xref: /linux/arch/sparc/mm/srmmu.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * srmmu.c:  SRMMU specific routines for memory management.
3  *
4  * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
5  * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
6  * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
7  * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
8  * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
9  */
10 
11 #include <linux/seq_file.h>
12 #include <linux/spinlock.h>
13 #include <linux/bootmem.h>
14 #include <linux/pagemap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/kdebug.h>
17 #include <linux/export.h>
18 #include <linux/kernel.h>
19 #include <linux/init.h>
20 #include <linux/log2.h>
21 #include <linux/gfp.h>
22 #include <linux/fs.h>
23 #include <linux/mm.h>
24 
25 #include <asm/mmu_context.h>
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
28 #include <asm/io-unit.h>
29 #include <asm/pgalloc.h>
30 #include <asm/pgtable.h>
31 #include <asm/bitext.h>
32 #include <asm/vaddrs.h>
33 #include <asm/cache.h>
34 #include <asm/traps.h>
35 #include <asm/oplib.h>
36 #include <asm/mbus.h>
37 #include <asm/page.h>
38 #include <asm/asi.h>
39 #include <asm/msi.h>
40 #include <asm/smp.h>
41 #include <asm/io.h>
42 
43 /* Now the cpu specific definitions. */
44 #include <asm/turbosparc.h>
45 #include <asm/tsunami.h>
46 #include <asm/viking.h>
47 #include <asm/swift.h>
48 #include <asm/leon.h>
49 #include <asm/mxcc.h>
50 #include <asm/ross.h>
51 
52 #include "mm_32.h"
53 
54 enum mbus_module srmmu_modtype;
55 static unsigned int hwbug_bitmask;
56 int vac_cache_size;
57 int vac_line_size;
58 
59 extern struct resource sparc_iomap;
60 
61 extern unsigned long last_valid_pfn;
62 
63 static pgd_t *srmmu_swapper_pg_dir;
64 
65 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
66 EXPORT_SYMBOL(sparc32_cachetlb_ops);
67 
68 #ifdef CONFIG_SMP
69 const struct sparc32_cachetlb_ops *local_ops;
70 
71 #define FLUSH_BEGIN(mm)
72 #define FLUSH_END
73 #else
74 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
75 #define FLUSH_END	}
76 #endif
77 
78 int flush_page_for_dma_global = 1;
79 
80 char *srmmu_name;
81 
82 ctxd_t *srmmu_ctx_table_phys;
83 static ctxd_t *srmmu_context_table;
84 
85 int viking_mxcc_present;
86 static DEFINE_SPINLOCK(srmmu_context_spinlock);
87 
88 static int is_hypersparc;
89 
90 static int srmmu_cache_pagetables;
91 
92 /* these will be initialized in srmmu_nocache_calcsize() */
93 static unsigned long srmmu_nocache_size;
94 static unsigned long srmmu_nocache_end;
95 
96 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
97 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
98 
99 /* The context table is a nocache user with the biggest alignment needs. */
100 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
101 
102 void *srmmu_nocache_pool;
103 static struct bit_map srmmu_nocache_map;
104 
105 static inline int srmmu_pmd_none(pmd_t pmd)
106 { return !(pmd_val(pmd) & 0xFFFFFFF); }
107 
108 /* XXX should we hyper_flush_whole_icache here - Anton */
109 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
110 { set_pte((pte_t *)ctxp, (SRMMU_ET_PTD | (__nocache_pa((unsigned long) pgdp) >> 4))); }
111 
112 void pmd_set(pmd_t *pmdp, pte_t *ptep)
113 {
114 	unsigned long ptp;	/* Physical address, shifted right by 4 */
115 	int i;
116 
117 	ptp = __nocache_pa((unsigned long) ptep) >> 4;
118 	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
119 		set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
120 		ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
121 	}
122 }
123 
124 void pmd_populate(struct mm_struct *mm, pmd_t *pmdp, struct page *ptep)
125 {
126 	unsigned long ptp;	/* Physical address, shifted right by 4 */
127 	int i;
128 
129 	ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);	/* watch for overflow */
130 	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
131 		set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
132 		ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
133 	}
134 }
135 
136 /* Find an entry in the third-level page table.. */
137 pte_t *pte_offset_kernel(pmd_t *dir, unsigned long address)
138 {
139 	void *pte;
140 
141 	pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
142 	return (pte_t *) pte +
143 	    ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
144 }
145 
146 /*
147  * size: bytes to allocate in the nocache area.
148  * align: bytes, number to align at.
149  * Returns the virtual address of the allocated area.
150  */
151 static void *__srmmu_get_nocache(int size, int align)
152 {
153 	int offset;
154 	unsigned long addr;
155 
156 	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
157 		printk(KERN_ERR "Size 0x%x too small for nocache request\n",
158 		       size);
159 		size = SRMMU_NOCACHE_BITMAP_SHIFT;
160 	}
161 	if (size & (SRMMU_NOCACHE_BITMAP_SHIFT - 1)) {
162 		printk(KERN_ERR "Size 0x%x unaligned int nocache request\n",
163 		       size);
164 		size += SRMMU_NOCACHE_BITMAP_SHIFT - 1;
165 	}
166 	BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
167 
168 	offset = bit_map_string_get(&srmmu_nocache_map,
169 				    size >> SRMMU_NOCACHE_BITMAP_SHIFT,
170 				    align >> SRMMU_NOCACHE_BITMAP_SHIFT);
171 	if (offset == -1) {
172 		printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
173 		       size, (int) srmmu_nocache_size,
174 		       srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
175 		return NULL;
176 	}
177 
178 	addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
179 	return (void *)addr;
180 }
181 
182 void *srmmu_get_nocache(int size, int align)
183 {
184 	void *tmp;
185 
186 	tmp = __srmmu_get_nocache(size, align);
187 
188 	if (tmp)
189 		memset(tmp, 0, size);
190 
191 	return tmp;
192 }
193 
194 void srmmu_free_nocache(void *addr, int size)
195 {
196 	unsigned long vaddr;
197 	int offset;
198 
199 	vaddr = (unsigned long)addr;
200 	if (vaddr < SRMMU_NOCACHE_VADDR) {
201 		printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
202 		    vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
203 		BUG();
204 	}
205 	if (vaddr + size > srmmu_nocache_end) {
206 		printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
207 		    vaddr, srmmu_nocache_end);
208 		BUG();
209 	}
210 	if (!is_power_of_2(size)) {
211 		printk("Size 0x%x is not a power of 2\n", size);
212 		BUG();
213 	}
214 	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
215 		printk("Size 0x%x is too small\n", size);
216 		BUG();
217 	}
218 	if (vaddr & (size - 1)) {
219 		printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
220 		BUG();
221 	}
222 
223 	offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
224 	size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
225 
226 	bit_map_clear(&srmmu_nocache_map, offset, size);
227 }
228 
229 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
230 						 unsigned long end);
231 
232 /* Return how much physical memory we have.  */
233 static unsigned long __init probe_memory(void)
234 {
235 	unsigned long total = 0;
236 	int i;
237 
238 	for (i = 0; sp_banks[i].num_bytes; i++)
239 		total += sp_banks[i].num_bytes;
240 
241 	return total;
242 }
243 
244 /*
245  * Reserve nocache dynamically proportionally to the amount of
246  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
247  */
248 static void __init srmmu_nocache_calcsize(void)
249 {
250 	unsigned long sysmemavail = probe_memory() / 1024;
251 	int srmmu_nocache_npages;
252 
253 	srmmu_nocache_npages =
254 		sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
255 
256  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
257 	// if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
258 	if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
259 		srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
260 
261 	/* anything above 1280 blows up */
262 	if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
263 		srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
264 
265 	srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
266 	srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
267 }
268 
269 static void __init srmmu_nocache_init(void)
270 {
271 	void *srmmu_nocache_bitmap;
272 	unsigned int bitmap_bits;
273 	pgd_t *pgd;
274 	pmd_t *pmd;
275 	pte_t *pte;
276 	unsigned long paddr, vaddr;
277 	unsigned long pteval;
278 
279 	bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
280 
281 	srmmu_nocache_pool = __alloc_bootmem(srmmu_nocache_size,
282 		SRMMU_NOCACHE_ALIGN_MAX, 0UL);
283 	memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
284 
285 	srmmu_nocache_bitmap =
286 		__alloc_bootmem(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
287 				SMP_CACHE_BYTES, 0UL);
288 	bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
289 
290 	srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
291 	memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
292 	init_mm.pgd = srmmu_swapper_pg_dir;
293 
294 	srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
295 
296 	paddr = __pa((unsigned long)srmmu_nocache_pool);
297 	vaddr = SRMMU_NOCACHE_VADDR;
298 
299 	while (vaddr < srmmu_nocache_end) {
300 		pgd = pgd_offset_k(vaddr);
301 		pmd = pmd_offset(__nocache_fix(pgd), vaddr);
302 		pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
303 
304 		pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
305 
306 		if (srmmu_cache_pagetables)
307 			pteval |= SRMMU_CACHE;
308 
309 		set_pte(__nocache_fix(pte), __pte(pteval));
310 
311 		vaddr += PAGE_SIZE;
312 		paddr += PAGE_SIZE;
313 	}
314 
315 	flush_cache_all();
316 	flush_tlb_all();
317 }
318 
319 pgd_t *get_pgd_fast(void)
320 {
321 	pgd_t *pgd = NULL;
322 
323 	pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
324 	if (pgd) {
325 		pgd_t *init = pgd_offset_k(0);
326 		memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
327 		memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
328 						(PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
329 	}
330 
331 	return pgd;
332 }
333 
334 /*
335  * Hardware needs alignment to 256 only, but we align to whole page size
336  * to reduce fragmentation problems due to the buddy principle.
