA support
670  * @dev: instance of PCI owned by the driver that's asking
671  * @mask:  number of address bits this PCI device can handle
672  *
673  * See Documentation/DMA-API-HOWTO.txt
674  */
675 static int sba_dma_supported( struct device *dev, u64 mask)
676 {
677 	struct ioc *ioc;
678 
679 	if (dev == NULL) {
680 		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
681 		BUG();
682 		return(0);
683 	}
684 
685 	/* Documentation/DMA-API-HOWTO.txt tells drivers to try 64-bit
686 	 * first, then fall back to 32-bit if that fails.
687 	 * We are just "encouraging" 32-bit DMA masks here since we can
688 	 * never allow IOMMU bypass unless we add special support for ZX1.
689 	 */
690 	if (mask > ~0U)
691 		return 0;
692 
693 	ioc = GET_IOC(dev);
694 
695 	/*
696 	 * check if mask is >= than the current max IO Virt Address
697 	 * The max IO Virt address will *always* < 30 bits.
698 	 */
699 	return((int)(mask >= (ioc->ibase - 1 +
700 			(ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
701 }
702 
703 
704 /**
705  * sba_map_single - map one buffer and return IOVA for DMA
706  * @dev: instance of PCI owned by the driver that's asking.
707  * @addr:  driver buffer to map.
708  * @size:  number of bytes to map in driver buffer.
709  * @direction:  R/W or both.
710  *
711  * See Documentation/DMA-API-HOWTO.txt
712  */
713 static dma_addr_t
714 sba_map_single(struct device *dev, void *addr, size_t size,
715 	       enum dma_data_direction direction)
716 {
717 	struct ioc *ioc;
718 	unsigned long flags;
719 	dma_addr_t iovp;
720 	dma_addr_t offset;
721 	u64 *pdir_start;
722 	int pide;
723 
724 	ioc = GET_IOC(dev);
725 
726 	/* save offset bits */
727 	offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
728 
729 	/* round up to nearest IOVP_SIZE */
730 	size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
731 
732 	spin_lock_irqsave(&ioc->res_lock, flags);
733 #ifdef ASSERT_PDIR_SANITY
734 	sba_check_pdir(ioc,"Check before sba_map_single()");
735 #endif
736 
737 #ifdef SBA_COLLECT_STATS
738 	ioc->msingle_calls++;
739 	ioc->msingle_pages += size >> IOVP_SHIFT;
740 #endif
741 	pide = sba_alloc_range(ioc, dev, size);
742 	iovp = (dma_addr_t) pide << IOVP_SHIFT;
743 
744 	DBG_RUN("%s() 0x%p -> 0x%lx\n",
745 		__func__, addr, (long) iovp | offset);
746 
747 	pdir_start = &(ioc->pdir_base[pide]);
748 
749 	while (size > 0) {
750 		sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
751 
752 		DBG_RUN("	pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
753 			pdir_start,
754 			(u8) (((u8 *) pdir_start)[7]),
755 			(u8) (((u8 *) pdir_start)[6]),
756 			(u8) (((u8 *) pdir_start)[5]),
757 			(u8) (((u8 *) pdir_start)[4]),
758 			(u8) (((u8 *) pdir_start)[3]),
759 			(u8) (((u8 *) pdir_start)[2]),
760 			(u8) (((u8 *) pdir_start)[1]),
761 			(u8) (((u8 *) pdir_start)[0])
762 			);
763 
764 		addr += IOVP_SIZE;
765 		size -= IOVP_SIZE;
766 		pdir_start++;
767 	}
768 
769 	/* force FDC ops in io_pdir_entry() to be visible to IOMMU */
770 	if (ioc_needs_fdc)
771 		asm volatile("sync" : : );
772 
773 #ifdef ASSERT_PDIR_SANITY
774 	sba_check_pdir(ioc,"Check after sba_map_single()");
775 #endif
776 	spin_unlock_irqrestore(&ioc->res_lock, flags);
777 
778 	/* form complete address */
779 	return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
780 }
781 
782 
783 static dma_addr_t
784 sba_map_page(struct device *dev, struct page *page, unsigned long offset,
785 		size_t size, enum dma_data_direction direction,
786 		struct dma_attrs *attrs)
787 {
788 	return sba_map_single(dev, page_address(page) + offset, size,
789 			direction);
790 }
791 
792 
793 /**
794  * sba_unmap_page - unmap one IOVA and free resources
795  * @dev: instance of PCI owned by the driver that's asking.
796  * @iova:  IOVA of driver buffer previously mapped.
797  * @size:  number of bytes mapped in driver buffer.
798  * @direction:  R/W or both.
799  *
800  * See Documentation/DMA-API-HOWTO.txt
801  */
802 static void
803 sba_unmap_page(struct device *dev, dma_addr_t iova, size_t size,
804 		enum dma_data_direction direction, struct dma_attrs *attrs)
805 {
806 	struct ioc *ioc;
807 #if DELAYED_RESOURCE_CNT > 0
808 	struct sba_dma_pair *d;
809 #endif
810 	unsigned long flags;
811 	dma_addr_t offset;
812 
813 	DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size);
814 
815 	ioc = GET_IOC(dev);
816 	offset = iova & ~IOVP_MASK;
817 	iova ^= offset;        /* clear offset bits */
818 	size += offset;
819 	size = ALIGN(size, IOVP_SIZE);
820 
821 	spin_lock_irqsave(&ioc->res_lock, flags);
822 
823 #ifdef SBA_COLLECT_STATS
824 	ioc->usingle_calls++;
825 	ioc->usingle_pages += size >> IOVP_SHIFT;
826 #endif
827 
828 	sba_mark_invalid(ioc, iova, size);
829 
830 #if DELAYED_RESOURCE_CNT > 0
831 	/* Delaying when we re-use a IO Pdir entry reduces the number
832 	 * of MMIO reads needed to flush writes to the PCOM register.
