xref: /openbmc/linux/drivers/parisc/sba_iommu.c (revision 479965a2)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 **  System Bus Adapter (SBA) I/O MMU manager
4 **
5 **	(c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
6 **	(c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
7 **	(c) Copyright 2000-2004 Hewlett-Packard Company
8 **
9 **	Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
10 **
11 **
12 **
13 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
14 ** J5000/J7000/N-class/L-class machines and their successors.
15 **
16 ** FIXME: add DMA hint support programming in both sba and lba modules.
17 */
18 
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/spinlock.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 
25 #include <linux/mm.h>
26 #include <linux/string.h>
27 #include <linux/pci.h>
28 #include <linux/dma-map-ops.h>
29 #include <linux/scatterlist.h>
30 #include <linux/iommu-helper.h>
31 /*
32  * The semantics of 64 register access on 32bit systems can't be guaranteed
33  * by the C standard, we hope the _lo_hi() macros defining readq and writeq
34  * here will behave as expected.
35  */
36 #include <linux/io-64-nonatomic-lo-hi.h>
37 
38 #include <asm/byteorder.h>
39 #include <asm/io.h>
40 #include <asm/dma.h>		/* for DMA_CHUNK_SIZE */
41 
42 #include <asm/hardware.h>	/* for register_parisc_driver() stuff */
43 
44 #include <linux/proc_fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/module.h>
47 
48 #include <asm/ropes.h>
49 #include <asm/mckinley.h>	/* for proc_mckinley_root */
50 #include <asm/runway.h>		/* for proc_runway_root */
51 #include <asm/page.h>		/* for PAGE0 */
52 #include <asm/pdc.h>		/* for PDC_MODEL_* */
53 #include <asm/pdcpat.h>		/* for is_pdc_pat() */
54 #include <asm/parisc-device.h>
55 
56 #include "iommu.h"
57 
58 #define MODULE_NAME "SBA"
59 
60 /*
61 ** The number of debug flags is a clue - this code is fragile.
62 ** Don't even think about messing with it unless you have
63 ** plenty of 710's to sacrifice to the computer gods. :^)
64 */
65 #undef DEBUG_SBA_INIT
66 #undef DEBUG_SBA_RUN
67 #undef DEBUG_SBA_RUN_SG
68 #undef DEBUG_SBA_RESOURCE
69 #undef ASSERT_PDIR_SANITY
70 #undef DEBUG_LARGE_SG_ENTRIES
71 #undef DEBUG_DMB_TRAP
72 
73 #ifdef DEBUG_SBA_INIT
74 #define DBG_INIT(x...)	printk(x)
75 #else
76 #define DBG_INIT(x...)
77 #endif
78 
79 #ifdef DEBUG_SBA_RUN
80 #define DBG_RUN(x...)	printk(x)
81 #else
82 #define DBG_RUN(x...)
83 #endif
84 
85 #ifdef DEBUG_SBA_RUN_SG
86 #define DBG_RUN_SG(x...)	printk(x)
87 #else
88 #define DBG_RUN_SG(x...)
89 #endif
90 
91 
92 #ifdef DEBUG_SBA_RESOURCE
93 #define DBG_RES(x...)	printk(x)
94 #else
95 #define DBG_RES(x...)
96 #endif
97 
98 #define DEFAULT_DMA_HINT_REG	0
99 
100 struct sba_device *sba_list;
101 EXPORT_SYMBOL_GPL(sba_list);
102 
103 static unsigned long ioc_needs_fdc = 0;
104 
105 /* global count of IOMMUs in the system */
106 static unsigned int global_ioc_cnt = 0;
107 
108 /* PA8700 (Piranha 2.2) bug workaround */
109 static unsigned long piranha_bad_128k = 0;
110 
111 /* Looks nice and keeps the compiler happy */
112 #define SBA_DEV(d) ((struct sba_device *) (d))
113 
114 #ifdef CONFIG_AGP_PARISC
115 #define SBA_AGP_SUPPORT
116 #endif /*CONFIG_AGP_PARISC*/
117 
118 #ifdef SBA_AGP_SUPPORT
119 static int sba_reserve_agpgart = 1;
120 module_param(sba_reserve_agpgart, int, 0444);
121 MODULE_PARM_DESC(sba_reserve_agpgart, "Reserve half of IO pdir as AGPGART");
122 #endif
123 
124 static struct proc_dir_entry *proc_runway_root __ro_after_init;
125 struct proc_dir_entry *proc_mckinley_root __ro_after_init;
126 
127 /************************************
128 ** SBA register read and write support
129 **
130 ** BE WARNED: register writes are posted.
131 **  (ie follow writes which must reach HW with a read)
132 **
133 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
134 */
135 #define READ_REG32(addr)	readl(addr)
136 #define READ_REG64(addr)	readq(addr)
137 #define WRITE_REG32(val, addr)	writel((val), (addr))
138 #define WRITE_REG64(val, addr)	writeq((val), (addr))
139 
140 #ifdef CONFIG_64BIT
141 #define READ_REG(addr)		READ_REG64(addr)
142 #define WRITE_REG(value, addr)	WRITE_REG64(value, addr)
143 #else
144 #define READ_REG(addr)		READ_REG32(addr)
145 #define WRITE_REG(value, addr)	WRITE_REG32(value, addr)
146 #endif
147 
148 #ifdef DEBUG_SBA_INIT
149 
150 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
151 
152 /**
153  * sba_dump_ranges - debugging only - print ranges assigned to this IOA
154  * @hpa: base address of the sba
155  *
156  * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
157  * IO Adapter (aka Bus Converter).
158  */
159 static void
160 sba_dump_ranges(void __iomem *hpa)
161 {
162 	DBG_INIT("SBA at 0x%p\n", hpa);
163 	DBG_INIT("IOS_DIST_BASE   : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
164 	DBG_INIT("IOS_DIST_MASK   : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
165 	DBG_INIT("IOS_DIST_ROUTE  : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
166 	DBG_INIT("\n");
167 	DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
168 	DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
169 	DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
170 }
171 
172 /**
173  * sba_dump_tlb - debugging only - print IOMMU operating parameters
174  * @hpa: base address of the IOMMU
175  *
176  * Print the size/location of the IO MMU PDIR.
177  */
178 static void sba_dump_tlb(void __iomem *hpa)
179 {
180 	DBG_INIT("IO TLB at 0x%p\n", hpa);
181 	DBG_INIT("IOC_IBASE    : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
182 	DBG_INIT("IOC_IMASK    : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
183 	DBG_INIT("IOC_TCNFG    : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
184 	DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
185 	DBG_INIT("\n");
186 }
187 #else
188 #define sba_dump_ranges(x)
189 #define sba_dump_tlb(x)
190 #endif	/* DEBUG_SBA_INIT */
191 
192 
193 #ifdef ASSERT_PDIR_SANITY
194 
195 /**
196  * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
197  * @ioc: IO MMU structure which owns the pdir we are interested in.
198  * @msg: text to print ont the output line.
199  * @pide: pdir index.
200  *
201  * Print one entry of the IO MMU PDIR in human readable form.
202  */
203 static void
204 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
205 {
206 	/* start printing from lowest pde in rval */
207 	u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
208 	unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
209 	uint rcnt;
210 
211 	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
212 		 msg,
213 		 rptr, pide & (BITS_PER_LONG - 1), *rptr);
214 
215 	rcnt = 0;
216 	while (rcnt < BITS_PER_LONG) {
217 		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
218 			(rcnt == (pide & (BITS_PER_LONG - 1)))
219 				? "    -->" : "       ",
220 			rcnt, ptr, *ptr );
221 		rcnt++;
222 		ptr++;
223 	}
224 	printk(KERN_DEBUG "%s", msg);
225 }
226 
227 
228 /**
229  * sba_check_pdir - debugging only - consistency checker
230  * @ioc: IO MMU structure which owns the pdir we are interested in.
231  * @msg: text to print ont the output line.
232  *
233  * Verify the resource map and pdir state is consistent
234  */
235 static int
236 sba_check_pdir(struct ioc *ioc, char *msg)
237 {
238 	u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
239 	u32 *rptr = (u32 *) ioc->res_map;	/* resource map ptr */
240 	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
241 	uint pide = 0;
242 
243 	while (rptr < rptr_end) {
244 		u32 rval = *rptr;
245 		int rcnt = 32;	/* number of bits we might check */
246 
247 		while (rcnt) {
248 			/* Get last byte and highest bit from that */
249 			u32 pde = ((u32) (((char *)pptr)[7])) << 24;
250 			if ((rval ^ pde) & 0x80000000)
251 			{
252 				/*
253 				** BUMMER!  -- res_map != pdir --
254 				** Dump rval and matching pdir entries
255 				*/
256 				sba_dump_pdir_entry(ioc, msg, pide);
257 				return(1);
258 			}
259 			rcnt--;
260 			rval <<= 1;	/* try the next bit */
261 			pptr++;
262 			pide++;
263 		}
264 		rptr++;	/* look at next word of res_map */
265 	}
266 	/* It'd be nice if we always got here :^) */
267 	return 0;
268 }
269 
270 
271 /**
272  * sba_dump_sg - debugging only - print Scatter-Gather list
273  * @ioc: IO MMU structure which owns the pdir we are interested in.
