xref: /linux/drivers/parisc/sba_iommu.c (revision de2fe5e07d58424bc286fff3fd3c1b0bf933cd58)
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/config.h>
23 #include <linux/types.h>
24 #include <linux/kernel.h>
25 #include <linux/spinlock.h>
26 #include <linux/slab.h>
27 #include <linux/init.h>
28 
29 #include <linux/mm.h>
30 #include <linux/string.h>
31 #include <linux/pci.h>
32 
33 #include <asm/byteorder.h>
34 #include <asm/io.h>
35 #include <asm/dma.h>		/* for DMA_CHUNK_SIZE */
36 
37 #include <asm/hardware.h>	/* for register_parisc_driver() stuff */
38 
39 #include <linux/proc_fs.h>
40 #include <linux/seq_file.h>
41 
42 #include <asm/runway.h>		/* for proc_runway_root */
43 #include <asm/pdc.h>		/* for PDC_MODEL_* */
44 #include <asm/pdcpat.h>		/* for is_pdc_pat() */
45 #include <asm/parisc-device.h>
46 
47 
48 /* declared in arch/parisc/kernel/setup.c */
49 extern struct proc_dir_entry * proc_mckinley_root;
50 
51 #define MODULE_NAME "SBA"
52 
53 #ifdef CONFIG_PROC_FS
54 /* depends on proc fs support. But costs CPU performance */
55 #undef SBA_COLLECT_STATS
56 #endif
57 
58 /*
59 ** The number of debug flags is a clue - this code is fragile.
60 ** Don't even think about messing with it unless you have
61 ** plenty of 710's to sacrifice to the computer gods. :^)
62 */
63 #undef DEBUG_SBA_INIT
64 #undef DEBUG_SBA_RUN
65 #undef DEBUG_SBA_RUN_SG
66 #undef DEBUG_SBA_RESOURCE
67 #undef ASSERT_PDIR_SANITY
68 #undef DEBUG_LARGE_SG_ENTRIES
69 #undef DEBUG_DMB_TRAP
70 
71 #ifdef DEBUG_SBA_INIT
72 #define DBG_INIT(x...)	printk(x)
73 #else
74 #define DBG_INIT(x...)
75 #endif
76 
77 #ifdef DEBUG_SBA_RUN
78 #define DBG_RUN(x...)	printk(x)
79 #else
80 #define DBG_RUN(x...)
81 #endif
82 
83 #ifdef DEBUG_SBA_RUN_SG
84 #define DBG_RUN_SG(x...)	printk(x)
85 #else
86 #define DBG_RUN_SG(x...)
87 #endif
88 
89 
90 #ifdef DEBUG_SBA_RESOURCE
91 #define DBG_RES(x...)	printk(x)
92 #else
93 #define DBG_RES(x...)
94 #endif
95 
96 #if defined(CONFIG_64BIT)
97 /* "low end" PA8800 machines use ZX1 chipset: PAT PDC and only run 64-bit */
98 #define ZX1_SUPPORT
99 #endif
100 
101 #define SBA_INLINE	__inline__
102 
103 
104 /*
105 ** The number of pdir entries to "free" before issueing
106 ** a read to PCOM register to flush out PCOM writes.
107 ** Interacts with allocation granularity (ie 4 or 8 entries
108 ** allocated and free'd/purged at a time might make this
109 ** less interesting).
110 */
111 #define DELAYED_RESOURCE_CNT	16
112 
113 #define DEFAULT_DMA_HINT_REG	0
114 
115 #define ASTRO_RUNWAY_PORT	0x582
116 #define IKE_MERCED_PORT		0x803
117 #define REO_MERCED_PORT		0x804
118 #define REOG_MERCED_PORT	0x805
119 #define PLUTO_MCKINLEY_PORT	0x880
120 
121 #define SBA_FUNC_ID	0x0000	/* function id */
122 #define SBA_FCLASS	0x0008	/* function class, bist, header, rev... */
123 
124 #define IS_ASTRO(id)		((id)->hversion == ASTRO_RUNWAY_PORT)
125 #define IS_IKE(id)		((id)->hversion == IKE_MERCED_PORT)
126 #define IS_PLUTO(id)		((id)->hversion == PLUTO_MCKINLEY_PORT)
127 
128 #define SBA_FUNC_SIZE 4096   /* SBA configuration function reg set */
129 
130 #define ASTRO_IOC_OFFSET	(32 * SBA_FUNC_SIZE)
131 #define PLUTO_IOC_OFFSET	(1 * SBA_FUNC_SIZE)
132 /* Ike's IOC's occupy functions 2 and 3 */
133 #define IKE_IOC_OFFSET(p)	((p+2) * SBA_FUNC_SIZE)
134 
135 #define IOC_CTRL          0x8	/* IOC_CTRL offset */
136 #define IOC_CTRL_TC       (1 << 0) /* TOC Enable */
137 #define IOC_CTRL_CE       (1 << 1) /* Coalesce Enable */
138 #define IOC_CTRL_DE       (1 << 2) /* Dillon Enable */
139 #define IOC_CTRL_RM       (1 << 8) /* Real Mode */
140 #define IOC_CTRL_NC       (1 << 9) /* Non Coherent Mode */
141 #define IOC_CTRL_D4       (1 << 11) /* Disable 4-byte coalescing */
142 #define IOC_CTRL_DD       (1 << 13) /* Disable distr. LMMIO range coalescing */
143 
144 #define MAX_IOC		2	/* per Ike. Pluto/Astro only have 1. */
145 
146 #define ROPES_PER_IOC	8	/* per Ike half or Pluto/Astro */
147 
148 
149 /*
150 ** Offsets into MBIB (Function 0 on Ike and hopefully Astro)
151 ** Firmware programs this stuff. Don't touch it.
152 */
153 #define LMMIO_DIRECT0_BASE  0x300
154 #define LMMIO_DIRECT0_MASK  0x308
155 #define LMMIO_DIRECT0_ROUTE 0x310
156 
157 #define LMMIO_DIST_BASE  0x360
158 #define LMMIO_DIST_MASK  0x368
159 #define LMMIO_DIST_ROUTE 0x370
160 
161 #define IOS_DIST_BASE	0x390
162 #define IOS_DIST_MASK	0x398
163 #define IOS_DIST_ROUTE	0x3A0
164 
165 #define IOS_DIRECT_BASE	0x3C0
166 #define IOS_DIRECT_MASK	0x3C8
167 #define IOS_DIRECT_ROUTE 0x3D0
168 
169 /*
170 ** Offsets into I/O TLB (Function 2 and 3 on Ike)
171 */
172 #define ROPE0_CTL	0x200  /* "regbus pci0" */
173 #define ROPE1_CTL	0x208
174 #define ROPE2_CTL	0x210
175 #define ROPE3_CTL	0x218
176 #define ROPE4_CTL	0x220
177 #define ROPE5_CTL	0x228
178 #define ROPE6_CTL	0x230
179 #define ROPE7_CTL	0x238
180 
181 #define HF_ENABLE	0x40
182 
183 
184 #define IOC_IBASE	0x300	/* IO TLB */
185 #define IOC_IMASK	0x308
186 #define IOC_PCOM	0x310
187 #define IOC_TCNFG	0x318
188 #define IOC_PDIR_BASE	0x320
189 
190 /* AGP GART driver looks for this */
191 #define SBA_IOMMU_COOKIE    0x0000badbadc0ffeeUL
192 
193 
194 /*
195 ** IOC supports 4/8/16/64KB page sizes (see TCNFG register)
196 ** It's safer (avoid memory corruption) to keep DMA page mappings
197 ** equivalently sized to VM PAGE_SIZE.
198 **
199 ** We really can't avoid generating a new mapping for each
200 ** page since the Virtual Coherence Index has to be generated
201 ** and updated for each page.
202 **
203 ** PAGE_SIZE could be greater than IOVP_SIZE. But not the inverse.
204 */
205 #define IOVP_SIZE	PAGE_SIZE
206 #define IOVP_SHIFT	PAGE_SHIFT
207 #define IOVP_MASK	PAGE_MASK
208 
209 #define SBA_PERF_CFG	0x708	/* Performance Counter stuff */
210 #define SBA_PERF_MASK1	0x718
211 #define SBA_PERF_MASK2	0x730
212 
213 
214 /*
215 ** Offsets into PCI Performance Counters (functions 12 and 13)
216 ** Controlled by PERF registers in function 2 & 3 respectively.
