xref: /openbmc/linux/arch/sparc/mm/srmmu.c (revision f7d84fa7)
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 EXPORT_SYMBOL(vac_cache_size);
58 int vac_line_size;
59 
60 extern struct resource sparc_iomap;
61 
62 extern unsigned long last_valid_pfn;
63 
64 static pgd_t *srmmu_swapper_pg_dir;
65 
66 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
67 EXPORT_SYMBOL(sparc32_cachetlb_ops);
68 
69 #ifdef CONFIG_SMP
70 const struct sparc32_cachetlb_ops *local_ops;
71 
72 #define FLUSH_BEGIN(mm)
73 #define FLUSH_END
74 #else
75 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
76 #define FLUSH_END	}
77 #endif
78 
79 int flush_page_for_dma_global = 1;
80 
81 char *srmmu_name;
82 
83 ctxd_t *srmmu_ctx_table_phys;
84 static ctxd_t *srmmu_context_table;
85 
86 int viking_mxcc_present;
87 static DEFINE_SPINLOCK(srmmu_context_spinlock);
88 
89 static int is_hypersparc;
90 
91 static int srmmu_cache_pagetables;
92 
93 /* these will be initialized in srmmu_nocache_calcsize() */
94 static unsigned long srmmu_nocache_size;
95 static unsigned long srmmu_nocache_end;
96 
97 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
98 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
99 
100 /* The context table is a nocache user with the biggest alignment needs. */
101 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
102 
103 void *srmmu_nocache_pool;
104 static struct bit_map srmmu_nocache_map;
105 
106 static inline int srmmu_pmd_none(pmd_t pmd)
107 { return !(pmd_val(pmd) & 0xFFFFFFF); }
108 
109 /* XXX should we hyper_flush_whole_icache here - Anton */
110 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
111 {
112 	pte_t pte;
113 
114 	pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4)));
115 	set_pte((pte_t *)ctxp, pte);
116 }
117 
118 void pmd_set(pmd_t *pmdp, pte_t *ptep)
119 {
120 	unsigned long ptp;	/* Physical address, shifted right by 4 */
121 	int i;
122 
123 	ptp = __nocache_pa(ptep) >> 4;
124 	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
125 		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
126 		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
127 	}
128 }
129 
130 void pmd_populate(struct mm_struct *mm, pmd_t *pmdp, struct page *ptep)
131 {
132 	unsigned long ptp;	/* Physical address, shifted right by 4 */
133 	int i;
134 
135 	ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);	/* watch for overflow */
136 	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
137 		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
138 		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
139 	}
140 }
141 
142 /* Find an entry in the third-level page table.. */
143 pte_t *pte_offset_kernel(pmd_t *dir, unsigned long address)
144 {
145 	void *pte;
146 
147 	pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
148 	return (pte_t *) pte +
149 	    ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
150 }
151 
152 /*
153  * size: bytes to allocate in the nocache area.
154  * align: bytes, number to align at.
155  * Returns the virtual address of the allocated area.
156  */
157 static void *__srmmu_get_nocache(int size, int align)
158 {
159 	int offset;
160 	unsigned long addr;
161 
162 	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
163 		printk(KERN_ERR "Size 0x%x too small for nocache request\n",
164 		       size);
165 		size = SRMMU_NOCACHE_BITMAP_SHIFT;
166 	}
167 	if (size & (SRMMU_NOCACHE_BITMAP_SHIFT - 1)) {
168 		printk(KERN_ERR "Size 0x%x unaligned int nocache request\n",
169 		       size);
170 		size += SRMMU_NOCACHE_BITMAP_SHIFT - 1;
171 	}
172 	BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
173 
174 	offset = bit_map_string_get(&srmmu_nocache_map,
175 				    size >> SRMMU_NOCACHE_BITMAP_SHIFT,
176 				    align >> SRMMU_NOCACHE_BITMAP_SHIFT);
177 	if (offset == -1) {
178 		printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
179 		       size, (int) srmmu_nocache_size,
180 		       srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
181 		return NULL;
182 	}
183 
184 	addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
185 	return (void *)addr;
186 }
187 
188 void *srmmu_get_nocache(int size, int align)
189 {
190 	void *tmp;
191 
192 	tmp = __srmmu_get_nocache(size, align);
193 
194 	if (tmp)
195 		memset(tmp, 0, size);
196 
197 	return tmp;
198 }
199 
200 void srmmu_free_nocache(void *addr, int size)
201 {
202 	unsigned long vaddr;
203 	int offset;
204 
205 	vaddr = (unsigned long)addr;
206 	if (vaddr < SRMMU_NOCACHE_VADDR) {
207 		printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
208 		    vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
209 		BUG();
210 	}
211 	if (vaddr + size > srmmu_nocache_end) {
212 		printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
213 		    vaddr, srmmu_nocache_end);
214 		BUG();
215 	}
216 	if (!is_power_of_2(size)) {
217 		printk("Size 0x%x is not a power of 2\n", size);
218 		BUG();
219 	}
220 	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
221 		printk("Size 0x%x is too small\n", size);
222 		BUG();
223 	}
224 	if (vaddr & (size - 1)) {
225 		printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
226 		BUG();
227 	}
228 
229 	offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
230 	size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
231 
232 	bit_map_clear(&srmmu_nocache_map, offset, size);
233 }
234 
235 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
236 						 unsigned long end);
237 
238 /* Return how much physical memory we have.  */
239 static unsigned long __init probe_memory(void)
240 {
241 	unsigned long total = 0;
242 	int i;
243 
244 	for (i = 0; sp_banks[i].num_bytes; i++)
245 		total += sp_banks[i].num_bytes;
246 
247 	return total;
248 }
249 
250 /*
251  * Reserve nocache dynamically proportionally to the amount of
252  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
253  */
254 static void __init srmmu_nocache_calcsize(void)
255 {
256 	unsigned long sysmemavail = probe_memory() / 1024;
257 	int srmmu_nocache_npages;
258 
259 	srmmu_nocache_npages =
260 		sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
261 
262  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
263 	// if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
264 	if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
265 		srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
266 
267 	/* anything above 1280 blows up */
268 	if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
269 		srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
270 
271 	srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
272 	srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
273 }
274 
275 static void __init srmmu_nocache_init(void)
276 {
277 	void *srmmu_nocache_bitmap;
278 	unsigned int bitmap_bits;
279 	pgd_t *pgd;
280 	pmd_t *pmd;
281 	pte_t *pte;
282 	unsigned long paddr, vaddr;
283 	unsigned long pteval;
284 
285 	bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
286 
287 	srmmu_nocache_pool = __alloc_bootmem(srmmu_nocache_size,
288 		SRMMU_NOCACHE_ALIGN_MAX, 0UL);
289 	memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
290 
291 	srmmu_nocache_bitmap =
292 		__alloc_bootmem(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
293 				SMP_CACHE_BYTES, 0UL);
294 	bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
295 
296 	srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
297 	memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
298 	init_mm.pgd = srmmu_swapper_pg_dir;
299 
300 	srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
301 
302 	paddr = __pa((unsigned long)srmmu_nocache_pool);
303 	vaddr = SRMMU_NOCACHE_VADDR;
304 
305 	while (vaddr < srmmu_nocache_end) {
306 		pgd = pgd_offset_k(vaddr);
307 		pmd = pmd_offset(__nocache_fix(pgd), vaddr);
308 		pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
309 
310 		pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
311 
312 		if (srmmu_cache_pagetables)
313 			pteval |= SRMMU_CACHE;
314 
315 		set_pte(__nocache_fix(pte), __pte(pteval));
316 
317 		vaddr += PAGE_SIZE;
318 		paddr += PAGE_SIZE;
319 	}
320 
321 	flush_cache_all();
322 	flush_tlb_all();
323 }
324 
325 pgd_t *get_pgd_fast(void)
326 {
327 	pgd_t *pgd = NULL;
328 
329 	pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
330 	if (pgd) {
331 		pgd_t *init = pgd_offset_k(0);
332 		memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
333 		memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
334 						(PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
335 	}
336 
337 	return pgd;
338 }
339 
340 /*
341  * Hardware needs alignment to 256 only, but we align to whole page size
342  * to reduce fragmentation problems due to the buddy principle.