274  * @startsg: head of the SG list
275  * @nents: number of entries in SG list
276  *
277  * print the SG list so we can verify it's correct by hand.
278  */
279 static void
280 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
281 {
282 	while (nents-- > 0) {
283 		printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
284 				nents,
285 				(unsigned long) sg_dma_address(startsg),
286 				sg_dma_len(startsg),
287 				sg_virt(startsg), startsg->length);
288 		startsg++;
289 	}
290 }
291 
292 #endif /* ASSERT_PDIR_SANITY */
293 
294 
295 
296 
297 /**************************************************************
298 *
299 *   I/O Pdir Resource Management
300 *
301 *   Bits set in the resource map are in use.
302 *   Each bit can represent a number of pages.
303 *   LSbs represent lower addresses (IOVA's).
304 *
305 ***************************************************************/
306 #define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
307 
308 /* Convert from IOVP to IOVA and vice versa. */
309 
310 #ifdef ZX1_SUPPORT
311 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
312 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
313 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
314 #else
315 /* only support Astro and ancestors. Saves a few cycles in key places */
316 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
317 #define SBA_IOVP(ioc,iova) (iova)
318 #endif
319 
320 #define PDIR_INDEX(iovp)   ((iovp)>>IOVP_SHIFT)
321 
322 #define RESMAP_MASK(n)    (~0UL << (BITS_PER_LONG - (n)))
323 #define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)
324 
325 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
326 				 unsigned int bitshiftcnt)
327 {
328 	return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
329 		+ bitshiftcnt;
330 }
331 
332 /**
333  * sba_search_bitmap - find free space in IO PDIR resource bitmap
334  * @ioc: IO MMU structure which owns the pdir we are interested in.
335  * @dev: device to query the bitmap for
336  * @bits_wanted: number of entries we need.
337  *
338  * Find consecutive free bits in resource bitmap.
339  * Each bit represents one entry in the IO Pdir.
340  * Cool perf optimization: search for log2(size) bits at a time.
341  */
342 static unsigned long
343 sba_search_bitmap(struct ioc *ioc, struct device *dev,
344 		  unsigned long bits_wanted)
345 {
346 	unsigned long *res_ptr = ioc->res_hint;
347 	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
348 	unsigned long pide = ~0UL, tpide;
349 	unsigned long boundary_size;
350 	unsigned long shift;
351 	int ret;
352 
353 	boundary_size = dma_get_seg_boundary_nr_pages(dev, IOVP_SHIFT);
354 
355 #if defined(ZX1_SUPPORT)
356 	BUG_ON(ioc->ibase & ~IOVP_MASK);
357 	shift = ioc->ibase >> IOVP_SHIFT;
358 #else
359 	shift = 0;
360 #endif
361 
362 	if (bits_wanted > (BITS_PER_LONG/2)) {
363 		/* Search word at a time - no mask needed */
364 		for(; res_ptr < res_end; ++res_ptr) {
365 			tpide = ptr_to_pide(ioc, res_ptr, 0);
366 			ret = iommu_is_span_boundary(tpide, bits_wanted,
367 						     shift,
368 						     boundary_size);
369 			if ((*res_ptr == 0) && !ret) {
370 				*res_ptr = RESMAP_MASK(bits_wanted);
371 				pide = tpide;
372 				break;
373 			}
374 		}
375 		/* point to the next word on next pass */
376 		res_ptr++;
377 		ioc->res_bitshift = 0;
378 	} else {
379 		/*
380 		** Search the resource bit map on well-aligned values.
381 		** "o" is the alignment.
382 		** We need the alignment to invalidate I/O TLB using
383 		** SBA HW features in the unmap path.
384 		*/
385 		unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
386 		uint bitshiftcnt = ALIGN(ioc->res_bitshift, o);
387 		unsigned long mask;
388 
389 		if (bitshiftcnt >= BITS_PER_LONG) {
390 			bitshiftcnt = 0;
391 			res_ptr++;
392 		}
393 		mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
394 
395 		DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
396 		while(res_ptr < res_end)
397 		{
398 			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
399 			WARN_ON(mask == 0);
400 			tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
401 			ret = iommu_is_span_boundary(tpide, bits_wanted,
402 						     shift,
403 						     boundary_size);
404 			if ((((*res_ptr) & mask) == 0) && !ret) {
405 				*res_ptr |= mask;     /* mark resources busy! */
406 				pide = tpide;
407 				break;
408 			}
409 			mask >>= o;
410 			bitshiftcnt += o;
411 			if (mask == 0) {
412 				mask = RESMAP_MASK(bits_wanted);
413 				bitshiftcnt=0;
414 				res_ptr++;
415 			}
416 		}
417 		/* look in the same word on the next pass */
418 		ioc->res_bitshift = bitshiftcnt + bits_wanted;
419 	}
420 
421 	/* wrapped ? */
422 	if (res_end <= res_ptr) {
423 		ioc->res_hint = (unsigned long *) ioc->res_map;
424 		ioc->res_bitshift = 0;
425 	} else {
426 		ioc->res_hint = res_ptr;
427 	}
428 	return (pide);
429 }
430 
431 
432 /**
433  * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
434  * @ioc: IO MMU structure which owns the pdir we are interested in.
435  * @dev: device for which pages should be alloced
436  * @size: number of bytes to create a mapping for
437  *
438  * Given a size, find consecutive unmarked and then mark those bits in the
439  * resource bit map.
440  */
441 static int
442 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
443 {
444 	unsigned int pages_needed = size >> IOVP_SHIFT;
445 #ifdef SBA_COLLECT_STATS
446 	unsigned long cr_start = mfctl(16);
447 #endif
448 	unsigned long pide;
449 
450 	pide = sba_search_bitmap(ioc, dev, pages_needed);
451 	if (pide >= (ioc->res_size << 3)) {
452 		pide = sba_search_bitmap(ioc, dev, pages_needed);
453 		if (pide >= (ioc->res_size << 3))
454 			panic("%s: I/O MMU @ %p is out of mapping resources\n",
455 			      __FILE__, ioc->ioc_hpa);
456 	}
457 
458 #ifdef ASSERT_PDIR_SANITY
459 	/* verify the first enable bit is clear */
460 	if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
461 		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
462 	}
463 #endif
464 
465 	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
466 		__func__, size, pages_needed, pide,
467 		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
468 		ioc->res_bitshift );
469 
470 #ifdef SBA_COLLECT_STATS
471 	{
472 		unsigned long cr_end = mfctl(16);
473 		unsigned long tmp = cr_end - cr_start;
474 		/* check for roll over */
475 		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
476 	}
477 	ioc->avg_search[ioc->avg_idx++] = cr_start;
478 	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
479 
480 	ioc->used_pages += pages_needed;
481 #endif
482 
483 	return (pide);
484 }
485 
486 
487 /**
488  * sba_free_range - unmark bits in IO PDIR resource bitmap
489  * @ioc: IO MMU structure which owns the pdir we are interested in.
490  * @iova: IO virtual address which was previously allocated.
491  * @size: number of bytes to create a mapping for
492  *
493  * clear bits in the ioc's resource map
494  */
495 static void
496 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
497 {
498 	unsigned long iovp = SBA_IOVP(ioc, iova);
499 	unsigned int pide = PDIR_INDEX(iovp);
500 	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
501 	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
502 
503 	int bits_not_wanted = size >> IOVP_SHIFT;
504 
505 	/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
506 	unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
507 
508 	DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
509 		__func__, (uint) iova, size,
510 		bits_not_wanted, m, pide, res_ptr, *res_ptr);
511 
512 #ifdef SBA_COLLECT_STATS
513 	ioc->used_pages -= bits_not_wanted;
514 #endif
515 
516 	*res_ptr &= ~m;
517 }
518 
519 
520 /**************************************************************
521 *
522 *   "Dynamic DMA Mapping" support (aka "Coherent I/O")
523 *
524 ***************************************************************/
525 
526 #ifdef SBA_HINT_SUPPORT
527 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
528 #endif
529 
530 typedef unsigned long space_t;
531 #define KERNEL_SPACE 0
532 
533 /**
534  * sba_io_pdir_entry - fill in one IO PDIR entry
535  * @pdir_ptr:  pointer to IO PDIR entry
536  * @sid: process Space ID - currently only support KERNEL_SPACE
537  * @vba: Virtual CPU address of buffer to map
538  * @hint: DMA hint set to use for this mapping
539  *
540  * SBA Mapping Routine
541  *
542  * Given a virtual address (vba, arg2) and space id, (sid, arg1)
543  * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
544  * pdir_ptr (arg0).