833 	 */
834 	d = &(ioc->saved[ioc->saved_cnt]);
835 	d->iova = iova;
836 	d->size = size;
837 	if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
838 		int cnt = ioc->saved_cnt;
839 		while (cnt--) {
840 			sba_free_range(ioc, d->iova, d->size);
841 			d--;
842 		}
843 		ioc->saved_cnt = 0;
844 
845 		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
846 	}
847 #else /* DELAYED_RESOURCE_CNT == 0 */
848 	sba_free_range(ioc, iova, size);
849 
850 	/* If fdc's were issued, force fdc's to be visible now */
851 	if (ioc_needs_fdc)
852 		asm volatile("sync" : : );
853 
854 	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
855 #endif /* DELAYED_RESOURCE_CNT == 0 */
856 
857 	spin_unlock_irqrestore(&ioc->res_lock, flags);
858 
859 	/* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
860 	** For Astro based systems this isn't a big deal WRT performance.
861 	** As long as 2.4 kernels copyin/copyout data from/to userspace,
862 	** we don't need the syncdma. The issue here is I/O MMU cachelines
863 	** are *not* coherent in all cases.  May be hwrev dependent.
864 	** Need to investigate more.
865 	asm volatile("syncdma");
866 	*/
867 }
868 
869 
870 /**
871  * sba_alloc - allocate/map shared mem for DMA
872  * @hwdev: instance of PCI owned by the driver that's asking.
873  * @size:  number of bytes mapped in driver buffer.
874  * @dma_handle:  IOVA of new buffer.
875  *
876  * See Documentation/DMA-API-HOWTO.txt
877  */
878 static void *sba_alloc(struct device *hwdev, size_t size, dma_addr_t *dma_handle,
879 		gfp_t gfp, struct dma_attrs *attrs)
880 {
881 	void *ret;
882 
883 	if (!hwdev) {
884 		/* only support PCI */
885 		*dma_handle = 0;
886 		return NULL;
887 	}
888 
889         ret = (void *) __get_free_pages(gfp, get_order(size));
890 
891 	if (ret) {
892 		memset(ret, 0, size);
893 		*dma_handle = sba_map_single(hwdev, ret, size, 0);
894 	}
895 
896 	return ret;
897 }
898 
899 
900 /**
901  * sba_free - free/unmap shared mem for DMA
902  * @hwdev: instance of PCI owned by the driver that's asking.
903  * @size:  number of bytes mapped in driver buffer.
904  * @vaddr:  virtual address IOVA of "consistent" buffer.
905  * @dma_handler:  IO virtual address of "consistent" buffer.
906  *
907  * See Documentation/DMA-API-HOWTO.txt
908  */
909 static void
910 sba_free(struct device *hwdev, size_t size, void *vaddr,
911 		    dma_addr_t dma_handle, struct dma_attrs *attrs)
912 {
913 	sba_unmap_page(hwdev, dma_handle, size, 0, NULL);
914 	free_pages((unsigned long) vaddr, get_order(size));
915 }
916 
917 
918 /*
919 ** Since 0 is a valid pdir_base index value, can't use that
920 ** to determine if a value is valid or not. Use a flag to indicate
921 ** the SG list entry contains a valid pdir index.
922 */
923 #define PIDE_FLAG 0x80000000UL
924 
925 #ifdef SBA_COLLECT_STATS
926 #define IOMMU_MAP_STATS
927 #endif
928 #include "iommu-helpers.h"
929 
930 #ifdef DEBUG_LARGE_SG_ENTRIES
931 int dump_run_sg = 0;
932 #endif
933 
934 
935 /**
936  * sba_map_sg - map Scatter/Gather list
937  * @dev: instance of PCI owned by the driver that's asking.
938  * @sglist:  array of buffer/length pairs
939  * @nents:  number of entries in list
940  * @direction:  R/W or both.
941  *
942  * See Documentation/DMA-API-HOWTO.txt
943  */
944 static int
945 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
946 	   enum dma_data_direction direction, struct dma_attrs *attrs)
947 {
948 	struct ioc *ioc;
949 	int coalesced, filled = 0;
950 	unsigned long flags;
951 
952 	DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
953 
954 	ioc = GET_IOC(dev);
955 
956 	/* Fast path single entry scatterlists. */
957 	if (nents == 1) {
958 		sg_dma_address(sglist) = sba_map_single(dev, sg_virt(sglist),
959 						sglist->length, direction);
960 		sg_dma_len(sglist)     = sglist->length;
961 		return 1;
962 	}
963 
964 	spin_lock_irqsave(&ioc->res_lock, flags);
965 
966 #ifdef ASSERT_PDIR_SANITY
967 	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
968 	{
969 		sba_dump_sg(ioc, sglist, nents);
970 		panic("Check before sba_map_sg()");
971 	}
972 #endif
973 
974 #ifdef SBA_COLLECT_STATS
975 	ioc->msg_calls++;
976 #endif
977 
978 	/*
979 	** First coalesce the chunks and allocate I/O pdir space
980 	**
981 	** If this is one DMA stream, we can properly map using the
982 	** correct virtual address associated with each DMA page.
983 	** w/o this association, we wouldn't have coherent DMA!
984 	** Access to the virtual address is what forces a two pass algorithm.
985 	*/
986 	coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, sba_alloc_range);
987 
988 	/*
989 	** Program the I/O Pdir
990 	**
991 	** map the virtual addresses to the I/O Pdir
992 	** o dma_address will contain the pdir index
993 	** o dma_len will contain the number of bytes to map
994 	** o address contains the virtual address.
995 	*/
996 	filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
997 
998 	/* force FDC ops in io_pdir_entry() to be visible to IOMMU */
999 	if (ioc_needs_fdc)
1000 		asm volatile("sync" : : );
1001 
1002 #ifdef ASSERT_PDIR_SANITY
1003 	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1004 	{
1005 		sba_dump_sg(ioc, sglist, nents);
1006 		panic("Check after sba_map_sg()\n");
1007 	}
1008 #endif
1009 
1010 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1011 
1012 	DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
1013 
1014 	return filled;
1015 }
1016 
1017 
1018 /**
1019  * sba_unmap_sg - unmap Scatter/Gather list
1020  * @dev: instance of PCI owned by the driver that's asking.
1021  * @sglist:  array of buffer/length pairs
1022  * @nents:  number of entries in list
1023  * @direction:  R/W or both.