337  * XXX Provide actual fragmentation statistics in /proc.
338  *
339  * Alignments up to the page size are the same for physical and virtual
340  * addresses of the nocache area.
341  */
342 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
343 {
344 	unsigned long pte;
345 	struct page *page;
346 
347 	if ((pte = (unsigned long)pte_alloc_one_kernel(mm, address)) == 0)
348 		return NULL;
349 	page = pfn_to_page(__nocache_pa(pte) >> PAGE_SHIFT);
350 	if (!pgtable_page_ctor(page)) {
351 		__free_page(page);
352 		return NULL;
353 	}
354 	return page;
355 }
356 
357 void pte_free(struct mm_struct *mm, pgtable_t pte)
358 {
359 	unsigned long p;
360 
361 	pgtable_page_dtor(pte);
362 	p = (unsigned long)page_address(pte);	/* Cached address (for test) */
363 	if (p == 0)
364 		BUG();
365 	p = page_to_pfn(pte) << PAGE_SHIFT;	/* Physical address */
366 
367 	/* free non cached virtual address*/
368 	srmmu_free_nocache(__nocache_va(p), PTE_SIZE);
369 }
370 
371 /* context handling - a dynamically sized pool is used */
372 #define NO_CONTEXT	-1
373 
374 struct ctx_list {
375 	struct ctx_list *next;
376 	struct ctx_list *prev;
377 	unsigned int ctx_number;
378 	struct mm_struct *ctx_mm;
379 };
380 
381 static struct ctx_list *ctx_list_pool;
382 static struct ctx_list ctx_free;
383 static struct ctx_list ctx_used;
384 
385 /* At boot time we determine the number of contexts */
386 static int num_contexts;
387 
388 static inline void remove_from_ctx_list(struct ctx_list *entry)
389 {
390 	entry->next->prev = entry->prev;
391 	entry->prev->next = entry->next;
392 }
393 
394 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
395 {
396 	entry->next = head;
397 	(entry->prev = head->prev)->next = entry;
398 	head->prev = entry;
399 }
400 #define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
401 #define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
402 
403 
404 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
405 {
406 	struct ctx_list *ctxp;
407 
408 	ctxp = ctx_free.next;
409 	if (ctxp != &ctx_free) {
410 		remove_from_ctx_list(ctxp);
411 		add_to_used_ctxlist(ctxp);
412 		mm->context = ctxp->ctx_number;
413 		ctxp->ctx_mm = mm;
414 		return;
415 	}
416 	ctxp = ctx_used.next;
417 	if (ctxp->ctx_mm == old_mm)
418 		ctxp = ctxp->next;
419 	if (ctxp == &ctx_used)
420 		panic("out of mmu contexts");
421 	flush_cache_mm(ctxp->ctx_mm);
422 	flush_tlb_mm(ctxp->ctx_mm);
423 	remove_from_ctx_list(ctxp);
424 	add_to_used_ctxlist(ctxp);
425 	ctxp->ctx_mm->context = NO_CONTEXT;
426 	ctxp->ctx_mm = mm;
427 	mm->context = ctxp->ctx_number;
428 }
429 
430 static inline void free_context(int context)
431 {
432 	struct ctx_list *ctx_old;
433 
434 	ctx_old = ctx_list_pool + context;
435 	remove_from_ctx_list(ctx_old);
436 	add_to_free_ctxlist(ctx_old);
437 }
438 
439 static void __init sparc_context_init(int numctx)
440 {
441 	int ctx;
442 	unsigned long size;
443 
444 	size = numctx * sizeof(struct ctx_list);
445 	ctx_list_pool = __alloc_bootmem(size, SMP_CACHE_BYTES, 0UL);
446 
447 	for (ctx = 0; ctx < numctx; ctx++) {
448 		struct ctx_list *clist;
449 
450 		clist = (ctx_list_pool + ctx);
451 		clist->ctx_number = ctx;
452 		clist->ctx_mm = NULL;
453 	}
454 	ctx_free.next = ctx_free.prev = &ctx_free;
455 	ctx_used.next = ctx_used.prev = &ctx_used;
456 	for (ctx = 0; ctx < numctx; ctx++)
457 		add_to_free_ctxlist(ctx_list_pool + ctx);
458 }
459 
460 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
461 	       struct task_struct *tsk)
462 {
463 	if (mm->context == NO_CONTEXT) {
464 		spin_lock(&srmmu_context_spinlock);
465 		alloc_context(old_mm, mm);
466 		spin_unlock(&srmmu_context_spinlock);
467 		srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
468 	}
469 
470 	if (sparc_cpu_model == sparc_leon)
471 		leon_switch_mm();
472 
473 	if (is_hypersparc)
474 		hyper_flush_whole_icache();
475 
476 	srmmu_set_context(mm->context);
477 }
478 
479 /* Low level IO area allocation on the SRMMU. */
480 static inline void srmmu_mapioaddr(unsigned long physaddr,
481 				   unsigned long virt_addr, int bus_type)
482 {
483 	pgd_t *pgdp;
484 	pmd_t *pmdp;
485 	pte_t *ptep;
486 	unsigned long tmp;
487 
488 	physaddr &= PAGE_MASK;
489 	pgdp = pgd_offset_k(virt_addr);
490 	pmdp = pmd_offset(pgdp, virt_addr);
491 	ptep = pte_offset_kernel(pmdp, virt_addr);
492 	tmp = (physaddr >> 4) | SRMMU_ET_PTE;
493 
494 	/* I need to test whether this is consistent over all
495 	 * sun4m's.  The bus_type represents the upper 4 bits of
496 	 * 36-bit physical address on the I/O space lines...