545  * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
546  * for Astro/Ike looks like:
547  *
548  *
549  *  0                    19                                 51   55       63
550  * +-+---------------------+----------------------------------+----+--------+
551  * |V|        U            |            PPN[43:12]            | U  |   VI   |
552  * +-+---------------------+----------------------------------+----+--------+
553  *
554  * Pluto is basically identical, supports fewer physical address bits:
555  *
556  *  0                       23                              51   55       63
557  * +-+------------------------+-------------------------------+----+--------+
558  * |V|        U               |         PPN[39:12]            | U  |   VI   |
559  * +-+------------------------+-------------------------------+----+--------+
560  *
561  *  V  == Valid Bit  (Most Significant Bit is bit 0)
562  *  U  == Unused
563  * PPN == Physical Page Number
564  * VI  == Virtual Index (aka Coherent Index)
565  *
566  * LPA instruction output is put into PPN field.
567  * LCI (Load Coherence Index) instruction provides the "VI" bits.
568  *
569  * We pre-swap the bytes since PCX-W is Big Endian and the
570  * IOMMU uses little endian for the pdir.
571  */
572 
573 static void
574 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
575 		  unsigned long hint)
576 {
577 	u64 pa; /* physical address */
578 	register unsigned ci; /* coherent index */
579 
580 	pa = lpa(vba);
581 	pa &= IOVP_MASK;
582 
583 	asm("lci 0(%1), %0" : "=r" (ci) : "r" (vba));
584 	pa |= (ci >> PAGE_SHIFT) & 0xff;  /* move CI (8 bits) into lowest byte */
585 
586 	pa |= SBA_PDIR_VALID_BIT;	/* set "valid" bit */
587 	*pdir_ptr = cpu_to_le64(pa);	/* swap and store into I/O Pdir */
588 
589 	/*
590 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
591 	 * (bit #61, big endian), we have to flush and sync every time
592 	 * IO-PDIR is changed in Ike/Astro.
593 	 */
594 	asm_io_fdc(pdir_ptr);
595 }
596 
597 
598 /**
599  * sba_mark_invalid - invalidate one or more IO PDIR entries
600  * @ioc: IO MMU structure which owns the pdir we are interested in.
601  * @iova:  IO Virtual Address mapped earlier
602  * @byte_cnt:  number of bytes this mapping covers.
603  *
604  * Marking the IO PDIR entry(ies) as Invalid and invalidate
605  * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
606  * is to purge stale entries in the IO TLB when unmapping entries.
607  *
608  * The PCOM register supports purging of multiple pages, with a minium
609  * of 1 page and a maximum of 2GB. Hardware requires the address be
610  * aligned to the size of the range being purged. The size of the range
611  * must be a power of 2. The "Cool perf optimization" in the
612  * allocation routine helps keep that true.
613  */
614 static void
615 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
616 {
617 	u32 iovp = (u32) SBA_IOVP(ioc,iova);
618 	u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
619 
620 #ifdef ASSERT_PDIR_SANITY
621 	/* Assert first pdir entry is set.
622 	**
623 	** Even though this is a big-endian machine, the entries
624 	** in the iopdir are little endian. That's why we look at
625 	** the byte at +7 instead of at +0.
626 	*/
627 	if (0x80 != (((u8 *) pdir_ptr)[7])) {
628 		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
629 	}
630 #endif
631 
632 	if (byte_cnt > IOVP_SIZE)
633 	{
634 #if 0
635 		unsigned long entries_per_cacheline = ioc_needs_fdc ?
636 				L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
637 					- (unsigned long) pdir_ptr;
638 				: 262144;
639 #endif
640 
641 		/* set "size" field for PCOM */
642 		iovp |= get_order(byte_cnt) + PAGE_SHIFT;
643 
644 		do {
645 			/* clear I/O Pdir entry "valid" bit first */
646 			((u8 *) pdir_ptr)[7] = 0;
647 			asm_io_fdc(pdir_ptr);
648 			if (ioc_needs_fdc) {
649 #if 0
650 				entries_per_cacheline = L1_CACHE_SHIFT - 3;
651 #endif
652 			}
653 			pdir_ptr++;
654 			byte_cnt -= IOVP_SIZE;
655 		} while (byte_cnt > IOVP_SIZE);
656 	} else
657 		iovp |= IOVP_SHIFT;     /* set "size" field for PCOM */
658 
659 	/*
660 	** clear I/O PDIR entry "valid" bit.
661 	** We have to R/M/W the cacheline regardless how much of the
662 	** pdir entry that we clobber.
663 	** The rest of the entry would be useful for debugging if we
664 	** could dump core on HPMC.
665 	*/
666 	((u8 *) pdir_ptr)[7] = 0;
667 	asm_io_fdc(pdir_ptr);
668 
669 	WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
670 }
671 
672 /**
673  * sba_dma_supported - PCI driver can query DMA support
674  * @dev: instance of PCI owned by the driver that's asking
675  * @mask:  number of address bits this PCI device can handle
676  *
677  * See Documentation/core-api/dma-api-howto.rst
678  */
679 static int sba_dma_supported( struct device *dev, u64 mask)
680 {
681 	struct ioc *ioc;
682 
683 	if (dev == NULL) {
684 		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
685 		BUG();
686 		return(0);
687 	}
688 
689 	ioc = GET_IOC(dev);
690 	if (!ioc)
691 		return 0;
692 
693 	/*
694 	 * check if mask is >= than the current max IO Virt Address
695 	 * The max IO Virt address will *always* < 30 bits.
696 	 */
697 	return((int)(mask >= (ioc->ibase - 1 +
698 			(ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
699 }
700 
701 
702 /**
703  * sba_map_single - map one buffer and return IOVA for DMA
704  * @dev: instance of PCI owned by the driver that's asking.
705  * @addr:  driver buffer to map.
706  * @size:  number of bytes to map in driver buffer.
707  * @direction:  R/W or both.
708  *
709  * See Documentation/core-api/dma-api-howto.rst
710  */
711 static dma_addr_t
712 sba_map_single(struct device *dev, void *addr, size_t size,
713 	       enum dma_data_direction direction)
714 {
715 	struct ioc *ioc;
716 	unsigned long flags;
717 	dma_addr_t iovp;
718 	dma_addr_t offset;
719 	u64 *pdir_start;
720 	int pide;
721 
722 	ioc = GET_IOC(dev);
723 	if (!ioc)
724 		return DMA_MAPPING_ERROR;
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 	asm_io_sync();
771 
772 #ifdef ASSERT_PDIR_SANITY
773 	sba_check_pdir(ioc,"Check after sba_map_single()");
774 #endif
775 	spin_unlock_irqrestore(&ioc->res_lock, flags);
776 
777 	/* form complete address */
778 	return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
779 }
780 
781 
782 static dma_addr_t
783 sba_map_page(struct device *dev, struct page *page, unsigned long offset,
784 		size_t size, enum dma_data_direction direction,
785 		unsigned long attrs)
786 {
787 	return sba_map_single(dev, page_address(page) + offset, size,
788 			direction);
789 }
790 
791 
792 /**
793  * sba_unmap_page - unmap one IOVA and free resources
794  * @dev: instance of PCI owned by the driver that's asking.
795  * @iova:  IOVA of driver buffer previously mapped.
796  * @size:  number of bytes mapped in driver buffer.
797  * @direction:  R/W or both.
798  * @attrs: attributes
799  *
800  * See Documentation/core-api/dma-api-howto.rst
801  */
802 static void
803 sba_unmap_page(struct device *dev, dma_addr_t iova, size_t size,
804 		enum dma_data_direction direction, unsigned long 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 	if (!ioc) {
817 		WARN_ON(!ioc);
818 		return;
819 	}
820 	offset = iova & ~IOVP_MASK;
821 	iova ^= offset;        /* clear offset bits */
822 	size += offset;
823 	size = ALIGN(size, IOVP_SIZE);
824 
825 	spin_lock_irqsave(&ioc->res_lock, flags);
826 
827 #ifdef SBA_COLLECT_STATS
828 	ioc->usingle_calls++;
829 	ioc->usingle_pages += size >> IOVP_SHIFT;
830 #endif
831 
832 	sba_mark_invalid(ioc, iova, size);
833 
834 #if DELAYED_RESOURCE_CNT > 0
835 	/* Delaying when we re-use a IO Pdir entry reduces the number
836 	 * of MMIO reads needed to flush writes to the PCOM register.