217 */
218 #define SBA_PERF_CNT1	0x200
219 #define SBA_PERF_CNT2	0x208
220 #define SBA_PERF_CNT3	0x210
221 
222 
223 struct ioc {
224 	void __iomem	*ioc_hpa;	/* I/O MMU base address */
225 	char		*res_map;	/* resource map, bit == pdir entry */
226 	u64		*pdir_base;	/* physical base address */
227 	unsigned long	ibase;	/* pdir IOV Space base - shared w/lba_pci */
228 	unsigned long	imask;	/* pdir IOV Space mask - shared w/lba_pci */
229 #ifdef ZX1_SUPPORT
230 	unsigned long	iovp_mask;	/* help convert IOVA to IOVP */
231 #endif
232 	unsigned long	*res_hint;	/* next avail IOVP - circular search */
233 	spinlock_t	res_lock;
234 	unsigned int	res_bitshift;	/* from the LEFT! */
235 	unsigned int	res_size;	/* size of resource map in bytes */
236 #ifdef SBA_HINT_SUPPORT
237 /* FIXME : DMA HINTs not used */
238 	unsigned long	hint_mask_pdir;	/* bits used for DMA hints */
239 	unsigned int	hint_shift_pdir;
240 #endif
241 #if DELAYED_RESOURCE_CNT > 0
242 	int saved_cnt;
243 	struct sba_dma_pair {
244 		dma_addr_t	iova;
245 		size_t		size;
246 	} saved[DELAYED_RESOURCE_CNT];
247 #endif
248 
249 #ifdef SBA_COLLECT_STATS
250 #define SBA_SEARCH_SAMPLE	0x100
251 	unsigned long avg_search[SBA_SEARCH_SAMPLE];
252 	unsigned long avg_idx;	/* current index into avg_search */
253 	unsigned long used_pages;
254 	unsigned long msingle_calls;
255 	unsigned long msingle_pages;
256 	unsigned long msg_calls;
257 	unsigned long msg_pages;
258 	unsigned long usingle_calls;
259 	unsigned long usingle_pages;
260 	unsigned long usg_calls;
261 	unsigned long usg_pages;
262 #endif
263 
264 	/* STUFF We don't need in performance path */
265 	unsigned int	pdir_size;	/* in bytes, determined by IOV Space size */
266 };
267 
268 struct sba_device {
269 	struct sba_device	*next;	/* list of SBA's in system */
270 	struct parisc_device	*dev;	/* dev found in bus walk */
271 	struct parisc_device_id	*iodc;	/* data about dev from firmware */
272 	const char 		*name;
273 	void __iomem		*sba_hpa; /* base address */
274 	spinlock_t		sba_lock;
275 	unsigned int		flags;  /* state/functionality enabled */
276 	unsigned int		hw_rev;  /* HW revision of chip */
277 
278 	struct resource		chip_resv; /* MMIO reserved for chip */
279 	struct resource		iommu_resv; /* MMIO reserved for iommu */
280 
281 	unsigned int		num_ioc;  /* number of on-board IOC's */
282 	struct ioc		ioc[MAX_IOC];
283 };
284 
285 
286 static struct sba_device *sba_list;
287 
288 static unsigned long ioc_needs_fdc = 0;
289 
290 /* global count of IOMMUs in the system */
291 static unsigned int global_ioc_cnt = 0;
292 
293 /* PA8700 (Piranha 2.2) bug workaround */
294 static unsigned long piranha_bad_128k = 0;
295 
296 /* Looks nice and keeps the compiler happy */
297 #define SBA_DEV(d) ((struct sba_device *) (d))
298 
299 #ifdef SBA_AGP_SUPPORT
300 static int reserve_sba_gart = 1;
301 #endif
302 
303 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
304 
305 
306 /************************************
307 ** SBA register read and write support
308 **
309 ** BE WARNED: register writes are posted.
310 **  (ie follow writes which must reach HW with a read)
311 **
312 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
313 */
314 #define READ_REG32(addr)	 le32_to_cpu(__raw_readl(addr))
315 #define READ_REG64(addr)	 le64_to_cpu(__raw_readq(addr))
316 #define WRITE_REG32(val, addr) __raw_writel(cpu_to_le32(val), addr)
317 #define WRITE_REG64(val, addr) __raw_writeq(cpu_to_le64(val), addr)
318 
319 #ifdef CONFIG_64BIT
320 #define READ_REG(addr)		READ_REG64(addr)
321 #define WRITE_REG(value, addr)	WRITE_REG64(value, addr)
322 #else
323 #define READ_REG(addr)		READ_REG32(addr)
324 #define WRITE_REG(value, addr)	WRITE_REG32(value, addr)
325 #endif
326 
327 #ifdef DEBUG_SBA_INIT
328 
329 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
330 
331 /**
332  * sba_dump_ranges - debugging only - print ranges assigned to this IOA
333  * @hpa: base address of the sba
334  *
335  * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
336  * IO Adapter (aka Bus Converter).
337  */
338 static void
339 sba_dump_ranges(void __iomem *hpa)
340 {
341 	DBG_INIT("SBA at 0x%p\n", hpa);
342 	DBG_INIT("IOS_DIST_BASE   : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
343 	DBG_INIT("IOS_DIST_MASK   : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
344 	DBG_INIT("IOS_DIST_ROUTE  : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
345 	DBG_INIT("\n");
346 	DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
347 	DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
348 	DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
349 }
350 
351 /**
352  * sba_dump_tlb - debugging only - print IOMMU operating parameters
353  * @hpa: base address of the IOMMU
354  *
355  * Print the size/location of the IO MMU PDIR.
356  */
357 static void sba_dump_tlb(void __iomem *hpa)
358 {
359 	DBG_INIT("IO TLB at 0x%p\n", hpa);
360 	DBG_INIT("IOC_IBASE    : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
361 	DBG_INIT("IOC_IMASK    : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
362 	DBG_INIT("IOC_TCNFG    : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
363 	DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
364 	DBG_INIT("\n");
365 }
366 #else
367 #define sba_dump_ranges(x)
368 #define sba_dump_tlb(x)
369 #endif	/* DEBUG_SBA_INIT */
370 
371 
372 #ifdef ASSERT_PDIR_SANITY
373 
374 /**
375  * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
376  * @ioc: IO MMU structure which owns the pdir we are interested in.
377  * @msg: text to print ont the output line.
378  * @pide: pdir index.
379  *
380  * Print one entry of the IO MMU PDIR in human readable form.
381  */
382 static void
383 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
384 {
385 	/* start printing from lowest pde in rval */
386 	u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
387 	unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
388 	uint rcnt;
389 
390 	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
391 		 msg,
392 		 rptr, pide & (BITS_PER_LONG - 1), *rptr);
393 
394 	rcnt = 0;
395 	while (rcnt < BITS_PER_LONG) {
396 		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
397 			(rcnt == (pide & (BITS_PER_LONG - 1)))
398 				? "    -->" : "       ",
399 			rcnt, ptr, *ptr );
400 		rcnt++;
401 		ptr++;
402 	}
403 	printk(KERN_DEBUG "%s", msg);
404 }
405 
406 
407 /**
408  * sba_check_pdir - debugging only - consistency checker
409  * @ioc: IO MMU structure which owns the pdir we are interested in.
410  * @msg: text to print ont the output line.
411  *
412  * Verify the resource map and pdir state is consistent
413  */
414 static int
415 sba_check_pdir(struct ioc *ioc, char *msg)
416 {
417 	u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
418 	u32 *rptr = (u32 *) ioc->res_map;	/* resource map ptr */
419 	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
420 	uint pide = 0;
421 
422 	while (rptr < rptr_end) {
423 		u32 rval = *rptr;
424 		int rcnt = 32;	/* number of bits we might check */
425 
426 		while (rcnt) {
427 			/* Get last byte and highest bit from that */
428 			u32 pde = ((u32) (((char *)pptr)[7])) << 24;
429 			if ((rval ^ pde) & 0x80000000)
430 			{
431 				/*
432 				** BUMMER!  -- res_map != pdir --
433 				** Dump rval and matching pdir entries
434 				*/
435 				sba_dump_pdir_entry(ioc, msg, pide);
436 				return(1);
437 			}
438 			rcnt--;
439 			rval <<= 1;	/* try the next bit */
440 			pptr++;
441 			pide++;
442 		}
443 		rptr++;	/* look at next word of res_map */
444 	}
445 	/* It'd be nice if we always got here :^) */
446 	return 0;
447 }
448 
449 
450 /**
451  * sba_dump_sg - debugging only - print Scatter-Gather list
452  * @ioc: IO MMU structure which owns the pdir we are interested in.
453  * @startsg: head of the SG list
454  * @nents: number of entries in SG list
455  *
456  * print the SG list so we can verify it's correct by hand.
457  */
458 static void
459 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
460 {
461 	while (nents-- > 0) {
462 		printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
463 				nents,
464 				(unsigned long) sg_dma_address(startsg),
465 				sg_dma_len(startsg),
466 				sg_virt_addr(startsg), startsg->length);
467 		startsg++;
468 	}
469 }
470 
471 #endif /* ASSERT_PDIR_SANITY */
472 
473 
474 
475 
476 /**************************************************************
477 *
478 *   I/O Pdir Resource Management
479 *
480 *   Bits set in the resource map are in use.
481 *   Each bit can represent a number of pages.
482 *   LSbs represent lower addresses (IOVA's).
483 *
484 ***************************************************************/
485 #define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
486 
487 /* Convert from IOVP to IOVA and vice versa. */
488 
489 #ifdef ZX1_SUPPORT
490 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
491 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
492 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
493 #else
494 /* only support Astro and ancestors. Saves a few cycles in key places */
495 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
496 #define SBA_IOVP(ioc,iova) (iova)
497 #endif
498 
499 #define PDIR_INDEX(iovp)   ((iovp)>>IOVP_SHIFT)
500 
501 #define RESMAP_MASK(n)    (~0UL << (BITS_PER_LONG - (n)))
502 #define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)
503 
504 
505 /**
506  * sba_search_bitmap - find free space in IO PDIR resource bitmap
507  * @ioc: IO MMU structure which owns the pdir we are interested in.
508  * @bits_wanted: number of entries we need.
509  *
510  * Find consecutive free bits in resource bitmap.
511  * Each bit represents one entry in the IO Pdir.
512  * Cool perf optimization: search for log2(size) bits at a time.
513  */
514 static SBA_INLINE unsigned long
515 sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted)
516 {
517 	unsigned long *res_ptr = ioc->res_hint;
518 	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
519 	unsigned long pide = ~0UL;
520 
521 	if (bits_wanted > (BITS_PER_LONG/2)) {
522 		/* Search word at a time - no mask needed */
523 		for(; res_ptr < res_end; ++res_ptr) {
524 			if (*res_ptr == 0) {
525 				*res_ptr = RESMAP_MASK(bits_wanted);
526 				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
527 				pide <<= 3;	/* convert to bit address */
528 				break;
529 			}
530 		}
531 		/* point to the next word on next pass */
532 		res_ptr++;
533 		ioc->res_bitshift = 0;
534 	} else {
535 		/*
536 		** Search the resource bit map on well-aligned values.
537 		** "o" is the alignment.