497 	 */
498 	tmp |= (bus_type << 28);
499 	tmp |= SRMMU_PRIV;
500 	__flush_page_to_ram(virt_addr);
501 	set_pte(ptep, __pte(tmp));
502 }
503 
504 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
505 		      unsigned long xva, unsigned int len)
506 {
507 	while (len != 0) {
508 		len -= PAGE_SIZE;
509 		srmmu_mapioaddr(xpa, xva, bus);
510 		xva += PAGE_SIZE;
511 		xpa += PAGE_SIZE;
512 	}
513 	flush_tlb_all();
514 }
515 
516 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
517 {
518 	pgd_t *pgdp;
519 	pmd_t *pmdp;
520 	pte_t *ptep;
521 
522 	pgdp = pgd_offset_k(virt_addr);
523 	pmdp = pmd_offset(pgdp, virt_addr);
524 	ptep = pte_offset_kernel(pmdp, virt_addr);
525 
526 	/* No need to flush uncacheable page. */
527 	__pte_clear(ptep);
528 }
529 
530 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
531 {
532 	while (len != 0) {
533 		len -= PAGE_SIZE;
534 		srmmu_unmapioaddr(virt_addr);
535 		virt_addr += PAGE_SIZE;
536 	}
537 	flush_tlb_all();
538 }
539 
540 /* tsunami.S */
541 extern void tsunami_flush_cache_all(void);
542 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
543 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
544 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
545 extern void tsunami_flush_page_to_ram(unsigned long page);
546 extern void tsunami_flush_page_for_dma(unsigned long page);
547 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
548 extern void tsunami_flush_tlb_all(void);
549 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
550 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
551 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
552 extern void tsunami_setup_blockops(void);
553 
554 /* swift.S */
555 extern void swift_flush_cache_all(void);
556 extern void swift_flush_cache_mm(struct mm_struct *mm);
557 extern void swift_flush_cache_range(struct vm_area_struct *vma,
558 				    unsigned long start, unsigned long end);
559 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
560 extern void swift_flush_page_to_ram(unsigned long page);
561 extern void swift_flush_page_for_dma(unsigned long page);
562 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
563 extern void swift_flush_tlb_all(void);
564 extern void swift_flush_tlb_mm(struct mm_struct *mm);
565 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
566 				  unsigned long start, unsigned long end);
567 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
568 
569 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
570 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
571 {
572 	int cctx, ctx1;
573 
574 	page &= PAGE_MASK;
575 	if ((ctx1 = vma->vm_mm->context) != -1) {
576 		cctx = srmmu_get_context();
577 /* Is context # ever different from current context? P3 */
578 		if (cctx != ctx1) {
579 			printk("flush ctx %02x curr %02x\n", ctx1, cctx);
580 			srmmu_set_context(ctx1);
581 			swift_flush_page(page);
582 			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
583 					"r" (page), "i" (ASI_M_FLUSH_PROBE));
584 			srmmu_set_context(cctx);
585 		} else {
586 			 /* Rm. prot. bits from virt. c. */
587 			/* swift_flush_cache_all(); */
588 			/* swift_flush_cache_page(vma, page); */
589 			swift_flush_page(page);
590 
591 			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
592 				"r" (page), "i" (ASI_M_FLUSH_PROBE));
593 			/* same as above: srmmu_flush_tlb_page() */
594 		}
595 	}
596 }
597 #endif
598 
599 /*
600  * The following are all MBUS based SRMMU modules, and therefore could
601  * be found in a multiprocessor configuration.  On the whole, these
602  * chips seems to be much more touchy about DVMA and page tables
603  * with respect to cache coherency.
604  */
605 
606 /* viking.S */
607 extern void viking_flush_cache_all(void);
608 extern void viking_flush_cache_mm(struct mm_struct *mm);
609 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
610 				     unsigned long end);
611 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
612 extern void viking_flush_page_to_ram(unsigned long page);
613 extern void viking_flush_page_for_dma(unsigned long page);
614 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
615 extern void viking_flush_page(unsigned long page);
616 extern void viking_mxcc_flush_page(unsigned long page);
617 extern void viking_flush_tlb_all(void);
618 extern void viking_flush_tlb_mm(struct mm_struct *mm);
619 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
620 				   unsigned long end);
621 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
622 				  unsigned long page);
623 extern void sun4dsmp_flush_tlb_all(void);
624 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
625 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
626 				   unsigned long end);
627 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
628 				  unsigned long page);
629 
630 /* hypersparc.S */
631 extern void hypersparc_flush_cache_all(void);
632 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
633 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
634 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
635 extern void hypersparc_flush_page_to_ram(unsigned long page);
636 extern void hypersparc_flush_page_for_dma(unsigned long page);
637 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
638 extern void hypersparc_flush_tlb_all(void);
639 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
640 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
641 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
642 extern void hypersparc_setup_blockops(void);
643 
644 /*
645  * NOTE: All of this startup code assumes the low 16mb (approx.) of
646  *       kernel mappings are done with one single contiguous chunk of
647  *       ram.  On small ram machines (classics mainly) we only get
648  *       around 8mb mapped for us.
649  */
650 
651 static void __init early_pgtable_allocfail(char *type)
652 {
653 	prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
654 	prom_halt();
655 }
656 
657 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
658 							unsigned long end)
659 {
660 	pgd_t *pgdp;
661 	pmd_t *pmdp;
662 	pte_t *ptep;
663 
664 	while (start < end) {
665 		pgdp = pgd_offset_k(start);
666 		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
667 			pmdp = __srmmu_get_nocache(
668 			    SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
669 			if (pmdp == NULL)
670 				early_pgtable_allocfail("pmd");
671 			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
672 			pgd_set(__nocache_fix(pgdp), pmdp);
673 		}
674 		pmdp = pmd_offset(__nocache_fix(pgdp), start);
675 		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
676 			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
677 			if (ptep == NULL)
678 				early_pgtable_allocfail("pte");
679 			memset(__nocache_fix(ptep), 0, PTE_SIZE);
680 			pmd_set(__nocache_fix(pmdp), ptep);
681 		}
682 		if (start > (0xffffffffUL - PMD_SIZE))
683 			break;
684 		start = (start + PMD_SIZE) & PMD_MASK;
685 	}
686 }
687 
688 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
689 						  unsigned long end)
690 {
691 	pgd_t *pgdp;
692 	pmd_t *pmdp;
693 	pte_t *ptep;
694 
695 	while (start < end) {
696 		pgdp = pgd_offset_k(start);
697 		if (pgd_none(*pgdp)) {
698 			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
699 			if (pmdp == NULL)
700 				early_pgtable_allocfail("pmd");
701 			memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
702 			pgd_set(pgdp, pmdp);
703 		}
704 		pmdp = pmd_offset(pgdp, start);
705 		if (srmmu_pmd_none(*pmdp)) {
706 			ptep = __srmmu_get_nocache(PTE_SIZE,
707 							     PTE_SIZE);
708 			if (ptep == NULL)
709 				early_pgtable_allocfail("pte");
710 			memset(ptep, 0, PTE_SIZE);
711 			pmd_set(pmdp, ptep);
712 		}
713 		if (start > (0xffffffffUL - PMD_SIZE))
714 			break;
715 		start = (start + PMD_SIZE) & PMD_MASK;
716 	}
717 }
718 
719 /* These flush types are not available on all chips... */
720 static inline unsigned long srmmu_probe(unsigned long vaddr)
721 {
722 	unsigned long retval;
723 
724 	if (sparc_cpu_model != sparc_leon) {
725 
726 		vaddr &= PAGE_MASK;
727 		__asm__ __volatile__("lda [%1] %2, %0\n\t" :
728 				     "=r" (retval) :
729 				     "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
730 	} else {
731 		retval = leon_swprobe(vaddr, NULL);
732 	}
733 	return retval;
734 }
735 
736 /*
737  * This is much cleaner than poking around physical address space
738  * looking at the prom's page table directly which is what most
739  * other OS's do.  Yuck... this is much better.