837 	 */
838 	d = &(ioc->saved[ioc->saved_cnt]);
839 	d->iova = iova;
840 	d->size = size;
841 	if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
842 		int cnt = ioc->saved_cnt;
843 		while (cnt--) {
844 			sba_free_range(ioc, d->iova, d->size);
845 			d--;
846 		}
847 		ioc->saved_cnt = 0;
848 
849 		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
850 	}
851 #else /* DELAYED_RESOURCE_CNT == 0 */
852 	sba_free_range(ioc, iova, size);
853 
854 	/* If fdc's were issued, force fdc's to be visible now */
855 	asm_io_sync();
856 
857 	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
858 #endif /* DELAYED_RESOURCE_CNT == 0 */
859 
860 	spin_unlock_irqrestore(&ioc->res_lock, flags);
861 
862 	/* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
863 	** For Astro based systems this isn't a big deal WRT performance.
864 	** As long as 2.4 kernels copyin/copyout data from/to userspace,
865 	** we don't need the syncdma. The issue here is I/O MMU cachelines
866 	** are *not* coherent in all cases.  May be hwrev dependent.
867 	** Need to investigate more.
868 	asm volatile("syncdma");
869 	*/
870 }
871 
872 
873 /**
874  * sba_alloc - allocate/map shared mem for DMA
875  * @hwdev: instance of PCI owned by the driver that's asking.
876  * @size:  number of bytes mapped in driver buffer.
877  * @dma_handle:  IOVA of new buffer.
878  * @gfp: allocation flags
879  * @attrs: attributes
880  *
881  * See Documentation/core-api/dma-api-howto.rst
882  */
883 static void *sba_alloc(struct device *hwdev, size_t size, dma_addr_t *dma_handle,
884 		gfp_t gfp, unsigned long attrs)
885 {
886 	void *ret;
887 
888 	if (!hwdev) {
889 		/* only support PCI */
890 		*dma_handle = 0;
891 		return NULL;
892 	}
893 
894         ret = (void *) __get_free_pages(gfp, get_order(size));
895 
896 	if (ret) {
897 		memset(ret, 0, size);
898 		*dma_handle = sba_map_single(hwdev, ret, size, 0);
899 	}
900 
901 	return ret;
902 }
903 
904 
905 /**
906  * sba_free - free/unmap shared mem for DMA
907  * @hwdev: instance of PCI owned by the driver that's asking.
908  * @size:  number of bytes mapped in driver buffer.
909  * @vaddr:  virtual address IOVA of "consistent" buffer.
910  * @dma_handle:  IO virtual address of "consistent" buffer.
911  * @attrs: attributes
912  *
913  * See Documentation/core-api/dma-api-howto.rst
914  */
915 static void
916 sba_free(struct device *hwdev, size_t size, void *vaddr,
917 		    dma_addr_t dma_handle, unsigned long attrs)
918 {
919 	sba_unmap_page(hwdev, dma_handle, size, 0, 0);
920 	free_pages((unsigned long) vaddr, get_order(size));
921 }
922 
923 
924 /*
925 ** Since 0 is a valid pdir_base index value, can't use that
926 ** to determine if a value is valid or not. Use a flag to indicate
927 ** the SG list entry contains a valid pdir index.
928 */
929 #define PIDE_FLAG 0x80000000UL
930 
931 #ifdef SBA_COLLECT_STATS
932 #define IOMMU_MAP_STATS
933 #endif
934 #include "iommu-helpers.h"
935 
936 #ifdef DEBUG_LARGE_SG_ENTRIES
937 int dump_run_sg = 0;
938 #endif
939 
940 
941 /**
942  * sba_map_sg - map Scatter/Gather list
943  * @dev: instance of PCI owned by the driver that's asking.
944  * @sglist:  array of buffer/length pairs
945  * @nents:  number of entries in list
946  * @direction:  R/W or both.
947  * @attrs: attributes
948  *
949  * See Documentation/core-api/dma-api-howto.rst
950  */
951 static int
952 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
953 	   enum dma_data_direction direction, unsigned long attrs)
954 {
955 	struct ioc *ioc;
956 	int filled = 0;
957 	unsigned long flags;
958 
959 	DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
960 
961 	ioc = GET_IOC(dev);
962 	if (!ioc)
963 		return -EINVAL;
964 
965 	/* Fast path single entry scatterlists. */
966 	if (nents == 1) {
967 		sg_dma_address(sglist) = sba_map_single(dev, sg_virt(sglist),
968 						sglist->length, direction);
969 		sg_dma_len(sglist)     = sglist->length;
970 		return 1;
971 	}
972 
973 	spin_lock_irqsave(&ioc->res_lock, flags);
974 
975 #ifdef ASSERT_PDIR_SANITY
976 	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
977 	{
978 		sba_dump_sg(ioc, sglist, nents);
979 		panic("Check before sba_map_sg()");
980 	}
981 #endif
982 
983 #ifdef SBA_COLLECT_STATS
984 	ioc->msg_calls++;
985 #endif
986 
987 	/*
988 	** First coalesce the chunks and allocate I/O pdir space
989 	**
990 	** If this is one DMA stream, we can properly map using the
991 	** correct virtual address associated with each DMA page.
992 	** w/o this association, we wouldn't have coherent DMA!
993 	** Access to the virtual address is what forces a two pass algorithm.
994 	*/
995 	iommu_coalesce_chunks(ioc, dev, sglist, nents, sba_alloc_range);
996 
997 	/*
998 	** Program the I/O Pdir
999 	**
1000 	** map the virtual addresses to the I/O Pdir
1001 	** o dma_address will contain the pdir index
1002 	** o dma_len will contain the number of bytes to map
1003 	** o address contains the virtual address.
1004 	*/
1005 	filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
1006 
1007 	/* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1008 	asm_io_sync();
1009 
1010 #ifdef ASSERT_PDIR_SANITY
1011 	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1012 	{
1013 		sba_dump_sg(ioc, sglist, nents);
1014 		panic("Check after sba_map_sg()\n");
1015 	}
1016 #endif
1017 
1018 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1019 
1020 	DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
1021 
1022 	return filled;
1023 }
1024 
1025 
1026 /**
1027  * sba_unmap_sg - unmap Scatter/Gather list
1028  * @dev: instance of PCI owned by the driver that's asking.
1029  * @sglist:  array of buffer/length pairs
1030  * @nents:  number of entries in list
1031  * @direction:  R/W or both.
1032  * @attrs: attributes
1033  *
1034  * See Documentation/core-api/dma-api-howto.rst
1035  */
1036 static void
1037 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1038 	     enum dma_data_direction direction, unsigned long attrs)
1039 {
1040 	struct ioc *ioc;
1041 #ifdef ASSERT_PDIR_SANITY
1042 	unsigned long flags;
1043 #endif
1044 
1045 	DBG_RUN_SG("%s() START %d entries,  %p,%x\n",
1046 		__func__, nents, sg_virt(sglist), sglist->length);
1047 
1048 	ioc = GET_IOC(dev);
1049 	if (!ioc) {
1050 		WARN_ON(!ioc);
1051 		return;
1052 	}
1053 
1054 #ifdef SBA_COLLECT_STATS
1055 	ioc->usg_calls++;
1056 #endif
1057 
1058 #ifdef ASSERT_PDIR_SANITY
1059 	spin_lock_irqsave(&ioc->res_lock, flags);
1060 	sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1061 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1062 #endif
1063 
1064 	while (nents && sg_dma_len(sglist)) {
1065 
1066 		sba_unmap_page(dev, sg_dma_address(sglist), sg_dma_len(sglist),
1067 				direction, 0);
1068 #ifdef SBA_COLLECT_STATS
1069 		ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1070 		ioc->usingle_calls--;	/* kluge since call is unmap_sg() */
1071 #endif
1072 		++sglist;
1073 		nents--;
1074 	}
1075 
1076 	DBG_RUN_SG("%s() DONE (nents %d)\n", __func__,  nents);
1077 
1078 #ifdef ASSERT_PDIR_SANITY
1079 	spin_lock_irqsave(&ioc->res_lock, flags);
1080 	sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1081 	spin_unlock_irqrestore(&ioc->res_lock, flags);
1082 #endif
1083 
1084 }
1085 
1086 static const struct dma_map_ops sba_ops = {
1087 	.dma_supported =	sba_dma_supported,
1088 	.alloc =		sba_alloc,
1089 	.free =			sba_free,
1090 	.map_page =		sba_map_page,
1091 	.unmap_page =		sba_unmap_page,
1092 	.map_sg =		sba_map_sg,
1093 	.unmap_sg =		sba_unmap_sg,
1094 	.get_sgtable =		dma_common_get_sgtable,
1095 	.alloc_pages =		dma_common_alloc_pages,
1096 	.free_pages =		dma_common_free_pages,
1097 };
1098 
1099 
1100 /**************************************************************************
1101 **
1102 **   SBA PAT PDC support
1103 **
1104 **   o call pdc_pat_cell_module()
1105 **   o store ranges in PCI "resource" structures
1106 **
1107 **************************************************************************/
1108 
1109 static void
1110 sba_get_pat_resources(struct sba_device *sba_dev)
1111 {
1112 #if 0
1113 /*
1114 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1115 **      PAT PDC to program the SBA/LBA directed range registers...this