343  * XXX Provide actual fragmentation statistics in /proc.
344  *
345  * Alignments up to the page size are the same for physical and virtual
346  * addresses of the nocache area.
347  */
348 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
349 {
350 	unsigned long pte;
351 	struct page *page;
352 
353 	if ((pte = (unsigned long)pte_alloc_one_kernel(mm, address)) == 0)
354 		return NULL;
355 	page = pfn_to_page(__nocache_pa(pte) >> PAGE_SHIFT);
356 	if (!pgtable_page_ctor(page)) {
357 		__free_page(page);
358 		return NULL;
359 	}
360 	return page;
361 }
362 
363 void pte_free(struct mm_struct *mm, pgtable_t pte)
364 {
365 	unsigned long p;
366 
367 	pgtable_page_dtor(pte);
368 	p = (unsigned long)page_address(pte);	/* Cached address (for test) */
369 	if (p == 0)
370 		BUG();
371 	p = page_to_pfn(pte) << PAGE_SHIFT;	/* Physical address */
372 
373 	/* free non cached virtual address*/
374 	srmmu_free_nocache(__nocache_va(p), PTE_SIZE);
375 }
376 
377 /* context handling - a dynamically sized pool is used */
378 #define NO_CONTEXT	-1
379 
380 struct ctx_list {
381 	struct ctx_list *next;
382 	struct ctx_list *prev;
383 	unsigned int ctx_number;
384 	struct mm_struct *ctx_mm;
385 };
386 
387 static struct ctx_list *ctx_list_pool;
388 static struct ctx_list ctx_free;
389 static struct ctx_list ctx_used;
390 
391 /* At boot time we determine the number of contexts */
392 static int num_contexts;
393 
394 static inline void remove_from_ctx_list(struct ctx_list *entry)
395 {
396 	entry->next->prev = entry->prev;
397 	entry->prev->next = entry->next;
398 }
399 
400 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
401 {
402 	entry->next = head;
403 	(entry->prev = head->prev)->next = entry;
404 	head->prev = entry;
405 }
406 #define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
407 #define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
408 
409 
410 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
411 {
412 	struct ctx_list *ctxp;
413 
414 	ctxp = ctx_free.next;
415 	if (ctxp != &ctx_free) {
416 		remove_from_ctx_list(ctxp);
417 		add_to_used_ctxlist(ctxp);
418 		mm->context = ctxp->ctx_number;
419 		ctxp->ctx_mm = mm;
420 		return;
421 	}
422 	ctxp = ctx_used.next;
423 	if (ctxp->ctx_mm == old_mm)
424 		ctxp = ctxp->next;
425 	if (ctxp == &ctx_used)
426 		panic("out of mmu contexts");
427 	flush_cache_mm(ctxp->ctx_mm);
428 	flush_tlb_mm(ctxp->ctx_mm);
429 	remove_from_ctx_list(ctxp);
430 	add_to_used_ctxlist(ctxp);
431 	ctxp->ctx_mm->context = NO_CONTEXT;
432 	ctxp->ctx_mm = mm;
433 	mm->context = ctxp->ctx_number;
434 }
435 
436 static inline void free_context(int context)
437 {
438 	struct ctx_list *ctx_old;
439 
440 	ctx_old = ctx_list_pool + context;
441 	remove_from_ctx_list(ctx_old);
442 	add_to_free_ctxlist(ctx_old);
443 }
444 
445 static void __init sparc_context_init(int numctx)
446 {
447 	int ctx;
448 	unsigned long size;
449 
450 	size = numctx * sizeof(struct ctx_list);
451 	ctx_list_pool = __alloc_bootmem(size, SMP_CACHE_BYTES, 0UL);
452 
453 	for (ctx = 0; ctx < numctx; ctx++) {
454 		struct ctx_list *clist;
455 
456 		clist = (ctx_list_pool + ctx);
457 		clist->ctx_number = ctx;
458 		clist->ctx_mm = NULL;
459 	}
460 	ctx_free.next = ctx_free.prev = &ctx_free;
461 	ctx_used.next = ctx_used.prev = &ctx_used;
462 	for (ctx = 0; ctx < numctx; ctx++)
463 		add_to_free_ctxlist(ctx_list_pool + ctx);
464 }
465 
466 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
467 	       struct task_struct *tsk)
468 {
469 	unsigned long flags;
470 
471 	if (mm->context == NO_CONTEXT) {
472 		spin_lock_irqsave(&srmmu_context_spinlock, flags);
473 		alloc_context(old_mm, mm);
474 		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
475 		srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
476 	}
477 
478 	if (sparc_cpu_model == sparc_leon)
479 		leon_switch_mm();
480 
481 	if (is_hypersparc)
482 		hyper_flush_whole_icache();
483 
484 	srmmu_set_context(mm->context);
485 }
486 
487 /* Low level IO area allocation on the SRMMU. */
488 static inline void srmmu_mapioaddr(unsigned long physaddr,
489 				   unsigned long virt_addr, int bus_type)
490 {
491 	pgd_t *pgdp;
492 	pmd_t *pmdp;
493 	pte_t *ptep;
494 	unsigned long tmp;
495 
496 	physaddr &= PAGE_MASK;
497 	pgdp = pgd_offset_k(virt_addr);
498 	pmdp = pmd_offset(pgdp, virt_addr);
499 	ptep = pte_offset_kernel(pmdp, virt_addr);
500 	tmp = (physaddr >> 4) | SRMMU_ET_PTE;
501 
502 	/* I need to test whether this is consistent over all
503 	 * sun4m's.  The bus_type represents the upper 4 bits of
504 	 * 36-bit physical address on the I/O space lines...