740  */
741 static void __init srmmu_inherit_prom_mappings(unsigned long start,
742 					       unsigned long end)
743 {
744 	unsigned long probed;
745 	unsigned long addr;
746 	pgd_t *pgdp;
747 	pmd_t *pmdp;
748 	pte_t *ptep;
749 	int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
750 
751 	while (start <= end) {
752 		if (start == 0)
753 			break; /* probably wrap around */
754 		if (start == 0xfef00000)
755 			start = KADB_DEBUGGER_BEGVM;
756 		probed = srmmu_probe(start);
757 		if (!probed) {
758 			/* continue probing until we find an entry */
759 			start += PAGE_SIZE;
760 			continue;
761 		}
762 
763 		/* A red snapper, see what it really is. */
764 		what = 0;
765 		addr = start - PAGE_SIZE;
766 
767 		if (!(start & ~(SRMMU_REAL_PMD_MASK))) {
768 			if (srmmu_probe(addr + SRMMU_REAL_PMD_SIZE) == probed)
769 				what = 1;
770 		}
771 
772 		if (!(start & ~(SRMMU_PGDIR_MASK))) {
773 			if (srmmu_probe(addr + SRMMU_PGDIR_SIZE) == probed)
774 				what = 2;
775 		}
776 
777 		pgdp = pgd_offset_k(start);
778 		if (what == 2) {
779 			*(pgd_t *)__nocache_fix(pgdp) = __pgd(probed);
780 			start += SRMMU_PGDIR_SIZE;
781 			continue;
782 		}
783 		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
784 			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
785 						   SRMMU_PMD_TABLE_SIZE);
786 			if (pmdp == NULL)
787 				early_pgtable_allocfail("pmd");
788 			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
789 			pgd_set(__nocache_fix(pgdp), pmdp);
790 		}
791 		pmdp = pmd_offset(__nocache_fix(pgdp), start);
792 		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
793 			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
794 			if (ptep == NULL)
795 				early_pgtable_allocfail("pte");
796 			memset(__nocache_fix(ptep), 0, PTE_SIZE);
797 			pmd_set(__nocache_fix(pmdp), ptep);
798 		}
799 		if (what == 1) {
800 			/* We bend the rule where all 16 PTPs in a pmd_t point
801 			 * inside the same PTE page, and we leak a perfectly
802 			 * good hardware PTE piece. Alternatives seem worse.
803 			 */
804 			unsigned int x;	/* Index of HW PMD in soft cluster */
805 			unsigned long *val;
806 			x = (start >> PMD_SHIFT) & 15;
807 			val = &pmdp->pmdv[x];
808 			*(unsigned long *)__nocache_fix(val) = probed;
809 			start += SRMMU_REAL_PMD_SIZE;
810 			continue;
811 		}
812 		ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
813 		*(pte_t *)__nocache_fix(ptep) = __pte(probed);
814 		start += PAGE_SIZE;
815 	}
816 }
817 
818 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
819 
820 /* Create a third-level SRMMU 16MB page mapping. */
821 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
822 {
823 	pgd_t *pgdp = pgd_offset_k(vaddr);
824 	unsigned long big_pte;
825 
826 	big_pte = KERNEL_PTE(phys_base >> 4);
827 	*(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
828 }
829 
830 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
831 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
832 {
833 	unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
834 	unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
835 	unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
836 	/* Map "low" memory only */
837 	const unsigned long min_vaddr = PAGE_OFFSET;
838 	const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
839 
840 	if (vstart < min_vaddr || vstart >= max_vaddr)
841 		return vstart;
842 
843 	if (vend > max_vaddr || vend < min_vaddr)
844 		vend = max_vaddr;
845 
846 	while (vstart < vend) {
847 		do_large_mapping(vstart, pstart);
848 		vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
849 	}
850 	return vstart;
851 }
852 
853 static void __init map_kernel(void)
854 {
855 	int i;
856 
857 	if (phys_base > 0) {
858 		do_large_mapping(PAGE_OFFSET, phys_base);
859 	}
860 
861 	for (i = 0; sp_banks[i].num_bytes != 0; i++) {
862 		map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
863 	}
864 }
865 
866 void (*poke_srmmu)(void) = NULL;
867 
868 void __init srmmu_paging_init(void)
869 {
870 	int i;
871 	phandle cpunode;
872 	char node_str[128];
873 	pgd_t *pgd;
874 	pmd_t *pmd;
875 	pte_t *pte;
876 	unsigned long pages_avail;
877 
878 	init_mm.context = (unsigned long) NO_CONTEXT;
879 	sparc_iomap.start = SUN4M_IOBASE_VADDR;	/* 16MB of IOSPACE on all sun4m's. */
880 
881 	if (sparc_cpu_model == sun4d)
882 		num_contexts = 65536; /* We know it is Viking */
883 	else {
884 		/* Find the number of contexts on the srmmu. */
885 		cpunode = prom_getchild(prom_root_node);
886 		num_contexts = 0;
887 		while (cpunode != 0) {
888 			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
889 			if (!strcmp(node_str, "cpu")) {
890 				num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
891 				break;
892 			}
893 			cpunode = prom_getsibling(cpunode);
894 		}
895 	}
896 
897 	if (!num_contexts) {
898 		prom_printf("Something wrong, can't find cpu node in paging_init.\n");
899 		prom_halt();
900 	}
901 
902 	pages_avail = 0;
903 	last_valid_pfn = bootmem_init(&pages_avail);
904 
905 	srmmu_nocache_calcsize();
906 	srmmu_nocache_init();
907 	srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
908 	map_kernel();
909 
910 	/* ctx table has to be physically aligned to its size */
911 	srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
912 	srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa((unsigned long)srmmu_context_table);
913 
914 	for (i = 0; i < num_contexts; i++)
915 		srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
916 
917 	flush_cache_all();
918 	srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
919 #ifdef CONFIG_SMP
920 	/* Stop from hanging here... */
921 	local_ops->tlb_all();
922 #else
923 	flush_tlb_all();
924 #endif
925 	poke_srmmu();
926 
927 	srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
928 	srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
929 
930 	srmmu_allocate_ptable_skeleton(
931 		__fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
932 	srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
933 
934 	pgd = pgd_offset_k(PKMAP_BASE);
935 	pmd = pmd_offset(pgd, PKMAP_BASE);
936 	pte = pte_offset_kernel(pmd, PKMAP_BASE);
937 	pkmap_page_table = pte;
938 
939 	flush_cache_all();
940 	flush_tlb_all();
941 
942 	sparc_context_init(num_contexts);
943 
944 	kmap_init();
945 
946 	{
947 		unsigned long zones_size[MAX_NR_ZONES];
948 		unsigned long zholes_size[MAX_NR_ZONES];
949 		unsigned long npages;
950 		int znum;
951 
952 		for (znum = 0; znum < MAX_NR_ZONES; znum++)
953 			zones_size[znum] = zholes_size[znum] = 0;
954 
955 		npages = max_low_pfn - pfn_base;
956 
957 		zones_size[ZONE_DMA] = npages;
958 		zholes_size[ZONE_DMA] = npages - pages_avail;
959 
960 		npages = highend_pfn - max_low_pfn;
961 		zones_size[ZONE_HIGHMEM] = npages;
962 		zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
963 
964 		free_area_init_node(0, zones_size, pfn_base, zholes_size);
965 	}
966 }
967 
968 void mmu_info(struct seq_file *m)
969 {
970 	seq_printf(m,
971 		   "MMU type\t: %s\n"
972 		   "contexts\t: %d\n"
973 		   "nocache total\t: %ld\n"
974 		   "nocache used\t: %d\n",
975 		   srmmu_name,
976 		   num_contexts,
977 		   srmmu_nocache_size,
978 		   srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
979 }
980 
981 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
982 {
983 	mm->context = NO_CONTEXT;
984 	return 0;
985 }
986 
987 void destroy_context(struct mm_struct *mm)
988 {
989 
990 	if (mm->context != NO_CONTEXT) {
991 		flush_cache_mm(mm);
992 		srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
993 		flush_tlb_mm(mm);
994 		spin_lock(&srmmu_context_spinlock);
995 		free_context(mm->context);
996 		spin_unlock(&srmmu_context_spinlock);
997 		mm->context = NO_CONTEXT;
998 	}
999 }
1000 
1001 /* Init various srmmu chip types. */
1002 static void __init srmmu_is_bad(void)
1003 {
1004 	prom_printf("Could not determine SRMMU chip type.\n");
1005 	prom_halt();
1006 }
1007 
1008 static void __init init_vac_layout(void)
1009 {
1010 	phandle nd;
1011 	int cache_lines;
1012 	char node_str[128];
1013 #ifdef CONFIG_SMP
1014 	int cpu = 0;
1015 	unsigned long max_size = 0;
1016 	unsigned long min_line_size = 0x10000000;
1017 #endif
1018 
1019 	nd = prom_getchild(prom_root_node);
1020 	while ((nd = prom_getsibling(nd)) != 0) {
1021 		prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1022 		if (!strcmp(node_str, "cpu")) {
1023 			vac_line_size = prom_getint(nd, "cache-line-size");
1024 			if (vac_line_size == -1) {
1025 				prom_printf("can't determine cache-line-size, halting.\n");
1026 				prom_halt();
1027 			}
1028 			cache_lines = prom_getint(nd, "cache-nlines");
1029 			if (cache_lines == -1) {
1030 				prom_printf("can't determine cache-nlines, halting.\n");
1031 				prom_halt();
1032 			}
1033 
1034 			vac_cache_size = cache_lines * vac_line_size;
1035 #ifdef CONFIG_SMP
1036 			if (vac_cache_size > max_size)
1037 				max_size = vac_cache_size;
1038 			if (vac_line_size < min_line_size)
1039 				min_line_size = vac_line_size;
1040 			//FIXME: cpus not contiguous!!