1116 **      burden may fall on the LBA code since it directly supports the
1117 **      PCI subsystem. It's not clear yet. - ggg
1118 */
1119 PAT_MOD(mod)->mod_info.mod_pages   = PAT_GET_MOD_PAGES(temp);
1120 	FIXME : ???
1121 PAT_MOD(mod)->mod_info.dvi         = PAT_GET_DVI(temp);
1122 	Tells where the dvi bits are located in the address.
1123 PAT_MOD(mod)->mod_info.ioc         = PAT_GET_IOC(temp);
1124 	FIXME : ???
1125 #endif
1126 }
1127 
1128 
1129 /**************************************************************
1130 *
1131 *   Initialization and claim
1132 *
1133 ***************************************************************/
1134 #define PIRANHA_ADDR_MASK	0x00160000UL /* bit 17,18,20 */
1135 #define PIRANHA_ADDR_VAL	0x00060000UL /* bit 17,18 on */
1136 static void *
1137 sba_alloc_pdir(unsigned int pdir_size)
1138 {
1139         unsigned long pdir_base;
1140 	unsigned long pdir_order = get_order(pdir_size);
1141 
1142 	pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1143 	if (NULL == (void *) pdir_base)	{
1144 		panic("%s() could not allocate I/O Page Table\n",
1145 			__func__);
1146 	}
1147 
1148 	/* If this is not PA8700 (PCX-W2)
1149 	**	OR newer than ver 2.2
1150 	**	OR in a system that doesn't need VINDEX bits from SBA,
1151 	**
1152 	** then we aren't exposed to the HW bug.
1153 	*/
1154 	if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1155 			|| (boot_cpu_data.pdc.versions > 0x202)
1156 			|| (boot_cpu_data.pdc.capabilities & 0x08L) )
1157 		return (void *) pdir_base;
1158 
1159 	/*
1160 	 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1161 	 *
1162 	 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1163 	 * Ike/Astro can cause silent data corruption. This is only
1164 	 * a problem if the I/O PDIR is located in memory such that
1165 	 * (little-endian)  bits 17 and 18 are on and bit 20 is off.
1166 	 *
1167 	 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1168 	 * right physical address, we can either avoid (IOPDIR <= 1MB)
1169 	 * or minimize (2MB IO Pdir) the problem if we restrict the
1170 	 * IO Pdir to a maximum size of 2MB-128K (1902K).
1171 	 *
1172 	 * Because we always allocate 2^N sized IO pdirs, either of the
1173 	 * "bad" regions will be the last 128K if at all. That's easy
1174 	 * to test for.
1175 	 *
1176 	 */
1177 	if (pdir_order <= (19-12)) {
1178 		if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1179 			/* allocate a new one on 512k alignment */
1180 			unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1181 			/* release original */
1182 			free_pages(pdir_base, pdir_order);
1183 
1184 			pdir_base = new_pdir;
1185 
1186 			/* release excess */
1187 			while (pdir_order < (19-12)) {
1188 				new_pdir += pdir_size;
1189 				free_pages(new_pdir, pdir_order);
1190 				pdir_order +=1;
1191 				pdir_size <<=1;
1192 			}
1193 		}
1194 	} else {
1195 		/*
1196 		** 1MB or 2MB Pdir
1197 		** Needs to be aligned on an "odd" 1MB boundary.
1198 		*/
1199 		unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1200 
1201 		/* release original */
1202 		free_pages( pdir_base, pdir_order);
1203 
1204 		/* release first 1MB */
1205 		free_pages(new_pdir, 20-12);
1206 
1207 		pdir_base = new_pdir + 1024*1024;
1208 
1209 		if (pdir_order > (20-12)) {
1210 			/*
1211 			** 2MB Pdir.
1212 			**
1213 			** Flag tells init_bitmap() to mark bad 128k as used
1214 			** and to reduce the size by 128k.
1215 			*/
1216 			piranha_bad_128k = 1;
1217 
1218 			new_pdir += 3*1024*1024;
1219 			/* release last 1MB */
1220 			free_pages(new_pdir, 20-12);
1221 
1222 			/* release unusable 128KB */
1223 			free_pages(new_pdir - 128*1024 , 17-12);
1224 
1225 			pdir_size -= 128*1024;
1226 		}
1227 	}
1228 
1229 	memset((void *) pdir_base, 0, pdir_size);
1230 	return (void *) pdir_base;
1231 }
1232 
1233 struct ibase_data_struct {
1234 	struct ioc *ioc;
1235 	int ioc_num;
1236 };
1237 
1238 static int setup_ibase_imask_callback(struct device *dev, void *data)
1239 {
1240 	struct parisc_device *lba = to_parisc_device(dev);
1241 	struct ibase_data_struct *ibd = data;
1242 	int rope_num = (lba->hpa.start >> 13) & 0xf;
1243 	if (rope_num >> 3 == ibd->ioc_num)
1244 		lba_set_iregs(lba, ibd->ioc->ibase, ibd->ioc->imask);
1245 	return 0;
1246 }
1247 
1248 /* setup Mercury or Elroy IBASE/IMASK registers. */
1249 static void
1250 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1251 {
1252 	struct ibase_data_struct ibase_data = {
1253 		.ioc		= ioc,
1254 		.ioc_num	= ioc_num,
1255 	};
1256 
1257 	device_for_each_child(&sba->dev, &ibase_data,
1258 			      setup_ibase_imask_callback);
1259 }
1260 
1261 #ifdef SBA_AGP_SUPPORT
1262 static int
1263 sba_ioc_find_quicksilver(struct device *dev, void *data)
1264 {
1265 	int *agp_found = data;
1266 	struct parisc_device *lba = to_parisc_device(dev);
1267 
1268 	if (IS_QUICKSILVER(lba))
1269 		*agp_found = 1;
1270 	return 0;
1271 }
1272 #endif
1273 
1274 static void
1275 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1276 {
1277 	u32 iova_space_mask;
1278 	u32 iova_space_size;
1279 	int iov_order, tcnfg;
1280 #ifdef SBA_AGP_SUPPORT
1281 	int agp_found = 0;
1282 #endif
1283 	/*
1284 	** Firmware programs the base and size of a "safe IOVA space"
1285 	** (one that doesn't overlap memory or LMMIO space) in the
1286 	** IBASE and IMASK registers.
1287 	*/
1288 	ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1fffffULL;
1289 	iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1290 
1291 	if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1292 		printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1293 		iova_space_size /= 2;
1294 	}
1295 
1296 	/*
1297 	** iov_order is always based on a 1GB IOVA space since we want to
1298 	** turn on the other half for AGP GART.
1299 	*/
1300 	iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1301 	ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1302 
1303 	DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1304 		__func__, ioc->ioc_hpa, iova_space_size >> 20,
1305 		iov_order + PAGE_SHIFT);
1306 
1307 	ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1308 						   get_order(ioc->pdir_size));
1309 	if (!ioc->pdir_base)
1310 		panic("Couldn't allocate I/O Page Table\n");
1311 
1312 	memset(ioc->pdir_base, 0, ioc->pdir_size);
1313 
1314 	DBG_INIT("%s() pdir %p size %x\n",
1315 			__func__, ioc->pdir_base, ioc->pdir_size);
1316 
1317 #ifdef SBA_HINT_SUPPORT
1318 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1319 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1320 
1321 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1322 		ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1323 #endif
1324 
1325 	WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1326 	WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1327 
1328 	/* build IMASK for IOC and Elroy */
1329 	iova_space_mask =  0xffffffff;
1330 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1331 	ioc->imask = iova_space_mask;
1332 #ifdef ZX1_SUPPORT
1333 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1334 #endif
1335 	sba_dump_tlb(ioc->ioc_hpa);
1336 
1337 	setup_ibase_imask(sba, ioc, ioc_num);
1338 
1339 	WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1340 
1341 #ifdef CONFIG_64BIT
1342 	/*
1343 	** Setting the upper bits makes checking for bypass addresses
1344 	** a little faster later on.
