Linux Audio

Check our new training course

Loading...
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * srmmu.c:  SRMMU specific routines for memory management.
   4 *
   5 * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
   6 * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
   7 * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
   8 * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   9 * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
  10 */
  11
  12#include <linux/seq_file.h>
  13#include <linux/spinlock.h>
  14#include <linux/bootmem.h>
  15#include <linux/pagemap.h>
  16#include <linux/vmalloc.h>
  17#include <linux/kdebug.h>
  18#include <linux/export.h>
  19#include <linux/kernel.h>
  20#include <linux/init.h>
  21#include <linux/log2.h>
  22#include <linux/gfp.h>
  23#include <linux/fs.h>
  24#include <linux/mm.h>
  25
  26#include <asm/mmu_context.h>
  27#include <asm/cacheflush.h>
  28#include <asm/tlbflush.h>
  29#include <asm/io-unit.h>
  30#include <asm/pgalloc.h>
  31#include <asm/pgtable.h>
  32#include <asm/bitext.h>
  33#include <asm/vaddrs.h>
  34#include <asm/cache.h>
  35#include <asm/traps.h>
  36#include <asm/oplib.h>
  37#include <asm/mbus.h>
  38#include <asm/page.h>
  39#include <asm/asi.h>
  40#include <asm/msi.h>
  41#include <asm/smp.h>
  42#include <asm/io.h>
  43
  44/* Now the cpu specific definitions. */
  45#include <asm/turbosparc.h>
  46#include <asm/tsunami.h>
  47#include <asm/viking.h>
  48#include <asm/swift.h>
  49#include <asm/leon.h>
  50#include <asm/mxcc.h>
  51#include <asm/ross.h>
  52
  53#include "mm_32.h"
  54
  55enum mbus_module srmmu_modtype;
  56static unsigned int hwbug_bitmask;
  57int vac_cache_size;
  58EXPORT_SYMBOL(vac_cache_size);
  59int vac_line_size;
  60
  61extern struct resource sparc_iomap;
  62
  63extern unsigned long last_valid_pfn;
  64
  65static pgd_t *srmmu_swapper_pg_dir;
  66
  67const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
  68EXPORT_SYMBOL(sparc32_cachetlb_ops);
  69
  70#ifdef CONFIG_SMP
  71const struct sparc32_cachetlb_ops *local_ops;
  72
  73#define FLUSH_BEGIN(mm)
  74#define FLUSH_END
  75#else
  76#define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
  77#define FLUSH_END	}
  78#endif
  79
  80int flush_page_for_dma_global = 1;
  81
  82char *srmmu_name;
  83
  84ctxd_t *srmmu_ctx_table_phys;
  85static ctxd_t *srmmu_context_table;
  86
  87int viking_mxcc_present;
  88static DEFINE_SPINLOCK(srmmu_context_spinlock);
  89
  90static int is_hypersparc;
  91
  92static int srmmu_cache_pagetables;
  93
  94/* these will be initialized in srmmu_nocache_calcsize() */
  95static unsigned long srmmu_nocache_size;
  96static unsigned long srmmu_nocache_end;
  97
  98/* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
  99#define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
 100
 101/* The context table is a nocache user with the biggest alignment needs. */
 102#define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
 103
 104void *srmmu_nocache_pool;
 105static struct bit_map srmmu_nocache_map;
 106
 107static inline int srmmu_pmd_none(pmd_t pmd)
 108{ return !(pmd_val(pmd) & 0xFFFFFFF); }
 109
 110/* XXX should we hyper_flush_whole_icache here - Anton */
 111static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
 112{
 113	pte_t pte;
 114
 115	pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4)));
 116	set_pte((pte_t *)ctxp, pte);
 117}
 118
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 119void pmd_set(pmd_t *pmdp, pte_t *ptep)
 120{
 121	unsigned long ptp;	/* Physical address, shifted right by 4 */
 122	int i;
 123
 124	ptp = __nocache_pa(ptep) >> 4;
 125	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 126		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
 127		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
 128	}
 129}
 130
 131void pmd_populate(struct mm_struct *mm, pmd_t *pmdp, struct page *ptep)
 132{
 133	unsigned long ptp;	/* Physical address, shifted right by 4 */
 134	int i;
 135
 136	ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);	/* watch for overflow */
 137	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 138		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
 139		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
 140	}
 141}
 142
 143/* Find an entry in the third-level page table.. */
 144pte_t *pte_offset_kernel(pmd_t *dir, unsigned long address)
 145{
 146	void *pte;
 147
 148	pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
 149	return (pte_t *) pte +
 150	    ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 151}
 152
 153/*
 154 * size: bytes to allocate in the nocache area.
 155 * align: bytes, number to align at.
 156 * Returns the virtual address of the allocated area.
 157 */
 158static void *__srmmu_get_nocache(int size, int align)
 159{
 160	int offset;
 161	unsigned long addr;
 162
 163	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 164		printk(KERN_ERR "Size 0x%x too small for nocache request\n",
 165		       size);
 166		size = SRMMU_NOCACHE_BITMAP_SHIFT;
 167	}
 168	if (size & (SRMMU_NOCACHE_BITMAP_SHIFT - 1)) {
 169		printk(KERN_ERR "Size 0x%x unaligned int nocache request\n",
 170		       size);
 171		size += SRMMU_NOCACHE_BITMAP_SHIFT - 1;
 172	}
 173	BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
 174
 175	offset = bit_map_string_get(&srmmu_nocache_map,
 176				    size >> SRMMU_NOCACHE_BITMAP_SHIFT,
 177				    align >> SRMMU_NOCACHE_BITMAP_SHIFT);
 178	if (offset == -1) {
 179		printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
 180		       size, (int) srmmu_nocache_size,
 181		       srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 182		return NULL;
 183	}
 184
 185	addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
 186	return (void *)addr;
 187}
 188
 189void *srmmu_get_nocache(int size, int align)
 190{
 191	void *tmp;
 192
 193	tmp = __srmmu_get_nocache(size, align);
 194
 195	if (tmp)
 196		memset(tmp, 0, size);
 197
 198	return tmp;
 199}
 200
 201void srmmu_free_nocache(void *addr, int size)
 202{
 203	unsigned long vaddr;
 204	int offset;
 205
 206	vaddr = (unsigned long)addr;
 207	if (vaddr < SRMMU_NOCACHE_VADDR) {
 208		printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
 209		    vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
 210		BUG();
 211	}
 212	if (vaddr + size > srmmu_nocache_end) {
 213		printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
 214		    vaddr, srmmu_nocache_end);
 215		BUG();
 216	}
 217	if (!is_power_of_2(size)) {
 218		printk("Size 0x%x is not a power of 2\n", size);
 219		BUG();
 220	}
 221	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 222		printk("Size 0x%x is too small\n", size);
 223		BUG();
 224	}
 225	if (vaddr & (size - 1)) {
 226		printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
 227		BUG();
 228	}
 229
 230	offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
 231	size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 232
 233	bit_map_clear(&srmmu_nocache_map, offset, size);
 234}
 235
 236static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
 237						 unsigned long end);
 238
 239/* Return how much physical memory we have.  */
 240static unsigned long __init probe_memory(void)
 241{
 242	unsigned long total = 0;
 243	int i;
 244
 245	for (i = 0; sp_banks[i].num_bytes; i++)
 246		total += sp_banks[i].num_bytes;
 247
 248	return total;
 249}
 250
 251/*
 252 * Reserve nocache dynamically proportionally to the amount of
 253 * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
 254 */
 255static void __init srmmu_nocache_calcsize(void)
 256{
 257	unsigned long sysmemavail = probe_memory() / 1024;
 258	int srmmu_nocache_npages;
 259
 260	srmmu_nocache_npages =
 261		sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
 262
 263 /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
 264	// if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
 265	if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
 266		srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
 267
 268	/* anything above 1280 blows up */
 269	if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
 270		srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
 271
 272	srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
 273	srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
 274}
 275
 276static void __init srmmu_nocache_init(void)
 277{
 278	void *srmmu_nocache_bitmap;
 279	unsigned int bitmap_bits;
 280	pgd_t *pgd;
 281	pmd_t *pmd;
 282	pte_t *pte;
 283	unsigned long paddr, vaddr;
 284	unsigned long pteval;
 285
 286	bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 287
 288	srmmu_nocache_pool = __alloc_bootmem(srmmu_nocache_size,
 289		SRMMU_NOCACHE_ALIGN_MAX, 0UL);
 
 
 
 290	memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
 291
 292	srmmu_nocache_bitmap =
 293		__alloc_bootmem(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
 294				SMP_CACHE_BYTES, 0UL);
 
 
 
 295	bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
 296
 297	srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 298	memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
 299	init_mm.pgd = srmmu_swapper_pg_dir;
 300
 301	srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
 302
 303	paddr = __pa((unsigned long)srmmu_nocache_pool);
 304	vaddr = SRMMU_NOCACHE_VADDR;
 305
 306	while (vaddr < srmmu_nocache_end) {
 307		pgd = pgd_offset_k(vaddr);
 308		pmd = pmd_offset(__nocache_fix(pgd), vaddr);
 309		pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
 310
 311		pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
 312
 313		if (srmmu_cache_pagetables)
 314			pteval |= SRMMU_CACHE;
 315
 316		set_pte(__nocache_fix(pte), __pte(pteval));
 317
 318		vaddr += PAGE_SIZE;
 319		paddr += PAGE_SIZE;
 320	}
 321
 322	flush_cache_all();
 323	flush_tlb_all();
 324}
 325
 326pgd_t *get_pgd_fast(void)
 327{
 328	pgd_t *pgd = NULL;
 329
 330	pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 331	if (pgd) {
 332		pgd_t *init = pgd_offset_k(0);
 333		memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
 334		memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
 335						(PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
 336	}
 337
 338	return pgd;
 339}
 340
 341/*
 342 * Hardware needs alignment to 256 only, but we align to whole page size
 343 * to reduce fragmentation problems due to the buddy principle.
 344 * XXX Provide actual fragmentation statistics in /proc.
 345 *
 346 * Alignments up to the page size are the same for physical and virtual
 347 * addresses of the nocache area.