1024  *
1025  * See Documentation/DMA-API-HOWTO.txt
1026  */
1027 static void
1028 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1029 	     enum dma_data_direction direction, struct dma_attrs *attrs)
1030 {
1031 	struct ioc *ioc;
1032 #ifdef ASSERT_PDIR_SANITY
1033 	unsigned long flags;
1034 #endif
1035 
1036 	DBG_RUN_SG("%s() START %d entries,  %p,%x\n",
1037 		__func__, nents, sg_virt(sglist), sglist->length);
1038 
1039 	ioc = GET_IOC(dev);
1040 
1041 #ifdef SBA_COLLECT_STATS
1042 	ioc->usg_calls++;
1043 #endif
1044 
1045 #ifdef ASSERT_PDIR_SANITY
1046 	spin_lock_irqsave(&ioc->res_lock, flags);
1047 	sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1048 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1049 #endif
1050 
1051 	while (sg_dma_len(sglist) && nents--) {
1052 
1053 		sba_unmap_page(dev, sg_dma_address(sglist), sg_dma_len(sglist),
1054 				direction, NULL);
1055 #ifdef SBA_COLLECT_STATS
1056 		ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1057 		ioc->usingle_calls--;	/* kluge since call is unmap_sg() */
1058 #endif
1059 		++sglist;
1060 	}
1061 
1062 	DBG_RUN_SG("%s() DONE (nents %d)\n", __func__,  nents);
1063 
1064 #ifdef ASSERT_PDIR_SANITY
1065 	spin_lock_irqsave(&ioc->res_lock, flags);
1066 	sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1067 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1068 #endif
1069 
1070 }
1071 
1072 static struct dma_map_ops sba_ops = {
1073 	.dma_supported =	sba_dma_supported,
1074 	.alloc =		sba_alloc,
1075 	.free =			sba_free,
1076 	.map_page =		sba_map_page,
1077 	.unmap_page =		sba_unmap_page,
1078 	.map_sg =		sba_map_sg,
1079 	.unmap_sg =		sba_unmap_sg,
1080 };
1081 
1082 
1083 /**************************************************************************
1084 **
1085 **   SBA PAT PDC support
1086 **
1087 **   o call pdc_pat_cell_module()
1088 **   o store ranges in PCI "resource" structures
1089 **
1090 **************************************************************************/
1091 
1092 static void
1093 sba_get_pat_resources(struct sba_device *sba_dev)
1094 {
1095 #if 0
1096 /*
1097 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1098 **      PAT PDC to program the SBA/LBA directed range registers...this
1099 **      burden may fall on the LBA code since it directly supports the
1100 **      PCI subsystem. It's not clear yet. - ggg
1101 */
1102 PAT_MOD(mod)->mod_info.mod_pages   = PAT_GET_MOD_PAGES(temp);
1103 	FIXME : ???
1104 PAT_MOD(mod)->mod_info.dvi         = PAT_GET_DVI(temp);
1105 	Tells where the dvi bits are located in the address.
1106 PAT_MOD(mod)->mod_info.ioc         = PAT_GET_IOC(temp);
1107 	FIXME : ???
1108 #endif
1109 }
1110 
1111 
1112 /**************************************************************
1113 *
1114 *   Initialization and claim
1115 *
1116 ***************************************************************/
1117 #define PIRANHA_ADDR_MASK	0x00160000UL /* bit 17,18,20 */
1118 #define PIRANHA_ADDR_VAL	0x00060000UL /* bit 17,18 on */
1119 static void *
1120 sba_alloc_pdir(unsigned int pdir_size)
1121 {
1122         unsigned long pdir_base;
1123 	unsigned long pdir_order = get_order(pdir_size);
1124 
1125 	pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1126 	if (NULL == (void *) pdir_base)	{
1127 		panic("%s() could not allocate I/O Page Table\n",
1128 			__func__);
1129 	}
1130 
1131 	/* If this is not PA8700 (PCX-W2)
1132 	**	OR newer than ver 2.2
1133 	**	OR in a system that doesn't need VINDEX bits from SBA,
1134 	**
1135 	** then we aren't exposed to the HW bug.
1136 	*/
1137 	if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1138 			|| (boot_cpu_data.pdc.versions > 0x202)
1139 			|| (boot_cpu_data.pdc.capabilities & 0x08L) )
1140 		return (void *) pdir_base;
1141 
1142 	/*
1143 	 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1144 	 *
1145 	 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1146 	 * Ike/Astro can cause silent data corruption. This is only
1147 	 * a problem if the I/O PDIR is located in memory such that
1148 	 * (little-endian)  bits 17 and 18 are on and bit 20 is off.
1149 	 *
1150 	 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1151 	 * right physical address, we can either avoid (IOPDIR <= 1MB)
1152 	 * or minimize (2MB IO Pdir) the problem if we restrict the
1153 	 * IO Pdir to a maximum size of 2MB-128K (1902K).
1154 	 *
1155 	 * Because we always allocate 2^N sized IO pdirs, either of the
1156 	 * "bad" regions will be the last 128K if at all. That's easy
1157 	 * to test for.
1158 	 *
1159 	 */
1160 	if (pdir_order <= (19-12)) {
1161 		if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1162 			/* allocate a new one on 512k alignment */
1163 			unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1164 			/* release original */
1165 			free_pages(pdir_base, pdir_order);
1166 
1167 			pdir_base = new_pdir;
1168 
1169 			/* release excess */
1170 			while (pdir_order < (19-12)) {
1171 				new_pdir += pdir_size;
1172 				free_pages(new_pdir, pdir_order);
1173 				pdir_order +=1;
1174 				pdir_size <<=1;
1175 			}
1176 		}
1177 	} else {
1178 		/*
1179 		** 1MB or 2MB Pdir
1180 		** Needs to be aligned on an "odd" 1MB boundary.
1181 		*/
1182 		unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1183 
1184 		/* release original */
1185 		free_pages( pdir_base, pdir_order);
1186 
1187 		/* release first 1MB */
1188 		free_pages(new_pdir, 20-12);
1189 
1190 		pdir_base = new_pdir + 1024*1024;
1191 
1192 		if (pdir_order > (20-12)) {
1193 			/*
1194 			** 2MB Pdir.
1195 			**
1196 			** Flag tells init_bitmap() to mark bad 128k as used
1197 			** and to reduce the size by 128k.