505 	 */
506 	tmp |= (bus_type << 28);
507 	tmp |= SRMMU_PRIV;
508 	__flush_page_to_ram(virt_addr);
509 	set_pte(ptep, __pte(tmp));
510 }
511 
512 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
513 		      unsigned long xva, unsigned int len)
514 {
515 	while (len != 0) {
516 		len -= PAGE_SIZE;
517 		srmmu_mapioaddr(xpa, xva, bus);
518 		xva += PAGE_SIZE;
519 		xpa += PAGE_SIZE;
520 	}
521 	flush_tlb_all();
522 }
523 
524 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
525 {
526 	pgd_t *pgdp;
527 	pmd_t *pmdp;
528 	pte_t *ptep;
529 
530 	pgdp = pgd_offset_k(virt_addr);
531 	pmdp = pmd_offset(pgdp, virt_addr);
532 	ptep = pte_offset_kernel(pmdp, virt_addr);
533 
534 	/* No need to flush uncacheable page. */
535 	__pte_clear(ptep);
536 }
537 
538 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
539 {
540 	while (len != 0) {
541 		len -= PAGE_SIZE;
542 		srmmu_unmapioaddr(virt_addr);
543 		virt_addr += PAGE_SIZE;
544 	}
545 	flush_tlb_all();
546 }
547 
548 /* tsunami.S */
549 extern void tsunami_flush_cache_all(void);
550 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
551 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
552 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
553 extern void tsunami_flush_page_to_ram(unsigned long page);
554 extern void tsunami_flush_page_for_dma(unsigned long page);
555 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
556 extern void tsunami_flush_tlb_all(void);
557 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
558 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
559 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
560 extern void tsunami_setup_blockops(void);
561 
562 /* swift.S */
563 extern void swift_flush_cache_all(void);
564 extern void swift_flush_cache_mm(struct mm_struct *mm);
565 extern void swift_flush_cache_range(struct vm_area_struct *vma,
566 				    unsigned long start, unsigned long end);
567 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
568 extern void swift_flush_page_to_ram(unsigned long page);
569 extern void swift_flush_page_for_dma(unsigned long page);
570 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
571 extern void swift_flush_tlb_all(void);
572 extern void swift_flush_tlb_mm(struct mm_struct *mm);
573 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
574 				  unsigned long start, unsigned long end);
575 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
576 
577 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
578 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
579 {
580 	int cctx, ctx1;
581 
582 	page &= PAGE_MASK;
583 	if ((ctx1 = vma->vm_mm->context) != -1) {
584 		cctx = srmmu_get_context();
585 /* Is context # ever different from current context? P3 */
586 		if (cctx != ctx1) {
587 			printk("flush ctx %02x curr %02x\n", ctx1, cctx);
588 			srmmu_set_context(ctx1);
589 			swift_flush_page(page);
590 			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
591 					"r" (page), "i" (ASI_M_FLUSH_PROBE));
592 			srmmu_set_context(cctx);
593 		} else {
594 			 /* Rm. prot. bits from virt. c. */
595 			/* swift_flush_cache_all(); */
596 			/* swift_flush_cache_page(vma, page); */
597 			swift_flush_page(page);
598 
599 			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
600 				"r" (page), "i" (ASI_M_FLUSH_PROBE));
601 			/* same as above: srmmu_flush_tlb_page() */
602 		}
603 	}
604 }
605 #endif
606 
607 /*
608  * The following are all MBUS based SRMMU modules, and therefore could
609  * be found in a multiprocessor configuration.  On the whole, these
610  * chips seems to be much more touchy about DVMA and page tables
611  * with respect to cache coherency.
612  */
613 
614 /* viking.S */
615 extern void viking_flush_cache_all(void);
616 extern void viking_flush_cache_mm(struct mm_struct *mm);
617 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
618 				     unsigned long end);
619 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
620 extern void viking_flush_page_to_ram(unsigned long page);
621 extern void viking_flush_page_for_dma(unsigned long page);
622 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
623 extern void viking_flush_page(unsigned long page);
624 extern void viking_mxcc_flush_page(unsigned long page);
625 extern void viking_flush_tlb_all(void);
626 extern void viking_flush_tlb_mm(struct mm_struct *mm);
627 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
628 				   unsigned long end);
629 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
630 				  unsigned long page);
631 extern void sun4dsmp_flush_tlb_all(void);
632 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
633 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
634 				   unsigned long end);
635 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
636 				  unsigned long page);
637 
638 /* hypersparc.S */
639 extern void hypersparc_flush_cache_all(void);
640 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
641 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
642 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
643 extern void hypersparc_flush_page_to_ram(unsigned long page);
644 extern void hypersparc_flush_page_for_dma(unsigned long page);
645 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
646 extern void hypersparc_flush_tlb_all(void);
647 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
648 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
649 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
650 extern void hypersparc_setup_blockops(void);
651 
652 /*
653  * NOTE: All of this startup code assumes the low 16mb (approx.) of
654  *       kernel mappings are done with one single contiguous chunk of
655  *       ram.  On small ram machines (classics mainly) we only get
656  *       around 8mb mapped for us.
657  */
658 
659 static void __init early_pgtable_allocfail(char *type)
660 {
661 	prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
662 	prom_halt();
663 }
664 
665 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
666 							unsigned long end)
667 {
668 	pgd_t *pgdp;
669 	pmd_t *pmdp;
670 	pte_t *ptep;
671 
672 	while (start < end) {
673 		pgdp = pgd_offset_k(start);
674 		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
675 			pmdp = __srmmu_get_nocache(
676 			    SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
677 			if (pmdp == NULL)
678 				early_pgtable_allocfail("pmd");
679 			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
680 			pgd_set(__nocache_fix(pgdp), pmdp);
681 		}
682 		pmdp = pmd_offset(__nocache_fix(pgdp), start);
683 		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
684 			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
685 			if (ptep == NULL)
686 				early_pgtable_allocfail("pte");
687 			memset(__nocache_fix(ptep), 0, PTE_SIZE);
688 			pmd_set(__nocache_fix(pmdp), ptep);
689 		}
690 		if (start > (0xffffffffUL - PMD_SIZE))
691 			break;
692 		start = (start + PMD_SIZE) & PMD_MASK;
693 	}
694 }
695 
696 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
697 						  unsigned long end)
698 {
699 	pgd_t *pgdp;
700 	pmd_t *pmdp;
701 	pte_t *ptep;
702 
703 	while (start < end) {
704 		pgdp = pgd_offset_k(start);
705 		if (pgd_none(*pgdp)) {
706 			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
707 			if (pmdp == NULL)
708 				early_pgtable_allocfail("pmd");
709 			memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
710 			pgd_set(pgdp, pmdp);
711 		}
712 		pmdp = pmd_offset(pgdp, start);
713 		if (srmmu_pmd_none(*pmdp)) {
714 			ptep = __srmmu_get_nocache(PTE_SIZE,
715 							     PTE_SIZE);
716 			if (ptep == NULL)
717 				early_pgtable_allocfail("pte");
718 			memset(ptep, 0, PTE_SIZE);
719 			pmd_set(pmdp, ptep);
720 		}
721 		if (start > (0xffffffffUL - PMD_SIZE))
722 			break;
723 		start = (start + PMD_SIZE) & PMD_MASK;
724 	}
725 }
726 
727 /* These flush types are not available on all chips... */
728 static inline unsigned long srmmu_probe(unsigned long vaddr)
729 {
730 	unsigned long retval;
731 
732 	if (sparc_cpu_model != sparc_leon) {
733 
734 		vaddr &= PAGE_MASK;
735 		__asm__ __volatile__("lda [%1] %2, %0\n\t" :
736 				     "=r" (retval) :
737 				     "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
738 	} else {
739 		retval = leon_swprobe(vaddr, NULL);
740 	}
741 	return retval;
742 }
743 
744 /*
745  * This is much cleaner than poking around physical address space
746  * looking at the prom's page table directly which is what most
747  * other OS's do.  Yuck... this is much better.