 348 */
 349pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
 350{
 351	unsigned long pte;
 352	struct page *page;
 353
 354	if ((pte = (unsigned long)pte_alloc_one_kernel(mm, address)) == 0)
 355		return NULL;
 356	page = pfn_to_page(__nocache_pa(pte) >> PAGE_SHIFT);
 357	if (!pgtable_page_ctor(page)) {
 358		__free_page(page);
 359		return NULL;
 360	}
 361	return page;
 362}
 363
 364void pte_free(struct mm_struct *mm, pgtable_t pte)
 365{
 366	unsigned long p;
 367
 368	pgtable_page_dtor(pte);
 369	p = (unsigned long)page_address(pte);	/* Cached address (for test) */
 370	if (p == 0)
 371		BUG();
 372	p = page_to_pfn(pte) << PAGE_SHIFT;	/* Physical address */
 373
 374	/* free non cached virtual address*/
 375	srmmu_free_nocache(__nocache_va(p), PTE_SIZE);
 376}
 377
 378/* context handling - a dynamically sized pool is used */
 379#define NO_CONTEXT	-1
 380
 381struct ctx_list {
 382	struct ctx_list *next;
 383	struct ctx_list *prev;
 384	unsigned int ctx_number;
 385	struct mm_struct *ctx_mm;
 386};
 387
 388static struct ctx_list *ctx_list_pool;
 389static struct ctx_list ctx_free;
 390static struct ctx_list ctx_used;
 391
 392/* At boot time we determine the number of contexts */
 393static int num_contexts;
 394
 395static inline void remove_from_ctx_list(struct ctx_list *entry)
 396{
 397	entry->next->prev = entry->prev;
 398	entry->prev->next = entry->next;
 399}
 400
 401static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
 402{
 403	entry->next = head;
 404	(entry->prev = head->prev)->next = entry;
 405	head->prev = entry;
 406}
 407#define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
 408#define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
 409
 410
 411static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
 412{
 413	struct ctx_list *ctxp;
 414
 415	ctxp = ctx_free.next;
 416	if (ctxp != &ctx_free) {
 417		remove_from_ctx_list(ctxp);
 418		add_to_used_ctxlist(ctxp);
 419		mm->context = ctxp->ctx_number;
 420		ctxp->ctx_mm = mm;
 421		return;
 422	}
 423	ctxp = ctx_used.next;
 424	if (ctxp->ctx_mm == old_mm)
 425		ctxp = ctxp->next;
 426	if (ctxp == &ctx_used)
 427		panic("out of mmu contexts");
 428	flush_cache_mm(ctxp->ctx_mm);
 429	flush_tlb_mm(ctxp->ctx_mm);
 430	remove_from_ctx_list(ctxp);
 431	add_to_used_ctxlist(ctxp);
 432	ctxp->ctx_mm->context = NO_CONTEXT;
 433	ctxp->ctx_mm = mm;
 434	mm->context = ctxp->ctx_number;
 435}
 436
 437static inline void free_context(int context)
 438{
 439	struct ctx_list *ctx_old;
 440
 441	ctx_old = ctx_list_pool + context;
 442	remove_from_ctx_list(ctx_old);
 443	add_to_free_ctxlist(ctx_old);
 444}
 445
 446static void __init sparc_context_init(int numctx)
 447{
 448	int ctx;
 449	unsigned long size;
 450
 451	size = numctx * sizeof(struct ctx_list);
 452	ctx_list_pool = __alloc_bootmem(size, SMP_CACHE_BYTES, 0UL);
 
 
 453
 454	for (ctx = 0; ctx < numctx; ctx++) {
 455		struct ctx_list *clist;
 456
 457		clist = (ctx_list_pool + ctx);
 458		clist->ctx_number = ctx;
 459		clist->ctx_mm = NULL;
 460	}
 461	ctx_free.next = ctx_free.prev = &ctx_free;
 462	ctx_used.next = ctx_used.prev = &ctx_used;
 463	for (ctx = 0; ctx < numctx; ctx++)
 464		add_to_free_ctxlist(ctx_list_pool + ctx);
 465}
 466
 467void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
 468	       struct task_struct *tsk)
 469{
 470	unsigned long flags;
 471
 472	if (mm->context == NO_CONTEXT) {
 473		spin_lock_irqsave(&srmmu_context_spinlock, flags);
 474		alloc_context(old_mm, mm);
 475		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
 476		srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
 477	}
 478
 479	if (sparc_cpu_model == sparc_leon)
 480		leon_switch_mm();
 481
 482	if (is_hypersparc)
 483		hyper_flush_whole_icache();
 484
 485	srmmu_set_context(mm->context);
 486}
 487
 488/* Low level IO area allocation on the SRMMU. */
 489static inline void srmmu_mapioaddr(unsigned long physaddr,
 490				   unsigned long virt_addr, int bus_type)
 491{
 492	pgd_t *pgdp;
 493	pmd_t *pmdp;
 494	pte_t *ptep;
 495	unsigned long tmp;
 496
 497	physaddr &= PAGE_MASK;
 498	pgdp = pgd_offset_k(virt_addr);
 499	pmdp = pmd_offset(pgdp, virt_addr);
 500	ptep = pte_offset_kernel(pmdp, virt_addr);
 501	tmp = (physaddr >> 4) | SRMMU_ET_PTE;
 502
 503	/* I need to test whether this is consistent over all
 504	 * sun4m's.  The bus_type represents the upper 4 bits of
 505	 * 36-bit physical address on the I/O space lines...
 506	 */
 507	tmp |= (bus_type << 28);
 508	tmp |= SRMMU_PRIV;
 509	__flush_page_to_ram(virt_addr);
 510	set_pte(ptep, __pte(tmp));
 511}
 512
 513void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
 514		      unsigned long xva, unsigned int len)
 515{
 516	while (len != 0) {
 517		len -= PAGE_SIZE;
 518		srmmu_mapioaddr(xpa, xva, bus);
 519		xva += PAGE_SIZE;
 520		xpa += PAGE_SIZE;
 521	}
 522	flush_tlb_all();
 523}
 524
 525static inline void srmmu_unmapioaddr(unsigned long virt_addr)
 526{
 527	pgd_t *pgdp;
 528	pmd_t *pmdp;
 529	pte_t *ptep;
 530
 531	pgdp = pgd_offset_k(virt_addr);
 532	pmdp = pmd_offset(pgdp, virt_addr);
 533	ptep = pte_offset_kernel(pmdp, virt_addr);
 534
 535	/* No need to flush uncacheable page. */
 536	__pte_clear(ptep);
 537}
 538
 539void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
 540{
 541	while (len != 0) {
 542		len -= PAGE_SIZE;
 543		srmmu_unmapioaddr(virt_addr);
 544		virt_addr += PAGE_SIZE;
 545	}
 546	flush_tlb_all();
 547}
 548
 549/* tsunami.S */
 550extern void tsunami_flush_cache_all(void);
 551extern void tsunami_flush_cache_mm(struct mm_struct *mm);
 552extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 553extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 554extern void tsunami_flush_page_to_ram(unsigned long page);
 555extern void tsunami_flush_page_for_dma(unsigned long page);
 556extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 557extern void tsunami_flush_tlb_all(void);
 558extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
 559extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 560extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 561extern void tsunami_setup_blockops(void);
 562
 563/* swift.S */
 564extern void swift_flush_cache_all(void);
 565extern void swift_flush_cache_mm(struct mm_struct *mm);
 566extern void swift_flush_cache_range(struct vm_area_struct *vma,
 567				    unsigned long start, unsigned long end);
 568extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 569extern void swift_flush_page_to_ram(unsigned long page);
 570extern void swift_flush_page_for_dma(unsigned long page);
 571extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 572extern void swift_flush_tlb_all(void);
 573extern void swift_flush_tlb_mm(struct mm_struct *mm);
 574extern void swift_flush_tlb_range(struct vm_area_struct *vma,
 575				  unsigned long start, unsigned long end);
 576extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 577
 578#if 0  /* P3: deadwood to debug precise flushes on Swift. */
 579void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 580{
 581	int cctx, ctx1;
 582
 583	page &= PAGE_MASK;
 584	if ((ctx1 = vma->vm_mm->context) != -1) {
 585		cctx = srmmu_get_context();
 586/* Is context # ever different from current context? P3 */
 587		if (cctx != ctx1) {
 588			printk("flush ctx %02x curr %02x\n", ctx1, cctx);
 589			srmmu_set_context(ctx1);
 590			swift_flush_page(page);
 591			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 592					"r" (page), "i" (ASI_M_FLUSH_PROBE));
 593			srmmu_set_context(cctx);
 594		} else {
 595			 /* Rm. prot. bits from virt. c. */
 596			/* swift_flush_cache_all(); */
 597			/* swift_flush_cache_page(vma, page); */
 598			swift_flush_page(page);
 599
 600			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 601				"r" (page), "i" (ASI_M_FLUSH_PROBE));
 602			/* same as above: srmmu_flush_tlb_page() */
 603		}
 604	}
 605}
 606#endif
 607
 608/*
 609 * The following are all MBUS based SRMMU modules, and therefore could
 610 * be found in a multiprocessor configuration.  On the whole, these
 611 * chips seems to be much more touchy about DVMA and page tables
 612 * with respect to cache coherency.
 613 */
 614
 615/* viking.S */
 616extern void viking_flush_cache_all(void);
 617extern void viking_flush_cache_mm(struct mm_struct *mm);
 618extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 619				     unsigned long end);
 620extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 621extern void viking_flush_page_to_ram(unsigned long page);
 622extern void viking_flush_page_for_dma(unsigned long page);
 623extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
 624extern void viking_flush_page(unsigned long page);
 625extern void viking_mxcc_flush_page(unsigned long page);
 626extern void viking_flush_tlb_all(void);
 627extern void viking_flush_tlb_mm(struct mm_struct *mm);
 628extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 629				   unsigned long end);
 630extern void viking_flush_tlb_page(struct vm_area_struct *vma,
 631				  unsigned long page);
 632extern void sun4dsmp_flush_tlb_all(void);
 633extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
 634extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 635				   unsigned long end);
 636extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
 637				  unsigned long page);
 638
 639/* hypersparc.S */
 640extern void hypersparc_flush_cache_all(void);
 641extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
 642extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 643extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 644extern void hypersparc_flush_page_to_ram(unsigned long page);
 645extern void hypersparc_flush_page_for_dma(unsigned long page);
 646extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 647extern void hypersparc_flush_tlb_all(void);
 648extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
 649extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 650extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 651extern void hypersparc_setup_blockops(void);
 652
 653/*
 654 * NOTE: All of this startup code assumes the low 16mb (approx.) of
 655 *       kernel mappings are done with one single contiguous chunk of
 656 *       ram.  On small ram machines (classics mainly) we only get
 657 *       around 8mb mapped for us.
 658 */
 659
 660static void __init early_pgtable_allocfail(char *type)
 661{
 662	prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
 663	prom_halt();
 664}
 665
 666static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
 667							unsigned long end)
 668{
 669	pgd_t *pgdp;
 670	pmd_t *pmdp;
 671	pte_t *ptep;
 672
 673	while (start < end) {
 674		pgdp = pgd_offset_k(start);
 675		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
 676			pmdp = __srmmu_get_nocache(
 677			    SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 678			if (pmdp == NULL)
 679				early_pgtable_allocfail("pmd");
 680			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 681			pgd_set(__nocache_fix(pgdp), pmdp);
 682		}
 683		pmdp = pmd_offset(__nocache_fix(pgdp), start);
 684		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 685			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 686			if (ptep == NULL)
 687				early_pgtable_allocfail("pte");
 688			memset(__nocache_fix(ptep), 0, PTE_SIZE);
 689			pmd_set(__nocache_fix(pmdp), ptep);
 690		}
 691		if (start > (0xffffffffUL - PMD_SIZE))
 692			break;
 693		start = (start + PMD_SIZE) & PMD_MASK;
 694	}
 695}
 696
 697static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
 698						  unsigned long end)
 699{
 700	pgd_t *pgdp;
 701	pmd_t *pmdp;
 702	pte_t *ptep;
 703
 704	while (start < end) {
 705		pgdp = pgd_offset_k(start);
 706		if (pgd_none(*pgdp)) {
 707			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 708			if (pmdp == NULL)
 709				early_pgtable_allocfail("pmd");
 710			memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
 711			pgd_set(pgdp, pmdp);
 712		}
 713		pmdp = pmd_offset(pgdp, start);
 714		if (srmmu_pmd_none(*pmdp)) {
 715			ptep = __srmmu_get_nocache(PTE_SIZE,
 716							     PTE_SIZE);
 717			if (ptep == NULL)
 718				early_pgtable_allocfail("pte");
 719			memset(ptep, 0, PTE_SIZE);
 720			pmd_set(pmdp, ptep);
 721		}
 722		if (start > (0xffffffffUL - PMD_SIZE))
 723			break;
 724		start = (start + PMD_SIZE) & PMD_MASK;
 725	}
 726}
 727
 728/* These flush types are not available on all chips... */
 729static inline unsigned long srmmu_probe(unsigned long vaddr)
 730{
 731	unsigned long retval;
 732
 733	if (sparc_cpu_model != sparc_leon) {
 734
 735		vaddr &= PAGE_MASK;
 736		__asm__ __volatile__("lda [%1] %2, %0\n\t" :
 737				     "=r" (retval) :
 738				     "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
 739	} else {
 740		retval = leon_swprobe(vaddr, NULL);
 741	}
 742	return retval;
 743}
 744
 745/*
 746 * This is much cleaner than poking around physical address space
 747 * looking at the prom's page table directly which is what most
 748 * other OS's do.  Yuck... this is much better.
 749 */
 750static void __init srmmu_inherit_prom_mappings(unsigned long start,
 751					       unsigned long end)
 752{
 753	unsigned long probed;
 754	unsigned long addr;
 755	pgd_t *pgdp;
 756	pmd_t *pmdp;
 757	pte_t *ptep;
 758	int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
 759
 760	while (start <= end) {
 761		if (start == 0)
 762			break; /* probably wrap around */
 763		if (start == 0xfef00000)
 764			start = KADB_DEBUGGER_BEGVM;
 765		probed = srmmu_probe(start);
 766		if (!probed) {
 767			/* continue probing until we find an entry */
 768			start += PAGE_SIZE;
 769			continue;
 770		}
 771
 772		/* A red snapper, see what it really is. */
 773		what = 0;
 774		addr = start - PAGE_SIZE;
 775
 776		if (!(start & ~(SRMMU_REAL_PMD_MASK))) {
 777			if (srmmu_probe(addr + SRMMU_REAL_PMD_SIZE) == probed)
 778				what = 1;
 779		}
 780
 781		if (!(start & ~(SRMMU_PGDIR_MASK))) {
 782			if (srmmu_probe(addr + SRMMU_PGDIR_SIZE) == probed)
 783				what = 2;
 784		}
 785
 786		pgdp = pgd_offset_k(start);
 787		if (what == 2) {
 788			*(pgd_t *)__nocache_fix(pgdp) = __pgd(probed);
 789			start += SRMMU_PGDIR_SIZE;
 790			continue;
 791		}
 792		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
 793			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
 794						   SRMMU_PMD_TABLE_SIZE);
 795			if (pmdp == NULL)
 796				early_pgtable_allocfail("pmd");
 797			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 798			pgd_set(__nocache_fix(pgdp), pmdp);
 799		}
 800		pmdp = pmd_offset(__nocache_fix(pgdp), start);
 801		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 802			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 803			if (ptep == NULL)
 804				early_pgtable_allocfail("pte");
 805			memset(__nocache_fix(ptep), 0, PTE_SIZE);
 806			pmd_set(__nocache_fix(pmdp), ptep);
 807		}
 808		if (what == 1) {
 809			/* We bend the rule where all 16 PTPs in a pmd_t point
 810			 * inside the same PTE page, and we leak a perfectly
 811			 * good hardware PTE piece. Alternatives seem worse.