1041 			cpu++;
1042 			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1043 				break;
1044 #else
1045 			break;
1046 #endif
1047 		}
1048 	}
1049 	if (nd == 0) {
1050 		prom_printf("No CPU nodes found, halting.\n");
1051 		prom_halt();
1052 	}
1053 #ifdef CONFIG_SMP
1054 	vac_cache_size = max_size;
1055 	vac_line_size = min_line_size;
1056 #endif
1057 	printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1058 	       (int)vac_cache_size, (int)vac_line_size);
1059 }
1060 
1061 static void poke_hypersparc(void)
1062 {
1063 	volatile unsigned long clear;
1064 	unsigned long mreg = srmmu_get_mmureg();
1065 
1066 	hyper_flush_unconditional_combined();
1067 
1068 	mreg &= ~(HYPERSPARC_CWENABLE);
1069 	mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1070 	mreg |= (HYPERSPARC_CMODE);
1071 
1072 	srmmu_set_mmureg(mreg);
1073 
1074 #if 0 /* XXX I think this is bad news... -DaveM */
1075 	hyper_clear_all_tags();
1076 #endif
1077 
1078 	put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1079 	hyper_flush_whole_icache();
1080 	clear = srmmu_get_faddr();
1081 	clear = srmmu_get_fstatus();
1082 }
1083 
1084 static const struct sparc32_cachetlb_ops hypersparc_ops = {
1085 	.cache_all	= hypersparc_flush_cache_all,
1086 	.cache_mm	= hypersparc_flush_cache_mm,
1087 	.cache_page	= hypersparc_flush_cache_page,
1088 	.cache_range	= hypersparc_flush_cache_range,
1089 	.tlb_all	= hypersparc_flush_tlb_all,
1090 	.tlb_mm		= hypersparc_flush_tlb_mm,
1091 	.tlb_page	= hypersparc_flush_tlb_page,
1092 	.tlb_range	= hypersparc_flush_tlb_range,
1093 	.page_to_ram	= hypersparc_flush_page_to_ram,
1094 	.sig_insns	= hypersparc_flush_sig_insns,
1095 	.page_for_dma	= hypersparc_flush_page_for_dma,
1096 };
1097 
1098 static void __init init_hypersparc(void)
1099 {
1100 	srmmu_name = "ROSS HyperSparc";
1101 	srmmu_modtype = HyperSparc;
1102 
1103 	init_vac_layout();
1104 
1105 	is_hypersparc = 1;
1106 	sparc32_cachetlb_ops = &hypersparc_ops;
1107 
1108 	poke_srmmu = poke_hypersparc;
1109 
1110 	hypersparc_setup_blockops();
1111 }
1112 
1113 static void poke_swift(void)
1114 {
1115 	unsigned long mreg;
1116 
1117 	/* Clear any crap from the cache or else... */
1118 	swift_flush_cache_all();
1119 
1120 	/* Enable I & D caches */
1121 	mreg = srmmu_get_mmureg();
1122 	mreg |= (SWIFT_IE | SWIFT_DE);
1123 	/*
1124 	 * The Swift branch folding logic is completely broken.  At
1125 	 * trap time, if things are just right, if can mistakenly
1126 	 * think that a trap is coming from kernel mode when in fact
1127 	 * it is coming from user mode (it mis-executes the branch in
1128 	 * the trap code).  So you see things like crashme completely
1129 	 * hosing your machine which is completely unacceptable.  Turn
1130 	 * this shit off... nice job Fujitsu.
1131 	 */
1132 	mreg &= ~(SWIFT_BF);
1133 	srmmu_set_mmureg(mreg);
1134 }
1135 
1136 static const struct sparc32_cachetlb_ops swift_ops = {
1137 	.cache_all	= swift_flush_cache_all,
1138 	.cache_mm	= swift_flush_cache_mm,
1139 	.cache_page	= swift_flush_cache_page,
1140 	.cache_range	= swift_flush_cache_range,
1141 	.tlb_all	= swift_flush_tlb_all,
1142 	.tlb_mm		= swift_flush_tlb_mm,
1143 	.tlb_page	= swift_flush_tlb_page,
1144 	.tlb_range	= swift_flush_tlb_range,
1145 	.page_to_ram	= swift_flush_page_to_ram,
1146 	.sig_insns	= swift_flush_sig_insns,
1147 	.page_for_dma	= swift_flush_page_for_dma,
1148 };
1149 
1150 #define SWIFT_MASKID_ADDR  0x10003018
1151 static void __init init_swift(void)
1152 {
1153 	unsigned long swift_rev;
1154 
1155 	__asm__ __volatile__("lda [%1] %2, %0\n\t"
1156 			     "srl %0, 0x18, %0\n\t" :
1157 			     "=r" (swift_rev) :
1158 			     "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1159 	srmmu_name = "Fujitsu Swift";
1160 	switch (swift_rev) {
1161 	case 0x11:
1162 	case 0x20:
1163 	case 0x23:
1164 	case 0x30:
1165 		srmmu_modtype = Swift_lots_o_bugs;
1166 		hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1167 		/*
1168 		 * Gee george, I wonder why Sun is so hush hush about
1169 		 * this hardware bug... really braindamage stuff going
1170 		 * on here.  However I think we can find a way to avoid
1171 		 * all of the workaround overhead under Linux.  Basically,
1172 		 * any page fault can cause kernel pages to become user
1173 		 * accessible (the mmu gets confused and clears some of
1174 		 * the ACC bits in kernel ptes).  Aha, sounds pretty
1175 		 * horrible eh?  But wait, after extensive testing it appears
1176 		 * that if you use pgd_t level large kernel pte's (like the
1177 		 * 4MB pages on the Pentium) the bug does not get tripped
1178 		 * at all.  This avoids almost all of the major overhead.