748  */
749 static void __init srmmu_inherit_prom_mappings(unsigned long start,
750 					       unsigned long end)
751 {
752 	unsigned long probed;
753 	unsigned long addr;
754 	pgd_t *pgdp;
755 	pmd_t *pmdp;
756 	pte_t *ptep;
757 	int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
758 
759 	while (start <= end) {
760 		if (start == 0)
761 			break; /* probably wrap around */
762 		if (start == 0xfef00000)
763 			start = KADB_DEBUGGER_BEGVM;
764 		probed = srmmu_probe(start);
765 		if (!probed) {
766 			/* continue probing until we find an entry */
767 			start += PAGE_SIZE;
768 			continue;
769 		}
770 
771 		/* A red snapper, see what it really is. */
772 		what = 0;
773 		addr = start - PAGE_SIZE;
774 
775 		if (!(start & ~(SRMMU_REAL_PMD_MASK))) {
776 			if (srmmu_probe(addr + SRMMU_REAL_PMD_SIZE) == probed)
777 				what = 1;
778 		}
779 
780 		if (!(start & ~(SRMMU_PGDIR_MASK))) {
781 			if (srmmu_probe(addr + SRMMU_PGDIR_SIZE) == probed)
782 				what = 2;
783 		}
784 
785 		pgdp = pgd_offset_k(start);
786 		if (what == 2) {
787 			*(pgd_t *)__nocache_fix(pgdp) = __pgd(probed);
788 			start += SRMMU_PGDIR_SIZE;
789 			continue;
790 		}
791 		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
792 			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
793 						   SRMMU_PMD_TABLE_SIZE);
794 			if (pmdp == NULL)
795 				early_pgtable_allocfail("pmd");
796 			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
797 			pgd_set(__nocache_fix(pgdp), pmdp);
798 		}
799 		pmdp = pmd_offset(__nocache_fix(pgdp), start);
800 		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
801 			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
802 			if (ptep == NULL)
803 				early_pgtable_allocfail("pte");
804 			memset(__nocache_fix(ptep), 0, PTE_SIZE);
805 			pmd_set(__nocache_fix(pmdp), ptep);
806 		}
807 		if (what == 1) {
808 			/* We bend the rule where all 16 PTPs in a pmd_t point
809 			 * inside the same PTE page, and we leak a perfectly
810 			 * good hardware PTE piece. Alternatives seem worse.
811 			 */
812 			unsigned int x;	/* Index of HW PMD in soft cluster */
813 			unsigned long *val;
814 			x = (start >> PMD_SHIFT) & 15;
815 			val = &pmdp->pmdv[x];
816 			*(unsigned long *)__nocache_fix(val) = probed;
817 			start += SRMMU_REAL_PMD_SIZE;
818 			continue;
819 		}
820 		ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
821 		*(pte_t *)__nocache_fix(ptep) = __pte(probed);
822 		start += PAGE_SIZE;
823 	}
824 }
825 
826 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
827 
828 /* Create a third-level SRMMU 16MB page mapping. */
829 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
830 {
831 	pgd_t *pgdp = pgd_offset_k(vaddr);
832 	unsigned long big_pte;
833 
834 	big_pte = KERNEL_PTE(phys_base >> 4);
835 	*(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
836 }
837 
838 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
839 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
840 {
841 	unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
842 	unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
843 	unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
844 	/* Map "low" memory only */
845 	const unsigned long min_vaddr = PAGE_OFFSET;
846 	const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
847 
848 	if (vstart < min_vaddr || vstart >= max_vaddr)
849 		return vstart;
850 
851 	if (vend > max_vaddr || vend < min_vaddr)
852 		vend = max_vaddr;
853 
854 	while (vstart < vend) {
855 		do_large_mapping(vstart, pstart);
856 		vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
857 	}
858 	return vstart;
859 }
860 
861 static void __init map_kernel(void)
862 {
863 	int i;
864 
865 	if (phys_base > 0) {
866 		do_large_mapping(PAGE_OFFSET, phys_base);
867 	}
868 
869 	for (i = 0; sp_banks[i].num_bytes != 0; i++) {
870 		map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
871 	}
872 }
873 
874 void (*poke_srmmu)(void) = NULL;
875 
876 void __init srmmu_paging_init(void)
877 {
878 	int i;
879 	phandle cpunode;
880 	char node_str[128];
881 	pgd_t *pgd;
882 	pmd_t *pmd;
883 	pte_t *pte;
884 	unsigned long pages_avail;
885 
886 	init_mm.context = (unsigned long) NO_CONTEXT;
887 	sparc_iomap.start = SUN4M_IOBASE_VADDR;	/* 16MB of IOSPACE on all sun4m's. */
888 
889 	if (sparc_cpu_model == sun4d)
890 		num_contexts = 65536; /* We know it is Viking */
891 	else {
892 		/* Find the number of contexts on the srmmu. */
893 		cpunode = prom_getchild(prom_root_node);
894 		num_contexts = 0;
895 		while (cpunode != 0) {
896 			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
897 			if (!strcmp(node_str, "cpu")) {
898 				num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
899 				break;
900 			}
901 			cpunode = prom_getsibling(cpunode);
902 		}
903 	}
904 
905 	if (!num_contexts) {
906 		prom_printf("Something wrong, can't find cpu node in paging_init.\n");
907 		prom_halt();
908 	}
909 
910 	pages_avail = 0;
911 	last_valid_pfn = bootmem_init(&pages_avail);
912 
913 	srmmu_nocache_calcsize();
914 	srmmu_nocache_init();
915 	srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
916 	map_kernel();
917 
918 	/* ctx table has to be physically aligned to its size */
919 	srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
920 	srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
921 
922 	for (i = 0; i < num_contexts; i++)
923 		srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
924 
925 	flush_cache_all();
926 	srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
927 #ifdef CONFIG_SMP
928 	/* Stop from hanging here... */
929 	local_ops->tlb_all();
930 #else
931 	flush_tlb_all();
932 #endif
933 	poke_srmmu();
934 
935 	srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
936 	srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
937 
938 	srmmu_allocate_ptable_skeleton(
939 		__fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
940 	srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
941 
942 	pgd = pgd_offset_k(PKMAP_BASE);
943 	pmd = pmd_offset(pgd, PKMAP_BASE);
944 	pte = pte_offset_kernel(pmd, PKMAP_BASE);
945 	pkmap_page_table = pte;
946 
947 	flush_cache_all();
948 	flush_tlb_all();
949 
950 	sparc_context_init(num_contexts);
951 
952 	kmap_init();
953 
954 	{
955 		unsigned long zones_size[MAX_NR_ZONES];
956 		unsigned long zholes_size[MAX_NR_ZONES];
957 		unsigned long npages;
958 		int znum;
959 
960 		for (znum = 0; znum < MAX_NR_ZONES; znum++)
961 			zones_size[znum] = zholes_size[znum] = 0;
962 
963 		npages = max_low_pfn - pfn_base;
964 
965 		zones_size[ZONE_DMA] = npages;
966 		zholes_size[ZONE_DMA] = npages - pages_avail;
967 
968 		npages = highend_pfn - max_low_pfn;
969 		zones_size[ZONE_HIGHMEM] = npages;
970 		zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
971 
972 		free_area_init_node(0, zones_size, pfn_base, zholes_size);
973 	}
974 }
975 
976 void mmu_info(struct seq_file *m)
977 {
978 	seq_printf(m,
979 		   "MMU type\t: %s\n"
980 		   "contexts\t: %d\n"
981 		   "nocache total\t: %ld\n"
982 		   "nocache used\t: %d\n",
983 		   srmmu_name,
984 		   num_contexts,
985 		   srmmu_nocache_size,
986 		   srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
987 }
988 
989 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
990 {
991 	mm->context = NO_CONTEXT;
992 	return 0;
993 }
994 
995 void destroy_context(struct mm_struct *mm)
996 {
997 	unsigned long flags;
998 
999 	if (mm->context != NO_CONTEXT) {
1000 		flush_cache_mm(mm);
1001 		srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1002 		flush_tlb_mm(mm);
1003 		spin_lock_irqsave(&srmmu_context_spinlock, flags);
1004 		free_context(mm->context);
1005 		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1006 		mm->context = NO_CONTEXT;
1007 	}
1008 }
1009 
1010 /* Init various srmmu chip types. */
1011 static void __init srmmu_is_bad(void)
1012 {
1013 	prom_printf("Could not determine SRMMU chip type.\n");
1014 	prom_halt();
1015 }
1016 
1017 static void __init init_vac_layout(void)
1018 {
1019 	phandle nd;
1020 	int cache_lines;
1021 	char node_str[128];
1022 #ifdef CONFIG_SMP
1023 	int cpu = 0;
1024 	unsigned long max_size = 0;
1025 	unsigned long min_line_size = 0x10000000;
1026 #endif
1027 
1028 	nd = prom_getchild(prom_root_node);
1029 	while ((nd = prom_getsibling(nd)) != 0) {
1030 		prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1031 		if (!strcmp(node_str, "cpu")) {
1032 			vac_line_size = prom_getint(nd, "cache-line-size");
1033 			if (vac_line_size == -1) {
1034 				prom_printf("can't determine cache-line-size, halting.\n");
1035 				prom_halt();
1036 			}
1037 			cache_lines = prom_getint(nd, "cache-nlines");
1038 			if (cache_lines == -1) {
1039 				prom_printf("can't determine cache-nlines, halting.\n");
1040 				prom_halt();
1041 			}
1042 
1043 			vac_cache_size = cache_lines * vac_line_size;
1044 #ifdef CONFIG_SMP
1045 			if (vac_cache_size > max_size)
1046 				max_size = vac_cache_size;
1047 			if (vac_line_size < min_line_size)
1048 				min_line_size = vac_line_size;
1049 			//FIXME: cpus not contiguous!!