 812			 */
 813			unsigned int x;	/* Index of HW PMD in soft cluster */
 814			unsigned long *val;
 815			x = (start >> PMD_SHIFT) & 15;
 816			val = &pmdp->pmdv[x];
 817			*(unsigned long *)__nocache_fix(val) = probed;
 818			start += SRMMU_REAL_PMD_SIZE;
 819			continue;
 820		}
 821		ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
 822		*(pte_t *)__nocache_fix(ptep) = __pte(probed);
 823		start += PAGE_SIZE;
 824	}
 825}
 826
 827#define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
 828
 829/* Create a third-level SRMMU 16MB page mapping. */
 830static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
 831{
 832	pgd_t *pgdp = pgd_offset_k(vaddr);
 833	unsigned long big_pte;
 834
 835	big_pte = KERNEL_PTE(phys_base >> 4);
 836	*(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
 837}
 838
 839/* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
 840static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
 841{
 842	unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
 843	unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
 844	unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
 845	/* Map "low" memory only */
 846	const unsigned long min_vaddr = PAGE_OFFSET;
 847	const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
 848
 849	if (vstart < min_vaddr || vstart >= max_vaddr)
 850		return vstart;
 851
 852	if (vend > max_vaddr || vend < min_vaddr)
 853		vend = max_vaddr;
 854
 855	while (vstart < vend) {
 856		do_large_mapping(vstart, pstart);
 857		vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
 858	}
 859	return vstart;
 860}
 861
 862static void __init map_kernel(void)
 863{
 864	int i;
 865
 866	if (phys_base > 0) {
 867		do_large_mapping(PAGE_OFFSET, phys_base);
 868	}
 869
 870	for (i = 0; sp_banks[i].num_bytes != 0; i++) {
 871		map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
 872	}
 873}
 874
 875void (*poke_srmmu)(void) = NULL;
 876
 877void __init srmmu_paging_init(void)
 878{
 879	int i;
 880	phandle cpunode;
 881	char node_str[128];
 882	pgd_t *pgd;
 883	pmd_t *pmd;
 884	pte_t *pte;
 885	unsigned long pages_avail;
 886
 887	init_mm.context = (unsigned long) NO_CONTEXT;
 888	sparc_iomap.start = SUN4M_IOBASE_VADDR;	/* 16MB of IOSPACE on all sun4m's. */
 889
 890	if (sparc_cpu_model == sun4d)
 891		num_contexts = 65536; /* We know it is Viking */
 892	else {
 893		/* Find the number of contexts on the srmmu. */
 894		cpunode = prom_getchild(prom_root_node);
 895		num_contexts = 0;
 896		while (cpunode != 0) {
 897			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
 898			if (!strcmp(node_str, "cpu")) {
 899				num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
 900				break;
 901			}
 902			cpunode = prom_getsibling(cpunode);
 903		}
 904	}
 905
 906	if (!num_contexts) {
 907		prom_printf("Something wrong, can't find cpu node in paging_init.\n");
 908		prom_halt();
 909	}
 910
 911	pages_avail = 0;
 912	last_valid_pfn = bootmem_init(&pages_avail);
 913
 914	srmmu_nocache_calcsize();
 915	srmmu_nocache_init();
 916	srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
 917	map_kernel();
 918
 919	/* ctx table has to be physically aligned to its size */
 920	srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
 921	srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
 922
 923	for (i = 0; i < num_contexts; i++)
 924		srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
 925
 926	flush_cache_all();
 927	srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
 928#ifdef CONFIG_SMP
 929	/* Stop from hanging here... */
 930	local_ops->tlb_all();
 931#else
 932	flush_tlb_all();
 933#endif
 934	poke_srmmu();
 935
 936	srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
 937	srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
 938
 939	srmmu_allocate_ptable_skeleton(
 940		__fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
 941	srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
 942
 943	pgd = pgd_offset_k(PKMAP_BASE);
 944	pmd = pmd_offset(pgd, PKMAP_BASE);
 945	pte = pte_offset_kernel(pmd, PKMAP_BASE);
 946	pkmap_page_table = pte;
 947
 948	flush_cache_all();
 949	flush_tlb_all();
 950
 951	sparc_context_init(num_contexts);
 952
 953	kmap_init();
 954
 955	{
 956		unsigned long zones_size[MAX_NR_ZONES];
 957		unsigned long zholes_size[MAX_NR_ZONES];
 958		unsigned long npages;
 959		int znum;
 960
 961		for (znum = 0; znum < MAX_NR_ZONES; znum++)
 962			zones_size[znum] = zholes_size[znum] = 0;
 963
 964		npages = max_low_pfn - pfn_base;
 965
 966		zones_size[ZONE_DMA] = npages;
 967		zholes_size[ZONE_DMA] = npages - pages_avail;
 968
 969		npages = highend_pfn - max_low_pfn;
 970		zones_size[ZONE_HIGHMEM] = npages;
 971		zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
 972
 973		free_area_init_node(0, zones_size, pfn_base, zholes_size);
 974	}
 975}
 976
 977void mmu_info(struct seq_file *m)
 978{
 979	seq_printf(m,
 980		   "MMU type\t: %s\n"
 981		   "contexts\t: %d\n"
 982		   "nocache total\t: %ld\n"
 983		   "nocache used\t: %d\n",
 984		   srmmu_name,
 985		   num_contexts,
 986		   srmmu_nocache_size,
 987		   srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 988}
 989
 990int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 991{
 992	mm->context = NO_CONTEXT;
 993	return 0;
 994}
 995
 996void destroy_context(struct mm_struct *mm)
 997{
 998	unsigned long flags;
 999
1000	if (mm->context != NO_CONTEXT) {
1001		flush_cache_mm(mm);
1002		srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1003		flush_tlb_mm(mm);
1004		spin_lock_irqsave(&srmmu_context_spinlock, flags);
1005		free_context(mm->context);
1006		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1007		mm->context = NO_CONTEXT;
1008	}
1009}
1010
1011/* Init various srmmu chip types. */
1012static void __init srmmu_is_bad(void)
1013{
1014	prom_printf("Could not determine SRMMU chip type.\n");
1015	prom_halt();
1016}
1017
1018static void __init init_vac_layout(void)
1019{
1020	phandle nd;
1021	int cache_lines;
1022	char node_str[128];
1023#ifdef CONFIG_SMP
1024	int cpu = 0;
1025	unsigned long max_size = 0;
1026	unsigned long min_line_size = 0x10000000;
1027#endif
1028
1029	nd = prom_getchild(prom_root_node);
1030	while ((nd = prom_getsibling(nd)) != 0) {
1031		prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1032		if (!strcmp(node_str, "cpu")) {
1033			vac_line_size = prom_getint(nd, "cache-line-size");
1034			if (vac_line_size == -1) {
1035				prom_printf("can't determine cache-line-size, halting.\n");
1036				prom_halt();
1037			}
1038			cache_lines = prom_getint(nd, "cache-nlines");
1039			if (cache_lines == -1) {
1040				prom_printf("can't determine cache-nlines, halting.\n");
1041				prom_halt();
1042			}
1043
1044			vac_cache_size = cache_lines * vac_line_size;
1045#ifdef CONFIG_SMP
1046			if (vac_cache_size > max_size)
1047				max_size = vac_cache_size;
1048			if (vac_line_size < min_line_size)
1049				min_line_size = vac_line_size;
1050			//FIXME: cpus not contiguous!!
1051			cpu++;
1052			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1053				break;
1054#else
1055			break;
1056#endif
1057		}
1058	}
1059	if (nd == 0) {
1060		prom_printf("No CPU nodes found, halting.\n");
1061		prom_halt();
1062	}
1063#ifdef CONFIG_SMP
1064	vac_cache_size = max_size;
1065	vac_line_size = min_line_size;
1066#endif
1067	printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1068	       (int)vac_cache_size, (int)vac_line_size);
1069}
1070
1071static void poke_hypersparc(void)
1072{
1073	volatile unsigned long clear;
1074	unsigned long mreg = srmmu_get_mmureg();
1075
1076	hyper_flush_unconditional_combined();
1077
1078	mreg &= ~(HYPERSPARC_CWENABLE);
1079	mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1080	mreg |= (HYPERSPARC_CMODE);
1081
1082	srmmu_set_mmureg(mreg);
1083
1084#if 0 /* XXX I think this is bad news... -DaveM */
1085	hyper_clear_all_tags();
1086#endif
1087
1088	put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1089	hyper_flush_whole_icache();
1090	clear = srmmu_get_faddr();
1091	clear = srmmu_get_fstatus();
1092}
1093
1094static const struct sparc32_cachetlb_ops hypersparc_ops = {
1095	.cache_all	= hypersparc_flush_cache_all,
1096	.cache_mm	= hypersparc_flush_cache_mm,
1097	.cache_page	= hypersparc_flush_cache_page,
1098	.cache_range	= hypersparc_flush_cache_range,
1099	.tlb_all	= hypersparc_flush_tlb_all,
1100	.tlb_mm		= hypersparc_flush_tlb_mm,
1101	.tlb_page	= hypersparc_flush_tlb_page,
1102	.tlb_range	= hypersparc_flush_tlb_range,
1103	.page_to_ram	= hypersparc_flush_page_to_ram,
1104	.sig_insns	= hypersparc_flush_sig_insns,
1105	.page_for_dma	= hypersparc_flush_page_for_dma,
1106};
1107
1108static void __init init_hypersparc(void)
1109{
1110	srmmu_name = "ROSS HyperSparc";
1111	srmmu_modtype = HyperSparc;
1112
1113	init_vac_layout();
1114
1115	is_hypersparc = 1;
1116	sparc32_cachetlb_ops = &hypersparc_ops;
1117
1118	poke_srmmu = poke_hypersparc;
1119
1120	hypersparc_setup_blockops();
1121}
1122
1123static void poke_swift(void)
1124{
1125	unsigned long mreg;
1126
1127	/* Clear any crap from the cache or else... */
1128	swift_flush_cache_all();
1129
1130	/* Enable I & D caches */
1131	mreg = srmmu_get_mmureg();
1132	mreg |= (SWIFT_IE | SWIFT_DE);
1133	/*
1134	 * The Swift branch folding logic is completely broken.  At
1135	 * trap time, if things are just right, if can mistakenly
1136	 * think that a trap is coming from kernel mode when in fact
1137	 * it is coming from user mode (it mis-executes the branch in
1138	 * the trap code).  So you see things like crashme completely
1139	 * hosing your machine which is completely unacceptable.  Turn
1140	 * this shit off... nice job Fujitsu.
1141	 */
1142	mreg &= ~(SWIFT_BF);
1143	srmmu_set_mmureg(mreg);
1144}
1145
1146static const struct sparc32_cachetlb_ops swift_ops = {
1147	.cache_all	= swift_flush_cache_all,
1148	.cache_mm	= swift_flush_cache_mm,
1149	.cache_page	= swift_flush_cache_page,
1150	.cache_range	= swift_flush_cache_range,
1151	.tlb_all	= swift_flush_tlb_all,
1152	.tlb_mm		= swift_flush_tlb_mm,
1153	.tlb_page	= swift_flush_tlb_page,
1154	.tlb_range	= swift_flush_tlb_range,
1155	.page_to_ram	= swift_flush_page_to_ram,
1156	.sig_insns	= swift_flush_sig_insns,
1157	.page_for_dma	= swift_flush_page_for_dma,
1158};
1159
1160#define SWIFT_MASKID_ADDR  0x10003018
1161static void __init init_swift(void)
1162{
1163	unsigned long swift_rev;
1164
1165	__asm__ __volatile__("lda [%1] %2, %0\n\t"
1166			     "srl %0, 0x18, %0\n\t" :
1167			     "=r" (swift_rev) :
1168			     "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1169	srmmu_name = "Fujitsu Swift";
1170	switch (swift_rev) {
1171	case 0x11:
1172	case 0x20:
1173	case 0x23:
1174	case 0x30:
1175		srmmu_modtype = Swift_lots_o_bugs;
1176		hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1177		/*
1178		 * Gee george, I wonder why Sun is so hush hush about
1179		 * this hardware bug... really braindamage stuff going
1180		 * on here.  However I think we can find a way to avoid
1181		 * all of the workaround overhead under Linux.  Basically,
1182		 * any page fault can cause kernel pages to become user
1183		 * accessible (the mmu gets confused and clears some of
1184		 * the ACC bits in kernel ptes).  Aha, sounds pretty
1185		 * horrible eh?  But wait, after extensive testing it appears
1186		 * that if you use pgd_t level large kernel pte's (like the
1187		 * 4MB pages on the Pentium) the bug does not get tripped
1188		 * at all.  This avoids almost all of the major overhead.