1345 	*/
1346 	ioc->imask |= 0xFFFFFFFF00000000UL;
1347 #endif
1348 
1349 	/* Set I/O PDIR Page size to system page size */
1350 	switch (PAGE_SHIFT) {
1351 		case 12: tcnfg = 0; break;	/*  4K */
1352 		case 13: tcnfg = 1; break;	/*  8K */
1353 		case 14: tcnfg = 2; break;	/* 16K */
1354 		case 16: tcnfg = 3; break;	/* 64K */
1355 		default:
1356 			panic(__FILE__ "Unsupported system page size %d",
1357 				1 << PAGE_SHIFT);
1358 			break;
1359 	}
1360 	WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1361 
1362 	/*
1363 	** Program the IOC's ibase and enable IOVA translation
1364 	** Bit zero == enable bit.
1365 	*/
1366 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1367 
1368 	/*
1369 	** Clear I/O TLB of any possible entries.
1370 	** (Yes. This is a bit paranoid...but so what)
1371 	*/
1372 	WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1373 
1374 #ifdef SBA_AGP_SUPPORT
1375 
1376 	/*
1377 	** If an AGP device is present, only use half of the IOV space
1378 	** for PCI DMA.  Unfortunately we can't know ahead of time
1379 	** whether GART support will actually be used, for now we
1380 	** can just key on any AGP device found in the system.
1381 	** We program the next pdir index after we stop w/ a key for
1382 	** the GART code to handshake on.
1383 	*/
1384 	device_for_each_child(&sba->dev, &agp_found, sba_ioc_find_quicksilver);
1385 
1386 	if (agp_found && sba_reserve_agpgart) {
1387 		printk(KERN_INFO "%s: reserving %dMb of IOVA space for agpgart\n",
1388 		       __func__, (iova_space_size/2) >> 20);
1389 		ioc->pdir_size /= 2;
1390 		ioc->pdir_base[PDIR_INDEX(iova_space_size/2)] = SBA_AGPGART_COOKIE;
1391 	}
1392 #endif /*SBA_AGP_SUPPORT*/
1393 }
1394 
1395 static void
1396 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1397 {
1398 	u32 iova_space_size, iova_space_mask;
1399 	unsigned int pdir_size, iov_order, tcnfg;
1400 
1401 	/*
1402 	** Determine IOVA Space size from memory size.
1403 	**
1404 	** Ideally, PCI drivers would register the maximum number
1405 	** of DMA they can have outstanding for each device they
1406 	** own.  Next best thing would be to guess how much DMA
1407 	** can be outstanding based on PCI Class/sub-class. Both
1408 	** methods still require some "extra" to support PCI
1409 	** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1410 	**
1411 	** While we have 32-bits "IOVA" space, top two 2 bits are used
1412 	** for DMA hints - ergo only 30 bits max.
1413 	*/
1414 
1415 	iova_space_size = (u32) (totalram_pages()/global_ioc_cnt);
1416 
1417 	/* limit IOVA space size to 1MB-1GB */
1418 	if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1419 		iova_space_size = 1 << (20 - PAGE_SHIFT);
1420 	}
1421 	else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1422 		iova_space_size = 1 << (30 - PAGE_SHIFT);
1423 	}
1424 
1425 	/*
1426 	** iova space must be log2() in size.
1427 	** thus, pdir/res_map will also be log2().
1428 	** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1429 	*/
1430 	iov_order = get_order(iova_space_size << PAGE_SHIFT);
1431 
1432 	/* iova_space_size is now bytes, not pages */
1433 	iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1434 
1435 	ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1436 
1437 	DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1438 			__func__,
1439 			ioc->ioc_hpa,
1440 			(unsigned long) totalram_pages() >> (20 - PAGE_SHIFT),
1441 			iova_space_size>>20,
1442 			iov_order + PAGE_SHIFT);
1443 
1444 	ioc->pdir_base = sba_alloc_pdir(pdir_size);
1445 
1446 	DBG_INIT("%s() pdir %p size %x\n",
1447 			__func__, ioc->pdir_base, pdir_size);
1448 
1449 #ifdef SBA_HINT_SUPPORT
1450 	/* FIXME : DMA HINTs not used */
1451 	ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1452 	ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1453 
1454 	DBG_INIT("	hint_shift_pdir %x hint_mask_pdir %lx\n",
1455 			ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1456 #endif
1457 
1458 	WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1459 
1460 	/* build IMASK for IOC and Elroy */
1461 	iova_space_mask =  0xffffffff;
1462 	iova_space_mask <<= (iov_order + PAGE_SHIFT);
1463 
1464 	/*
1465 	** On C3000 w/512MB mem, HP-UX 10.20 reports:
1466 	**     ibase=0, imask=0xFE000000, size=0x2000000.
1467 	*/
1468 	ioc->ibase = 0;
1469 	ioc->imask = iova_space_mask;	/* save it */
1470 #ifdef ZX1_SUPPORT
1471 	ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1472 #endif
1473 
1474 	DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1475 		__func__, ioc->ibase, ioc->imask);
1476 
1477 	/*
1478 	** FIXME: Hint registers are programmed with default hint
1479 	** values during boot, so hints should be sane even if we
1480 	** can't reprogram them the way drivers want.
1481 	*/
1482 
1483 	setup_ibase_imask(sba, ioc, ioc_num);
1484 
1485 	/*
1486 	** Program the IOC's ibase and enable IOVA translation
1487 	*/
1488 	WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1489 	WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1490 
1491 	/* Set I/O PDIR Page size to system page size */
1492 	switch (PAGE_SHIFT) {
1493 		case 12: tcnfg = 0; break;	/*  4K */
1494 		case 13: tcnfg = 1; break;	/*  8K */
1495 		case 14: tcnfg = 2; break;	/* 16K */
1496 		case 16: tcnfg = 3; break;	/* 64K */
1497 		default:
1498 			panic(__FILE__ "Unsupported system page size %d",
1499 				1 << PAGE_SHIFT);
1500 			break;
1501 	}
1502 	/* Set I/O PDIR Page size to PAGE_SIZE (4k/16k/...) */
1503 	WRITE_REG(tcnfg, ioc->ioc_hpa+IOC_TCNFG);
1504 
1505 	/*
1506 	** Clear I/O TLB of any possible entries.
1507 	** (Yes. This is a bit paranoid...but so what)
1508 	*/
1509 	WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1510 
1511 	ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1512 
1513 	DBG_INIT("%s() DONE\n", __func__);
1514 }
1515 
1516 
1517 
1518 /**************************************************************************
1519 **
1520 **   SBA initialization code (HW and SW)
1521 **
1522 **   o identify SBA chip itself
1523 **   o initialize SBA chip modes (HardFail)
1524 **   o initialize SBA chip modes (HardFail)
1525 **   o FIXME: initialize DMA hints for reasonable defaults
1526 **
1527 **************************************************************************/
1528 
1529 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1530 {
1531 	return ioremap(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1532 }
1533 
1534 static void sba_hw_init(struct sba_device *sba_dev)
1535 {
1536 	int i;
1537 	int num_ioc;
1538 	u64 ioc_ctl;
1539 
1540 	if (!is_pdc_pat()) {
1541 		/* Shutdown the USB controller on Astro-based workstations.
1542 		** Once we reprogram the IOMMU, the next DMA performed by
1543 		** USB will HPMC the box. USB is only enabled if a
1544 		** keyboard is present and found.
1545 		**
1546 		** With serial console, j6k v5.0 firmware says:
1547 		**   mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1548 		**
1549 		** FIXME: Using GFX+USB console at power up but direct
1550 		**	linux to serial console is still broken.
1551 		**	USB could generate DMA so we must reset USB.
1552 		**	The proper sequence would be:
1553 		**	o block console output
1554 		**	o reset USB device
1555 		**	o reprogram serial port
1556 		**	o unblock console output
1557 		*/
1558 		if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1559 			pdc_io_reset_devices();
1560 		}
1561 
1562 	}
1563 
1564 
1565 #if 0
1566 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1567 	PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1568 
1569 	/*
1570 	** Need to deal with DMA from LAN.
1571 	**	Maybe use page zero boot device as a handle to talk
1572 	**	to PDC about which device to shutdown.
1573 	**
1574 	** Netbooting, j6k v5.0 firmware says:
1575 	** 	mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1576 	** ARGH! invalid class.