1050 			cpu++;
1051 			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1052 				break;
1053 #else
1054 			break;
1055 #endif
1056 		}
1057 	}
1058 	if (nd == 0) {
1059 		prom_printf("No CPU nodes found, halting.\n");
1060 		prom_halt();
1061 	}
1062 #ifdef CONFIG_SMP
1063 	vac_cache_size = max_size;
1064 	vac_line_size = min_line_size;
1065 #endif
1066 	printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1067 	       (int)vac_cache_size, (int)vac_line_size);
1068 }
1069 
1070 static void poke_hypersparc(void)
1071 {
1072 	volatile unsigned long clear;
1073 	unsigned long mreg = srmmu_get_mmureg();
1074 
1075 	hyper_flush_unconditional_combined();
1076 
1077 	mreg &= ~(HYPERSPARC_CWENABLE);
1078 	mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1079 	mreg |= (HYPERSPARC_CMODE);
1080 
1081 	srmmu_set_mmureg(mreg);
1082 
1083 #if 0 /* XXX I think this is bad news... -DaveM */
1084 	hyper_clear_all_tags();
1085 #endif
1086 
1087 	put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1088 	hyper_flush_whole_icache();
1089 	clear = srmmu_get_faddr();
1090 	clear = srmmu_get_fstatus();
1091 }
1092 
1093 static const struct sparc32_cachetlb_ops hypersparc_ops = {
1094 	.cache_all	= hypersparc_flush_cache_all,
1095 	.cache_mm	= hypersparc_flush_cache_mm,
1096 	.cache_page	= hypersparc_flush_cache_page,
1097 	.cache_range	= hypersparc_flush_cache_range,
1098 	.tlb_all	= hypersparc_flush_tlb_all,
1099 	.tlb_mm		= hypersparc_flush_tlb_mm,
1100 	.tlb_page	= hypersparc_flush_tlb_page,
1101 	.tlb_range	= hypersparc_flush_tlb_range,
1102 	.page_to_ram	= hypersparc_flush_page_to_ram,
1103 	.sig_insns	= hypersparc_flush_sig_insns,
1104 	.page_for_dma	= hypersparc_flush_page_for_dma,
1105 };
1106 
1107 static void __init init_hypersparc(void)
1108 {
1109 	srmmu_name = "ROSS HyperSparc";
1110 	srmmu_modtype = HyperSparc;
1111 
1112 	init_vac_layout();
1113 
1114 	is_hypersparc = 1;
1115 	sparc32_cachetlb_ops = &hypersparc_ops;
1116 
1117 	poke_srmmu = poke_hypersparc;
1118 
1119 	hypersparc_setup_blockops();
1120 }
1121 
1122 static void poke_swift(void)
1123 {
1124 	unsigned long mreg;
1125 
1126 	/* Clear any crap from the cache or else... */
1127 	swift_flush_cache_all();
1128 
1129 	/* Enable I & D caches */
1130 	mreg = srmmu_get_mmureg();
1131 	mreg |= (SWIFT_IE | SWIFT_DE);
1132 	/*
1133 	 * The Swift branch folding logic is completely broken.  At
1134 	 * trap time, if things are just right, if can mistakenly
1135 	 * think that a trap is coming from kernel mode when in fact
1136 	 * it is coming from user mode (it mis-executes the branch in
1137 	 * the trap code).  So you see things like crashme completely
1138 	 * hosing your machine which is completely unacceptable.  Turn
1139 	 * this shit off... nice job Fujitsu.
1140 	 */
1141 	mreg &= ~(SWIFT_BF);
1142 	srmmu_set_mmureg(mreg);
1143 }
1144 
1145 static const struct sparc32_cachetlb_ops swift_ops = {
1146 	.cache_all	= swift_flush_cache_all,
1147 	.cache_mm	= swift_flush_cache_mm,
1148 	.cache_page	= swift_flush_cache_page,
1149 	.cache_range	= swift_flush_cache_range,
1150 	.tlb_all	= swift_flush_tlb_all,
1151 	.tlb_mm		= swift_flush_tlb_mm,
1152 	.tlb_page	= swift_flush_tlb_page,
1153 	.tlb_range	= swift_flush_tlb_range,
1154 	.page_to_ram	= swift_flush_page_to_ram,
1155 	.sig_insns	= swift_flush_sig_insns,
1156 	.page_for_dma	= swift_flush_page_for_dma,
1157 };
1158 
1159 #define SWIFT_MASKID_ADDR  0x10003018
1160 static void __init init_swift(void)
1161 {
1162 	unsigned long swift_rev;
1163 
1164 	__asm__ __volatile__("lda [%1] %2, %0\n\t"
1165 			     "srl %0, 0x18, %0\n\t" :
1166 			     "=r" (swift_rev) :
1167 			     "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1168 	srmmu_name = "Fujitsu Swift";
1169 	switch (swift_rev) {
1170 	case 0x11:
1171 	case 0x20:
1172 	case 0x23:
1173 	case 0x30:
1174 		srmmu_modtype = Swift_lots_o_bugs;
1175 		hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1176 		/*
1177 		 * Gee george, I wonder why Sun is so hush hush about
1178 		 * this hardware bug... really braindamage stuff going
1179 		 * on here.  However I think we can find a way to avoid
1180 		 * all of the workaround overhead under Linux.  Basically,
1181 		 * any page fault can cause kernel pages to become user
1182 		 * accessible (the mmu gets confused and clears some of
1183 		 * the ACC bits in kernel ptes).  Aha, sounds pretty
1184 		 * horrible eh?  But wait, after extensive testing it appears
1185 		 * that if you use pgd_t level large kernel pte's (like the
1186 		 * 4MB pages on the Pentium) the bug does not get tripped
1187 		 * at all.  This avoids almost all of the major overhead.
1188 		 * Welcome to a world where your vendor tells you to,
1189 		 * "apply this kernel patch" instead of "sorry for the
1190 		 * broken hardware, send it back and we'll give you
1191 		 * properly functioning parts"
1192 		 */
1193 		break;
1194 	case 0x25:
1195 	case 0x31:
1196 		srmmu_modtype = Swift_bad_c;
1197 		hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1198 		/*
1199 		 * You see Sun allude to this hardware bug but never
1200 		 * admit things directly, they'll say things like,
1201 		 * "the Swift chip cache problems" or similar.