1189		 * Welcome to a world where your vendor tells you to,
1190		 * "apply this kernel patch" instead of "sorry for the
1191		 * broken hardware, send it back and we'll give you
1192		 * properly functioning parts"
1193		 */
1194		break;
1195	case 0x25:
1196	case 0x31:
1197		srmmu_modtype = Swift_bad_c;
1198		hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1199		/*
1200		 * You see Sun allude to this hardware bug but never
1201		 * admit things directly, they'll say things like,
1202		 * "the Swift chip cache problems" or similar.
1203		 */
1204		break;
1205	default:
1206		srmmu_modtype = Swift_ok;
1207		break;
1208	}
1209
1210	sparc32_cachetlb_ops = &swift_ops;
1211	flush_page_for_dma_global = 0;
1212
1213	/*
1214	 * Are you now convinced that the Swift is one of the
1215	 * biggest VLSI abortions of all time?  Bravo Fujitsu!
1216	 * Fujitsu, the !#?!%$'d up processor people.  I bet if
1217	 * you examined the microcode of the Swift you'd find
1218	 * XXX's all over the place.
1219	 */
1220	poke_srmmu = poke_swift;
1221}
1222
1223static void turbosparc_flush_cache_all(void)
1224{
1225	flush_user_windows();
1226	turbosparc_idflash_clear();
1227}
1228
1229static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1230{
1231	FLUSH_BEGIN(mm)
1232	flush_user_windows();
1233	turbosparc_idflash_clear();
1234	FLUSH_END
1235}
1236
1237static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1238{
1239	FLUSH_BEGIN(vma->vm_mm)
1240	flush_user_windows();
1241	turbosparc_idflash_clear();
1242	FLUSH_END
1243}
1244
1245static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1246{
1247	FLUSH_BEGIN(vma->vm_mm)
1248	flush_user_windows();
1249	if (vma->vm_flags & VM_EXEC)
1250		turbosparc_flush_icache();
1251	turbosparc_flush_dcache();
1252	FLUSH_END
1253}
1254
1255/* TurboSparc is copy-back, if we turn it on, but this does not work. */
1256static void turbosparc_flush_page_to_ram(unsigned long page)
1257{
1258#ifdef TURBOSPARC_WRITEBACK
1259	volatile unsigned long clear;
1260
1261	if (srmmu_probe(page))
1262		turbosparc_flush_page_cache(page);
1263	clear = srmmu_get_fstatus();
1264#endif
1265}
1266
1267static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1268{
1269}
1270
1271static void turbosparc_flush_page_for_dma(unsigned long page)
1272{
1273	turbosparc_flush_dcache();
1274}
1275
1276static void turbosparc_flush_tlb_all(void)
1277{
1278	srmmu_flush_whole_tlb();
1279}
1280
1281static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1282{
1283	FLUSH_BEGIN(mm)
1284	srmmu_flush_whole_tlb();
1285	FLUSH_END
1286}
1287
1288static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1289{
1290	FLUSH_BEGIN(vma->vm_mm)
1291	srmmu_flush_whole_tlb();
1292	FLUSH_END
1293}
1294
1295static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1296{
1297	FLUSH_BEGIN(vma->vm_mm)
1298	srmmu_flush_whole_tlb();
1299	FLUSH_END
1300}
1301
1302
1303static void poke_turbosparc(void)
1304{
1305	unsigned long mreg = srmmu_get_mmureg();
1306	unsigned long ccreg;
1307
1308	/* Clear any crap from the cache or else... */
1309	turbosparc_flush_cache_all();
1310	/* Temporarily disable I & D caches */
1311	mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1312	mreg &= ~(TURBOSPARC_PCENABLE);		/* Don't check parity */
1313	srmmu_set_mmureg(mreg);
1314
1315	ccreg = turbosparc_get_ccreg();
1316
1317#ifdef TURBOSPARC_WRITEBACK
1318	ccreg |= (TURBOSPARC_SNENABLE);		/* Do DVMA snooping in Dcache */
1319	ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1320			/* Write-back D-cache, emulate VLSI
1321			 * abortion number three, not number one */
1322#else
1323	/* For now let's play safe, optimize later */
1324	ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1325			/* Do DVMA snooping in Dcache, Write-thru D-cache */
1326	ccreg &= ~(TURBOSPARC_uS2);
1327			/* Emulate VLSI abortion number three, not number one */
1328#endif
1329
1330	switch (ccreg & 7) {
1331	case 0: /* No SE cache */
1332	case 7: /* Test mode */
1333		break;
1334	default:
1335		ccreg |= (TURBOSPARC_SCENABLE);
1336	}
1337	turbosparc_set_ccreg(ccreg);
1338
1339	mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1340	mreg |= (TURBOSPARC_ICSNOOP);		/* Icache snooping on */
1341	srmmu_set_mmureg(mreg);
1342}
1343
1344static const struct sparc32_cachetlb_ops turbosparc_ops = {
1345	.cache_all	= turbosparc_flush_cache_all,
1346	.cache_mm	= turbosparc_flush_cache_mm,
1347	.cache_page	= turbosparc_flush_cache_page,
1348	.cache_range	= turbosparc_flush_cache_range,
1349	.tlb_all	= turbosparc_flush_tlb_all,
1350	.tlb_mm		= turbosparc_flush_tlb_mm,
1351	.tlb_page	= turbosparc_flush_tlb_page,
1352	.tlb_range	= turbosparc_flush_tlb_range,
1353	.page_to_ram	= turbosparc_flush_page_to_ram,
1354	.sig_insns	= turbosparc_flush_sig_insns,
1355	.page_for_dma	= turbosparc_flush_page_for_dma,
1356};
1357
1358static void __init init_turbosparc(void)
1359{
1360	srmmu_name = "Fujitsu TurboSparc";
1361	srmmu_modtype = TurboSparc;
1362	sparc32_cachetlb_ops = &turbosparc_ops;
1363	poke_srmmu = poke_turbosparc;
1364}
1365
1366static void poke_tsunami(void)
1367{
1368	unsigned long mreg = srmmu_get_mmureg();
1369
1370	tsunami_flush_icache();
1371	tsunami_flush_dcache();
1372	mreg &= ~TSUNAMI_ITD;
1373	mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1374	srmmu_set_mmureg(mreg);
1375}
1376
1377static const struct sparc32_cachetlb_ops tsunami_ops = {
1378	.cache_all	= tsunami_flush_cache_all,
1379	.cache_mm	= tsunami_flush_cache_mm,
1380	.cache_page	= tsunami_flush_cache_page,
1381	.cache_range	= tsunami_flush_cache_range,
1382	.tlb_all	= tsunami_flush_tlb_all,
1383	.tlb_mm		= tsunami_flush_tlb_mm,
1384	.tlb_page	= tsunami_flush_tlb_page,
1385	.tlb_range	= tsunami_flush_tlb_range,
1386	.page_to_ram	= tsunami_flush_page_to_ram,
1387	.sig_insns	= tsunami_flush_sig_insns,
1388	.page_for_dma	= tsunami_flush_page_for_dma,
1389};
1390
1391static void __init init_tsunami(void)
1392{
1393	/*
1394	 * Tsunami's pretty sane, Sun and TI actually got it
1395	 * somewhat right this time.  Fujitsu should have
1396	 * taken some lessons from them.
1397	 */
1398
1399	srmmu_name = "TI Tsunami";
1400	srmmu_modtype = Tsunami;
1401	sparc32_cachetlb_ops = &tsunami_ops;
1402	poke_srmmu = poke_tsunami;
1403
1404	tsunami_setup_blockops();
1405}
1406
1407static void poke_viking(void)
1408{
1409	unsigned long mreg = srmmu_get_mmureg();
1410	static int smp_catch;
1411
1412	if (viking_mxcc_present) {
1413		unsigned long mxcc_control = mxcc_get_creg();
1414
1415		mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1416		mxcc_control &= ~(MXCC_CTL_RRC);
1417		mxcc_set_creg(mxcc_control);
1418
1419		/*
1420		 * We don't need memory parity checks.
1421		 * XXX This is a mess, have to dig out later. ecd.
1422		viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1423		 */
1424
1425		/* We do cache ptables on MXCC. */
1426		mreg |= VIKING_TCENABLE;
1427	} else {
1428		unsigned long bpreg;
1429
1430		mreg &= ~(VIKING_TCENABLE);
1431		if (smp_catch++) {
1432			/* Must disable mixed-cmd mode here for other cpu's. */
1433			bpreg = viking_get_bpreg();
1434			bpreg &= ~(VIKING_ACTION_MIX);
1435			viking_set_bpreg(bpreg);
1436
1437			/* Just in case PROM does something funny. */
1438			msi_set_sync();
1439		}
1440	}
1441
1442	mreg |= VIKING_SPENABLE;
1443	mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1444	mreg |= VIKING_SBENABLE;
1445	mreg &= ~(VIKING_ACENABLE);
1446	srmmu_set_mmureg(mreg);
1447}
1448
1449static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1450	.cache_all	= viking_flush_cache_all,
1451	.cache_mm	= viking_flush_cache_mm,
1452	.cache_page	= viking_flush_cache_page,
1453	.cache_range	= viking_flush_cache_range,
1454	.tlb_all	= viking_flush_tlb_all,
1455	.tlb_mm		= viking_flush_tlb_mm,
1456	.tlb_page	= viking_flush_tlb_page,
1457	.tlb_range	= viking_flush_tlb_range,
1458	.page_to_ram	= viking_flush_page_to_ram,
1459	.sig_insns	= viking_flush_sig_insns,
1460	.page_for_dma	= viking_flush_page_for_dma,
1461};
1462
1463#ifdef CONFIG_SMP
1464/* On sun4d the cpu broadcasts local TLB flushes, so we can just
1465 * perform the local TLB flush and all the other cpus will see it.
1466 * But, unfortunately, there is a bug in the sun4d XBUS backplane
1467 * that requires that we add some synchronization to these flushes.
1468 *
1469 * The bug is that the fifo which keeps track of all the pending TLB
1470 * broadcasts in the system is an entry or two too small, so if we
1471 * have too many going at once we'll overflow that fifo and lose a TLB
1472 * flush resulting in corruption.
1473 *
1474 * Our workaround is to take a global spinlock around the TLB flushes,
1475 * which guarentees we won't ever have too many pending.  It's a big
1476 * hammer, but a semaphore like system to make sure we only have N TLB
1477 * flushes going at once will require SMP locking anyways so there's
1478 * no real value in trying any harder than this.
1479 */
1480static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1481	.cache_all	= viking_flush_cache_all,
1482	.cache_mm	= viking_flush_cache_mm,
1483	.cache_page	= viking_flush_cache_page,
1484	.cache_range	= viking_flush_cache_range,
1485	.tlb_all	= sun4dsmp_flush_tlb_all,
1486	.tlb_mm		= sun4dsmp_flush_tlb_mm,
1487	.tlb_page	= sun4dsmp_flush_tlb_page,
1488	.tlb_range	= sun4dsmp_flush_tlb_range,
1489	.page_to_ram	= viking_flush_page_to_ram,
1490	.sig_insns	= viking_flush_sig_insns,
1491	.page_for_dma	= viking_flush_page_for_dma,
1492};
1493#endif
1494
1495static void __init init_viking(void)
1496{
1497	unsigned long mreg = srmmu_get_mmureg();
1498
1499	/* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1500	if (mreg & VIKING_MMODE) {
1501		srmmu_name = "TI Viking";
1502		viking_mxcc_present = 0;
1503		msi_set_sync();
1504
1505		/*
1506		 * We need this to make sure old viking takes no hits
1507		 * on it's cache for dma snoops to workaround the
1508		 * "load from non-cacheable memory" interrupt bug.
1509		 * This is only necessary because of the new way in
1510		 * which we use the IOMMU.