538 		** We need the alignment to invalidate I/O TLB using
539 		** SBA HW features in the unmap path.
540 		*/
541 		unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
542 		uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
543 		unsigned long mask;
544 
545 		if (bitshiftcnt >= BITS_PER_LONG) {
546 			bitshiftcnt = 0;
547 			res_ptr++;
548 		}
549 		mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
550 
551 		DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
552 		while(res_ptr < res_end)
553 		{
554 			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
555 			WARN_ON(mask == 0);
556 			if(((*res_ptr) & mask) == 0) {
557 				*res_ptr |= mask;     /* mark resources busy! */
558 				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
559 				pide <<= 3;	/* convert to bit address */
560 				pide += bitshiftcnt;
561 				break;
562 			}
563 			mask >>= o;
564 			bitshiftcnt += o;
565 			if (mask == 0) {
566 				mask = RESMAP_MASK(bits_wanted);
567 				bitshiftcnt=0;
568 				res_ptr++;
569 			}
570 		}
571 		/* look in the same word on the next pass */
572 		ioc->res_bitshift = bitshiftcnt + bits_wanted;
573 	}
574 
575 	/* wrapped ? */
576 	if (res_end <= res_ptr) {
577 		ioc->res_hint = (unsigned long *) ioc->res_map;
578 		ioc->res_bitshift = 0;
579 	} else {
580 		ioc->res_hint = res_ptr;
581 	}
582 	return (pide);
583 }
584 
585 
586 /**
587  * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
588  * @ioc: IO MMU structure which owns the pdir we are interested in.
589  * @size: number of bytes to create a mapping for
590  *
591  * Given a size, find consecutive unmarked and then mark those bits in the
592  * resource bit map.
593  */
594 static int
595 sba_alloc_range(struct ioc *ioc, size_t size)
596 {
597 	unsigned int pages_needed = size >> IOVP_SHIFT;
598 #ifdef SBA_COLLECT_STATS
599 	unsigned long cr_start = mfctl(16);
600 #endif
601 	unsigned long pide;
602 
603 	pide = sba_search_bitmap(ioc, pages_needed);
604 	if (pide >= (ioc->res_size << 3)) {
605 		pide = sba_search_bitmap(ioc, pages_needed);
606 		if (pide >= (ioc->res_size << 3))
607 			panic("%s: I/O MMU @ %p is out of mapping resources\n",
608 			      __FILE__, ioc->ioc_hpa);
609 	}
610 
611 #ifdef ASSERT_PDIR_SANITY
612 	/* verify the first enable bit is clear */
613 	if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
614 		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
615 	}
616 #endif
617 
618 	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
619 		__FUNCTION__, size, pages_needed, pide,
620 		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
621 		ioc->res_bitshift );
622 
623 #ifdef SBA_COLLECT_STATS
624 	{
625 		unsigned long cr_end = mfctl(16);
626 		unsigned long tmp = cr_end - cr_start;
627 		/* check for roll over */
628 		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
629 	}
630 	ioc->avg_search[ioc->avg_idx++] = cr_start;
631 	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
632 
633 	ioc->used_pages += pages_needed;
634 #endif
635 
636 	return (pide);
637 }
638 
639 
640 /**
641  * sba_free_range - unmark bits in IO PDIR resource bitmap
642  * @ioc: IO MMU structure which owns the pdir we are interested in.
643  * @iova: IO virtual address which was previously allocated.
644  * @size: number of bytes to create a mapping for
645  *
646  * clear bits in the ioc's resource map
647  */
648 static SBA_INLINE void
649 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
650 {
651 	unsigned long iovp = SBA_IOVP(ioc, iova);
652 	unsigned int pide = PDIR_INDEX(iovp);
653 	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
654 	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
655 
656 	int bits_not_wanted = size >> IOVP_SHIFT;
657 
658 	/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
659 	unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
660 
661 	DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
662 		__FUNCTION__, (uint) iova, size,
663 		bits_not_wanted, m, pide, res_ptr, *res_ptr);
664 
665 #ifdef SBA_COLLECT_STATS
666 	ioc->used_pages -= bits_not_wanted;
667 #endif
668 
669 	*res_ptr &= ~m;
670 }
671 
672 
673 /**************************************************************
674 *
675 *   "Dynamic DMA Mapping" support (aka "Coherent I/O")
676 *
677 ***************************************************************/
678 
679 #ifdef SBA_HINT_SUPPORT
680 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
681 #endif
682 
683 typedef unsigned long space_t;
684 #define KERNEL_SPACE 0
685 
686 /**
687  * sba_io_pdir_entry - fill in one IO PDIR entry
688  * @pdir_ptr:  pointer to IO PDIR entry
689  * @sid: process Space ID - currently only support KERNEL_SPACE
690  * @vba: Virtual CPU address of buffer to map
691  * @hint: DMA hint set to use for this mapping
692  *
693  * SBA Mapping Routine
694  *
695  * Given a virtual address (vba, arg2) and space id, (sid, arg1)
696  * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
697  * pdir_ptr (arg0).
698  * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
699  * for Astro/Ike looks like:
700  *
701  *
702  *  0                    19                                 51   55       63
703  * +-+---------------------+----------------------------------+----+--------+
704  * |V|        U            |            PPN[43:12]            | U  |   VI   |
705  * +-+---------------------+----------------------------------+----+--------+
706  *
707  * Pluto is basically identical, supports fewer physical address bits:
708  *
709  *  0                       23                              51   55       63
710  * +-+------------------------+-------------------------------+----+--------+
711  * |V|        U               |         PPN[39:12]            | U  |   VI   |
712  * +-+------------------------+-------------------------------+----+--------+
713  *
714  *  V  == Valid Bit  (Most Significant Bit is bit 0)
715  *  U  == Unused
716  * PPN == Physical Page Number
717  * VI  == Virtual Index (aka Coherent Index)
718  *
719  * LPA instruction output is put into PPN field.
720  * LCI (Load Coherence Index) instruction provides the "VI" bits.
721  *
722  * We pre-swap the bytes since PCX-W is Big Endian and the
723  * IOMMU uses little endian for the pdir.
724  */
725 
726 void SBA_INLINE
727 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
728 		  unsigned long hint)
729 {
730 	u64 pa; /* physical address */
731 	register unsigned ci; /* coherent index */
732 
733 	pa = virt_to_phys(vba);
734 	pa &= IOVP_MASK;
735 
736 	mtsp(sid,1);
737 	asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
738 	pa |= (ci >> 12) & 0xff;  /* move CI (8 bits) into lowest byte */
739 
740 	pa |= 0x8000000000000000ULL;	/* set "valid" bit */
741 	*pdir_ptr = cpu_to_le64(pa);	/* swap and store into I/O Pdir */
742 
743 	/*
744 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
745 	 * (bit #61, big endian), we have to flush and sync every time
746 	 * IO-PDIR is changed in Ike/Astro.
747 	 */
748 	if (ioc_needs_fdc)
749 		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
750 }
751 
752 
753 /**
754  * sba_mark_invalid - invalidate one or more IO PDIR entries
755  * @ioc: IO MMU structure which owns the pdir we are interested in.
756  * @iova:  IO Virtual Address mapped earlier
757  * @byte_cnt:  number of bytes this mapping covers.
758  *
759  * Marking the IO PDIR entry(ies) as Invalid and invalidate
760  * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
761  * is to purge stale entries in the IO TLB when unmapping entries.
762  *
763  * The PCOM register supports purging of multiple pages, with a minium
764  * of 1 page and a maximum of 2GB. Hardware requires the address be
765  * aligned to the size of the range being purged. The size of the range
766  * must be a power of 2. The "Cool perf optimization" in the
767  * allocation routine helps keep that true.
768  */
769 static SBA_INLINE void
770 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
771 {
772 	u32 iovp = (u32) SBA_IOVP(ioc,iova);
773 	u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
774 
775 #ifdef ASSERT_PDIR_SANITY
776 	/* Assert first pdir entry is set.
777 	**
778 	** Even though this is a big-endian machine, the entries
779 	** in the iopdir are little endian. That's why we look at
780 	** the byte at +7 instead of at +0.
781 	*/
782 	if (0x80 != (((u8 *) pdir_ptr)[7])) {
783 		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
784 	}
785 #endif
786 
787 	if (byte_cnt > IOVP_SIZE)
788 	{
789 #if 0
790 		unsigned long entries_per_cacheline = ioc_needs_fdc ?
791 				L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
792 					- (unsigned long) pdir_ptr;
793 				: 262144;
794 #endif
795 
796 		/* set "size" field for PCOM */
797 		iovp |= get_order(byte_cnt) + PAGE_SHIFT;
798 
799 		do {
800 			/* clear I/O Pdir entry "valid" bit first */
801 			((u8 *) pdir_ptr)[7] = 0;
802 			if (ioc_needs_fdc) {
803 				asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
804 #if 0
805 				entries_per_cacheline = L1_CACHE_SHIFT - 3;
806 #endif
807 			}
808 			pdir_ptr++;
809 			byte_cnt -= IOVP_SIZE;
810 		} while (byte_cnt > IOVP_SIZE);
811 	} else
812 		iovp |= IOVP_SHIFT;     /* set "size" field for PCOM */
813 
814 	/*
815 	** clear I/O PDIR entry "valid" bit.
816 	** We have to R/M/W the cacheline regardless how much of the
817 	** pdir entry that we clobber.
818 	** The rest of the entry would be useful for debugging if we
819 	** could dump core on HPMC.
820 	*/
821 	((u8 *) pdir_ptr)[7] = 0;
822 	if (ioc_needs_fdc)
823 		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
824 
825 	WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
826 }
827 
828 /**
829  * sba_dma_supported - PCI driver can query DMA support
830  * @dev: instance of PCI owned by the driver that's asking
831  * @mask:  number of address bits this PCI device can handle
832  *
833  * See Documentation/DMA-mapping.txt
834  */
835 static int sba_dma_supported( struct device *dev, u64 mask)
836 {
837 	struct ioc *ioc;
838 
839 	if (dev == NULL) {
840 		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
841 		BUG();
842 		return(0);
843 	}
844 
845 	/* Documentation/DMA-mapping.txt tells drivers to try 64-bit first,
846 	 * then fall back to 32-bit if that fails.