1198 			*/
1199 			piranha_bad_128k = 1;
1200 
1201 			new_pdir += 3*1024*1024;
1202 			/* release last 1MB */
1203 			free_pages(new_pdir, 20-12);
1204 
1205 			/* release unusable 128KB */
1206 			free_pages(new_pdir - 128*1024 , 17-12);
1207 
1208 			pdir_size -= 128*1024;
1209 		}
1210 	}
1211 
1212 	memset((void *) pdir_base, 0, pdir_size);
1213 	return (void *) pdir_base;
1214 }
1215 
1216 struct ibase_data_struct {
1217 	struct ioc *ioc;
1218 	int ioc_num;
1219 };
1220 
1221 static int setup_ibase_imask_callback(struct device *dev, void *data)
1222 {
1223 	/* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1224         extern void lba_set_iregs(struct parisc_device *, u32, u32);
1225 	struct parisc_device *lba = to_parisc_device(dev);
1226 	struct ibase_data_struct *ibd = data;
1227 	int rope_num = (lba->hpa.start >> 13) & 0xf;
1228 	if (rope_num >> 3 == ibd->ioc_num)
1229 		lba_set_iregs(lba, ibd->ioc->ibase, ibd->ioc->imask);
1230 	return 0;
1231 }
1232 
1233 /* setup Mercury or Elroy IBASE/IMASK registers. */
1234 static void
1235 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1236 {
1237 	struct ibase_data_struct ibase_data = {
1238 		.ioc		= ioc,
1239 		.ioc_num	= ioc_num,
1240 	};
1241 
1242 	device_for_each_child(&sba->dev, &ibase_data,
1243 			      setup_ibase_imask_callback);
1244 }
1245 
1246 #ifdef SBA_AGP_SUPPORT
1247 static int
1248 sba_ioc_find_quicksilver(struct device *dev, void *data)
1249 {
1250 	int *agp_found = data;
1251 	struct parisc_device *lba = to_parisc_device(dev);
1252 
1253 	if (IS_QUICKSILVER(lba))
1254 		*agp_found = 1;
1255 	return 0;
1256 }
1257 #endif
1258 
1259 static void
1260 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1261 {
1262 	u32 iova_space_mask;
1263 	u32 iova_space_size;
1264 	int iov_order, tcnfg;
1265 #ifdef SBA_AGP_SUPPORT
1266 	int agp_found = 0;
1267 #endif
1268 	/*
1269 	** Firmware programs the base and size of a "safe IOVA space"
1270 	** (one that doesn't overlap memory or LMMIO space) in the
1271 	** IBASE and IMASK registers.
1272 	*/
1273 	ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE);
1274 	iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1275 
1276 	if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1277 		printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1278 		iova_space_size /= 2;
1279 	}
1280 
1281 	/*
1282 	** iov_order is always based on a 1GB IOVA space since we want to
1283 	** turn on the other half for AGP GART.
1284 	*/
1285 	iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1286 	ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1287 
1288 	DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1289 		__func__, ioc->ioc_hpa, iova_space_size >> 20,
1290 		iov_order + PAGE_SHIFT);
1291 
1292 	ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1293 						   get_order(ioc->pdir_size));
1294 	if (!ioc->pdir_base)
1295 		panic("Couldn't allocate I/O Page Table\n");
1296 
1297 	memset(ioc->pdir_base, 0, ioc->pdir_size);
1298 
1299 	DBG_INIT("%s() pdir %p size %x\n",
1300 			__func__, ioc->pdir_base, ioc->pdir_size);
1301 
1302 #ifdef SBA_HINT_SUPPORT
1303 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1304 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1305 
1306 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1307 		ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1308 #endif
1309 
1310 	WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1311 	WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1312 
1313 	/* build IMASK for IOC and Elroy */
1314 	iova_space_mask =  0xffffffff;
1315 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1316 	ioc->imask = iova_space_mask;
1317 #ifdef ZX1_SUPPORT
1318 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1319 #endif
1320 	sba_dump_tlb(ioc->ioc_hpa);
1321 
1322 	setup_ibase_imask(sba, ioc, ioc_num);
1323 
1324 	WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1325 
1326 #ifdef CONFIG_64BIT
1327 	/*
1328 	** Setting the upper bits makes checking for bypass addresses
1329 	** a little faster later on.
1330 	*/
1331 	ioc->imask |= 0xFFFFFFFF00000000UL;
1332 #endif
1333 
1334 	/* Set I/O PDIR Page size to system page size */
1335 	switch (PAGE_SHIFT) {
1336 		case 12: tcnfg = 0; break;	/*  4K */
1337 		case 13: tcnfg = 1; break;	/*  8K */
1338 		case 14: tcnfg = 2; break;	/* 16K */
1339 		case 16: tcnfg = 3; break;	/* 64K */
1340 		default:
1341 			panic(__FILE__ "Unsupported system page size %d",
1342 				1 << PAGE_SHIFT);
1343 			break;
1344 	}
1345 	WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1346 
1347 	/*
1348 	** Program the IOC's ibase and enable IOVA translation
1349 	** Bit zero == enable bit.
1350 	*/
1351 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1352 
1353 	/*
1354 	** Clear I/O TLB of any possible entries.
1355 	** (Yes. This is a bit paranoid...but so what)
1356 	*/
1357 	WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1358 
1359 #ifdef SBA_AGP_SUPPORT
1360 
1361 	/*
1362 	** If an AGP device is present, only use half of the IOV space
1363 	** for PCI DMA.  Unfortunately we can't know ahead of time
1364 	** whether GART support will actually be used, for now we
1365 	** can just key on any AGP device found in the system.
1366 	** We program the next pdir index after we stop w/ a key for
1367 	** the GART code to handshake on.
1368 	*/
1369 	device_for_each_child(&sba->dev, &agp_found, sba_ioc_find_quicksilver);
1370 
1371 	if (agp_found && sba_reserve_agpgart) {
1372 		printk(KERN_INFO "%s: reserving %dMb of IOVA space for agpgart\n",
1373 		       __func__, (iova_space_size/2) >> 20);
1374 		ioc->pdir_size /= 2;
1375 		ioc->pdir_base[PDIR_INDEX(iova_space_size/2)] = SBA_AGPGART_COOKIE;
1376 	}
1377 #endif /*SBA_AGP_SUPPORT*/
1378 }
1379 
1380 static void
1381 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1382 {
1383 	u32 iova_space_size, iova_space_mask;
1384 	unsigned int pdir_size, iov_order, tcnfg;
1385 
1386 	/*
1387 	** Determine IOVA Space size from memory size.