1577 	*/
1578 	if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1579 		&& (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1580 			pdc_io_reset();
1581 	}
1582 #endif
1583 
1584 	if (!IS_PLUTO(sba_dev->dev)) {
1585 		ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1586 		DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1587 			__func__, sba_dev->sba_hpa, ioc_ctl);
1588 		ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1589 		ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1590 			/* j6700 v1.6 firmware sets 0x294f */
1591 			/* A500 firmware sets 0x4d */
1592 
1593 		WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1594 
1595 #ifdef DEBUG_SBA_INIT
1596 		ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1597 		DBG_INIT(" 0x%Lx\n", ioc_ctl);
1598 #endif
1599 	} /* if !PLUTO */
1600 
1601 	if (IS_ASTRO(sba_dev->dev)) {
1602 		int err;
1603 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1604 		num_ioc = 1;
1605 
1606 		sba_dev->chip_resv.name = "Astro Intr Ack";
1607 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1608 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff000000UL - 1) ;
1609 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1610 		BUG_ON(err < 0);
1611 
1612 	} else if (IS_PLUTO(sba_dev->dev)) {
1613 		int err;
1614 
1615 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1616 		num_ioc = 1;
1617 
1618 		sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1619 		sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1620 		sba_dev->chip_resv.end   = PCI_F_EXTEND | (0xff200000UL - 1);
1621 		err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1622 		WARN_ON(err < 0);
1623 
1624 		sba_dev->iommu_resv.name = "IOVA Space";
1625 		sba_dev->iommu_resv.start = 0x40000000UL;
1626 		sba_dev->iommu_resv.end   = 0x50000000UL - 1;
1627 		err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1628 		WARN_ON(err < 0);
1629 	} else {
1630 		/* IKE, REO */
1631 		sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1632 		sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1633 		num_ioc = 2;
1634 
1635 		/* TODO - LOOKUP Ike/Stretch chipset mem map */
1636 	}
1637 	/* XXX: What about Reo Grande? */
1638 
1639 	sba_dev->num_ioc = num_ioc;
1640 	for (i = 0; i < num_ioc; i++) {
1641 		void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1642 		unsigned int j;
1643 
1644 		for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1645 
1646 			/*
1647 			 * Clear ROPE(N)_CONFIG AO bit.
1648 			 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1649 			 * Overrides bit 1 in DMA Hint Sets.
1650 			 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1651 			 */
1652 			if (IS_PLUTO(sba_dev->dev)) {
1653 				void __iomem *rope_cfg;
1654 				unsigned long cfg_val;
1655 
1656 				rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1657 				cfg_val = READ_REG(rope_cfg);
1658 				cfg_val &= ~IOC_ROPE_AO;
1659 				WRITE_REG(cfg_val, rope_cfg);
1660 			}
1661 
1662 			/*
1663 			** Make sure the box crashes on rope errors.
1664 			*/
1665 			WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1666 		}
1667 
1668 		/* flush out the last writes */
1669 		READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1670 
1671 		DBG_INIT("	ioc[%d] ROPE_CFG 0x%Lx  ROPE_DBG 0x%Lx\n",
1672 				i,
1673 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1674 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1675 			);
1676 		DBG_INIT("	STATUS_CONTROL 0x%Lx  FLUSH_CTRL 0x%Lx\n",
1677 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1678 				READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1679 			);
1680 
1681 		if (IS_PLUTO(sba_dev->dev)) {
1682 			sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1683 		} else {
1684 			sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1685 		}
1686 	}
1687 }
1688 
1689 static void
1690 sba_common_init(struct sba_device *sba_dev)
1691 {
1692 	int i;
1693 
1694 	/* add this one to the head of the list (order doesn't matter)
1695 	** This will be useful for debugging - especially if we get coredumps
1696 	*/
1697 	sba_dev->next = sba_list;
1698 	sba_list = sba_dev;
1699 
1700 	for(i=0; i< sba_dev->num_ioc; i++) {
1701 		int res_size;
1702 #ifdef DEBUG_DMB_TRAP
1703 		extern void iterate_pages(unsigned long , unsigned long ,
1704 					  void (*)(pte_t * , unsigned long),
1705 					  unsigned long );
1706 		void set_data_memory_break(pte_t * , unsigned long);
1707 #endif
1708 		/* resource map size dictated by pdir_size */
1709 		res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1710 
1711 		/* Second part of PIRANHA BUG */
1712 		if (piranha_bad_128k) {
1713 			res_size -= (128*1024)/sizeof(u64);
1714 		}
1715 
1716 		res_size >>= 3;  /* convert bit count to byte count */
1717 		DBG_INIT("%s() res_size 0x%x\n",
1718 			__func__, res_size);
1719 
1720 		sba_dev->ioc[i].res_size = res_size;
1721 		sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1722 
1723 #ifdef DEBUG_DMB_TRAP
1724 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1725 				set_data_memory_break, 0);
1726 #endif
1727 
1728 		if (NULL == sba_dev->ioc[i].res_map)
1729 		{
1730 			panic("%s:%s() could not allocate resource map\n",
1731 			      __FILE__, __func__ );
1732 		}
1733 
1734 		memset(sba_dev->ioc[i].res_map, 0, res_size);
1735 		/* next available IOVP - circular search */
1736 		sba_dev->ioc[i].res_hint = (unsigned long *)
1737 				&(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1738 
1739 #ifdef ASSERT_PDIR_SANITY
1740 		/* Mark first bit busy - ie no IOVA 0 */
1741 		sba_dev->ioc[i].res_map[0] = 0x80;
1742 		sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1743 #endif
1744 
1745 		/* Third (and last) part of PIRANHA BUG */
1746 		if (piranha_bad_128k) {
1747 			/* region from +1408K to +1536 is un-usable. */
1748 
1749 			int idx_start = (1408*1024/sizeof(u64)) >> 3;
1750 			int idx_end   = (1536*1024/sizeof(u64)) >> 3;
1751 			long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1752 			long *p_end   = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1753 
1754 			/* mark that part of the io pdir busy */
1755 			while (p_start < p_end)
1756 				*p_start++ = -1;
1757 
1758 		}
1759 
1760 #ifdef DEBUG_DMB_TRAP
1761 		iterate_pages( sba_dev->ioc[i].res_map, res_size,
1762 				set_data_memory_break, 0);
1763 		iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1764 				set_data_memory_break, 0);
1765 #endif
1766 
1767 		DBG_INIT("%s() %d res_map %x %p\n",
1768 			__func__, i, res_size, sba_dev->ioc[i].res_map);
1769 	}
1770 
1771 	spin_lock_init(&sba_dev->sba_lock);
1772 	ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1773 
1774 #ifdef DEBUG_SBA_INIT
1775 	/*
1776 	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1777 	 * (bit #61, big endian), we have to flush and sync every time
1778 	 * IO-PDIR is changed in Ike/Astro.
1779 	 */
1780 	if (ioc_needs_fdc) {
1781 		printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1782 	} else {
1783 		printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1784 	}
1785 #endif
1786 }
1787 
1788 #ifdef CONFIG_PROC_FS
1789 static int sba_proc_info(struct seq_file *m, void *p)
1790 {
1791 	struct sba_device *sba_dev = sba_list;
1792 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1793 	int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1794 #ifdef SBA_COLLECT_STATS
1795 	unsigned long avg = 0, min, max;
1796 #endif
1797 	int i;
1798 
1799 	seq_printf(m, "%s rev %d.%d\n",
1800 		   sba_dev->name,
1801 		   (sba_dev->hw_rev & 0x7) + 1,
1802 		   (sba_dev->hw_rev & 0x18) >> 3);
1803 	seq_printf(m, "IO PDIR size    : %d bytes (%d entries)\n",
1804 		   (int)((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1805 		   total_pages);
1806 
1807 	seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1808 		   ioc->res_size, ioc->res_size << 3);   /* 8 bits per byte */
1809 
1810 	seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1811 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1812 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1813 		   READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE));
1814 
1815 	for (i=0; i<4; i++)
1816 		seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n",
1817 			   i,
1818 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE  + i*0x18),
1819 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK  + i*0x18),
1820 			   READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18));
1821 
1822 #ifdef SBA_COLLECT_STATS
1823 	seq_printf(m, "IO PDIR entries : %ld free  %ld used (%d%%)\n",
1824 		   total_pages - ioc->used_pages, ioc->used_pages,
1825 		   (int)(ioc->used_pages * 100 / total_pages));
1826 
1827 	min = max = ioc->avg_search[0];
1828 	for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1829 		avg += ioc->avg_search[i];
1830 		if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1831 		if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1832 	}
1833 	avg /= SBA_SEARCH_SAMPLE;
1834 	seq_printf(m, "  Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1835 		   min, avg, max);
1836 
1837 	seq_printf(m, "pci_map_single(): %12ld calls  %12ld pages (avg %d/1000)\n",
1838 		   ioc->msingle_calls, ioc->msingle_pages,
1839 		   (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1840 
1841 	/* KLUGE - unmap_sg calls unmap_single for each mapped page */
1842 	min = ioc->usingle_calls;
1843 	max = ioc->usingle_pages - ioc->usg_pages;
1844 	seq_printf(m, "pci_unmap_single: %12ld calls  %12ld pages (avg %d/1000)\n",
1845 		   min, max, (int)((max * 1000)/min));
1846 
1847 	seq_printf(m, "pci_map_sg()    : %12ld calls  %12ld pages (avg %d/1000)\n",
1848 		   ioc->msg_calls, ioc->msg_pages,
1849 		   (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1850 
1851 	seq_printf(m, "pci_unmap_sg()  : %12ld calls  %12ld pages (avg %d/1000)\n",
1852 		   ioc->usg_calls, ioc->usg_pages,
1853 		   (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1854 #endif
1855 
1856 	return 0;
1857 }
1858 
1859 static int
1860 sba_proc_bitmap_info(struct seq_file *m, void *p)
1861 {
1862 	struct sba_device *sba_dev = sba_list;
1863 	struct ioc *ioc = &sba_dev->ioc[0];	/* FIXME: Multi-IOC support! */
1864 
1865 	seq_hex_dump(m, "   ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map,
1866 		     ioc->res_size, false);
1867 	seq_putc(m, '\n');
1868 
1869 	return 0;
1870 }
1871 #endif /* CONFIG_PROC_FS */
1872 
1873 static const struct parisc_device_id sba_tbl[] __initconst = {
1874 	{ HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
1875 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
1876 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
1877 	{ HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
1878 	{ HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
1879 	{ 0, }
1880 };
1881 
1882 static int sba_driver_callback(struct parisc_device *);
1883 
1884 static struct parisc_driver sba_driver __refdata = {
1885 	.name =		MODULE_NAME,
1886 	.id_table =	sba_tbl,
1887 	.probe =	sba_driver_callback,
1888 };
1889 
1890 /*
1891 ** Determine if sba should claim this chip (return 0) or not (return 1).