1179 		 * Welcome to a world where your vendor tells you to,
1180 		 * "apply this kernel patch" instead of "sorry for the
1181 		 * broken hardware, send it back and we'll give you
1182 		 * properly functioning parts"
1183 		 */
1184 		break;
1185 	case 0x25:
1186 	case 0x31:
1187 		srmmu_modtype = Swift_bad_c;
1188 		hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1189 		/*
1190 		 * You see Sun allude to this hardware bug but never
1191 		 * admit things directly, they'll say things like,
1192 		 * "the Swift chip cache problems" or similar.
1193 		 */
1194 		break;
1195 	default:
1196 		srmmu_modtype = Swift_ok;
1197 		break;
1198 	}
1199 
1200 	sparc32_cachetlb_ops = &swift_ops;
1201 	flush_page_for_dma_global = 0;
1202 
1203 	/*
1204 	 * Are you now convinced that the Swift is one of the
1205 	 * biggest VLSI abortions of all time?  Bravo Fujitsu!
1206 	 * Fujitsu, the !#?!%$'d up processor people.  I bet if
1207 	 * you examined the microcode of the Swift you'd find
1208 	 * XXX's all over the place.
1209 	 */
1210 	poke_srmmu = poke_swift;
1211 }
1212 
1213 static void turbosparc_flush_cache_all(void)
1214 {
1215 	flush_user_windows();
1216 	turbosparc_idflash_clear();
1217 }
1218 
1219 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1220 {
1221 	FLUSH_BEGIN(mm)
1222 	flush_user_windows();
1223 	turbosparc_idflash_clear();
1224 	FLUSH_END
1225 }
1226 
1227 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1228 {
1229 	FLUSH_BEGIN(vma->vm_mm)
1230 	flush_user_windows();
1231 	turbosparc_idflash_clear();
1232 	FLUSH_END
1233 }
1234 
1235 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1236 {
1237 	FLUSH_BEGIN(vma->vm_mm)
1238 	flush_user_windows();
1239 	if (vma->vm_flags & VM_EXEC)
1240 		turbosparc_flush_icache();
1241 	turbosparc_flush_dcache();
1242 	FLUSH_END
1243 }
1244 
1245 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1246 static void turbosparc_flush_page_to_ram(unsigned long page)
1247 {
1248 #ifdef TURBOSPARC_WRITEBACK
1249 	volatile unsigned long clear;
1250 
1251 	if (srmmu_probe(page))
1252 		turbosparc_flush_page_cache(page);
1253 	clear = srmmu_get_fstatus();
1254 #endif
1255 }
1256 
1257 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1258 {
1259 }
1260 
1261 static void turbosparc_flush_page_for_dma(unsigned long page)
1262 {
1263 	turbosparc_flush_dcache();
1264 }
1265 
1266 static void turbosparc_flush_tlb_all(void)
1267 {
1268 	srmmu_flush_whole_tlb();
1269 }
1270 
1271 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1272 {
1273 	FLUSH_BEGIN(mm)
1274 	srmmu_flush_whole_tlb();
1275 	FLUSH_END
1276 }
1277 
1278 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1279 {
1280 	FLUSH_BEGIN(vma->vm_mm)
1281 	srmmu_flush_whole_tlb();
1282 	FLUSH_END
1283 }
1284 
1285 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1286 {
1287 	FLUSH_BEGIN(vma->vm_mm)
1288 	srmmu_flush_whole_tlb();
1289 	FLUSH_END
1290 }
1291 
1292 
1293 static void poke_turbosparc(void)
1294 {
1295 	unsigned long mreg = srmmu_get_mmureg();
1296 	unsigned long ccreg;
1297 
1298 	/* Clear any crap from the cache or else... */
1299 	turbosparc_flush_cache_all();
1300 	/* Temporarily disable I & D caches */
1301 	mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1302 	mreg &= ~(TURBOSPARC_PCENABLE);		/* Don't check parity */
1303 	srmmu_set_mmureg(mreg);
1304 
1305 	ccreg = turbosparc_get_ccreg();
1306 
1307 #ifdef TURBOSPARC_WRITEBACK
1308 	ccreg |= (TURBOSPARC_SNENABLE);		/* Do DVMA snooping in Dcache */
1309 	ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1310 			/* Write-back D-cache, emulate VLSI
1311 			 * abortion number three, not number one */
1312 #else
1313 	/* For now let's play safe, optimize later */
1314 	ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1315 			/* Do DVMA snooping in Dcache, Write-thru D-cache */
1316 	ccreg &= ~(TURBOSPARC_uS2);
1317 			/* Emulate VLSI abortion number three, not number one */
1318 #endif
1319 
1320 	switch (ccreg & 7) {
1321 	case 0: /* No SE cache */
1322 	case 7: /* Test mode */
1323 		break;
1324 	default:
1325 		ccreg |= (TURBOSPARC_SCENABLE);
1326 	}
1327 	turbosparc_set_ccreg(ccreg);
1328 
1329 	mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1330 	mreg |= (TURBOSPARC_ICSNOOP);		/* Icache snooping on */
1331 	srmmu_set_mmureg(mreg);
1332 }
1333 
1334 static const struct sparc32_cachetlb_ops turbosparc_ops = {
1335 	.cache_all	= turbosparc_flush_cache_all,
1336 	.cache_mm	= turbosparc_flush_cache_mm,
1337 	.cache_page	= turbosparc_flush_cache_page,
1338 	.cache_range	= turbosparc_flush_cache_range,
1339 	.tlb_all	= turbosparc_flush_tlb_all,
1340 	.tlb_mm		= turbosparc_flush_tlb_mm,
1341 	.tlb_page	= turbosparc_flush_tlb_page,
1342 	.tlb_range	= turbosparc_flush_tlb_range,
1343 	.page_to_ram	= turbosparc_flush_page_to_ram,
1344 	.sig_insns	= turbosparc_flush_sig_insns,
1345 	.page_for_dma	= turbosparc_flush_page_for_dma,
1346 };
1347 
1348 static void __init init_turbosparc(void)
1349 {
1350 	srmmu_name = "Fujitsu TurboSparc";
1351 	srmmu_modtype = TurboSparc;
1352 	sparc32_cachetlb_ops = &turbosparc_ops;
1353 	poke_srmmu = poke_turbosparc;
1354 }
1355 
1356 static void poke_tsunami(void)
1357 {
1358 	unsigned long mreg = srmmu_get_mmureg();
1359 
1360 	tsunami_flush_icache();
1361 	tsunami_flush_dcache();
1362 	mreg &= ~TSUNAMI_ITD;
1363 	mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1364 	srmmu_set_mmureg(mreg);
1365 }
1366 
1367 static const struct sparc32_cachetlb_ops tsunami_ops = {
1368 	.cache_all	= tsunami_flush_cache_all,
1369 	.cache_mm	= tsunami_flush_cache_mm,
1370 	.cache_page	= tsunami_flush_cache_page,
1371 	.cache_range	= tsunami_flush_cache_range,
1372 	.tlb_all	= tsunami_flush_tlb_all,
1373 	.tlb_mm		= tsunami_flush_tlb_mm,
1374 	.tlb_page	= tsunami_flush_tlb_page,
1375 	.tlb_range	= tsunami_flush_tlb_range,
1376 	.page_to_ram	= tsunami_flush_page_to_ram,
1377 	.sig_insns	= tsunami_flush_sig_insns,
1378 	.page_for_dma	= tsunami_flush_page_for_dma,
1379 };
1380 
1381 static void __init init_tsunami(void)
1382 {
1383 	/*
1384 	 * Tsunami's pretty sane, Sun and TI actually got it
1385 	 * somewhat right this time.  Fujitsu should have
1386 	 * taken some lessons from them.