1388 	**
1389 	** Ideally, PCI drivers would register the maximum number
1390 	** of DMA they can have outstanding for each device they
1391 	** own.  Next best thing would be to guess how much DMA
1392 	** can be outstanding based on PCI Class/sub-class. Both
1393 	** methods still require some "extra" to support PCI
1394 	** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1395 	**
1396 	** While we have 32-bits "IOVA" space, top two 2 bits are used
1397 	** for DMA hints - ergo only 30 bits max.
1398 	*/
1399 
1400 	iova_space_size = (u32) (totalram_pages/global_ioc_cnt);
1401 
1402 	/* limit IOVA space size to 1MB-1GB */
1403 	if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1404 		iova_space_size = 1 << (20 - PAGE_SHIFT);
1405 	}
1406 	else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1407 		iova_space_size = 1 << (30 - PAGE_SHIFT);
1408 	}
1409 
1410 	/*
1411 	** iova space must be log2() in size.
1412 	** thus, pdir/res_map will also be log2().
1413 	** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1414 	*/
1415 	iov_order = get_order(iova_space_size << PAGE_SHIFT);
1416 
1417 	/* iova_space_size is now bytes, not pages */
1418 	iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1419 
1420 	ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1421 
1422 	DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1423 			__func__,
1424 			ioc->ioc_hpa,
1425 			(unsigned long) totalram_pages >> (20 - PAGE_SHIFT),
1426 			iova_space_size>>20,
1427 			iov_order + PAGE_SHIFT);
1428 
1429 	ioc->pdir_base = sba_alloc_pdir(pdir_size);
1430 
1431 	DBG_INIT("%s() pdir %p size %x\n",
1432 			__func__, ioc->pdir_base, pdir_size);
1433 
1434 #ifdef SBA_HINT_SUPPORT
1435 	/* FIXME : DMA HINTs not used */
1436 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1437 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1438 
1439 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1440 			ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1441 #endif
1442 
1443 	WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1444 
1445 	/* build IMASK for IOC and Elroy */
1446 	iova_space_mask =  0xffffffff;
1447 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1448 
1449 	/*
1450 	** On C3000 w/512MB mem, HP-UX 10.20 reports:
1451 	**     ibase=0, imask=0xFE000000, size=0x2000000.
1452 	*/
1453 	ioc->ibase = 0;
1454 	ioc->imask = iova_space_mask;	/* save it */
1455 #ifdef ZX1_SUPPORT
1456 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1457 #endif
1458 
1459 	DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1460 		__func__, ioc->ibase, ioc->imask);
1461 
1462 	/*
1463 	** FIXME: Hint registers are programmed with default hint
1464 	** values during boot, so hints should be sane even if we
1465 	** can't reprogram them the way drivers want.
1466 	*/
1467 
1468 	setup_ibase_imask(sba, ioc, ioc_num);
1469 
1470 	/*
1471 	** Program the IOC's ibase and enable IOVA translation
1472 	*/
1473 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1474 	WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1475 
1476 	/* Set I/O PDIR Page size to system page size */
1477 	switch (PAGE_SHIFT) {
1478 		case 12: tcnfg = 0; break;	/*  4K */
1479 		case 13: tcnfg = 1; break;	/*  8K */
1480 		case 14: tcnfg = 2; break;	/* 16K */
1481 		case 16: tcnfg = 3; break;	/* 64K */
1482 		default:
1483 			panic(__FILE__ "Unsupported system page size %d",
1484 				1 << PAGE_SHIFT);
1485 			break;
1486 	}
1487 	/* Set I/O PDIR Page size to PAGE_SIZE (4k/16k/...) */
1488 	WRITE_REG(tcnfg, ioc->ioc_hpa+IOC_TCNFG);
1489 
1490 	/*
1491 	** Clear I/O TLB of any possible entries.
1492 	** (Yes. This is a bit paranoid...but so what)
1493 	*/
1494 	WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1495 
1496 	ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1497 
1498 	DBG_INIT("%s() DONE\n", __func__);
1499 }
1500 
1501 
1502 
1503 /**************************************************************************
1504 **
1505 **   SBA initialization code (HW and SW)
1506 **
1507 **   o identify SBA chip itself
1508 **   o initialize SBA chip modes (HardFail)
1509 **   o initialize SBA chip modes (HardFail)
1510 **   o FIXME: initialize DMA hints for reasonable defaults
1511 **
1512 **************************************************************************/
1513 
1514 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1515 {
1516 	return ioremap_nocache(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1517 }
1518 
1519 static void sba_hw_init(struct sba_device *sba_dev)
1520 {
1521 	int i;
1522 	int num_ioc;
1523 	u64 ioc_ctl;
1524 
1525 	if (!is_pdc_pat()) {
1526 		/* Shutdown the USB controller on Astro-based workstations.
1527 		** Once we reprogram the IOMMU, the next DMA performed by
1528 		** USB will HPMC the box. USB is only enabled if a
1529 		** keyboard is present and found.
1530 		**
1531 		** With serial console, j6k v5.0 firmware says:
1532 		**   mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1533 		**
1534 		** FIXME: Using GFX+USB console at power up but direct
1535 		**	linux to serial console is still broken.
1536 		**	USB could generate DMA so we must reset USB.
1537 		**	The proper sequence would be:
1538 		**	o block console output
1539 		**	o reset USB device
1540 		**	o reprogram serial port
1541 		**	o unblock console output
1542 		*/
1543 		if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1544 			pdc_io_reset_devices();
1545 		}
1546 
1547 	}
1548 
1549 
1550 #if 0
1551 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1552 	PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1553 
1554 	/*
1555 	** Need to deal with DMA from LAN.
1556 	**	Maybe use page zero boot device as a handle to talk
1557 	**	to PDC about which device to shutdown.
1558 	**
1559 	** Netbooting, j6k v5.0 firmware says:
1560 	** 	mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1561 	** ARGH! invalid class.