847 	 * We are just "encouraging" 32-bit DMA masks here since we can
848 	 * never allow IOMMU bypass unless we add special support for ZX1.
849 	 */
850 	if (mask > ~0U)
851 		return 0;
852 
853 	ioc = GET_IOC(dev);
854 
855 	/*
856 	 * check if mask is >= than the current max IO Virt Address
857 	 * The max IO Virt address will *always* < 30 bits.
858 	 */
859 	return((int)(mask >= (ioc->ibase - 1 +
860 			(ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
861 }
862 
863 
864 /**
865  * sba_map_single - map one buffer and return IOVA for DMA
866  * @dev: instance of PCI owned by the driver that's asking.
867  * @addr:  driver buffer to map.
868  * @size:  number of bytes to map in driver buffer.
869  * @direction:  R/W or both.
870  *
871  * See Documentation/DMA-mapping.txt
872  */
873 static dma_addr_t
874 sba_map_single(struct device *dev, void *addr, size_t size,
875 	       enum dma_data_direction direction)
876 {
877 	struct ioc *ioc;
878 	unsigned long flags;
879 	dma_addr_t iovp;
880 	dma_addr_t offset;
881 	u64 *pdir_start;
882 	int pide;
883 
884 	ioc = GET_IOC(dev);
885 
886 	/* save offset bits */
887 	offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
888 
889 	/* round up to nearest IOVP_SIZE */
890 	size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
891 
892 	spin_lock_irqsave(&ioc->res_lock, flags);
893 #ifdef ASSERT_PDIR_SANITY
894 	sba_check_pdir(ioc,"Check before sba_map_single()");
895 #endif
896 
897 #ifdef SBA_COLLECT_STATS
898 	ioc->msingle_calls++;
899 	ioc->msingle_pages += size >> IOVP_SHIFT;
900 #endif
901 	pide = sba_alloc_range(ioc, size);
902 	iovp = (dma_addr_t) pide << IOVP_SHIFT;
903 
904 	DBG_RUN("%s() 0x%p -> 0x%lx\n",
905 		__FUNCTION__, addr, (long) iovp | offset);
906 
907 	pdir_start = &(ioc->pdir_base[pide]);
908 
909 	while (size > 0) {
910 		sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
911 
912 		DBG_RUN("	pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
913 			pdir_start,
914 			(u8) (((u8 *) pdir_start)[7]),
915 			(u8) (((u8 *) pdir_start)[6]),
916 			(u8) (((u8 *) pdir_start)[5]),
917 			(u8) (((u8 *) pdir_start)[4]),
918 			(u8) (((u8 *) pdir_start)[3]),
919 			(u8) (((u8 *) pdir_start)[2]),
920 			(u8) (((u8 *) pdir_start)[1]),
921 			(u8) (((u8 *) pdir_start)[0])
922 			);
923 
924 		addr += IOVP_SIZE;
925 		size -= IOVP_SIZE;
926 		pdir_start++;
927 	}
928 
929 	/* force FDC ops in io_pdir_entry() to be visible to IOMMU */
930 	if (ioc_needs_fdc)
931 		asm volatile("sync" : : );
932 
933 #ifdef ASSERT_PDIR_SANITY
934 	sba_check_pdir(ioc,"Check after sba_map_single()");
935 #endif
936 	spin_unlock_irqrestore(&ioc->res_lock, flags);
937 
938 	/* form complete address */
939 	return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
940 }
941 
942 
943 /**
944  * sba_unmap_single - unmap one IOVA and free resources
945  * @dev: instance of PCI owned by the driver that's asking.
946  * @iova:  IOVA of driver buffer previously mapped.
947  * @size:  number of bytes mapped in driver buffer.
948  * @direction:  R/W or both.
949  *
950  * See Documentation/DMA-mapping.txt
951  */
952 static void
953 sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size,
954 		 enum dma_data_direction direction)
955 {
956 	struct ioc *ioc;
957 #if DELAYED_RESOURCE_CNT > 0
958 	struct sba_dma_pair *d;
959 #endif
960 	unsigned long flags;
961 	dma_addr_t offset;
962 
963 	DBG_RUN("%s() iovp 0x%lx/%x\n", __FUNCTION__, (long) iova, size);
964 
965 	ioc = GET_IOC(dev);
966 	offset = iova & ~IOVP_MASK;
967 	iova ^= offset;        /* clear offset bits */
968 	size += offset;
969 	size = ROUNDUP(size, IOVP_SIZE);
970 
971 	spin_lock_irqsave(&ioc->res_lock, flags);
972 
973 #ifdef SBA_COLLECT_STATS
974 	ioc->usingle_calls++;
975 	ioc->usingle_pages += size >> IOVP_SHIFT;
976 #endif
977 
978 	sba_mark_invalid(ioc, iova, size);
979 
980 #if DELAYED_RESOURCE_CNT > 0
981 	/* Delaying when we re-use a IO Pdir entry reduces the number
982 	 * of MMIO reads needed to flush writes to the PCOM register.
983 	 */
984 	d = &(ioc->saved[ioc->saved_cnt]);
985 	d->iova = iova;
986 	d->size = size;
987 	if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
988 		int cnt = ioc->saved_cnt;
989 		while (cnt--) {
990 			sba_free_range(ioc, d->iova, d->size);
991 			d--;
992 		}
993 		ioc->saved_cnt = 0;
994 
995 		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
996 	}
997 #else /* DELAYED_RESOURCE_CNT == 0 */
998 	sba_free_range(ioc, iova, size);
999 
1000 	/* If fdc's were issued, force fdc's to be visible now */
1001 	if (ioc_needs_fdc)
1002 		asm volatile("sync" : : );
1003 
1004 	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
1005 #endif /* DELAYED_RESOURCE_CNT == 0 */
1006 
1007 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1008 
1009 	/* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
1010 	** For Astro based systems this isn't a big deal WRT performance.
1011 	** As long as 2.4 kernels copyin/copyout data from/to userspace,
1012 	** we don't need the syncdma. The issue here is I/O MMU cachelines
1013 	** are *not* coherent in all cases.  May be hwrev dependent.
1014 	** Need to investigate more.
1015 	asm volatile("syncdma");
1016 	*/
1017 }
1018 
1019 
1020 /**
1021  * sba_alloc_consistent - allocate/map shared mem for DMA
1022  * @hwdev: instance of PCI owned by the driver that's asking.
1023  * @size:  number of bytes mapped in driver buffer.
1024  * @dma_handle:  IOVA of new buffer.
1025  *
1026  * See Documentation/DMA-mapping.txt
1027  */
1028 static void *sba_alloc_consistent(struct device *hwdev, size_t size,
1029 					dma_addr_t *dma_handle, gfp_t gfp)
1030 {
1031 	void *ret;
1032 
1033 	if (!hwdev) {
1034 		/* only support PCI */
1035 		*dma_handle = 0;
1036 		return 0;
1037 	}
1038 
1039         ret = (void *) __get_free_pages(gfp, get_order(size));
1040 
1041 	if (ret) {
1042 		memset(ret, 0, size);
1043 		*dma_handle = sba_map_single(hwdev, ret, size, 0);
1044 	}
1045 
1046 	return ret;
1047 }
1048 
1049 
1050 /**
1051  * sba_free_consistent - free/unmap shared mem for DMA
1052  * @hwdev: instance of PCI owned by the driver that's asking.
1053  * @size:  number of bytes mapped in driver buffer.
1054  * @vaddr:  virtual address IOVA of "consistent" buffer.
1055  * @dma_handler:  IO virtual address of "consistent" buffer.
1056  *
1057  * See Documentation/DMA-mapping.txt
1058  */
1059 static void
1060 sba_free_consistent(struct device *hwdev, size_t size, void *vaddr,
1061 		    dma_addr_t dma_handle)
1062 {
1063 	sba_unmap_single(hwdev, dma_handle, size, 0);
1064 	free_pages((unsigned long) vaddr, get_order(size));
1065 }
1066 
1067 
1068 /*
1069 ** Since 0 is a valid pdir_base index value, can't use that
1070 ** to determine if a value is valid or not. Use a flag to indicate
1071 ** the SG list entry contains a valid pdir index.
1072 */
1073 #define PIDE_FLAG 0x80000000UL
1074 
1075 #ifdef SBA_COLLECT_STATS
1076 #define IOMMU_MAP_STATS
1077 #endif
1078 #include "iommu-helpers.h"
1079 
1080 #ifdef DEBUG_LARGE_SG_ENTRIES
1081 int dump_run_sg = 0;
1082 #endif
1083 
1084 
1085 /**
1086  * sba_map_sg - map Scatter/Gather list
1087  * @dev: instance of PCI owned by the driver that's asking.
1088  * @sglist:  array of buffer/length pairs
1089  * @nents:  number of entries in list
1090  * @direction:  R/W or both.