1202 		 */
1203 		break;
1204 	default:
1205 		srmmu_modtype = Swift_ok;
1206 		break;
1207 	}
1208 
1209 	sparc32_cachetlb_ops = &swift_ops;
1210 	flush_page_for_dma_global = 0;
1211 
1212 	/*
1213 	 * Are you now convinced that the Swift is one of the
1214 	 * biggest VLSI abortions of all time?  Bravo Fujitsu!
1215 	 * Fujitsu, the !#?!%$'d up processor people.  I bet if
1216 	 * you examined the microcode of the Swift you'd find
1217 	 * XXX's all over the place.
1218 	 */
1219 	poke_srmmu = poke_swift;
1220 }
1221 
1222 static void turbosparc_flush_cache_all(void)
1223 {
1224 	flush_user_windows();
1225 	turbosparc_idflash_clear();
1226 }
1227 
1228 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1229 {
1230 	FLUSH_BEGIN(mm)
1231 	flush_user_windows();
1232 	turbosparc_idflash_clear();
1233 	FLUSH_END
1234 }
1235 
1236 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1237 {
1238 	FLUSH_BEGIN(vma->vm_mm)
1239 	flush_user_windows();
1240 	turbosparc_idflash_clear();
1241 	FLUSH_END
1242 }
1243 
1244 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1245 {
1246 	FLUSH_BEGIN(vma->vm_mm)
1247 	flush_user_windows();
1248 	if (vma->vm_flags & VM_EXEC)
1249 		turbosparc_flush_icache();
1250 	turbosparc_flush_dcache();
1251 	FLUSH_END
1252 }
1253 
1254 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1255 static void turbosparc_flush_page_to_ram(unsigned long page)
1256 {
1257 #ifdef TURBOSPARC_WRITEBACK
1258 	volatile unsigned long clear;
1259 
1260 	if (srmmu_probe(page))
1261 		turbosparc_flush_page_cache(page);
1262 	clear = srmmu_get_fstatus();
1263 #endif
1264 }
1265 
1266 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1267 {
1268 }
1269 
1270 static void turbosparc_flush_page_for_dma(unsigned long page)
1271 {
1272 	turbosparc_flush_dcache();
1273 }
1274 
1275 static void turbosparc_flush_tlb_all(void)
1276 {
1277 	srmmu_flush_whole_tlb();
1278 }
1279 
1280 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1281 {
1282 	FLUSH_BEGIN(mm)
1283 	srmmu_flush_whole_tlb();
1284 	FLUSH_END
1285 }
1286 
1287 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1288 {
1289 	FLUSH_BEGIN(vma->vm_mm)
1290 	srmmu_flush_whole_tlb();
1291 	FLUSH_END
1292 }
1293 
1294 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1295 {
1296 	FLUSH_BEGIN(vma->vm_mm)
1297 	srmmu_flush_whole_tlb();
1298 	FLUSH_END
1299 }
1300 
1301 
1302 static void poke_turbosparc(void)
1303 {
1304 	unsigned long mreg = srmmu_get_mmureg();
1305 	unsigned long ccreg;
1306 
1307 	/* Clear any crap from the cache or else... */
1308 	turbosparc_flush_cache_all();
1309 	/* Temporarily disable I & D caches */
1310 	mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1311 	mreg &= ~(TURBOSPARC_PCENABLE);		/* Don't check parity */
1312 	srmmu_set_mmureg(mreg);
1313 
1314 	ccreg = turbosparc_get_ccreg();
1315 
1316 #ifdef TURBOSPARC_WRITEBACK
1317 	ccreg |= (TURBOSPARC_SNENABLE);		/* Do DVMA snooping in Dcache */
1318 	ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1319 			/* Write-back D-cache, emulate VLSI
1320 			 * abortion number three, not number one */
1321 #else
1322 	/* For now let's play safe, optimize later */
1323 	ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1324 			/* Do DVMA snooping in Dcache, Write-thru D-cache */
1325 	ccreg &= ~(TURBOSPARC_uS2);
1326 			/* Emulate VLSI abortion number three, not number one */
1327 #endif
1328 
1329 	switch (ccreg & 7) {
1330 	case 0: /* No SE cache */
1331 	case 7: /* Test mode */
1332 		break;
1333 	default:
1334 		ccreg |= (TURBOSPARC_SCENABLE);
1335 	}
1336 	turbosparc_set_ccreg(ccreg);
1337 
1338 	mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1339 	mreg |= (TURBOSPARC_ICSNOOP);		/* Icache snooping on */
1340 	srmmu_set_mmureg(mreg);
1341 }
1342 
1343 static const struct sparc32_cachetlb_ops turbosparc_ops = {
1344 	.cache_all	= turbosparc_flush_cache_all,
1345 	.cache_mm	= turbosparc_flush_cache_mm,
1346 	.cache_page	= turbosparc_flush_cache_page,
1347 	.cache_range	= turbosparc_flush_cache_range,
1348 	.tlb_all	= turbosparc_flush_tlb_all,
1349 	.tlb_mm		= turbosparc_flush_tlb_mm,
1350 	.tlb_page	= turbosparc_flush_tlb_page,
1351 	.tlb_range	= turbosparc_flush_tlb_range,
1352 	.page_to_ram	= turbosparc_flush_page_to_ram,
1353 	.sig_insns	= turbosparc_flush_sig_insns,
1354 	.page_for_dma	= turbosparc_flush_page_for_dma,
1355 };
1356 
1357 static void __init init_turbosparc(void)
1358 {
1359 	srmmu_name = "Fujitsu TurboSparc";
1360 	srmmu_modtype = TurboSparc;
1361 	sparc32_cachetlb_ops = &turbosparc_ops;
1362 	poke_srmmu = poke_turbosparc;
1363 }
1364 
1365 static void poke_tsunami(void)
1366 {
1367 	unsigned long mreg = srmmu_get_mmureg();
1368 
1369 	tsunami_flush_icache();
1370 	tsunami_flush_dcache();
1371 	mreg &= ~TSUNAMI_ITD;
1372 	mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1373 	srmmu_set_mmureg(mreg);
1374 }
1375 
1376 static const struct sparc32_cachetlb_ops tsunami_ops = {
1377 	.cache_all	= tsunami_flush_cache_all,
1378 	.cache_mm	= tsunami_flush_cache_mm,
1379 	.cache_page	= tsunami_flush_cache_page,
1380 	.cache_range	= tsunami_flush_cache_range,
1381 	.tlb_all	= tsunami_flush_tlb_all,
1382 	.tlb_mm		= tsunami_flush_tlb_mm,
1383 	.tlb_page	= tsunami_flush_tlb_page,
1384 	.tlb_range	= tsunami_flush_tlb_range,
1385 	.page_to_ram	= tsunami_flush_page_to_ram,
1386 	.sig_insns	= tsunami_flush_sig_insns,
1387 	.page_for_dma	= tsunami_flush_page_for_dma,
1388 };
1389 
1390 static void __init init_tsunami(void)
1391 {
1392 	/*
1393 	 * Tsunami's pretty sane, Sun and TI actually got it
1394 	 * somewhat right this time.  Fujitsu should have
1395 	 * taken some lessons from them.