1511		 */
1512		viking_ops.page_for_dma = viking_flush_page;
1513#ifdef CONFIG_SMP
1514		viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1515#endif
1516		flush_page_for_dma_global = 0;
1517	} else {
1518		srmmu_name = "TI Viking/MXCC";
1519		viking_mxcc_present = 1;
1520		srmmu_cache_pagetables = 1;
1521	}
1522
1523	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1524		&viking_ops;
1525#ifdef CONFIG_SMP
1526	if (sparc_cpu_model == sun4d)
1527		sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1528			&viking_sun4d_smp_ops;
1529#endif
1530
1531	poke_srmmu = poke_viking;
1532}
1533
1534/* Probe for the srmmu chip version. */
1535static void __init get_srmmu_type(void)
1536{
1537	unsigned long mreg, psr;
1538	unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1539
1540	srmmu_modtype = SRMMU_INVAL_MOD;
1541	hwbug_bitmask = 0;
1542
1543	mreg = srmmu_get_mmureg(); psr = get_psr();
1544	mod_typ = (mreg & 0xf0000000) >> 28;
1545	mod_rev = (mreg & 0x0f000000) >> 24;
1546	psr_typ = (psr >> 28) & 0xf;
1547	psr_vers = (psr >> 24) & 0xf;
1548
1549	/* First, check for sparc-leon. */
1550	if (sparc_cpu_model == sparc_leon) {
1551		init_leon();
1552		return;
1553	}
1554
1555	/* Second, check for HyperSparc or Cypress. */
1556	if (mod_typ == 1) {
1557		switch (mod_rev) {
1558		case 7:
1559			/* UP or MP Hypersparc */
1560			init_hypersparc();
1561			break;
1562		case 0:
1563		case 2:
1564		case 10:
1565		case 11:
1566		case 12:
1567		case 13:
1568		case 14:
1569		case 15:
1570		default:
1571			prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1572			prom_halt();
1573			break;
1574		}
1575		return;
1576	}
1577
1578	/* Now Fujitsu TurboSparc. It might happen that it is
1579	 * in Swift emulation mode, so we will check later...
1580	 */
1581	if (psr_typ == 0 && psr_vers == 5) {
1582		init_turbosparc();
1583		return;
1584	}
1585
1586	/* Next check for Fujitsu Swift. */
1587	if (psr_typ == 0 && psr_vers == 4) {
1588		phandle cpunode;
1589		char node_str[128];
1590
1591		/* Look if it is not a TurboSparc emulating Swift... */
1592		cpunode = prom_getchild(prom_root_node);
1593		while ((cpunode = prom_getsibling(cpunode)) != 0) {
1594			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1595			if (!strcmp(node_str, "cpu")) {
1596				if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1597				    prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1598					init_turbosparc();
1599					return;
1600				}
1601				break;
1602			}
1603		}
1604
1605		init_swift();
1606		return;
1607	}
1608
1609	/* Now the Viking family of srmmu. */
1610	if (psr_typ == 4 &&
1611	   ((psr_vers == 0) ||
1612	    ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1613		init_viking();
1614		return;
1615	}
1616
1617	/* Finally the Tsunami. */
1618	if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1619		init_tsunami();
1620		return;
1621	}
1622
1623	/* Oh well */
1624	srmmu_is_bad();
1625}
1626
1627#ifdef CONFIG_SMP
1628/* Local cross-calls. */
1629static void smp_flush_page_for_dma(unsigned long page)
1630{
1631	xc1((smpfunc_t) local_ops->page_for_dma, page);
1632	local_ops->page_for_dma(page);
1633}
1634
1635static void smp_flush_cache_all(void)
1636{
1637	xc0((smpfunc_t) local_ops->cache_all);
1638	local_ops->cache_all();
1639}
1640
1641static void smp_flush_tlb_all(void)
1642{
1643	xc0((smpfunc_t) local_ops->tlb_all);
1644	local_ops->tlb_all();
1645}
1646
1647static void smp_flush_cache_mm(struct mm_struct *mm)
1648{
1649	if (mm->context != NO_CONTEXT) {
1650		cpumask_t cpu_mask;
1651		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1652		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1653		if (!cpumask_empty(&cpu_mask))
1654			xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1655		local_ops->cache_mm(mm);
1656	}
1657}
1658
1659static void smp_flush_tlb_mm(struct mm_struct *mm)
1660{
1661	if (mm->context != NO_CONTEXT) {
1662		cpumask_t cpu_mask;
1663		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1664		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1665		if (!cpumask_empty(&cpu_mask)) {
1666			xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1667			if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1668				cpumask_copy(mm_cpumask(mm),
1669					     cpumask_of(smp_processor_id()));
1670		}
1671		local_ops->tlb_mm(mm);
1672	}
1673}
1674
1675static void smp_flush_cache_range(struct vm_area_struct *vma,
1676				  unsigned long start,
1677				  unsigned long end)
1678{
1679	struct mm_struct *mm = vma->vm_mm;
1680
1681	if (mm->context != NO_CONTEXT) {
1682		cpumask_t cpu_mask;
1683		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1684		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1685		if (!cpumask_empty(&cpu_mask))
1686			xc3((smpfunc_t) local_ops->cache_range,
1687			    (unsigned long) vma, start, end);
1688		local_ops->cache_range(vma, start, end);
1689	}
1690}
1691
1692static void smp_flush_tlb_range(struct vm_area_struct *vma,
1693				unsigned long start,
1694				unsigned long end)
1695{
1696	struct mm_struct *mm = vma->vm_mm;
1697
1698	if (mm->context != NO_CONTEXT) {
1699		cpumask_t cpu_mask;
1700		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1701		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1702		if (!cpumask_empty(&cpu_mask))
1703			xc3((smpfunc_t) local_ops->tlb_range,
1704			    (unsigned long) vma, start, end);
1705		local_ops->tlb_range(vma, start, end);
1706	}
1707}
1708
1709static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1710{
1711	struct mm_struct *mm = vma->vm_mm;
1712
1713	if (mm->context != NO_CONTEXT) {
1714		cpumask_t cpu_mask;
1715		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1716		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1717		if (!cpumask_empty(&cpu_mask))
1718			xc2((smpfunc_t) local_ops->cache_page,
1719			    (unsigned long) vma, page);
1720		local_ops->cache_page(vma, page);
1721	}
1722}
1723
1724static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1725{
1726	struct mm_struct *mm = vma->vm_mm;
1727
1728	if (mm->context != NO_CONTEXT) {
1729		cpumask_t cpu_mask;
1730		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1731		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1732		if (!cpumask_empty(&cpu_mask))
1733			xc2((smpfunc_t) local_ops->tlb_page,
1734			    (unsigned long) vma, page);
1735		local_ops->tlb_page(vma, page);
1736	}
1737}
1738
1739static void smp_flush_page_to_ram(unsigned long page)
1740{
1741	/* Current theory is that those who call this are the one's
1742	 * who have just dirtied their cache with the pages contents
1743	 * in kernel space, therefore we only run this on local cpu.
1744	 *
1745	 * XXX This experiment failed, research further... -DaveM
1746	 */
1747#if 1
1748	xc1((smpfunc_t) local_ops->page_to_ram, page);
1749#endif
1750	local_ops->page_to_ram(page);
1751}
1752
1753static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1754{
1755	cpumask_t cpu_mask;
1756	cpumask_copy(&cpu_mask, mm_cpumask(mm));
1757	cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1758	if (!cpumask_empty(&cpu_mask))
1759		xc2((smpfunc_t) local_ops->sig_insns,
1760		    (unsigned long) mm, insn_addr);
1761	local_ops->sig_insns(mm, insn_addr);
1762}
1763
1764static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1765	.cache_all	= smp_flush_cache_all,
1766	.cache_mm	= smp_flush_cache_mm,
1767	.cache_page	= smp_flush_cache_page,
1768	.cache_range	= smp_flush_cache_range,
1769	.tlb_all	= smp_flush_tlb_all,
1770	.tlb_mm		= smp_flush_tlb_mm,
1771	.tlb_page	= smp_flush_tlb_page,
1772	.tlb_range	= smp_flush_tlb_range,
1773	.page_to_ram	= smp_flush_page_to_ram,
1774	.sig_insns	= smp_flush_sig_insns,
1775	.page_for_dma	= smp_flush_page_for_dma,
1776};
1777#endif
1778
1779/* Load up routines and constants for sun4m and sun4d mmu */
1780void __init load_mmu(void)
1781{
1782	/* Functions */
1783	get_srmmu_type();
1784
1785#ifdef CONFIG_SMP
1786	/* El switcheroo... */
1787	local_ops = sparc32_cachetlb_ops;
1788
1789	if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1790		smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1791		smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1792		smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1793		smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1794	}
1795
1796	if (poke_srmmu == poke_viking) {
1797		/* Avoid unnecessary cross calls. */
1798		smp_cachetlb_ops.cache_all = local_ops->cache_all;
1799		smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1800		smp_cachetlb_ops.cache_range = local_ops->cache_range;
1801		smp_cachetlb_ops.cache_page = local_ops->cache_page;
1802
1803		smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1804		smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1805		smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1806	}
1807
1808	/* It really is const after this point. */
1809	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1810		&smp_cachetlb_ops;
1811#endif
1812
1813	if (sparc_cpu_model == sun4d)
1814		ld_mmu_iounit();
1815	else
1816		ld_mmu_iommu();
1817#ifdef CONFIG_SMP
1818	if (sparc_cpu_model == sun4d)
1819		sun4d_init_smp();
1820	else if (sparc_cpu_model == sparc_leon)
1821		leon_init_smp();
1822	else
1823		sun4m_init_smp();
1824#endif
1825}
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * srmmu.c:  SRMMU specific routines for memory management.
   4 *
   5 * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
   6 * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
   7 * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
   8 * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   9 * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
  10 */
  11
  12#include <linux/seq_file.h>
  13#include <linux/spinlock.h>
  14#include <linux/memblock.h>
  15#include <linux/pagemap.h>
  16#include <linux/vmalloc.h>
  17#include <linux/kdebug.h>
  18#include <linux/export.h>
  19#include <linux/kernel.h>
  20#include <linux/init.h>
  21#include <linux/log2.h>
  22#include <linux/gfp.h>
  23#include <linux/fs.h>
  24#include <linux/mm.h>
  25
  26#include <asm/mmu_context.h>
  27#include <asm/cacheflush.h>
  28#include <asm/tlbflush.h>
  29#include <asm/io-unit.h>
  30#include <asm/pgalloc.h>
  31#include <asm/pgtable.h>
  32#include <asm/bitext.h>
  33#include <asm/vaddrs.h>
  34#include <asm/cache.h>
  35#include <asm/traps.h>
  36#include <asm/oplib.h>
  37#include <asm/mbus.h>
  38#include <asm/page.h>
  39#include <asm/asi.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
  54enum mbus_module srmmu_modtype;
  55static unsigned int hwbug_bitmask;
  56int vac_cache_size;
  57EXPORT_SYMBOL(vac_cache_size);
  58int vac_line_size;
  59
  60extern struct resource sparc_iomap;
  61
  62extern unsigned long last_valid_pfn;
  63
  64static pgd_t *srmmu_swapper_pg_dir;
  65
  66const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
  67EXPORT_SYMBOL(sparc32_cachetlb_ops);
  68
  69#ifdef CONFIG_SMP
  70const 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
  79int flush_page_for_dma_global = 1;
  80
  81char *srmmu_name;
  82
  83ctxd_t *srmmu_ctx_table_phys;
  84static ctxd_t *srmmu_context_table;
  85
  86int viking_mxcc_present;
  87static DEFINE_SPINLOCK(srmmu_context_spinlock);
  88
  89static int is_hypersparc;
  90
  91static int srmmu_cache_pagetables;
  92
  93/* these will be initialized in srmmu_nocache_calcsize() */
  94static unsigned long srmmu_nocache_size;
  95static 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
 103void *srmmu_nocache_pool;
 104static struct bit_map srmmu_nocache_map;
 105
 106static 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 */
 110static 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/*
 119 * Locations of MSI Registers.
 120 */
 121#define MSI_MBUS_ARBEN	0xe0001008	/* MBus Arbiter Enable register */
 122
 123/*
 124 * Useful bits in the MSI Registers.