1387 	 */
1388 
1389 	srmmu_name = "TI Tsunami";
1390 	srmmu_modtype = Tsunami;
1391 	sparc32_cachetlb_ops = &tsunami_ops;
1392 	poke_srmmu = poke_tsunami;
1393 
1394 	tsunami_setup_blockops();
1395 }
1396 
1397 static void poke_viking(void)
1398 {
1399 	unsigned long mreg = srmmu_get_mmureg();
1400 	static int smp_catch;
1401 
1402 	if (viking_mxcc_present) {
1403 		unsigned long mxcc_control = mxcc_get_creg();
1404 
1405 		mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1406 		mxcc_control &= ~(MXCC_CTL_RRC);
1407 		mxcc_set_creg(mxcc_control);
1408 
1409 		/*
1410 		 * We don't need memory parity checks.
1411 		 * XXX This is a mess, have to dig out later. ecd.
1412 		viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1413 		 */
1414 
1415 		/* We do cache ptables on MXCC. */
1416 		mreg |= VIKING_TCENABLE;
1417 	} else {
1418 		unsigned long bpreg;
1419 
1420 		mreg &= ~(VIKING_TCENABLE);
1421 		if (smp_catch++) {
1422 			/* Must disable mixed-cmd mode here for other cpu's. */
1423 			bpreg = viking_get_bpreg();
1424 			bpreg &= ~(VIKING_ACTION_MIX);
1425 			viking_set_bpreg(bpreg);
1426 
1427 			/* Just in case PROM does something funny. */
1428 			msi_set_sync();
1429 		}
1430 	}
1431 
1432 	mreg |= VIKING_SPENABLE;
1433 	mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1434 	mreg |= VIKING_SBENABLE;
1435 	mreg &= ~(VIKING_ACENABLE);
1436 	srmmu_set_mmureg(mreg);
1437 }
1438 
1439 static struct sparc32_cachetlb_ops viking_ops = {
1440 	.cache_all	= viking_flush_cache_all,
1441 	.cache_mm	= viking_flush_cache_mm,
1442 	.cache_page	= viking_flush_cache_page,
1443 	.cache_range	= viking_flush_cache_range,
1444 	.tlb_all	= viking_flush_tlb_all,
1445 	.tlb_mm		= viking_flush_tlb_mm,
1446 	.tlb_page	= viking_flush_tlb_page,
1447 	.tlb_range	= viking_flush_tlb_range,
1448 	.page_to_ram	= viking_flush_page_to_ram,
1449 	.sig_insns	= viking_flush_sig_insns,
1450 	.page_for_dma	= viking_flush_page_for_dma,
1451 };
1452 
1453 #ifdef CONFIG_SMP
1454 /* On sun4d the cpu broadcasts local TLB flushes, so we can just
1455  * perform the local TLB flush and all the other cpus will see it.
1456  * But, unfortunately, there is a bug in the sun4d XBUS backplane
1457  * that requires that we add some synchronization to these flushes.
1458  *
1459  * The bug is that the fifo which keeps track of all the pending TLB
1460  * broadcasts in the system is an entry or two too small, so if we
1461  * have too many going at once we'll overflow that fifo and lose a TLB
1462  * flush resulting in corruption.
1463  *
1464  * Our workaround is to take a global spinlock around the TLB flushes,
1465  * which guarentees we won't ever have too many pending.  It's a big
1466  * hammer, but a semaphore like system to make sure we only have N TLB
1467  * flushes going at once will require SMP locking anyways so there's
1468  * no real value in trying any harder than this.
1469  */
1470 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops = {
1471 	.cache_all	= viking_flush_cache_all,
1472 	.cache_mm	= viking_flush_cache_mm,
1473 	.cache_page	= viking_flush_cache_page,
1474 	.cache_range	= viking_flush_cache_range,
1475 	.tlb_all	= sun4dsmp_flush_tlb_all,
1476 	.tlb_mm		= sun4dsmp_flush_tlb_mm,
1477 	.tlb_page	= sun4dsmp_flush_tlb_page,
1478 	.tlb_range	= sun4dsmp_flush_tlb_range,
1479 	.page_to_ram	= viking_flush_page_to_ram,
1480 	.sig_insns	= viking_flush_sig_insns,
1481 	.page_for_dma	= viking_flush_page_for_dma,
1482 };
1483 #endif
1484 
1485 static void __init init_viking(void)
1486 {
1487 	unsigned long mreg = srmmu_get_mmureg();
1488 
1489 	/* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1490 	if (mreg & VIKING_MMODE) {
1491 		srmmu_name = "TI Viking";
1492 		viking_mxcc_present = 0;
1493 		msi_set_sync();
1494 
1495 		/*
1496 		 * We need this to make sure old viking takes no hits
1497 		 * on it's cache for dma snoops to workaround the
1498 		 * "load from non-cacheable memory" interrupt bug.
1499 		 * This is only necessary because of the new way in
1500 		 * which we use the IOMMU.
1501 		 */
1502 		viking_ops.page_for_dma = viking_flush_page;
1503 #ifdef CONFIG_SMP
1504 		viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1505 #endif
1506 		flush_page_for_dma_global = 0;
1507 	} else {
1508 		srmmu_name = "TI Viking/MXCC";
1509 		viking_mxcc_present = 1;
1510 		srmmu_cache_pagetables = 1;
1511 	}
1512 
1513 	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1514 		&viking_ops;
1515 #ifdef CONFIG_SMP
1516 	if (sparc_cpu_model == sun4d)
1517 		sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1518 			&viking_sun4d_smp_ops;
1519 #endif
1520 
1521 	poke_srmmu = poke_viking;
1522 }
1523 
1524 /* Probe for the srmmu chip version. */
1525 static void __init get_srmmu_type(void)
1526 {
1527 	unsigned long mreg, psr;
1528 	unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1529 
1530 	srmmu_modtype = SRMMU_INVAL_MOD;
1531 	hwbug_bitmask = 0;
1532 
1533 	mreg = srmmu_get_mmureg(); psr = get_psr();
1534 	mod_typ = (mreg & 0xf0000000) >> 28;
1535 	mod_rev = (mreg & 0x0f000000) >> 24;
1536 	psr_typ = (psr >> 28) & 0xf;
1537 	psr_vers = (psr >> 24) & 0xf;
1538 
1539 	/* First, check for sparc-leon. */
1540 	if (sparc_cpu_model == sparc_leon) {
1541 		init_leon();
1542 		return;
1543 	}
1544 
1545 	/* Second, check for HyperSparc or Cypress. */
1546 	if (mod_typ == 1) {
1547 		switch (mod_rev) {
1548 		case 7:
1549 			/* UP or MP Hypersparc */
1550 			init_hypersparc();
1551 			break;
1552 		case 0:
1553 		case 2:
1554 		case 10:
1555 		case 11:
1556 		case 12:
1557 		case 13:
1558 		case 14:
1559 		case 15:
1560 		default:
1561 			prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1562 			prom_halt();
1563 			break;
1564 		}
1565 		return;
1566 	}
1567 
1568 	/* Now Fujitsu TurboSparc. It might happen that it is
1569 	 * in Swift emulation mode, so we will check later...