1892 ** If so, initialize the chip and tell other partners in crime they
1893 ** have work to do.
1894 */
1895 static int __init sba_driver_callback(struct parisc_device *dev)
1896 {
1897 	struct sba_device *sba_dev;
1898 	u32 func_class;
1899 	int i;
1900 	char *version;
1901 	void __iomem *sba_addr = ioremap(dev->hpa.start, SBA_FUNC_SIZE);
1902 #ifdef CONFIG_PROC_FS
1903 	struct proc_dir_entry *root;
1904 #endif
1905 
1906 	sba_dump_ranges(sba_addr);
1907 
1908 	/* Read HW Rev First */
1909 	func_class = READ_REG(sba_addr + SBA_FCLASS);
1910 
1911 	if (IS_ASTRO(dev)) {
1912 		unsigned long fclass;
1913 		static char astro_rev[]="Astro ?.?";
1914 
1915 		/* Astro is broken...Read HW Rev First */
1916 		fclass = READ_REG(sba_addr);
1917 
1918 		astro_rev[6] = '1' + (char) (fclass & 0x7);
1919 		astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
1920 		version = astro_rev;
1921 
1922 	} else if (IS_IKE(dev)) {
1923 		static char ike_rev[] = "Ike rev ?";
1924 		ike_rev[8] = '0' + (char) (func_class & 0xff);
1925 		version = ike_rev;
1926 	} else if (IS_PLUTO(dev)) {
1927 		static char pluto_rev[]="Pluto ?.?";
1928 		pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
1929 		pluto_rev[8] = '0' + (char) (func_class & 0x0f);
1930 		version = pluto_rev;
1931 	} else {
1932 		static char reo_rev[] = "REO rev ?";
1933 		reo_rev[8] = '0' + (char) (func_class & 0xff);
1934 		version = reo_rev;
1935 	}
1936 
1937 	if (!global_ioc_cnt) {
1938 		global_ioc_cnt = count_parisc_driver(&sba_driver);
1939 
1940 		/* Astro and Pluto have one IOC per SBA */
1941 		if ((!IS_ASTRO(dev)) || (!IS_PLUTO(dev)))
1942 			global_ioc_cnt *= 2;
1943 	}
1944 
1945 	printk(KERN_INFO "%s found %s at 0x%llx\n",
1946 		MODULE_NAME, version, (unsigned long long)dev->hpa.start);
1947 
1948 	sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
1949 	if (!sba_dev) {
1950 		printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
1951 		return -ENOMEM;
1952 	}
1953 
1954 	parisc_set_drvdata(dev, sba_dev);
1955 
1956 	for(i=0; i<MAX_IOC; i++)
1957 		spin_lock_init(&(sba_dev->ioc[i].res_lock));
1958 
1959 	sba_dev->dev = dev;
1960 	sba_dev->hw_rev = func_class;
1961 	sba_dev->name = dev->name;
1962 	sba_dev->sba_hpa = sba_addr;
1963 
1964 	sba_get_pat_resources(sba_dev);
1965 	sba_hw_init(sba_dev);
1966 	sba_common_init(sba_dev);
1967 
1968 	hppa_dma_ops = &sba_ops;
1969 
1970 #ifdef CONFIG_PROC_FS
1971 	switch (dev->id.hversion) {
1972 	case PLUTO_MCKINLEY_PORT:
1973 		if (!proc_mckinley_root)
1974 			proc_mckinley_root = proc_mkdir("bus/mckinley", NULL);
1975 		root = proc_mckinley_root;
1976 		break;
1977 	case ASTRO_RUNWAY_PORT:
1978 	case IKE_MERCED_PORT:
1979 	default:
1980 		if (!proc_runway_root)
1981 			proc_runway_root = proc_mkdir("bus/runway", NULL);
1982 		root = proc_runway_root;
1983 		break;
1984 	}
1985 
1986 	proc_create_single("sba_iommu", 0, root, sba_proc_info);
1987 	proc_create_single("sba_iommu-bitmap", 0, root, sba_proc_bitmap_info);
1988 #endif
1989 	return 0;
1990 }
1991 
1992 /*
1993 ** One time initialization to let the world know the SBA was found.
1994 ** This is the only routine which is NOT static.
1995 ** Must be called exactly once before pci_init().
1996 */
1997 static int __init sba_init(void)
1998 {
1999 	return register_parisc_driver(&sba_driver);
2000 }
2001 arch_initcall(sba_init);
2002 
2003 
2004 /**
2005  * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2006  * @pci_hba: The parisc device.
2007  *
2008  * Returns the appropriate IOMMU data for the given parisc PCI controller.
2009  * This is cached and used later for PCI DMA Mapping.
2010  */
2011 void * sba_get_iommu(struct parisc_device *pci_hba)
2012 {
2013 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2014 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2015 	char t = sba_dev->id.hw_type;
2016 	int iocnum = (pci_hba->hw_path >> 3);	/* rope # */
2017 
2018 	WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2019 
2020 	return &(sba->ioc[iocnum]);
2021 }
2022 
2023 
2024 /**
2025  * sba_directed_lmmio - return first directed LMMIO range routed to rope
2026  * @pci_hba: The parisc device.
2027  * @r: resource PCI host controller wants start/end fields assigned.
2028  *
2029  * For the given parisc PCI controller, determine if any direct ranges
2030  * are routed down the corresponding rope.
2031  */
2032 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2033 {
2034 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2035 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2036 	char t = sba_dev->id.hw_type;
2037 	int i;
2038 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2039 
2040 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2041 
2042 	r->start = r->end = 0;
2043 
2044 	/* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2045 	for (i=0; i<4; i++) {
2046 		int base, size;
2047 		void __iomem *reg = sba->sba_hpa + i*0x18;
2048 
2049 		base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2050 		if ((base & 1) == 0)
2051 			continue;	/* not enabled */
2052 
2053 		size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2054 
2055 		if ((size & (ROPES_PER_IOC-1)) != rope)
2056 			continue;	/* directed down different rope */
2057 
2058 		r->start = (base & ~1UL) | PCI_F_EXTEND;
2059 		size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2060 		r->end = r->start + size;
2061 		r->flags = IORESOURCE_MEM;
2062 	}
2063 }
2064 
2065 
2066 /**
2067  * sba_distributed_lmmio - return portion of distributed LMMIO range
2068  * @pci_hba: The parisc device.
2069  * @r: resource PCI host controller wants start/end fields assigned.
2070  *
2071  * For the given parisc PCI controller, return portion of distributed LMMIO
2072  * range. The distributed LMMIO is always present and it's just a question
2073  * of the base address and size of the range.
2074  */
2075 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2076 {
2077 	struct parisc_device *sba_dev = parisc_parent(pci_hba);
2078 	struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2079 	char t = sba_dev->id.hw_type;
2080 	int base, size;
2081 	int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1));  /* rope # */
2082 
2083 	BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2084 
2085 	r->start = r->end = 0;
2086 
2087 	base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2088 	if ((base & 1) == 0) {
2089 		BUG();	/* Gah! Distr Range wasn't enabled! */
2090 		return;
2091 	}
2092 
2093 	r->start = (base & ~1UL) | PCI_F_EXTEND;
2094 
2095 	size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2096 	r->start += rope * (size + 1);	/* adjust base for this rope */
2097 	r->end = r->start + size;
2098 	r->flags = IORESOURCE_MEM;
2099 }
2100