 125 */
 126#define MSI_ASYNC_MODE  0x80000000	/* Operate the MSI asynchronously */
 127
 128static void msi_set_sync(void)
 129{
 130	__asm__ __volatile__ ("lda [%0] %1, %%g3\n\t"
 131			      "andn %%g3, %2, %%g3\n\t"
 132			      "sta %%g3, [%0] %1\n\t" : :
 133			      "r" (MSI_MBUS_ARBEN),
 134			      "i" (ASI_M_CTL), "r" (MSI_ASYNC_MODE) : "g3");
 135}
 136
 137void pmd_set(pmd_t *pmdp, pte_t *ptep)
 138{
 139	unsigned long ptp;	/* Physical address, shifted right by 4 */
 140	int i;
 141
 142	ptp = __nocache_pa(ptep) >> 4;
 143	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 144		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
 145		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
 146	}
 147}
 148
 149void pmd_populate(struct mm_struct *mm, pmd_t *pmdp, struct page *ptep)
 150{
 151	unsigned long ptp;	/* Physical address, shifted right by 4 */
 152	int i;
 153
 154	ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);	/* watch for overflow */
 155	for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 156		set_pte((pte_t *)&pmdp->pmdv[i], __pte(SRMMU_ET_PTD | ptp));
 157		ptp += (SRMMU_REAL_PTRS_PER_PTE * sizeof(pte_t) >> 4);
 158	}
 159}
 160
 161/* Find an entry in the third-level page table.. */
 162pte_t *pte_offset_kernel(pmd_t *dir, unsigned long address)
 163{
 164	void *pte;
 165
 166	pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
 167	return (pte_t *) pte +
 168	    ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 169}
 170
 171/*
 172 * size: bytes to allocate in the nocache area.
 173 * align: bytes, number to align at.
 174 * Returns the virtual address of the allocated area.
 175 */
 176static void *__srmmu_get_nocache(int size, int align)
 177{
 178	int offset;
 179	unsigned long addr;
 180
 181	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 182		printk(KERN_ERR "Size 0x%x too small for nocache request\n",
 183		       size);
 184		size = SRMMU_NOCACHE_BITMAP_SHIFT;
 185	}
 186	if (size & (SRMMU_NOCACHE_BITMAP_SHIFT - 1)) {
 187		printk(KERN_ERR "Size 0x%x unaligned int nocache request\n",
 188		       size);
 189		size += SRMMU_NOCACHE_BITMAP_SHIFT - 1;
 190	}
 191	BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
 192
 193	offset = bit_map_string_get(&srmmu_nocache_map,
 194				    size >> SRMMU_NOCACHE_BITMAP_SHIFT,
 195				    align >> SRMMU_NOCACHE_BITMAP_SHIFT);
 196	if (offset == -1) {
 197		printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
 198		       size, (int) srmmu_nocache_size,
 199		       srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 200		return NULL;
 201	}
 202
 203	addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
 204	return (void *)addr;
 205}
 206
 207void *srmmu_get_nocache(int size, int align)
 208{
 209	void *tmp;
 210
 211	tmp = __srmmu_get_nocache(size, align);
 212
 213	if (tmp)
 214		memset(tmp, 0, size);
 215
 216	return tmp;
 217}
 218
 219void srmmu_free_nocache(void *addr, int size)
 220{
 221	unsigned long vaddr;
 222	int offset;
 223
 224	vaddr = (unsigned long)addr;
 225	if (vaddr < SRMMU_NOCACHE_VADDR) {
 226		printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
 227		    vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
 228		BUG();
 229	}
 230	if (vaddr + size > srmmu_nocache_end) {
 231		printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
 232		    vaddr, srmmu_nocache_end);
 233		BUG();
 234	}
 235	if (!is_power_of_2(size)) {
 236		printk("Size 0x%x is not a power of 2\n", size);
 237		BUG();
 238	}
 239	if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 240		printk("Size 0x%x is too small\n", size);
 241		BUG();
 242	}
 243	if (vaddr & (size - 1)) {
 244		printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
 245		BUG();
 246	}
 247
 248	offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
 249	size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 250
 251	bit_map_clear(&srmmu_nocache_map, offset, size);
 252}
 253
 254static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
 255						 unsigned long end);
 256
 257/* Return how much physical memory we have.  */
 258static unsigned long __init probe_memory(void)
 259{
 260	unsigned long total = 0;
 261	int i;
 262
 263	for (i = 0; sp_banks[i].num_bytes; i++)
 264		total += sp_banks[i].num_bytes;
 265
 266	return total;
 267}
 268
 269/*
 270 * Reserve nocache dynamically proportionally to the amount of
 271 * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
 272 */
 273static void __init srmmu_nocache_calcsize(void)
 274{
 275	unsigned long sysmemavail = probe_memory() / 1024;
 276	int srmmu_nocache_npages;
 277
 278	srmmu_nocache_npages =
 279		sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
 280
 281 /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
 282	// if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
 283	if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
 284		srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
 285
 286	/* anything above 1280 blows up */
 287	if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
 288		srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
 289
 290	srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
 291	srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
 292}
 293
 294static void __init srmmu_nocache_init(void)
 295{
 296	void *srmmu_nocache_bitmap;
 297	unsigned int bitmap_bits;
 298	pgd_t *pgd;
 299	pmd_t *pmd;
 300	pte_t *pte;
 301	unsigned long paddr, vaddr;
 302	unsigned long pteval;
 303
 304	bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 305
 306	srmmu_nocache_pool = memblock_alloc(srmmu_nocache_size,
 307					    SRMMU_NOCACHE_ALIGN_MAX);
 308	if (!srmmu_nocache_pool)
 309		panic("%s: Failed to allocate %lu bytes align=0x%x\n",
 310		      __func__, srmmu_nocache_size, SRMMU_NOCACHE_ALIGN_MAX);
 311	memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
 312
 313	srmmu_nocache_bitmap =
 314		memblock_alloc(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
 315			       SMP_CACHE_BYTES);
 316	if (!srmmu_nocache_bitmap)
 317		panic("%s: Failed to allocate %zu bytes\n", __func__,
 318		      BITS_TO_LONGS(bitmap_bits) * sizeof(long));
 319	bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
 320
 321	srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 322	memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
 323	init_mm.pgd = srmmu_swapper_pg_dir;
 324
 325	srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
 326
 327	paddr = __pa((unsigned long)srmmu_nocache_pool);
 328	vaddr = SRMMU_NOCACHE_VADDR;
 329
 330	while (vaddr < srmmu_nocache_end) {
 331		pgd = pgd_offset_k(vaddr);
 332		pmd = pmd_offset(__nocache_fix(pgd), vaddr);
 333		pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
 334
 335		pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
 336
 337		if (srmmu_cache_pagetables)
 338			pteval |= SRMMU_CACHE;
 339
 340		set_pte(__nocache_fix(pte), __pte(pteval));
 341
 342		vaddr += PAGE_SIZE;
 343		paddr += PAGE_SIZE;
 344	}
 345
 346	flush_cache_all();
 347	flush_tlb_all();
 348}
 349
 350pgd_t *get_pgd_fast(void)
 351{
 352	pgd_t *pgd = NULL;
 353
 354	pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 355	if (pgd) {
 356		pgd_t *init = pgd_offset_k(0);
 357		memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
 358		memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
 359						(PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
 360	}
 361
 362	return pgd;
 363}
 364
 365/*
 366 * Hardware needs alignment to 256 only, but we align to whole page size
 367 * to reduce fragmentation problems due to the buddy principle.
 368 * XXX Provide actual fragmentation statistics in /proc.
 369 *
 370 * Alignments up to the page size are the same for physical and virtual
 371 * addresses of the nocache area.
 372 */
 373pgtable_t pte_alloc_one(struct mm_struct *mm)
 374{
 375	unsigned long pte;
 376	struct page *page;
 377
 378	if ((pte = (unsigned long)pte_alloc_one_kernel(mm)) == 0)
 379		return NULL;
 380	page = pfn_to_page(__nocache_pa(pte) >> PAGE_SHIFT);
 381	if (!pgtable_pte_page_ctor(page)) {
 382		__free_page(page);
 383		return NULL;
 384	}
 385	return page;
 386}
 387
 388void pte_free(struct mm_struct *mm, pgtable_t pte)
 389{
 390	unsigned long p;
 391
 392	pgtable_pte_page_dtor(pte);
 393	p = (unsigned long)page_address(pte);	/* Cached address (for test) */
 394	if (p == 0)
 395		BUG();
 396	p = page_to_pfn(pte) << PAGE_SHIFT;	/* Physical address */
 397
 398	/* free non cached virtual address*/
 399	srmmu_free_nocache(__nocache_va(p), PTE_SIZE);
 400}
 401
 402/* context handling - a dynamically sized pool is used */
 403#define NO_CONTEXT	-1
 404
 405struct ctx_list {
 406	struct ctx_list *next;
 407	struct ctx_list *prev;
 408	unsigned int ctx_number;
 409	struct mm_struct *ctx_mm;
 410};
 411
 412static struct ctx_list *ctx_list_pool;
 413static struct ctx_list ctx_free;
 414static struct ctx_list ctx_used;
 415
 416/* At boot time we determine the number of contexts */
 417static int num_contexts;
 418
 419static inline void remove_from_ctx_list(struct ctx_list *entry)
 420{
 421	entry->next->prev = entry->prev;
 422	entry->prev->next = entry->next;
 423}
 424
 425static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
 426{
 427	entry->next = head;
 428	(entry->prev = head->prev)->next = entry;
 429	head->prev = entry;
 430}
 431#define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
 432#define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
 433
 434
 435static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
 436{
 437	struct ctx_list *ctxp;
 438
 439	ctxp = ctx_free.next;
 440	if (ctxp != &ctx_free) {
 441		remove_from_ctx_list(ctxp);
 442		add_to_used_ctxlist(ctxp);
 443		mm->context = ctxp->ctx_number;
 444		ctxp->ctx_mm = mm;
 445		return;
 446	}
 447	ctxp = ctx_used.next;
 448	if (ctxp->ctx_mm == old_mm)
 449		ctxp = ctxp->next;
 450	if (ctxp == &ctx_used)
 451		panic("out of mmu contexts");
 452	flush_cache_mm(ctxp->ctx_mm);
 453	flush_tlb_mm(ctxp->ctx_mm);
 454	remove_from_ctx_list(ctxp);
 455	add_to_used_ctxlist(ctxp);
 456	ctxp->ctx_mm->context = NO_CONTEXT;
 457	ctxp->ctx_mm = mm;
 458	mm->context = ctxp->ctx_number;
 459}
 460
 461static inline void free_context(int context)
 462{
 463	struct ctx_list *ctx_old;
 464
 465	ctx_old = ctx_list_pool + context;
 466	remove_from_ctx_list(ctx_old);
 467	add_to_free_ctxlist(ctx_old);
 468}
 469
 470static void __init sparc_context_init(int numctx)
 471{
 472	int ctx;
 473	unsigned long size;
 474
 475	size = numctx * sizeof(struct ctx_list);
 476	ctx_list_pool = memblock_alloc(size, SMP_CACHE_BYTES);
 477	if (!ctx_list_pool)
 478		panic("%s: Failed to allocate %lu bytes\n", __func__, size);
 479
 480	for (ctx = 0; ctx < numctx; ctx++) {
 481		struct ctx_list *clist;
 482
 483		clist = (ctx_list_pool + ctx);
 484		clist->ctx_number = ctx;
 485		clist->ctx_mm = NULL;
 486	}
 487	ctx_free.next = ctx_free.prev = &ctx_free;
 488	ctx_used.next = ctx_used.prev = &ctx_used;
 489	for (ctx = 0; ctx < numctx; ctx++)
 490		add_to_free_ctxlist(ctx_list_pool + ctx);
 491}
 492
 493void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
 494	       struct task_struct *tsk)
 495{
 496	unsigned long flags;
 497
 498	if (mm->context == NO_CONTEXT) {
 499		spin_lock_irqsave(&srmmu_context_spinlock, flags);
 500		alloc_context(old_mm, mm);
 501		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
 502		srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
 503	}
 504
 505	if (sparc_cpu_model == sparc_leon)
 506		leon_switch_mm();
 507
 508	if (is_hypersparc)
 509		hyper_flush_whole_icache();
 510
 511	srmmu_set_context(mm->context);
 512}
 513
 514/* Low level IO area allocation on the SRMMU. */
 515static inline void srmmu_mapioaddr(unsigned long physaddr,
 516				   unsigned long virt_addr, int bus_type)
 517{
 518	pgd_t *pgdp;
 519	pmd_t *pmdp;
 520	pte_t *ptep;
 521	unsigned long tmp;
 522
 523	physaddr &= PAGE_MASK;
 524	pgdp = pgd_offset_k(virt_addr);
 525	pmdp = pmd_offset(pgdp, virt_addr);
 526	ptep = pte_offset_kernel(pmdp, virt_addr);
 527	tmp = (physaddr >> 4) | SRMMU_ET_PTE;
 528
 529	/* I need to test whether this is consistent over all
 530	 * sun4m's.  The bus_type represents the upper 4 bits of
 531	 * 36-bit physical address on the I/O space lines...