1570 	 */
1571 	if (psr_typ == 0 && psr_vers == 5) {
1572 		init_turbosparc();
1573 		return;
1574 	}
1575 
1576 	/* Next check for Fujitsu Swift. */
1577 	if (psr_typ == 0 && psr_vers == 4) {
1578 		phandle cpunode;
1579 		char node_str[128];
1580 
1581 		/* Look if it is not a TurboSparc emulating Swift... */
1582 		cpunode = prom_getchild(prom_root_node);
1583 		while ((cpunode = prom_getsibling(cpunode)) != 0) {
1584 			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1585 			if (!strcmp(node_str, "cpu")) {
1586 				if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1587 				    prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1588 					init_turbosparc();
1589 					return;
1590 				}
1591 				break;
1592 			}
1593 		}
1594 
1595 		init_swift();
1596 		return;
1597 	}
1598 
1599 	/* Now the Viking family of srmmu. */
1600 	if (psr_typ == 4 &&
1601 	   ((psr_vers == 0) ||
1602 	    ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1603 		init_viking();
1604 		return;
1605 	}
1606 
1607 	/* Finally the Tsunami. */
1608 	if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1609 		init_tsunami();
1610 		return;
1611 	}
1612 
1613 	/* Oh well */
1614 	srmmu_is_bad();
1615 }
1616 
1617 #ifdef CONFIG_SMP
1618 /* Local cross-calls. */
1619 static void smp_flush_page_for_dma(unsigned long page)
1620 {
1621 	xc1((smpfunc_t) local_ops->page_for_dma, page);
1622 	local_ops->page_for_dma(page);
1623 }
1624 
1625 static void smp_flush_cache_all(void)
1626 {
1627 	xc0((smpfunc_t) local_ops->cache_all);
1628 	local_ops->cache_all();
1629 }
1630 
1631 static void smp_flush_tlb_all(void)
1632 {
1633 	xc0((smpfunc_t) local_ops->tlb_all);
1634 	local_ops->tlb_all();
1635 }
1636 
1637 static void smp_flush_cache_mm(struct mm_struct *mm)
1638 {
1639 	if (mm->context != NO_CONTEXT) {
1640 		cpumask_t cpu_mask;
1641 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1642 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1643 		if (!cpumask_empty(&cpu_mask))
1644 			xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1645 		local_ops->cache_mm(mm);
1646 	}
1647 }
1648 
1649 static void smp_flush_tlb_mm(struct mm_struct *mm)
1650 {
1651 	if (mm->context != NO_CONTEXT) {
1652 		cpumask_t cpu_mask;
1653 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1654 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1655 		if (!cpumask_empty(&cpu_mask)) {
1656 			xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1657 			if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1658 				cpumask_copy(mm_cpumask(mm),
1659 					     cpumask_of(smp_processor_id()));
1660 		}
1661 		local_ops->tlb_mm(mm);
1662 	}
1663 }
1664 
1665 static void smp_flush_cache_range(struct vm_area_struct *vma,
1666 				  unsigned long start,
1667 				  unsigned long end)
1668 {
1669 	struct mm_struct *mm = vma->vm_mm;
1670 
1671 	if (mm->context != NO_CONTEXT) {
1672 		cpumask_t cpu_mask;
1673 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1674 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1675 		if (!cpumask_empty(&cpu_mask))
1676 			xc3((smpfunc_t) local_ops->cache_range,
1677 			    (unsigned long) vma, start, end);
1678 		local_ops->cache_range(vma, start, end);
1679 	}
1680 }
1681 
1682 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1683 				unsigned long start,
1684 				unsigned long end)
1685 {
1686 	struct mm_struct *mm = vma->vm_mm;
1687 
1688 	if (mm->context != NO_CONTEXT) {
1689 		cpumask_t cpu_mask;
1690 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1691 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1692 		if (!cpumask_empty(&cpu_mask))
1693 			xc3((smpfunc_t) local_ops->tlb_range,
1694 			    (unsigned long) vma, start, end);
1695 		local_ops->tlb_range(vma, start, end);
1696 	}
1697 }
1698 
1699 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1700 {
1701 	struct mm_struct *mm = vma->vm_mm;
1702 
1703 	if (mm->context != NO_CONTEXT) {
1704 		cpumask_t cpu_mask;
1705 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1706 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1707 		if (!cpumask_empty(&cpu_mask))
1708 			xc2((smpfunc_t) local_ops->cache_page,
1709 			    (unsigned long) vma, page);
1710 		local_ops->cache_page(vma, page);
1711 	}
1712 }
1713 
1714 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1715 {
1716 	struct mm_struct *mm = vma->vm_mm;
1717 
1718 	if (mm->context != NO_CONTEXT) {
1719 		cpumask_t cpu_mask;
1720 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1721 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1722 		if (!cpumask_empty(&cpu_mask))
1723 			xc2((smpfunc_t) local_ops->tlb_page,
1724 			    (unsigned long) vma, page);
1725 		local_ops->tlb_page(vma, page);
1726 	}
1727 }
1728 
1729 static void smp_flush_page_to_ram(unsigned long page)
1730 {
1731 	/* Current theory is that those who call this are the one's
1732 	 * who have just dirtied their cache with the pages contents
1733 	 * in kernel space, therefore we only run this on local cpu.
1734 	 *
1735 	 * XXX This experiment failed, research further... -DaveM
1736 	 */
1737 #if 1
1738 	xc1((smpfunc_t) local_ops->page_to_ram, page);
1739 #endif
1740 	local_ops->page_to_ram(page);
1741 }
1742 
1743 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1744 {
1745 	cpumask_t cpu_mask;
1746 	cpumask_copy(&cpu_mask, mm_cpumask(mm));
1747 	cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1748 	if (!cpumask_empty(&cpu_mask))
1749 		xc2((smpfunc_t) local_ops->sig_insns,
1750 		    (unsigned long) mm, insn_addr);
1751 	local_ops->sig_insns(mm, insn_addr);
1752 }
1753 
1754 static struct sparc32_cachetlb_ops smp_cachetlb_ops = {
1755 	.cache_all	= smp_flush_cache_all,
1756 	.cache_mm	= smp_flush_cache_mm,
1757 	.cache_page	= smp_flush_cache_page,
1758 	.cache_range	= smp_flush_cache_range,
1759 	.tlb_all	= smp_flush_tlb_all,
1760 	.tlb_mm		= smp_flush_tlb_mm,
1761 	.tlb_page	= smp_flush_tlb_page,
1762 	.tlb_range	= smp_flush_tlb_range,
1763 	.page_to_ram	= smp_flush_page_to_ram,
1764 	.sig_insns	= smp_flush_sig_insns,
1765 	.page_for_dma	= smp_flush_page_for_dma,
1766 };
1767 #endif
1768 
1769 /* Load up routines and constants for sun4m and sun4d mmu */
1770 void __init load_mmu(void)
1771 {
1772 	/* Functions */
1773 	get_srmmu_type();
1774 
1775 #ifdef CONFIG_SMP
1776 	/* El switcheroo... */
1777 	local_ops = sparc32_cachetlb_ops;
1778 
1779 	if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1780 		smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1781 		smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1782 		smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1783 		smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1784 	}
1785 
1786 	if (poke_srmmu == poke_viking) {
1787 		/* Avoid unnecessary cross calls. */
1788 		smp_cachetlb_ops.cache_all = local_ops->cache_all;
1789 		smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1790 		smp_cachetlb_ops.cache_range = local_ops->cache_range;
1791 		smp_cachetlb_ops.cache_page = local_ops->cache_page;
1792 
1793 		smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1794 		smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1795 		smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1796 	}
1797 
1798 	/* It really is const after this point. */
1799 	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1800 		&smp_cachetlb_ops;
1801 #endif
1802 
1803 	if (sparc_cpu_model == sun4d)
1804 		ld_mmu_iounit();
1805 	else
1806 		ld_mmu_iommu();
1807 #ifdef CONFIG_SMP
1808 	if (sparc_cpu_model == sun4d)
1809 		sun4d_init_smp();
1810 	else if (sparc_cpu_model == sparc_leon)
1811 		leon_init_smp();
1812 	else
1813 		sun4m_init_smp();
1814 #endif
1815 }
1816