1091  *
1092  * See Documentation/DMA-mapping.txt
1093  */
1094 static int
1095 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
1096 	   enum dma_data_direction direction)
1097 {
1098 	struct ioc *ioc;
1099 	int coalesced, filled = 0;
1100 	unsigned long flags;
1101 
1102 	DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents);
1103 
1104 	ioc = GET_IOC(dev);
1105 
1106 	/* Fast path single entry scatterlists. */
1107 	if (nents == 1) {
1108 		sg_dma_address(sglist) = sba_map_single(dev,
1109 						(void *)sg_virt_addr(sglist),
1110 						sglist->length, direction);
1111 		sg_dma_len(sglist)     = sglist->length;
1112 		return 1;
1113 	}
1114 
1115 	spin_lock_irqsave(&ioc->res_lock, flags);
1116 
1117 #ifdef ASSERT_PDIR_SANITY
1118 	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
1119 	{
1120 		sba_dump_sg(ioc, sglist, nents);
1121 		panic("Check before sba_map_sg()");
1122 	}
1123 #endif
1124 
1125 #ifdef SBA_COLLECT_STATS
1126 	ioc->msg_calls++;
1127 #endif
1128 
1129 	/*
1130 	** First coalesce the chunks and allocate I/O pdir space
1131 	**
1132 	** If this is one DMA stream, we can properly map using the
1133 	** correct virtual address associated with each DMA page.
1134 	** w/o this association, we wouldn't have coherent DMA!
1135 	** Access to the virtual address is what forces a two pass algorithm.
1136 	*/
1137 	coalesced = iommu_coalesce_chunks(ioc, sglist, nents, sba_alloc_range);
1138 
1139 	/*
1140 	** Program the I/O Pdir
1141 	**
1142 	** map the virtual addresses to the I/O Pdir
1143 	** o dma_address will contain the pdir index
1144 	** o dma_len will contain the number of bytes to map
1145 	** o address contains the virtual address.
1146 	*/
1147 	filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
1148 
1149 	/* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1150 	if (ioc_needs_fdc)
1151 		asm volatile("sync" : : );
1152 
1153 #ifdef ASSERT_PDIR_SANITY
1154 	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1155 	{
1156 		sba_dump_sg(ioc, sglist, nents);
1157 		panic("Check after sba_map_sg()\n");
1158 	}
1159 #endif
1160 
1161 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1162 
1163 	DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled);
1164 
1165 	return filled;
1166 }
1167 
1168 
1169 /**
1170  * sba_unmap_sg - unmap Scatter/Gather list
1171  * @dev: instance of PCI owned by the driver that's asking.
1172  * @sglist:  array of buffer/length pairs
1173  * @nents:  number of entries in list
1174  * @direction:  R/W or both.
1175  *
1176  * See Documentation/DMA-mapping.txt
1177  */
1178 static void
1179 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1180 	     enum dma_data_direction direction)
1181 {
1182 	struct ioc *ioc;
1183 #ifdef ASSERT_PDIR_SANITY
1184 	unsigned long flags;
1185 #endif
1186 
1187 	DBG_RUN_SG("%s() START %d entries,  %p,%x\n",
1188 		__FUNCTION__, nents, sg_virt_addr(sglist), sglist->length);
1189 
1190 	ioc = GET_IOC(dev);
1191 
1192 #ifdef SBA_COLLECT_STATS
1193 	ioc->usg_calls++;
1194 #endif
1195 
1196 #ifdef ASSERT_PDIR_SANITY
1197 	spin_lock_irqsave(&ioc->res_lock, flags);
1198 	sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1199 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1200 #endif
1201 
1202 	while (sg_dma_len(sglist) && nents--) {
1203 
1204 		sba_unmap_single(dev, sg_dma_address(sglist), sg_dma_len(sglist), direction);
1205 #ifdef SBA_COLLECT_STATS
1206 		ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1207 		ioc->usingle_calls--;	/* kluge since call is unmap_sg() */
1208 #endif
1209 		++sglist;
1210 	}
1211 
1212 	DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__,  nents);
1213 
1214 #ifdef ASSERT_PDIR_SANITY
1215 	spin_lock_irqsave(&ioc->res_lock, flags);
1216 	sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1217 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1218 #endif
1219 
1220 }
1221 
1222 static struct hppa_dma_ops sba_ops = {
1223 	.dma_supported =	sba_dma_supported,
1224 	.alloc_consistent =	sba_alloc_consistent,
1225 	.alloc_noncoherent =	sba_alloc_consistent,
1226 	.free_consistent =	sba_free_consistent,
1227 	.map_single =		sba_map_single,
1228 	.unmap_single =		sba_unmap_single,
1229 	.map_sg =		sba_map_sg,
1230 	.unmap_sg =		sba_unmap_sg,
1231 	.dma_sync_single_for_cpu =	NULL,
1232 	.dma_sync_single_for_device =	NULL,
1233 	.dma_sync_sg_for_cpu =		NULL,
1234 	.dma_sync_sg_for_device =	NULL,
1235 };
1236 
1237 
1238 /**************************************************************************
1239 **
1240 **   SBA PAT PDC support
1241 **
1242 **   o call pdc_pat_cell_module()
1243 **   o store ranges in PCI "resource" structures
1244 **
1245 **************************************************************************/
1246 
1247 static void
1248 sba_get_pat_resources(struct sba_device *sba_dev)
1249 {
1250 #if 0
1251 /*
1252 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1253 **      PAT PDC to program the SBA/LBA directed range registers...this
1254 **      burden may fall on the LBA code since it directly supports the
1255 **      PCI subsystem. It's not clear yet. - ggg
1256 */
1257 PAT_MOD(mod)->mod_info.mod_pages   = PAT_GET_MOD_PAGES(temp);
1258 	FIXME : ???
1259 PAT_MOD(mod)->mod_info.dvi         = PAT_GET_DVI(temp);
1260 	Tells where the dvi bits are located in the address.
1261 PAT_MOD(mod)->mod_info.ioc         = PAT_GET_IOC(temp);
1262 	FIXME : ???
1263 #endif
1264 }
1265 
1266 
1267 /**************************************************************
1268 *
1269 *   Initialization and claim
1270 *
1271 ***************************************************************/
1272 #define PIRANHA_ADDR_MASK	0x00160000UL /* bit 17,18,20 */
1273 #define PIRANHA_ADDR_VAL	0x00060000UL /* bit 17,18 on */
1274 static void *
1275 sba_alloc_pdir(unsigned int pdir_size)
1276 {
1277         unsigned long pdir_base;
1278 	unsigned long pdir_order = get_order(pdir_size);
1279 
1280 	pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1281 	if (NULL == (void *) pdir_base)	{
1282 		panic("%s() could not allocate I/O Page Table\n",
1283 			__FUNCTION__);
1284 	}
1285 
1286 	/* If this is not PA8700 (PCX-W2)
1287 	**	OR newer than ver 2.2
1288 	**	OR in a system that doesn't need VINDEX bits from SBA,
1289 	**
1290 	** then we aren't exposed to the HW bug.
1291 	*/
1292 	if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1293 			|| (boot_cpu_data.pdc.versions > 0x202)
1294 			|| (boot_cpu_data.pdc.capabilities & 0x08L) )
1295 		return (void *) pdir_base;
1296 
1297 	/*
1298 	 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1299 	 *
1300 	 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1301 	 * Ike/Astro can cause silent data corruption. This is only
1302 	 * a problem if the I/O PDIR is located in memory such that
1303 	 * (little-endian)  bits 17 and 18 are on and bit 20 is off.
1304 	 *
1305 	 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1306 	 * right physical address, we can either avoid (IOPDIR <= 1MB)
1307 	 * or minimize (2MB IO Pdir) the problem if we restrict the
1308 	 * IO Pdir to a maximum size of 2MB-128K (1902K).
1309 	 *
1310 	 * Because we always allocate 2^N sized IO pdirs, either of the
1311 	 * "bad" regions will be the last 128K if at all. That's easy
1312 	 * to test for.
1313 	 *
1314 	 */
1315 	if (pdir_order <= (19-12)) {
1316 		if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1317 			/* allocate a new one on 512k alignment */
1318 			unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1319 			/* release original */
1320 			free_pages(pdir_base, pdir_order);
1321 
1322 			pdir_base = new_pdir;
1323 
1324 			/* release excess */
1325 			while (pdir_order < (19-12)) {
1326 				new_pdir += pdir_size;
1327 				free_pages(new_pdir, pdir_order);
1328 				pdir_order +=1;
1329 				pdir_size <<=1;
1330 			}
1331 		}
1332 	} else {
1333 		/*
1334 		** 1MB or 2MB Pdir
1335 		** Needs to be aligned on an "odd" 1MB boundary.
1336 		*/
1337 		unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1338 
1339 		/* release original */
1340 		free_pages( pdir_base, pdir_order);
1341 
1342 		/* release first 1MB */
1343 		free_pages(new_pdir, 20-12);
1344 
1345 		pdir_base = new_pdir + 1024*1024;
1346 
1347 		if (pdir_order > (20-12)) {
1348 			/*
1349 			** 2MB Pdir.
1350 			**
1351 			** Flag tells init_bitmap() to mark bad 128k as used
1352 			** and to reduce the size by 128k.
1353 			*/
1354 			piranha_bad_128k = 1;
1355 
1356 			new_pdir += 3*1024*1024;
1357 			/* release last 1MB */
1358 			free_pages(new_pdir, 20-12);
1359 
1360 			/* release unusable 128KB */
1361 			free_pages(new_pdir - 128*1024 , 17-12);
1362 
1363 			pdir_size -= 128*1024;
1364 		}
1365 	}
1366 
1367 	memset((void *) pdir_base, 0, pdir_size);
1368 	return (void *) pdir_base;
1369 }
1370 
1371 static struct device *next_device(struct klist_iter *i)
1372 {
1373         struct klist_node * n = klist_next(i);
1374         return n ? container_of(n, struct device, knode_parent) : NULL;
1375 }
1376 
1377 /* setup Mercury or Elroy IBASE/IMASK registers. */
1378 static void
1379 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1380 {
1381 	/* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1382         extern void lba_set_iregs(struct parisc_device *, u32, u32);
1383 	struct device *dev;
1384 	struct klist_iter i;
1385 
1386 	klist_iter_init(&sba->dev.klist_children, &i);
1387 	while ((dev = next_device(&i))) {
1388 		struct parisc_device *lba = to_parisc_device(dev);
1389 		int rope_num = (lba->hpa.start >> 13) & 0xf;
1390 		if (rope_num >> 3 == ioc_num)
1391 			lba_set_iregs(lba, ioc->ibase, ioc->imask);
1392 	}
1393 	klist_iter_exit(&i);
1394 }
1395 
1396 static void
1397 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1398 {
1399 	u32 iova_space_mask;
1400 	u32 iova_space_size;
1401 	int iov_order, tcnfg;
1402 #ifdef SBA_AGP_SUPPORT
1403 	int agp_found = 0;
1404 #endif
1405 	/*
1406 	** Firmware programs the base and size of a "safe IOVA space"
1407 	** (one that doesn't overlap memory or LMMIO space) in the
1408 	** IBASE and IMASK registers.