 532	 */
 533	tmp |= (bus_type << 28);
 534	tmp |= SRMMU_PRIV;
 535	__flush_page_to_ram(virt_addr);
 536	set_pte(ptep, __pte(tmp));
 537}
 538
 539void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
 540		      unsigned long xva, unsigned int len)
 541{
 542	while (len != 0) {
 543		len -= PAGE_SIZE;
 544		srmmu_mapioaddr(xpa, xva, bus);
 545		xva += PAGE_SIZE;
 546		xpa += PAGE_SIZE;
 547	}
 548	flush_tlb_all();
 549}
 550
 551static inline void srmmu_unmapioaddr(unsigned long virt_addr)
 552{
 553	pgd_t *pgdp;
 554	pmd_t *pmdp;
 555	pte_t *ptep;
 556
 557	pgdp = pgd_offset_k(virt_addr);
 558	pmdp = pmd_offset(pgdp, virt_addr);
 559	ptep = pte_offset_kernel(pmdp, virt_addr);
 560
 561	/* No need to flush uncacheable page. */
 562	__pte_clear(ptep);
 563}
 564
 565void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
 566{
 567	while (len != 0) {
 568		len -= PAGE_SIZE;
 569		srmmu_unmapioaddr(virt_addr);
 570		virt_addr += PAGE_SIZE;
 571	}
 572	flush_tlb_all();
 573}
 574
 575/* tsunami.S */
 576extern void tsunami_flush_cache_all(void);
 577extern void tsunami_flush_cache_mm(struct mm_struct *mm);
 578extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 579extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 580extern void tsunami_flush_page_to_ram(unsigned long page);
 581extern void tsunami_flush_page_for_dma(unsigned long page);
 582extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 583extern void tsunami_flush_tlb_all(void);
 584extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
 585extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 586extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 587extern void tsunami_setup_blockops(void);
 588
 589/* swift.S */
 590extern void swift_flush_cache_all(void);
 591extern void swift_flush_cache_mm(struct mm_struct *mm);
 592extern void swift_flush_cache_range(struct vm_area_struct *vma,
 593				    unsigned long start, unsigned long end);
 594extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 595extern void swift_flush_page_to_ram(unsigned long page);
 596extern void swift_flush_page_for_dma(unsigned long page);
 597extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 598extern void swift_flush_tlb_all(void);
 599extern void swift_flush_tlb_mm(struct mm_struct *mm);
 600extern void swift_flush_tlb_range(struct vm_area_struct *vma,
 601				  unsigned long start, unsigned long end);
 602extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 603
 604#if 0  /* P3: deadwood to debug precise flushes on Swift. */
 605void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 606{
 607	int cctx, ctx1;
 608
 609	page &= PAGE_MASK;
 610	if ((ctx1 = vma->vm_mm->context) != -1) {
 611		cctx = srmmu_get_context();
 612/* Is context # ever different from current context? P3 */
 613		if (cctx != ctx1) {
 614			printk("flush ctx %02x curr %02x\n", ctx1, cctx);
 615			srmmu_set_context(ctx1);
 616			swift_flush_page(page);
 617			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 618					"r" (page), "i" (ASI_M_FLUSH_PROBE));
 619			srmmu_set_context(cctx);
 620		} else {
 621			 /* Rm. prot. bits from virt. c. */
 622			/* swift_flush_cache_all(); */
 623			/* swift_flush_cache_page(vma, page); */
 624			swift_flush_page(page);
 625
 626			__asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 627				"r" (page), "i" (ASI_M_FLUSH_PROBE));
 628			/* same as above: srmmu_flush_tlb_page() */
 629		}
 630	}
 631}
 632#endif
 633
 634/*
 635 * The following are all MBUS based SRMMU modules, and therefore could
 636 * be found in a multiprocessor configuration.  On the whole, these
 637 * chips seems to be much more touchy about DVMA and page tables
 638 * with respect to cache coherency.
 639 */
 640
 641/* viking.S */
 642extern void viking_flush_cache_all(void);
 643extern void viking_flush_cache_mm(struct mm_struct *mm);
 644extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 645				     unsigned long end);
 646extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 647extern void viking_flush_page_to_ram(unsigned long page);
 648extern void viking_flush_page_for_dma(unsigned long page);
 649extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
 650extern void viking_flush_page(unsigned long page);
 651extern void viking_mxcc_flush_page(unsigned long page);
 652extern void viking_flush_tlb_all(void);
 653extern void viking_flush_tlb_mm(struct mm_struct *mm);
 654extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 655				   unsigned long end);
 656extern void viking_flush_tlb_page(struct vm_area_struct *vma,
 657				  unsigned long page);
 658extern void sun4dsmp_flush_tlb_all(void);
 659extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
 660extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 661				   unsigned long end);
 662extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
 663				  unsigned long page);
 664
 665/* hypersparc.S */
 666extern void hypersparc_flush_cache_all(void);
 667extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
 668extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 669extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 670extern void hypersparc_flush_page_to_ram(unsigned long page);
 671extern void hypersparc_flush_page_for_dma(unsigned long page);
 672extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 673extern void hypersparc_flush_tlb_all(void);
 674extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
 675extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 676extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 677extern void hypersparc_setup_blockops(void);
 678
 679/*
 680 * NOTE: All of this startup code assumes the low 16mb (approx.) of
 681 *       kernel mappings are done with one single contiguous chunk of
 682 *       ram.  On small ram machines (classics mainly) we only get
 683 *       around 8mb mapped for us.
 684 */
 685
 686static void __init early_pgtable_allocfail(char *type)
 687{
 688	prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
 689	prom_halt();
 690}
 691
 692static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
 693							unsigned long end)
 694{
 695	pgd_t *pgdp;
 696	pmd_t *pmdp;
 697	pte_t *ptep;
 698
 699	while (start < end) {
 700		pgdp = pgd_offset_k(start);
 701		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
 702			pmdp = __srmmu_get_nocache(
 703			    SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 704			if (pmdp == NULL)
 705				early_pgtable_allocfail("pmd");
 706			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 707			pgd_set(__nocache_fix(pgdp), pmdp);
 708		}
 709		pmdp = pmd_offset(__nocache_fix(pgdp), start);
 710		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 711			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 712			if (ptep == NULL)
 713				early_pgtable_allocfail("pte");
 714			memset(__nocache_fix(ptep), 0, PTE_SIZE);
 715			pmd_set(__nocache_fix(pmdp), ptep);
 716		}
 717		if (start > (0xffffffffUL - PMD_SIZE))
 718			break;
 719		start = (start + PMD_SIZE) & PMD_MASK;
 720	}
 721}
 722
 723static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
 724						  unsigned long end)
 725{
 726	pgd_t *pgdp;
 727	pmd_t *pmdp;
 728	pte_t *ptep;
 729
 730	while (start < end) {
 731		pgdp = pgd_offset_k(start);
 732		if (pgd_none(*pgdp)) {
 733			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 734			if (pmdp == NULL)
 735				early_pgtable_allocfail("pmd");
 736			memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
 737			pgd_set(pgdp, pmdp);
 738		}
 739		pmdp = pmd_offset(pgdp, start);
 740		if (srmmu_pmd_none(*pmdp)) {
 741			ptep = __srmmu_get_nocache(PTE_SIZE,
 742							     PTE_SIZE);
 743			if (ptep == NULL)
 744				early_pgtable_allocfail("pte");
 745			memset(ptep, 0, PTE_SIZE);
 746			pmd_set(pmdp, ptep);
 747		}
 748		if (start > (0xffffffffUL - PMD_SIZE))
 749			break;
 750		start = (start + PMD_SIZE) & PMD_MASK;
 751	}
 752}
 753
 754/* These flush types are not available on all chips... */
 755static inline unsigned long srmmu_probe(unsigned long vaddr)
 756{
 757	unsigned long retval;
 758
 759	if (sparc_cpu_model != sparc_leon) {
 760
 761		vaddr &= PAGE_MASK;
 762		__asm__ __volatile__("lda [%1] %2, %0\n\t" :
 763				     "=r" (retval) :
 764				     "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
 765	} else {
 766		retval = leon_swprobe(vaddr, NULL);
 767	}
 768	return retval;
 769}
 770
 771/*
 772 * This is much cleaner than poking around physical address space
 773 * looking at the prom's page table directly which is what most
 774 * other OS's do.  Yuck... this is much better.
 775 */
 776static void __init srmmu_inherit_prom_mappings(unsigned long start,
 777					       unsigned long end)
 778{
 779	unsigned long probed;
 780	unsigned long addr;
 781	pgd_t *pgdp;
 782	pmd_t *pmdp;
 783	pte_t *ptep;
 784	int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
 785
 786	while (start <= end) {
 787		if (start == 0)
 788			break; /* probably wrap around */
 789		if (start == 0xfef00000)
 790			start = KADB_DEBUGGER_BEGVM;
 791		probed = srmmu_probe(start);
 792		if (!probed) {
 793			/* continue probing until we find an entry */
 794			start += PAGE_SIZE;
 795			continue;
 796		}
 797
 798		/* A red snapper, see what it really is. */
 799		what = 0;
 800		addr = start - PAGE_SIZE;
 801
 802		if (!(start & ~(SRMMU_REAL_PMD_MASK))) {
 803			if (srmmu_probe(addr + SRMMU_REAL_PMD_SIZE) == probed)
 804				what = 1;
 805		}
 806
 807		if (!(start & ~(SRMMU_PGDIR_MASK))) {
 808			if (srmmu_probe(addr + SRMMU_PGDIR_SIZE) == probed)
 809				what = 2;
 810		}
 811
 812		pgdp = pgd_offset_k(start);
 813		if (what == 2) {
 814			*(pgd_t *)__nocache_fix(pgdp) = __pgd(probed);
 815			start += SRMMU_PGDIR_SIZE;
 816			continue;
 817		}
 818		if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
 819			pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
 820						   SRMMU_PMD_TABLE_SIZE);
 821			if (pmdp == NULL)
 822				early_pgtable_allocfail("pmd");
 823			memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 824			pgd_set(__nocache_fix(pgdp), pmdp);
 825		}
 826		pmdp = pmd_offset(__nocache_fix(pgdp), start);
 827		if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 828			ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 829			if (ptep == NULL)
 830				early_pgtable_allocfail("pte");
 831			memset(__nocache_fix(ptep), 0, PTE_SIZE);
 832			pmd_set(__nocache_fix(pmdp), ptep);
 833		}
 834		if (what == 1) {
 835			/* We bend the rule where all 16 PTPs in a pmd_t point
 836			 * inside the same PTE page, and we leak a perfectly
 837			 * good hardware PTE piece. Alternatives seem worse.