1562 	*/
1563 	if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1564 		&& (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1565 			pdc_io_reset();
1566 	}
1567 #endif
1568 
1569 	if (!IS_PLUTO(sba_dev->dev)) {
1570 		ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1571 		DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1572 			__func__, sba_dev->sba_hpa, ioc_ctl);
1573 		ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1574 		ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1575 			/* j6700 v1.6 firmware sets 0x294f */
1576 			/* A500 firmware sets 0x4d */
1577 
1578 		WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1579 
1580 #ifdef DEBUG_SBA_INIT
1581 		ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1582 		DBG_INIT(" 0x%Lx\n", ioc_ctl);
1583 #endif
1584 	} /* if !PLUTO */
1585 
1586 	if (IS_ASTRO(sba_dev->dev)) {
1587 		int err;
1588 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1589 		num_ioc = 1;
1590 
1591 		sba_dev->chip_resv.name = "Astro Intr Ack";
1592 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1593 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff000000UL - 1) ;
1594 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1595 		BUG_ON(err < 0);
1596 
1597 	} else if (IS_PLUTO(sba_dev->dev)) {
1598 		int err;
1599 
1600 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1601 		num_ioc = 1;
1602 
1603 		sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1604 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1605 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff200000UL - 1);
1606 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1607 		WARN_ON(err < 0);
1608 
1609 		sba_dev->iommu_resv.name = "IOVA Space";
1610 		sba_dev->iommu_resv.start = 0x40000000UL;
1611 		sba_dev->iommu_resv.end   = 0x50000000UL - 1;
1612 		err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1613 		WARN_ON(err < 0);
1614 	} else {
1615 		/* IKE, REO */
1616 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1617 		sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1618 		num_ioc = 2;
1619 
1620 		/* TODO - LOOKUP Ike/Stretch chipset mem map */
1621 	}
1622 	/* XXX: What about Reo Grande? */
1623 
1624 	sba_dev->num_ioc = num_ioc;
1625 	for (i = 0; i < num_ioc; i++) {
1626 		void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1627 		unsigned int j;
1628 
1629 		for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1630 
1631 			/*
1632 			 * Clear ROPE(N)_CONFIG AO bit.
1633 			 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1634 			 * Overrides bit 1 in DMA Hint Sets.
1635 			 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1636 			 */
1637 			if (IS_PLUTO(sba_dev->dev)) {
1638 				void __iomem *rope_cfg;
1639 				unsigned long cfg_val;
1640 
1641 				rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1642 				cfg_val = READ_REG(rope_cfg);
1643 				cfg_val &= ~IOC_ROPE_AO;
1644 				WRITE_REG(cfg_val, rope_cfg);
1645 			}
1646 
1647 			/*
1648 			** Make sure the box crashes on rope errors.
1649 			*/
1650 			WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1651 		}
1652 
1653 		/* flush out the last writes */
1654 		READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1655 
1656 		DBG_INIT("	ioc[%d] ROPE_CFG 0x%Lx  ROPE_DBG 0x%Lx\n",
1657 				i,
1658 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1659 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1660 			);
1661 		DBG_INIT("	STATUS_CONTROL 0x%Lx  FLUSH_CTRL 0x%Lx\n",
1662 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1663 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1664 			);
1665 
1666 		if (IS_PLUTO(sba_dev->dev)) {
1667 			sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1668 		} else {
1669 			sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1670 		}
1671 	}
1672 }
1673 
1674 static void
1675 sba_common_init(struct sba_device *sba_dev)
1676 {
1677 	int i;
1678 
1679 	/* add this one to the head of the list (order doesn't matter)
1680 	** This will be useful for debugging - especially if we get coredumps
1681 	*/
1682 	sba_dev->next = sba_list;
1683 	sba_list = sba_dev;
1684 
1685 	for(i=0; i< sba_dev->num_ioc; i++) {
1686 		int res_size;
1687 #ifdef DEBUG_DMB_TRAP
1688 		extern void iterate_pages(unsigned long , unsigned long ,
1689 					  void (*)(pte_t * , unsigned long),
1690 					  unsigned long );
1691 		void set_data_memory_break(pte_t * , unsigned long);
1692 #endif
1693 		/* resource map size dictated by pdir_size */
1694 		res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1695 
1696 		/* Second part of PIRANHA BUG */
1697 		if (piranha_bad_128k) {
1698 			res_size -= (128*1024)/sizeof(u64);
1699 		}
1700 
1701 		res_size >>= 3;  /* convert bit count to byte count */
1702 		DBG_INIT("%s() res_size 0x%x\n",
1703 			__func__, res_size);
1704 
1705 		sba_dev->ioc[i].res_size = res_size;
1706 		sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1707 
1708 #ifdef DEBUG_DMB_TRAP
1709 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1710 				set_data_memory_break, 0);
1711 #endif
1712 
1713 		if (NULL == sba_dev->ioc[i].res_map)
1714 		{
1715 			panic("%s:%s() could not allocate resource map\n",
1716 			      __FILE__, __func__ );
1717 		}
1718 
1719 		memset(sba_dev->ioc[i].res_map, 0, res_size);
1720 		/* next available IOVP - circular search */
1721 		sba_dev->ioc[i].res_hint = (unsigned long *)
1722 				&(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1723 
1724 #ifdef ASSERT_PDIR_SANITY
1725 		/* Mark first bit busy - ie no IOVA 0 */
1726 		sba_dev->ioc[i].res_map[0] = 0x80;
1727 		sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1728 #endif
1729 
1730 		/* Third (and last) part of PIRANHA BUG */
1731 		if (piranha_bad_128k) {
1732 			/* region from +1408K to +1536 is un-usable. */
1733 
1734 			int idx_start = (1408*1024/sizeof(u64)) >> 3;
1735 			int idx_end   = (1536*1024/sizeof(u64)) >> 3;
1736 			long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1737 			long *p_end   = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1738 
1739 			/* mark that part of the io pdir busy */
1740 			while (p_start < p_end)
1741 				*p_start++ = -1;
1742 
1743 		}
1744 
1745 #ifdef DEBUG_DMB_TRAP
1746 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1747 				set_data_memory_break, 0);
1748 		iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1749 				set_data_memory_break, 0);
1750 #endif
1751 
1752 		DBG_INIT("%s() %d res_map %x %p\n",
1753 			__func__, i, res_size, sba_dev->ioc[i].res_map);
1754 	}
1755 
1756 	spin_lock_init(&sba_dev->sba_lock);
1757 	ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1758 
1759 #ifdef DEBUG_SBA_INIT
1760 	/*
1761 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1762 	 * (bit #61, big endian), we have to flush and sync every time
1763 	 * IO-PDIR is changed in Ike/Astro.