1409 	*/
1410 	ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE);
1411 	iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1412 
1413 	if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1414 		printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1415 		iova_space_size /= 2;
1416 	}
1417 
1418 	/*
1419 	** iov_order is always based on a 1GB IOVA space since we want to
1420 	** turn on the other half for AGP GART.
1421 	*/
1422 	iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1423 	ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1424 
1425 	DBG_INIT("%s() hpa 0x%lx IOV %dMB (%d bits)\n",
1426 		__FUNCTION__, ioc->ioc_hpa, iova_space_size >> 20,
1427 		iov_order + PAGE_SHIFT);
1428 
1429 	ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1430 						   get_order(ioc->pdir_size));
1431 	if (!ioc->pdir_base)
1432 		panic("Couldn't allocate I/O Page Table\n");
1433 
1434 	memset(ioc->pdir_base, 0, ioc->pdir_size);
1435 
1436 	DBG_INIT("%s() pdir %p size %x\n",
1437 			__FUNCTION__, ioc->pdir_base, ioc->pdir_size);
1438 
1439 #ifdef SBA_HINT_SUPPORT
1440 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1441 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1442 
1443 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1444 		ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1445 #endif
1446 
1447 	WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1448 	WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1449 
1450 	/* build IMASK for IOC and Elroy */
1451 	iova_space_mask =  0xffffffff;
1452 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1453 	ioc->imask = iova_space_mask;
1454 #ifdef ZX1_SUPPORT
1455 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1456 #endif
1457 	sba_dump_tlb(ioc->ioc_hpa);
1458 
1459 	setup_ibase_imask(sba, ioc, ioc_num);
1460 
1461 	WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1462 
1463 #ifdef CONFIG_64BIT
1464 	/*
1465 	** Setting the upper bits makes checking for bypass addresses
1466 	** a little faster later on.
1467 	*/
1468 	ioc->imask |= 0xFFFFFFFF00000000UL;
1469 #endif
1470 
1471 	/* Set I/O PDIR Page size to system page size */
1472 	switch (PAGE_SHIFT) {
1473 		case 12: tcnfg = 0; break;	/*  4K */
1474 		case 13: tcnfg = 1; break;	/*  8K */
1475 		case 14: tcnfg = 2; break;	/* 16K */
1476 		case 16: tcnfg = 3; break;	/* 64K */
1477 		default:
1478 			panic(__FILE__ "Unsupported system page size %d",
1479 				1 << PAGE_SHIFT);
1480 			break;
1481 	}
1482 	WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1483 
1484 	/*
1485 	** Program the IOC's ibase and enable IOVA translation
1486 	** Bit zero == enable bit.
1487 	*/
1488 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
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(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1495 
1496 #ifdef SBA_AGP_SUPPORT
1497 	/*
1498 	** If an AGP device is present, only use half of the IOV space
1499 	** for PCI DMA.  Unfortunately we can't know ahead of time
1500 	** whether GART support will actually be used, for now we
1501 	** can just key on any AGP device found in the system.
1502 	** We program the next pdir index after we stop w/ a key for
1503 	** the GART code to handshake on.
1504 	*/
1505 	device=NULL;
1506 	for (lba = sba->child; lba; lba = lba->sibling) {
1507 		if (IS_QUICKSILVER(lba))
1508 			break;
1509 	}
1510 
1511 	if (lba) {
1512 		DBG_INIT("%s: Reserving half of IOVA space for AGP GART support\n", __FUNCTION__);
1513 		ioc->pdir_size /= 2;
1514 		((u64 *)ioc->pdir_base)[PDIR_INDEX(iova_space_size/2)] = SBA_IOMMU_COOKIE;
1515 	} else {
1516 		DBG_INIT("%s: No GART needed - no AGP controller found\n", __FUNCTION__);
1517 	}
1518 #endif /* 0 */
1519 
1520 }
1521 
1522 static void
1523 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1524 {
1525 	u32 iova_space_size, iova_space_mask;
1526 	unsigned int pdir_size, iov_order;
1527 
1528 	/*
1529 	** Determine IOVA Space size from memory size.
1530 	**
1531 	** Ideally, PCI drivers would register the maximum number
1532 	** of DMA they can have outstanding for each device they
1533 	** own.  Next best thing would be to guess how much DMA
1534 	** can be outstanding based on PCI Class/sub-class. Both
1535 	** methods still require some "extra" to support PCI
1536 	** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1537 	**
1538 	** While we have 32-bits "IOVA" space, top two 2 bits are used
1539 	** for DMA hints - ergo only 30 bits max.
1540 	*/
1541 
1542 	iova_space_size = (u32) (num_physpages/global_ioc_cnt);
1543 
1544 	/* limit IOVA space size to 1MB-1GB */
1545 	if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1546 		iova_space_size = 1 << (20 - PAGE_SHIFT);
1547 	}
1548 	else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1549 		iova_space_size = 1 << (30 - PAGE_SHIFT);
1550 	}
1551 
1552 	/*
1553 	** iova space must be log2() in size.
1554 	** thus, pdir/res_map will also be log2().
1555 	** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1556 	*/
1557 	iov_order = get_order(iova_space_size << PAGE_SHIFT);
1558 
1559 	/* iova_space_size is now bytes, not pages */
1560 	iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1561 
1562 	ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1563 
1564 	DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1565 			__FUNCTION__,
1566 			ioc->ioc_hpa,
1567 			(unsigned long) num_physpages >> (20 - PAGE_SHIFT),
1568 			iova_space_size>>20,
1569 			iov_order + PAGE_SHIFT);
1570 
1571 	ioc->pdir_base = sba_alloc_pdir(pdir_size);
1572 
1573 	DBG_INIT("%s() pdir %p size %x\n",
1574 			__FUNCTION__, ioc->pdir_base, pdir_size);
1575 
1576 #ifdef SBA_HINT_SUPPORT
1577 	/* FIXME : DMA HINTs not used */
1578 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1579 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1580 
1581 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1582 			ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1583 #endif
1584 
1585 	WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1586 
1587 	/* build IMASK for IOC and Elroy */
1588 	iova_space_mask =  0xffffffff;
1589 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1590 
1591 	/*
1592 	** On C3000 w/512MB mem, HP-UX 10.20 reports:
1593 	**     ibase=0, imask=0xFE000000, size=0x2000000.
1594 	*/
1595 	ioc->ibase = 0;
1596 	ioc->imask = iova_space_mask;	/* save it */
1597 #ifdef ZX1_SUPPORT
1598 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1599 #endif
1600 
1601 	DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1602 		__FUNCTION__, ioc->ibase, ioc->imask);
1603 
1604 	/*
1605 	** FIXME: Hint registers are programmed with default hint
1606 	** values during boot, so hints should be sane even if we
1607 	** can't reprogram them the way drivers want.
1608 	*/
1609 
1610 	setup_ibase_imask(sba, ioc, ioc_num);
1611 
1612 	/*
1613 	** Program the IOC's ibase and enable IOVA translation
1614 	*/
1615 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1616 	WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1617 
1618 	/* Set I/O PDIR Page size to 4K */
1619 	WRITE_REG(0, ioc->ioc_hpa+IOC_TCNFG);
1620 
1621 	/*
1622 	** Clear I/O TLB of any possible entries.
1623 	** (Yes. This is a bit paranoid...but so what)
1624 	*/
1625 	WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1626 
1627 	ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1628 
1629 	DBG_INIT("%s() DONE\n", __FUNCTION__);
1630 }
1631 
1632 
1633 
1634 /**************************************************************************
1635 **
1636 **   SBA initialization code (HW and SW)
1637 **
1638 **   o identify SBA chip itself
1639 **   o initialize SBA chip modes (HardFail)
1640 **   o initialize SBA chip modes (HardFail)
1641 **   o FIXME: initialize DMA hints for reasonable defaults
1642 **
1643 **************************************************************************/
1644 
1645 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1646 {
1647 	return ioremap_nocache(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1648 }
1649 
1650 static void sba_hw_init(struct sba_device *sba_dev)
1651 {
1652 	int i;
1653 	int num_ioc;
1654 	u64 ioc_ctl;
1655 
1656 	if (!is_pdc_pat()) {
1657 		/* Shutdown the USB controller on Astro-based workstations.
1658 		** Once we reprogram the IOMMU, the next DMA performed by
1659 		** USB will HPMC the box. USB is only enabled if a
1660 		** keyboard is present and found.
1661 		**
1662 		** With serial console, j6k v5.0 firmware says:
1663 		**   mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1664 		**
1665 		** FIXME: Using GFX+USB console at power up but direct
1666 		**	linux to serial console is still broken.
1667 		**	USB could generate DMA so we must reset USB.
1668 		**	The proper sequence would be:
1669 		**	o block console output
1670 		**	o reset USB device
1671 		**	o reprogram serial port
1672 		**	o unblock console output
1673 		*/
1674 		if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1675 			pdc_io_reset_devices();
1676 		}
1677 
1678 	}
1679 
1680 
1681 #if 0
1682 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1683 	PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1684 
1685 	/*
1686 	** Need to deal with DMA from LAN.