1396 	 */
1397 
1398 	srmmu_name = "TI Tsunami";
1399 	srmmu_modtype = Tsunami;
1400 	sparc32_cachetlb_ops = &tsunami_ops;
1401 	poke_srmmu = poke_tsunami;
1402 
1403 	tsunami_setup_blockops();
1404 }
1405 
1406 static void poke_viking(void)
1407 {
1408 	unsigned long mreg = srmmu_get_mmureg();
1409 	static int smp_catch;
1410 
1411 	if (viking_mxcc_present) {
1412 		unsigned long mxcc_control = mxcc_get_creg();
1413 
1414 		mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1415 		mxcc_control &= ~(MXCC_CTL_RRC);
1416 		mxcc_set_creg(mxcc_control);
1417 
1418 		/*
1419 		 * We don't need memory parity checks.
1420 		 * XXX This is a mess, have to dig out later. ecd.
1421 		viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1422 		 */
1423 
1424 		/* We do cache ptables on MXCC. */
1425 		mreg |= VIKING_TCENABLE;
1426 	} else {
1427 		unsigned long bpreg;
1428 
1429 		mreg &= ~(VIKING_TCENABLE);
1430 		if (smp_catch++) {
1431 			/* Must disable mixed-cmd mode here for other cpu's. */
1432 			bpreg = viking_get_bpreg();
1433 			bpreg &= ~(VIKING_ACTION_MIX);
1434 			viking_set_bpreg(bpreg);
1435 
1436 			/* Just in case PROM does something funny. */
1437 			msi_set_sync();
1438 		}
1439 	}
1440 
1441 	mreg |= VIKING_SPENABLE;
1442 	mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1443 	mreg |= VIKING_SBENABLE;
1444 	mreg &= ~(VIKING_ACENABLE);
1445 	srmmu_set_mmureg(mreg);
1446 }
1447 
1448 static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1449 	.cache_all	= viking_flush_cache_all,
1450 	.cache_mm	= viking_flush_cache_mm,
1451 	.cache_page	= viking_flush_cache_page,
1452 	.cache_range	= viking_flush_cache_range,
1453 	.tlb_all	= viking_flush_tlb_all,
1454 	.tlb_mm		= viking_flush_tlb_mm,
1455 	.tlb_page	= viking_flush_tlb_page,
1456 	.tlb_range	= viking_flush_tlb_range,
1457 	.page_to_ram	= viking_flush_page_to_ram,
1458 	.sig_insns	= viking_flush_sig_insns,
1459 	.page_for_dma	= viking_flush_page_for_dma,
1460 };
1461 
1462 #ifdef CONFIG_SMP
1463 /* On sun4d the cpu broadcasts local TLB flushes, so we can just
1464  * perform the local TLB flush and all the other cpus will see it.
1465  * But, unfortunately, there is a bug in the sun4d XBUS backplane
1466  * that requires that we add some synchronization to these flushes.
1467  *
1468  * The bug is that the fifo which keeps track of all the pending TLB
1469  * broadcasts in the system is an entry or two too small, so if we
1470  * have too many going at once we'll overflow that fifo and lose a TLB
1471  * flush resulting in corruption.
1472  *
1473  * Our workaround is to take a global spinlock around the TLB flushes,
1474  * which guarentees we won't ever have too many pending.  It's a big
1475  * hammer, but a semaphore like system to make sure we only have N TLB
1476  * flushes going at once will require SMP locking anyways so there's
1477  * no real value in trying any harder than this.
1478  */
1479 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1480 	.cache_all	= viking_flush_cache_all,
1481 	.cache_mm	= viking_flush_cache_mm,
1482 	.cache_page	= viking_flush_cache_page,
1483 	.cache_range	= viking_flush_cache_range,
1484 	.tlb_all	= sun4dsmp_flush_tlb_all,
1485 	.tlb_mm		= sun4dsmp_flush_tlb_mm,
1486 	.tlb_page	= sun4dsmp_flush_tlb_page,
1487 	.tlb_range	= sun4dsmp_flush_tlb_range,
1488 	.page_to_ram	= viking_flush_page_to_ram,
1489 	.sig_insns	= viking_flush_sig_insns,
1490 	.page_for_dma	= viking_flush_page_for_dma,
1491 };
1492 #endif
1493 
1494 static void __init init_viking(void)
1495 {
1496 	unsigned long mreg = srmmu_get_mmureg();
1497 
1498 	/* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1499 	if (mreg & VIKING_MMODE) {
1500 		srmmu_name = "TI Viking";
1501 		viking_mxcc_present = 0;
1502 		msi_set_sync();
1503 
1504 		/*
1505 		 * We need this to make sure old viking takes no hits
1506 		 * on it's cache for dma snoops to workaround the
1507 		 * "load from non-cacheable memory" interrupt bug.
1508 		 * This is only necessary because of the new way in
1509 		 * which we use the IOMMU.
1510 		 */
1511 		viking_ops.page_for_dma = viking_flush_page;
1512 #ifdef CONFIG_SMP
1513 		viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1514 #endif
1515 		flush_page_for_dma_global = 0;
1516 	} else {
1517 		srmmu_name = "TI Viking/MXCC";
1518 		viking_mxcc_present = 1;
1519 		srmmu_cache_pagetables = 1;
1520 	}
1521 
1522 	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1523 		&viking_ops;
1524 #ifdef CONFIG_SMP
1525 	if (sparc_cpu_model == sun4d)
1526 		sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1527 			&viking_sun4d_smp_ops;
1528 #endif
1529 
1530 	poke_srmmu = poke_viking;
1531 }
1532 
1533 /* Probe for the srmmu chip version. */
1534 static void __init get_srmmu_type(void)
1535 {
1536 	unsigned long mreg, psr;
1537 	unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1538 
1539 	srmmu_modtype = SRMMU_INVAL_MOD;
1540 	hwbug_bitmask = 0;
1541 
1542 	mreg = srmmu_get_mmureg(); psr = get_psr();
1543 	mod_typ = (mreg & 0xf0000000) >> 28;
1544 	mod_rev = (mreg & 0x0f000000) >> 24;
1545 	psr_typ = (psr >> 28) & 0xf;
1546 	psr_vers = (psr >> 24) & 0xf;
1547 
1548 	/* First, check for sparc-leon. */
1549 	if (sparc_cpu_model == sparc_leon) {
1550 		init_leon();
1551 		return;
1552 	}
1553 
1554 	/* Second, check for HyperSparc or Cypress. */
1555 	if (mod_typ == 1) {
1556 		switch (mod_rev) {
1557 		case 7:
1558 			/* UP or MP Hypersparc */
1559 			init_hypersparc();
1560 			break;
1561 		case 0:
1562 		case 2:
1563 		case 10:
1564 		case 11:
1565 		case 12:
1566 		case 13:
1567 		case 14:
1568 		case 15:
1569 		default:
1570 			prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1571 			prom_halt();
1572 			break;
1573 		}
1574 		return;
1575 	}
1576 
1577 	/* Now Fujitsu TurboSparc. It might happen that it is
1578 	 * in Swift emulation mode, so we will check later...