1764 	 */
1765 	if (ioc_needs_fdc) {
1766 		printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1767 	} else {
1768 		printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1769 	}
1770 #endif
1771 }
1772 
1773 #ifdef CONFIG_PROC_FS
1774 static int sba_proc_info(struct seq_file *m, void *p)
1775 {
1776 	struct sba_device *sba_dev = sba_list;
1777 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1778 	int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1779 #ifdef SBA_COLLECT_STATS
1780 	unsigned long avg = 0, min, max;
1781 #endif
1782 	int i;
1783 
1784 	seq_printf(m, "%s rev %d.%d\n",
1785 		   sba_dev->name,
1786 		   (sba_dev->hw_rev & 0x7) + 1,
1787 		   (sba_dev->hw_rev & 0x18) >> 3);
1788 	seq_printf(m, "IO PDIR size    : %d bytes (%d entries)\n",
1789 		   (int)((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1790 		   total_pages);
1791 
1792 	seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1793 		   ioc->res_size, ioc->res_size << 3);   /* 8 bits per byte */
1794 
1795 	seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1796 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1797 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1798 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE));
1799 
1800 	for (i=0; i<4; i++)
1801 		seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n",
1802 			   i,
1803 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE  + i*0x18),
1804 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK  + i*0x18),
1805 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18));
1806 
1807 #ifdef SBA_COLLECT_STATS
1808 	seq_printf(m, "IO PDIR entries : %ld free  %ld used (%d%%)\n",
1809 		   total_pages - ioc->used_pages, ioc->used_pages,
1810 		   (int)(ioc->used_pages * 100 / total_pages));
1811 
1812 	min = max = ioc->avg_search[0];
1813 	for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1814 		avg += ioc->avg_search[i];
1815 		if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1816 		if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1817 	}
1818 	avg /= SBA_SEARCH_SAMPLE;
1819 	seq_printf(m, "  Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1820 		   min, avg, max);
1821 
1822 	seq_printf(m, "pci_map_single(): %12ld calls  %12ld pages (avg %d/1000)\n",
1823 		   ioc->msingle_calls, ioc->msingle_pages,
1824 		   (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1825 
1826 	/* KLUGE - unmap_sg calls unmap_single for each mapped page */
1827 	min = ioc->usingle_calls;
1828 	max = ioc->usingle_pages - ioc->usg_pages;
1829 	seq_printf(m, "pci_unmap_single: %12ld calls  %12ld pages (avg %d/1000)\n",
1830 		   min, max, (int)((max * 1000)/min));
1831 
1832 	seq_printf(m, "pci_map_sg()    : %12ld calls  %12ld pages (avg %d/1000)\n",
1833 		   ioc->msg_calls, ioc->msg_pages,
1834 		   (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1835 
1836 	seq_printf(m, "pci_unmap_sg()  : %12ld calls  %12ld pages (avg %d/1000)\n",
1837 		   ioc->usg_calls, ioc->usg_pages,
1838 		   (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1839 #endif
1840 
1841 	return 0;
1842 }
1843 
1844 static int
1845 sba_proc_open(struct inode *i, struct file *f)
1846 {
1847 	return single_open(f, &sba_proc_info, NULL);
1848 }
1849 
1850 static const struct file_operations sba_proc_fops = {
1851 	.owner = THIS_MODULE,
1852 	.open = sba_proc_open,
1853 	.read = seq_read,
1854 	.llseek = seq_lseek,
1855 	.release = single_release,
1856 };
1857 
1858 static int
1859 sba_proc_bitmap_info(struct seq_file *m, void *p)
1860 {
1861 	struct sba_device *sba_dev = sba_list;
1862 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1863 
1864 	seq_hex_dump(m, "   ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map,
1865 		     ioc->res_size, false);
1866 	seq_putc(m, '\n');
1867 
1868 	return 0;
1869 }
1870 
1871 static int
1872 sba_proc_bitmap_open(struct inode *i, struct file *f)
1873 {
1874 	return single_open(f, &sba_proc_bitmap_info, NULL);
1875 }
1876 
1877 static const struct file_operations sba_proc_bitmap_fops = {
1878 	.owner = THIS_MODULE,
1879 	.open = sba_proc_bitmap_open,
1880 	.read = seq_read,
1881 	.llseek = seq_lseek,
1882 	.release = single_release,
1883 };
1884 #endif /* CONFIG_PROC_FS */
1885 
1886 static struct parisc_device_id sba_tbl[] = {
1887 	{ HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
1888 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
1889 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
1890 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
1891 	{ HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
1892 	{ 0, }
1893 };
1894 
1895 static int sba_driver_callback(struct parisc_device *);
1896 
1897 static struct parisc_driver sba_driver = {
1898 	.name =		MODULE_NAME,
1899 	.id_table =	sba_tbl,
1900 	.probe =	sba_driver_callback,
1901 };
1902 
1903 /*
1904 ** Determine if sba should claim this chip (return 0) or not (return 1).
1905 ** If so, initialize the chip and tell other partners in crime they
1906 ** have work to do.