1687 	**	Maybe use page zero boot device as a handle to talk
1688 	**	to PDC about which device to shutdown.
1689 	**
1690 	** Netbooting, j6k v5.0 firmware says:
1691 	** 	mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1692 	** ARGH! invalid class.
1693 	*/
1694 	if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1695 		&& (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1696 			pdc_io_reset();
1697 	}
1698 #endif
1699 
1700 	if (!IS_PLUTO(sba_dev->iodc)) {
1701 		ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1702 		DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1703 			__FUNCTION__, sba_dev->sba_hpa, ioc_ctl);
1704 		ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1705 		ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1706 			/* j6700 v1.6 firmware sets 0x294f */
1707 			/* A500 firmware sets 0x4d */
1708 
1709 		WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1710 
1711 #ifdef DEBUG_SBA_INIT
1712 		ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1713 		DBG_INIT(" 0x%Lx\n", ioc_ctl);
1714 #endif
1715 	} /* if !PLUTO */
1716 
1717 	if (IS_ASTRO(sba_dev->iodc)) {
1718 		int err;
1719 		/* PAT_PDC (L-class) also reports the same goofy base */
1720 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1721 		num_ioc = 1;
1722 
1723 		sba_dev->chip_resv.name = "Astro Intr Ack";
1724 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1725 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff000000UL - 1) ;
1726 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1727 		BUG_ON(err < 0);
1728 
1729 	} else if (IS_PLUTO(sba_dev->iodc)) {
1730 		int err;
1731 
1732 		/* We use a negative value for IOC HPA so it gets
1733                  * corrected when we add it with IKE's IOC offset.
1734 		 * Doesnt look clean, but fewer code.
1735                  */
1736 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1737 		num_ioc = 1;
1738 
1739 		sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1740 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1741 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff200000UL - 1);
1742 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1743 		WARN_ON(err < 0);
1744 
1745 		sba_dev->iommu_resv.name = "IOVA Space";
1746 		sba_dev->iommu_resv.start = 0x40000000UL;
1747 		sba_dev->iommu_resv.end   = 0x50000000UL - 1;
1748 		err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1749 		WARN_ON(err < 0);
1750 	} else {
1751 		/* IS_IKE (ie N-class, L3000, L1500) */
1752 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1753 		sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1754 		num_ioc = 2;
1755 
1756 		/* TODO - LOOKUP Ike/Stretch chipset mem map */
1757 	}
1758 	/* XXX: What about Reo? */
1759 
1760 	sba_dev->num_ioc = num_ioc;
1761 	for (i = 0; i < num_ioc; i++) {
1762 		/*
1763 		** Make sure the box crashes if we get any errors on a rope.
1764 		*/
1765 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE0_CTL);
1766 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE1_CTL);
1767 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE2_CTL);
1768 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE3_CTL);
1769 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE4_CTL);
1770 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE5_CTL);
1771 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE6_CTL);
1772 		WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1773 
1774 		/* flush out the writes */
1775 		READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1776 
1777 		DBG_INIT("	ioc[%d] ROPE_CFG 0x%Lx  ROPE_DBG 0x%Lx\n",
1778 				i,
1779 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1780 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1781 			);
1782 		DBG_INIT("	STATUS_CONTROL 0x%Lx  FLUSH_CTRL 0x%Lx\n",
1783 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1784 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1785 			);
1786 
1787 		if (IS_PLUTO(sba_dev->iodc)) {
1788 			sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1789 		} else {
1790 			sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1791 		}
1792 	}
1793 }
1794 
1795 static void
1796 sba_common_init(struct sba_device *sba_dev)
1797 {
1798 	int i;
1799 
1800 	/* add this one to the head of the list (order doesn't matter)
1801 	** This will be useful for debugging - especially if we get coredumps
1802 	*/
1803 	sba_dev->next = sba_list;
1804 	sba_list = sba_dev;
1805 
1806 	for(i=0; i< sba_dev->num_ioc; i++) {
1807 		int res_size;
1808 #ifdef DEBUG_DMB_TRAP
1809 		extern void iterate_pages(unsigned long , unsigned long ,
1810 					  void (*)(pte_t * , unsigned long),
1811 					  unsigned long );
1812 		void set_data_memory_break(pte_t * , unsigned long);
1813 #endif
1814 		/* resource map size dictated by pdir_size */
1815 		res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1816 
1817 		/* Second part of PIRANHA BUG */
1818 		if (piranha_bad_128k) {
1819 			res_size -= (128*1024)/sizeof(u64);
1820 		}
1821 
1822 		res_size >>= 3;  /* convert bit count to byte count */
1823 		DBG_INIT("%s() res_size 0x%x\n",
1824 			__FUNCTION__, res_size);
1825 
1826 		sba_dev->ioc[i].res_size = res_size;
1827 		sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1828 
1829 #ifdef DEBUG_DMB_TRAP
1830 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1831 				set_data_memory_break, 0);
1832 #endif
1833 
1834 		if (NULL == sba_dev->ioc[i].res_map)
1835 		{
1836 			panic("%s:%s() could not allocate resource map\n",
1837 			      __FILE__, __FUNCTION__ );
1838 		}
1839 
1840 		memset(sba_dev->ioc[i].res_map, 0, res_size);
1841 		/* next available IOVP - circular search */
1842 		sba_dev->ioc[i].res_hint = (unsigned long *)
1843 				&(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1844 
1845 #ifdef ASSERT_PDIR_SANITY
1846 		/* Mark first bit busy - ie no IOVA 0 */
1847 		sba_dev->ioc[i].res_map[0] = 0x80;
1848 		sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1849 #endif
1850 
1851 		/* Third (and last) part of PIRANHA BUG */
1852 		if (piranha_bad_128k) {
1853 			/* region from +1408K to +1536 is un-usable. */
1854 
1855 			int idx_start = (1408*1024/sizeof(u64)) >> 3;
1856 			int idx_end   = (1536*1024/sizeof(u64)) >> 3;
1857 			long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1858 			long *p_end   = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1859 
1860 			/* mark that part of the io pdir busy */
1861 			while (p_start < p_end)
1862 				*p_start++ = -1;
1863 
1864 		}
1865 
1866 #ifdef DEBUG_DMB_TRAP
1867 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1868 				set_data_memory_break, 0);
1869 		iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1870 				set_data_memory_break, 0);
1871 #endif
1872 
1873 		DBG_INIT("%s() %d res_map %x %p\n",
1874 			__FUNCTION__, i, res_size, sba_dev->ioc[i].res_map);
1875 	}
1876 
1877 	spin_lock_init(&sba_dev->sba_lock);
1878 	ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1879 
1880 #ifdef DEBUG_SBA_INIT
1881 	/*
1882 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1883 	 * (bit #61, big endian), we have to flush and sync every time
1884 	 * IO-PDIR is changed in Ike/Astro.
1885 	 */
1886 	if (boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC) {
1887 		printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1888 	} else {
1889 		printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1890 	}
1891 #endif
1892 }
1893 
1894 #ifdef CONFIG_PROC_FS
1895 static int sba_proc_info(struct seq_file *m, void *p)
1896 {
1897 	struct sba_device *sba_dev = sba_list;
1898 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1899 	int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1900 #ifdef SBA_COLLECT_STATS
1901 	unsigned long avg = 0, min, max;
1902 #endif
1903 	int i, len = 0;
1904 
1905 	len += seq_printf(m, "%s rev %d.%d\n",
1906 		sba_dev->name,
1907 		(sba_dev->hw_rev & 0x7) + 1,
1908 		(sba_dev->hw_rev & 0x18) >> 3
1909 		);
1910 	len += seq_printf(m, "IO PDIR size    : %d bytes (%d entries)\n",
1911 		(int) ((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1912 		total_pages);
1913 
1914 	len += seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1915 		ioc->res_size, ioc->res_size << 3);   /* 8 bits per byte */
1916 
1917 	len += seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1918 		READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1919 		READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1920 		READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE)
1921 		);
1922 
1923 	for (i=0; i<4; i++)
1924 		len += seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n", i,
1925 			READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE  + i*0x18),
1926 			READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK  + i*0x18),
1927 			READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18)
1928 		);
1929 
1930 #ifdef SBA_COLLECT_STATS
1931 	len += seq_printf(m, "IO PDIR entries : %ld free  %ld used (%d%%)\n",
1932 		total_pages - ioc->used_pages, ioc->used_pages,
1933 		(int) (ioc->used_pages * 100 / total_pages));
1934 
1935 	min = max = ioc->avg_search[0];
1936 	for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1937 		avg += ioc->avg_search[i];
1938 		if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1939 		if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1940 	}
1941 	avg /= SBA_SEARCH_SAMPLE;
1942 	len += seq_printf(m, "  Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1943 		min, avg, max);
1944 
1945 	len += seq_printf(m, "pci_map_single(): %12ld calls  %12ld pages (avg %d/1000)\n",
1946 		ioc->msingle_calls, ioc->msingle_pages,
1947 		(int) ((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1948 
1949 	/* KLUGE - unmap_sg calls unmap_single for each mapped page */
1950 	min = ioc->usingle_calls;
1951 	max = ioc->usingle_pages - ioc->usg_pages;
1952 	len += seq_printf(m, "pci_unmap_single: %12ld calls  %12ld pages (avg %d/1000)\n",
1953 		min, max, (int) ((max * 1000)/min));
1954 
1955 	len += seq_printf(m, "pci_map_sg()    : %12ld calls  %12ld pages (avg %d/1000)\n",
1956 		ioc->msg_calls, ioc->msg_pages,
1957 		(int) ((ioc->msg_pages * 1000)/ioc->msg_calls));
1958 
1959 	len += seq_printf(m, "pci_unmap_sg()  : %12ld calls  %12ld pages (avg %d/1000)\n",
1960 		ioc->usg_calls, ioc->usg_pages,
1961 		(int) ((ioc->usg_pages * 1000)/ioc->usg_calls));
1962 #endif
1963 
1964 	return 0;
1965 }
1966 
1967 static int
1968 sba_proc_open(struct inode *i, struct file *f)
1969 {
1970 	return single_open(f, &sba_proc_info, NULL);
1971 }
1972 
1973 static struct file_operations sba_proc_fops = {
1974 	.owner = THIS_MODULE,
1975 	.open = sba_proc_open,
1976 	.read = seq_read,
1977 	.llseek = seq_lseek,
1978 	.release = single_release,
1979 };
1980 
1981 static int
1982 sba_proc_bitmap_info(struct seq_file *m, void *p)
1983 {
1984 	struct sba_device *sba_dev = sba_list;
1985 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1986 	unsigned int *res_ptr = (unsigned int *)ioc->res_map;
1987 	int i, len = 0;
1988 
1989 	for (i = 0; i < (ioc->res_size/sizeof(unsigned int)); ++i, ++res_ptr) {
1990 		if ((i & 7) == 0)
1991 			len += seq_printf(m, "\n   ");
1992 		len += seq_printf(m, " %08x", *res_ptr);
1993 	}
1994 	len += seq_printf(m, "\n");
1995 
1996 	return 0;
1997 }
1998 
1999 static int
2000 sba_proc_bitmap_open(struct inode *i, struct file *f)
2001 {
2002 	return single_open(f, &sba_proc_bitmap_info, NULL);
2003 }
2004 
2005 static struct file_operations sba_proc_bitmap_fops = {
2006 	.owner = THIS_MODULE,
2007 	.open = sba_proc_bitmap_open,
2008 	.read = seq_read,
2009 	.llseek = seq_lseek,
2010 	.release = single_release,
2011 };
2012 #endif /* CONFIG_PROC_FS */
2013 
2014 static struct parisc_device_id sba_tbl[] = {
2015 	{ HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
2016 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
2017 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
2018 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
2019 	{ HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
2020 	{ 0, }
2021 };
2022 
2023 int sba_driver_callback(struct parisc_device *);
2024 
2025 static struct parisc_driver sba_driver = {
2026 	.name =		MODULE_NAME,
2027 	.id_table =	sba_tbl,
2028 	.probe =	sba_driver_callback,
2029 };
2030 
2031 /*
2032 ** Determine if sba should claim this chip (return 0) or not (return 1).