1579 	 */
1580 	if (psr_typ == 0 && psr_vers == 5) {
1581 		init_turbosparc();
1582 		return;
1583 	}
1584 
1585 	/* Next check for Fujitsu Swift. */
1586 	if (psr_typ == 0 && psr_vers == 4) {
1587 		phandle cpunode;
1588 		char node_str[128];
1589 
1590 		/* Look if it is not a TurboSparc emulating Swift... */
1591 		cpunode = prom_getchild(prom_root_node);
1592 		while ((cpunode = prom_getsibling(cpunode)) != 0) {
1593 			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1594 			if (!strcmp(node_str, "cpu")) {
1595 				if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1596 				    prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1597 					init_turbosparc();
1598 					return;
1599 				}
1600 				break;
1601 			}
1602 		}
1603 
1604 		init_swift();
1605 		return;
1606 	}
1607 
1608 	/* Now the Viking family of srmmu. */
1609 	if (psr_typ == 4 &&
1610 	   ((psr_vers == 0) ||
1611 	    ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1612 		init_viking();
1613 		return;
1614 	}
1615 
1616 	/* Finally the Tsunami. */
1617 	if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1618 		init_tsunami();
1619 		return;
1620 	}
1621 
1622 	/* Oh well */
1623 	srmmu_is_bad();
1624 }
1625 
1626 #ifdef CONFIG_SMP
1627 /* Local cross-calls. */
1628 static void smp_flush_page_for_dma(unsigned long page)
1629 {
1630 	xc1((smpfunc_t) local_ops->page_for_dma, page);
1631 	local_ops->page_for_dma(page);
1632 }
1633 
1634 static void smp_flush_cache_all(void)
1635 {
1636 	xc0((smpfunc_t) local_ops->cache_all);
1637 	local_ops->cache_all();
1638 }
1639 
1640 static void smp_flush_tlb_all(void)
1641 {
1642 	xc0((smpfunc_t) local_ops->tlb_all);
1643 	local_ops->tlb_all();
1644 }
1645 
1646 static void smp_flush_cache_mm(struct mm_struct *mm)
1647 {
1648 	if (mm->context != NO_CONTEXT) {
1649 		cpumask_t cpu_mask;
1650 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1651 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1652 		if (!cpumask_empty(&cpu_mask))
1653 			xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1654 		local_ops->cache_mm(mm);
1655 	}
1656 }
1657 
1658 static void smp_flush_tlb_mm(struct mm_struct *mm)
1659 {
1660 	if (mm->context != NO_CONTEXT) {
1661 		cpumask_t cpu_mask;
1662 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1663 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1664 		if (!cpumask_empty(&cpu_mask)) {
1665 			xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1666 			if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1667 				cpumask_copy(mm_cpumask(mm),
1668 					     cpumask_of(smp_processor_id()));
1669 		}
1670 		local_ops->tlb_mm(mm);
1671 	}
1672 }
1673 
1674 static void smp_flush_cache_range(struct vm_area_struct *vma,
1675 				  unsigned long start,
1676 				  unsigned long end)
1677 {
1678 	struct mm_struct *mm = vma->vm_mm;
1679 
1680 	if (mm->context != NO_CONTEXT) {
1681 		cpumask_t cpu_mask;
1682 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1683 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1684 		if (!cpumask_empty(&cpu_mask))
1685 			xc3((smpfunc_t) local_ops->cache_range,
1686 			    (unsigned long) vma, start, end);
1687 		local_ops->cache_range(vma, start, end);
1688 	}
1689 }
1690 
1691 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1692 				unsigned long start,
1693 				unsigned long end)
1694 {
1695 	struct mm_struct *mm = vma->vm_mm;
1696 
1697 	if (mm->context != NO_CONTEXT) {
1698 		cpumask_t cpu_mask;
1699 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1700 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1701 		if (!cpumask_empty(&cpu_mask))
1702 			xc3((smpfunc_t) local_ops->tlb_range,
1703 			    (unsigned long) vma, start, end);
1704 		local_ops->tlb_range(vma, start, end);
1705 	}
1706 }
1707 
1708 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1709 {
1710 	struct mm_struct *mm = vma->vm_mm;
1711 
1712 	if (mm->context != NO_CONTEXT) {
1713 		cpumask_t cpu_mask;
1714 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1715 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1716 		if (!cpumask_empty(&cpu_mask))
1717 			xc2((smpfunc_t) local_ops->cache_page,
1718 			    (unsigned long) vma, page);
1719 		local_ops->cache_page(vma, page);
1720 	}
1721 }
1722 
1723 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1724 {
1725 	struct mm_struct *mm = vma->vm_mm;
1726 
1727 	if (mm->context != NO_CONTEXT) {
1728 		cpumask_t cpu_mask;
1729 		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1730 		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1731 		if (!cpumask_empty(&cpu_mask))
1732 			xc2((smpfunc_t) local_ops->tlb_page,
1733 			    (unsigned long) vma, page);
1734 		local_ops->tlb_page(vma, page);
1735 	}
1736 }
1737 
1738 static void smp_flush_page_to_ram(unsigned long page)
1739 {
1740 	/* Current theory is that those who call this are the one's
1741 	 * who have just dirtied their cache with the pages contents
1742 	 * in kernel space, therefore we only run this on local cpu.
1743 	 *
1744 	 * XXX This experiment failed, research further... -DaveM
1745 	 */
1746 #if 1
1747 	xc1((smpfunc_t) local_ops->page_to_ram, page);
1748 #endif
1749 	local_ops->page_to_ram(page);
1750 }
1751 
1752 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1753 {
1754 	cpumask_t cpu_mask;
1755 	cpumask_copy(&cpu_mask, mm_cpumask(mm));
1756 	cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1757 	if (!cpumask_empty(&cpu_mask))
1758 		xc2((smpfunc_t) local_ops->sig_insns,
1759 		    (unsigned long) mm, insn_addr);
1760 	local_ops->sig_insns(mm, insn_addr);
1761 }
1762 
1763 static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1764 	.cache_all	= smp_flush_cache_all,
1765 	.cache_mm	= smp_flush_cache_mm,
1766 	.cache_page	= smp_flush_cache_page,
1767 	.cache_range	= smp_flush_cache_range,
1768 	.tlb_all	= smp_flush_tlb_all,
1769 	.tlb_mm		= smp_flush_tlb_mm,
1770 	.tlb_page	= smp_flush_tlb_page,
1771 	.tlb_range	= smp_flush_tlb_range,
1772 	.page_to_ram	= smp_flush_page_to_ram,
1773 	.sig_insns	= smp_flush_sig_insns,
1774 	.page_for_dma	= smp_flush_page_for_dma,
1775 };
1776 #endif
1777 
1778 /* Load up routines and constants for sun4m and sun4d mmu */
1779 void __init load_mmu(void)
1780 {
1781 	/* Functions */
1782 	get_srmmu_type();
1783 
1784 #ifdef CONFIG_SMP
1785 	/* El switcheroo... */
1786 	local_ops = sparc32_cachetlb_ops;
1787 
1788 	if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1789 		smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1790 		smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1791 		smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1792 		smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1793 	}
1794 
1795 	if (poke_srmmu == poke_viking) {
1796 		/* Avoid unnecessary cross calls. */
1797 		smp_cachetlb_ops.cache_all = local_ops->cache_all;
1798 		smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1799 		smp_cachetlb_ops.cache_range = local_ops->cache_range;
1800 		smp_cachetlb_ops.cache_page = local_ops->cache_page;
1801 
1802 		smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1803 		smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1804 		smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1805 	}
1806 
1807 	/* It really is const after this point. */
1808 	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1809 		&smp_cachetlb_ops;
1810 #endif
1811 
1812 	if (sparc_cpu_model == sun4d)
1813 		ld_mmu_iounit();
1814 	else
1815 		ld_mmu_iommu();
1816 #ifdef CONFIG_SMP
1817 	if (sparc_cpu_model == sun4d)
1818 		sun4d_init_smp();
1819 	else if (sparc_cpu_model == sparc_leon)
1820 		leon_init_smp();
1821 	else
1822 		sun4m_init_smp();
1823 #endif
1824 }
1825