1907 */
1908 static int sba_driver_callback(struct parisc_device *dev)
1909 {
1910 	struct sba_device *sba_dev;
1911 	u32 func_class;
1912 	int i;
1913 	char *version;
1914 	void __iomem *sba_addr = ioremap_nocache(dev->hpa.start, SBA_FUNC_SIZE);
1915 #ifdef CONFIG_PROC_FS
1916 	struct proc_dir_entry *root;
1917 #endif
1918 
1919 	sba_dump_ranges(sba_addr);
1920 
1921 	/* Read HW Rev First */
1922 	func_class = READ_REG(sba_addr + SBA_FCLASS);
1923 
1924 	if (IS_ASTRO(dev)) {
1925 		unsigned long fclass;
1926 		static char astro_rev[]="Astro ?.?";
1927 
1928 		/* Astro is broken...Read HW Rev First */
1929 		fclass = READ_REG(sba_addr);
1930 
1931 		astro_rev[6] = '1' + (char) (fclass & 0x7);
1932 		astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
1933 		version = astro_rev;
1934 
1935 	} else if (IS_IKE(dev)) {
1936 		static char ike_rev[] = "Ike rev ?";
1937 		ike_rev[8] = '0' + (char) (func_class & 0xff);
1938 		version = ike_rev;
1939 	} else if (IS_PLUTO(dev)) {
1940 		static char pluto_rev[]="Pluto ?.?";
1941 		pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
1942 		pluto_rev[8] = '0' + (char) (func_class & 0x0f);
1943 		version = pluto_rev;
1944 	} else {
1945 		static char reo_rev[] = "REO rev ?";
1946 		reo_rev[8] = '0' + (char) (func_class & 0xff);
1947 		version = reo_rev;
1948 	}
1949 
1950 	if (!global_ioc_cnt) {
1951 		global_ioc_cnt = count_parisc_driver(&sba_driver);
1952 
1953 		/* Astro and Pluto have one IOC per SBA */
1954 		if ((!IS_ASTRO(dev)) || (!IS_PLUTO(dev)))
1955 			global_ioc_cnt *= 2;
1956 	}
1957 
1958 	printk(KERN_INFO "%s found %s at 0x%llx\n",
1959 		MODULE_NAME, version, (unsigned long long)dev->hpa.start);
1960 
1961 	sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
1962 	if (!sba_dev) {
1963 		printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
1964 		return -ENOMEM;
1965 	}
1966 
1967 	parisc_set_drvdata(dev, sba_dev);
1968 
1969 	for(i=0; i<MAX_IOC; i++)
1970 		spin_lock_init(&(sba_dev->ioc[i].res_lock));
1971 
1972 	sba_dev->dev = dev;
1973 	sba_dev->hw_rev = func_class;
1974 	sba_dev->name = dev->name;
1975 	sba_dev->sba_hpa = sba_addr;
1976 
1977 	sba_get_pat_resources(sba_dev);
1978 	sba_hw_init(sba_dev);
1979 	sba_common_init(sba_dev);
1980 
1981 	hppa_dma_ops = &sba_ops;
1982 
1983 #ifdef CONFIG_PROC_FS
1984 	switch (dev->id.hversion) {
1985 	case PLUTO_MCKINLEY_PORT:
1986 		root = proc_mckinley_root;
1987 		break;
1988 	case ASTRO_RUNWAY_PORT:
1989 	case IKE_MERCED_PORT:
1990 	default:
1991 		root = proc_runway_root;
1992 		break;
1993 	}
1994 
1995 	proc_create("sba_iommu", 0, root, &sba_proc_fops);
1996 	proc_create("sba_iommu-bitmap", 0, root, &sba_proc_bitmap_fops);
1997 #endif
1998 
1999 	parisc_has_iommu();
2000 	return 0;
2001 }
2002 
2003 /*
2004 ** One time initialization to let the world know the SBA was found.
2005 ** This is the only routine which is NOT static.
2006 ** Must be called exactly once before pci_init().
2007 */
2008 void __init sba_init(void)
2009 {
2010 	register_parisc_driver(&sba_driver);
2011 }
2012 
2013 
2014 /**
2015  * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2016  * @dev: The parisc device.
2017  *
2018  * Returns the appropriate IOMMU data for the given parisc PCI controller.
2019  * This is cached and used later for PCI DMA Mapping.
2020  */
2021 void * sba_get_iommu(struct parisc_device *pci_hba)
2022 {
2023 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2024 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2025 	char t = sba_dev->id.hw_type;
2026 	int iocnum = (pci_hba->hw_path >> 3);	/* rope # */
2027 
2028 	WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2029 
2030 	return &(sba->ioc[iocnum]);
2031 }
2032 
2033 
2034 /**
2035  * sba_directed_lmmio - return first directed LMMIO range routed to rope
2036  * @pa_dev: The parisc device.
2037  * @r: resource PCI host controller wants start/end fields assigned.
2038  *
2039  * For the given parisc PCI controller, determine if any direct ranges
2040  * are routed down the corresponding rope.
2041  */
2042 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2043 {
2044 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2045 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2046 	char t = sba_dev->id.hw_type;
2047 	int i;
2048 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2049 
2050 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2051 
2052 	r->start = r->end = 0;
2053 
2054 	/* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2055 	for (i=0; i<4; i++) {
2056 		int base, size;
2057 		void __iomem *reg = sba->sba_hpa + i*0x18;
2058 
2059 		base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2060 		if ((base & 1) == 0)
2061 			continue;	/* not enabled */
2062 
2063 		size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2064 
2065 		if ((size & (ROPES_PER_IOC-1)) != rope)
2066 			continue;	/* directed down different rope */
2067 
2068 		r->start = (base & ~1UL) | PCI_F_EXTEND;
2069 		size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2070 		r->end = r->start + size;
2071 		r->flags = IORESOURCE_MEM;
2072 	}
2073 }
2074 
2075 
2076 /**
2077  * sba_distributed_lmmio - return portion of distributed LMMIO range
2078  * @pa_dev: The parisc device.
2079  * @r: resource PCI host controller wants start/end fields assigned.
2080  *
2081  * For the given parisc PCI controller, return portion of distributed LMMIO
2082  * range. The distributed LMMIO is always present and it's just a question
2083  * of the base address and size of the range.
2084  */
2085 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2086 {
2087 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2088 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2089 	char t = sba_dev->id.hw_type;
2090 	int base, size;
2091 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2092 
2093 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2094 
2095 	r->start = r->end = 0;
2096 
2097 	base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2098 	if ((base & 1) == 0) {
2099 		BUG();	/* Gah! Distr Range wasn't enabled! */
2100 		return;
2101 	}
2102 
2103 	r->start = (base & ~1UL) | PCI_F_EXTEND;
2104 
2105 	size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2106 	r->start += rope * (size + 1);	/* adjust base for this rope */
2107 	r->end = r->start + size;
2108 	r->flags = IORESOURCE_MEM;
2109 }
2110