2033 ** If so, initialize the chip and tell other partners in crime they
2034 ** have work to do.
2035 */
2036 int
2037 sba_driver_callback(struct parisc_device *dev)
2038 {
2039 	struct sba_device *sba_dev;
2040 	u32 func_class;
2041 	int i;
2042 	char *version;
2043 	void __iomem *sba_addr = ioremap_nocache(dev->hpa.start, SBA_FUNC_SIZE);
2044 	struct proc_dir_entry *info_entry, *bitmap_entry, *root;
2045 
2046 	sba_dump_ranges(sba_addr);
2047 
2048 	/* Read HW Rev First */
2049 	func_class = READ_REG(sba_addr + SBA_FCLASS);
2050 
2051 	if (IS_ASTRO(&dev->id)) {
2052 		unsigned long fclass;
2053 		static char astro_rev[]="Astro ?.?";
2054 
2055 		/* Astro is broken...Read HW Rev First */
2056 		fclass = READ_REG(sba_addr);
2057 
2058 		astro_rev[6] = '1' + (char) (fclass & 0x7);
2059 		astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
2060 		version = astro_rev;
2061 
2062 	} else if (IS_IKE(&dev->id)) {
2063 		static char ike_rev[] = "Ike rev ?";
2064 		ike_rev[8] = '0' + (char) (func_class & 0xff);
2065 		version = ike_rev;
2066 	} else if (IS_PLUTO(&dev->id)) {
2067 		static char pluto_rev[]="Pluto ?.?";
2068 		pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
2069 		pluto_rev[8] = '0' + (char) (func_class & 0x0f);
2070 		version = pluto_rev;
2071 	} else {
2072 		static char reo_rev[] = "REO rev ?";
2073 		reo_rev[8] = '0' + (char) (func_class & 0xff);
2074 		version = reo_rev;
2075 	}
2076 
2077 	if (!global_ioc_cnt) {
2078 		global_ioc_cnt = count_parisc_driver(&sba_driver);
2079 
2080 		/* Astro and Pluto have one IOC per SBA */
2081 		if ((!IS_ASTRO(&dev->id)) || (!IS_PLUTO(&dev->id)))
2082 			global_ioc_cnt *= 2;
2083 	}
2084 
2085 	printk(KERN_INFO "%s found %s at 0x%lx\n",
2086 		MODULE_NAME, version, dev->hpa.start);
2087 
2088 	sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
2089 	if (!sba_dev) {
2090 		printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
2091 		return -ENOMEM;
2092 	}
2093 
2094 	parisc_set_drvdata(dev, sba_dev);
2095 
2096 	for(i=0; i<MAX_IOC; i++)
2097 		spin_lock_init(&(sba_dev->ioc[i].res_lock));
2098 
2099 	sba_dev->dev = dev;
2100 	sba_dev->hw_rev = func_class;
2101 	sba_dev->iodc = &dev->id;
2102 	sba_dev->name = dev->name;
2103 	sba_dev->sba_hpa = sba_addr;
2104 
2105 	sba_get_pat_resources(sba_dev);
2106 	sba_hw_init(sba_dev);
2107 	sba_common_init(sba_dev);
2108 
2109 	hppa_dma_ops = &sba_ops;
2110 
2111 #ifdef CONFIG_PROC_FS
2112 	switch (dev->id.hversion) {
2113 	case PLUTO_MCKINLEY_PORT:
2114 		root = proc_mckinley_root;
2115 		break;
2116 	case ASTRO_RUNWAY_PORT:
2117 	case IKE_MERCED_PORT:
2118 	default:
2119 		root = proc_runway_root;
2120 		break;
2121 	}
2122 
2123 	info_entry = create_proc_entry("sba_iommu", 0, root);
2124 	bitmap_entry = create_proc_entry("sba_iommu-bitmap", 0, root);
2125 
2126 	if (info_entry)
2127 		info_entry->proc_fops = &sba_proc_fops;
2128 
2129 	if (bitmap_entry)
2130 		bitmap_entry->proc_fops = &sba_proc_bitmap_fops;
2131 #endif
2132 
2133 	parisc_vmerge_boundary = IOVP_SIZE;
2134 	parisc_vmerge_max_size = IOVP_SIZE * BITS_PER_LONG;
2135 	parisc_has_iommu();
2136 	return 0;
2137 }
2138 
2139 /*
2140 ** One time initialization to let the world know the SBA was found.
2141 ** This is the only routine which is NOT static.
2142 ** Must be called exactly once before pci_init().
2143 */
2144 void __init sba_init(void)
2145 {
2146 	register_parisc_driver(&sba_driver);
2147 }
2148 
2149 
2150 /**
2151  * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2152  * @dev: The parisc device.
2153  *
2154  * Returns the appropriate IOMMU data for the given parisc PCI controller.
2155  * This is cached and used later for PCI DMA Mapping.
2156  */
2157 void * sba_get_iommu(struct parisc_device *pci_hba)
2158 {
2159 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2160 	struct sba_device *sba = sba_dev->dev.driver_data;
2161 	char t = sba_dev->id.hw_type;
2162 	int iocnum = (pci_hba->hw_path >> 3);	/* rope # */
2163 
2164 	WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2165 
2166 	return &(sba->ioc[iocnum]);
2167 }
2168 
2169 
2170 /**
2171  * sba_directed_lmmio - return first directed LMMIO range routed to rope
2172  * @pa_dev: The parisc device.
2173  * @r: resource PCI host controller wants start/end fields assigned.
2174  *
2175  * For the given parisc PCI controller, determine if any direct ranges
2176  * are routed down the corresponding rope.
2177  */
2178 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2179 {
2180 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2181 	struct sba_device *sba = sba_dev->dev.driver_data;
2182 	char t = sba_dev->id.hw_type;
2183 	int i;
2184 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2185 
2186 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2187 
2188 	r->start = r->end = 0;
2189 
2190 	/* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2191 	for (i=0; i<4; i++) {
2192 		int base, size;
2193 		void __iomem *reg = sba->sba_hpa + i*0x18;
2194 
2195 		base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2196 		if ((base & 1) == 0)
2197 			continue;	/* not enabled */
2198 
2199 		size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2200 
2201 		if ((size & (ROPES_PER_IOC-1)) != rope)
2202 			continue;	/* directed down different rope */
2203 
2204 		r->start = (base & ~1UL) | PCI_F_EXTEND;
2205 		size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2206 		r->end = r->start + size;
2207 	}
2208 }
2209 
2210 
2211 /**
2212  * sba_distributed_lmmio - return portion of distributed LMMIO range
2213  * @pa_dev: The parisc device.
2214  * @r: resource PCI host controller wants start/end fields assigned.
2215  *
2216  * For the given parisc PCI controller, return portion of distributed LMMIO
2217  * range. The distributed LMMIO is always present and it's just a question
2218  * of the base address and size of the range.
2219  */
2220 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2221 {
2222 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2223 	struct sba_device *sba = sba_dev->dev.driver_data;
2224 	char t = sba_dev->id.hw_type;
2225 	int base, size;
2226 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2227 
2228 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2229 
2230 	r->start = r->end = 0;
2231 
2232 	base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2233 	if ((base & 1) == 0) {
2234 		BUG();	/* Gah! Distr Range wasn't enabled! */
2235 		return;
2236 	}
2237 
2238 	r->start = (base & ~1UL) | PCI_F_EXTEND;
2239 
2240 	size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2241 	r->start += rope * (size + 1);	/* adjust base for this rope */
2242 	r->end = r->start + size;
2243 }
2244