 838			 */
 839			unsigned int x;	/* Index of HW PMD in soft cluster */
 840			unsigned long *val;
 841			x = (start >> PMD_SHIFT) & 15;
 842			val = &pmdp->pmdv[x];
 843			*(unsigned long *)__nocache_fix(val) = probed;
 844			start += SRMMU_REAL_PMD_SIZE;
 845			continue;
 846		}
 847		ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
 848		*(pte_t *)__nocache_fix(ptep) = __pte(probed);
 849		start += PAGE_SIZE;
 850	}
 851}
 852
 853#define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
 854
 855/* Create a third-level SRMMU 16MB page mapping. */
 856static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
 857{
 858	pgd_t *pgdp = pgd_offset_k(vaddr);
 859	unsigned long big_pte;
 860
 861	big_pte = KERNEL_PTE(phys_base >> 4);
 862	*(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
 863}
 864
 865/* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
 866static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
 867{
 868	unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
 869	unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
 870	unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
 871	/* Map "low" memory only */
 872	const unsigned long min_vaddr = PAGE_OFFSET;
 873	const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
 874
 875	if (vstart < min_vaddr || vstart >= max_vaddr)
 876		return vstart;
 877
 878	if (vend > max_vaddr || vend < min_vaddr)
 879		vend = max_vaddr;
 880
 881	while (vstart < vend) {
 882		do_large_mapping(vstart, pstart);
 883		vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
 884	}
 885	return vstart;
 886}
 887
 888static void __init map_kernel(void)
 889{
 890	int i;
 891
 892	if (phys_base > 0) {
 893		do_large_mapping(PAGE_OFFSET, phys_base);
 894	}
 895
 896	for (i = 0; sp_banks[i].num_bytes != 0; i++) {
 897		map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
 898	}
 899}
 900
 901void (*poke_srmmu)(void) = NULL;
 902
 903void __init srmmu_paging_init(void)
 904{
 905	int i;
 906	phandle cpunode;
 907	char node_str[128];
 908	pgd_t *pgd;
 909	pmd_t *pmd;
 910	pte_t *pte;
 911	unsigned long pages_avail;
 912
 913	init_mm.context = (unsigned long) NO_CONTEXT;
 914	sparc_iomap.start = SUN4M_IOBASE_VADDR;	/* 16MB of IOSPACE on all sun4m's. */
 915
 916	if (sparc_cpu_model == sun4d)
 917		num_contexts = 65536; /* We know it is Viking */
 918	else {
 919		/* Find the number of contexts on the srmmu. */
 920		cpunode = prom_getchild(prom_root_node);
 921		num_contexts = 0;
 922		while (cpunode != 0) {
 923			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
 924			if (!strcmp(node_str, "cpu")) {
 925				num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
 926				break;
 927			}
 928			cpunode = prom_getsibling(cpunode);
 929		}
 930	}
 931
 932	if (!num_contexts) {
 933		prom_printf("Something wrong, can't find cpu node in paging_init.\n");
 934		prom_halt();
 935	}
 936
 937	pages_avail = 0;
 938	last_valid_pfn = bootmem_init(&pages_avail);
 939
 940	srmmu_nocache_calcsize();
 941	srmmu_nocache_init();
 942	srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
 943	map_kernel();
 944
 945	/* ctx table has to be physically aligned to its size */
 946	srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
 947	srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
 948
 949	for (i = 0; i < num_contexts; i++)
 950		srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
 951
 952	flush_cache_all();
 953	srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
 954#ifdef CONFIG_SMP
 955	/* Stop from hanging here... */
 956	local_ops->tlb_all();
 957#else
 958	flush_tlb_all();
 959#endif
 960	poke_srmmu();
 961
 962	srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
 963	srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
 964
 965	srmmu_allocate_ptable_skeleton(
 966		__fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
 967	srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
 968
 969	pgd = pgd_offset_k(PKMAP_BASE);
 970	pmd = pmd_offset(pgd, PKMAP_BASE);
 971	pte = pte_offset_kernel(pmd, PKMAP_BASE);
 972	pkmap_page_table = pte;
 973
 974	flush_cache_all();
 975	flush_tlb_all();
 976
 977	sparc_context_init(num_contexts);
 978
 979	kmap_init();
 980
 981	{
 982		unsigned long zones_size[MAX_NR_ZONES];
 983		unsigned long zholes_size[MAX_NR_ZONES];
 984		unsigned long npages;
 985		int znum;
 986
 987		for (znum = 0; znum < MAX_NR_ZONES; znum++)
 988			zones_size[znum] = zholes_size[znum] = 0;
 989
 990		npages = max_low_pfn - pfn_base;
 991
 992		zones_size[ZONE_DMA] = npages;
 993		zholes_size[ZONE_DMA] = npages - pages_avail;
 994
 995		npages = highend_pfn - max_low_pfn;
 996		zones_size[ZONE_HIGHMEM] = npages;
 997		zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
 998
 999		free_area_init_node(0, zones_size, pfn_base, zholes_size);
1000	}
1001}
1002
1003void mmu_info(struct seq_file *m)
1004{
1005	seq_printf(m,
1006		   "MMU type\t: %s\n"
1007		   "contexts\t: %d\n"
1008		   "nocache total\t: %ld\n"
1009		   "nocache used\t: %d\n",
1010		   srmmu_name,
1011		   num_contexts,
1012		   srmmu_nocache_size,
1013		   srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
1014}
1015
1016int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
1017{
1018	mm->context = NO_CONTEXT;
1019	return 0;
1020}
1021
1022void destroy_context(struct mm_struct *mm)
1023{
1024	unsigned long flags;
1025
1026	if (mm->context != NO_CONTEXT) {
1027		flush_cache_mm(mm);
1028		srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1029		flush_tlb_mm(mm);
1030		spin_lock_irqsave(&srmmu_context_spinlock, flags);
1031		free_context(mm->context);
1032		spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1033		mm->context = NO_CONTEXT;
1034	}
1035}
1036
1037/* Init various srmmu chip types. */
1038static void __init srmmu_is_bad(void)
1039{
1040	prom_printf("Could not determine SRMMU chip type.\n");
1041	prom_halt();
1042}
1043
1044static void __init init_vac_layout(void)
1045{
1046	phandle nd;
1047	int cache_lines;
1048	char node_str[128];
1049#ifdef CONFIG_SMP
1050	int cpu = 0;
1051	unsigned long max_size = 0;
1052	unsigned long min_line_size = 0x10000000;
1053#endif
1054
1055	nd = prom_getchild(prom_root_node);
1056	while ((nd = prom_getsibling(nd)) != 0) {
1057		prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1058		if (!strcmp(node_str, "cpu")) {
1059			vac_line_size = prom_getint(nd, "cache-line-size");
1060			if (vac_line_size == -1) {
1061				prom_printf("can't determine cache-line-size, halting.\n");
1062				prom_halt();
1063			}
1064			cache_lines = prom_getint(nd, "cache-nlines");
1065			if (cache_lines == -1) {
1066				prom_printf("can't determine cache-nlines, halting.\n");
1067				prom_halt();
1068			}
1069
1070			vac_cache_size = cache_lines * vac_line_size;
1071#ifdef CONFIG_SMP
1072			if (vac_cache_size > max_size)
1073				max_size = vac_cache_size;
1074			if (vac_line_size < min_line_size)
1075				min_line_size = vac_line_size;
1076			//FIXME: cpus not contiguous!!
1077			cpu++;
1078			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1079				break;
1080#else
1081			break;
1082#endif
1083		}
1084	}
1085	if (nd == 0) {
1086		prom_printf("No CPU nodes found, halting.\n");
1087		prom_halt();
1088	}
1089#ifdef CONFIG_SMP
1090	vac_cache_size = max_size;
1091	vac_line_size = min_line_size;
1092#endif
1093	printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1094	       (int)vac_cache_size, (int)vac_line_size);
1095}
1096
1097static void poke_hypersparc(void)
1098{
1099	volatile unsigned long clear;
1100	unsigned long mreg = srmmu_get_mmureg();
1101
1102	hyper_flush_unconditional_combined();
1103
1104	mreg &= ~(HYPERSPARC_CWENABLE);
1105	mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1106	mreg |= (HYPERSPARC_CMODE);
1107
1108	srmmu_set_mmureg(mreg);
1109
1110#if 0 /* XXX I think this is bad news... -DaveM */
1111	hyper_clear_all_tags();
1112#endif
1113
1114	put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1115	hyper_flush_whole_icache();
1116	clear = srmmu_get_faddr();
1117	clear = srmmu_get_fstatus();
1118}
1119
1120static const struct sparc32_cachetlb_ops hypersparc_ops = {
1121	.cache_all	= hypersparc_flush_cache_all,
1122	.cache_mm	= hypersparc_flush_cache_mm,
1123	.cache_page	= hypersparc_flush_cache_page,
1124	.cache_range	= hypersparc_flush_cache_range,
1125	.tlb_all	= hypersparc_flush_tlb_all,
1126	.tlb_mm		= hypersparc_flush_tlb_mm,
1127	.tlb_page	= hypersparc_flush_tlb_page,
1128	.tlb_range	= hypersparc_flush_tlb_range,
1129	.page_to_ram	= hypersparc_flush_page_to_ram,
1130	.sig_insns	= hypersparc_flush_sig_insns,
1131	.page_for_dma	= hypersparc_flush_page_for_dma,
1132};
1133
1134static void __init init_hypersparc(void)
1135{
1136	srmmu_name = "ROSS HyperSparc";
1137	srmmu_modtype = HyperSparc;
1138
1139	init_vac_layout();
1140
1141	is_hypersparc = 1;
1142	sparc32_cachetlb_ops = &hypersparc_ops;
1143
1144	poke_srmmu = poke_hypersparc;
1145
1146	hypersparc_setup_blockops();
1147}
1148
1149static void poke_swift(void)
1150{
1151	unsigned long mreg;
1152
1153	/* Clear any crap from the cache or else... */
1154	swift_flush_cache_all();
1155
1156	/* Enable I & D caches */
1157	mreg = srmmu_get_mmureg();
1158	mreg |= (SWIFT_IE | SWIFT_DE);
1159	/*
1160	 * The Swift branch folding logic is completely broken.  At
1161	 * trap time, if things are just right, if can mistakenly
1162	 * think that a trap is coming from kernel mode when in fact
1163	 * it is coming from user mode (it mis-executes the branch in
1164	 * the trap code).  So you see things like crashme completely
1165	 * hosing your machine which is completely unacceptable.  Turn
1166	 * this shit off... nice job Fujitsu.
1167	 */
1168	mreg &= ~(SWIFT_BF);
1169	srmmu_set_mmureg(mreg);
1170}
1171
1172static const struct sparc32_cachetlb_ops swift_ops = {
1173	.cache_all	= swift_flush_cache_all,
1174	.cache_mm	= swift_flush_cache_mm,
1175	.cache_page	= swift_flush_cache_page,
1176	.cache_range	= swift_flush_cache_range,
1177	.tlb_all	= swift_flush_tlb_all,
1178	.tlb_mm		= swift_flush_tlb_mm,
1179	.tlb_page	= swift_flush_tlb_page,
1180	.tlb_range	= swift_flush_tlb_range,
1181	.page_to_ram	= swift_flush_page_to_ram,
1182	.sig_insns	= swift_flush_sig_insns,
1183	.page_for_dma	= swift_flush_page_for_dma,
1184};
1185
1186#define SWIFT_MASKID_ADDR  0x10003018
1187static void __init init_swift(void)
1188{
1189	unsigned long swift_rev;
1190
1191	__asm__ __volatile__("lda [%1] %2, %0\n\t"
1192			     "srl %0, 0x18, %0\n\t" :
1193			     "=r" (swift_rev) :
1194			     "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1195	srmmu_name = "Fujitsu Swift";
1196	switch (swift_rev) {
1197	case 0x11:
1198	case 0x20:
1199	case 0x23:
1200	case 0x30:
1201		srmmu_modtype = Swift_lots_o_bugs;
1202		hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1203		/*
1204		 * Gee george, I wonder why Sun is so hush hush about
1205		 * this hardware bug... really braindamage stuff going
1206		 * on here.  However I think we can find a way to avoid
1207		 * all of the workaround overhead under Linux.  Basically,
1208		 * any page fault can cause kernel pages to become user
1209		 * accessible (the mmu gets confused and clears some of
1210		 * the ACC bits in kernel ptes).  Aha, sounds pretty
1211		 * horrible eh?  But wait, after extensive testing it appears
1212		 * that if you use pgd_t level large kernel pte's (like the
1213		 * 4MB pages on the Pentium) the bug does not get tripped
1214		 * at all.  This avoids almost all of the major overhead.
1215		 * Welcome to a world where your vendor tells you to,
1216		 * "apply this kernel patch" instead of "sorry for the
1217		 * broken hardware, send it back and we'll give you
1218		 * properly functioning parts"
1219		 */
1220		break;
1221	case 0x25:
1222	case 0x31:
1223		srmmu_modtype = Swift_bad_c;
1224		hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1225		/*
1226		 * You see Sun allude to this hardware bug but never
1227		 * admit things directly, they'll say things like,
1228		 * "the Swift chip cache problems" or similar.
1229		 */
1230		break;
1231	default:
1232		srmmu_modtype = Swift_ok;
1233		break;
1234	}
1235
1236	sparc32_cachetlb_ops = &swift_ops;
1237	flush_page_for_dma_global = 0;
1238
1239	/*
1240	 * Are you now convinced that the Swift is one of the
1241	 * biggest VLSI abortions of all time?  Bravo Fujitsu!
1242	 * Fujitsu, the !#?!%$'d up processor people.  I bet if
1243	 * you examined the microcode of the Swift you'd find
1244	 * XXX's all over the place.
1245	 */
1246	poke_srmmu = poke_swift;
1247}
1248
1249static void turbosparc_flush_cache_all(void)
1250{
1251	flush_user_windows();
1252	turbosparc_idflash_clear();
1253}
1254
1255static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1256{
1257	FLUSH_BEGIN(mm)
1258	flush_user_windows();
1259	turbosparc_idflash_clear();
1260	FLUSH_END
1261}
1262
1263static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1264{
1265	FLUSH_BEGIN(vma->vm_mm)
1266	flush_user_windows();
1267	turbosparc_idflash_clear();
1268	FLUSH_END
1269}
1270
1271static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1272{
1273	FLUSH_BEGIN(vma->vm_mm)
1274	flush_user_windows();
1275	if (vma->vm_flags & VM_EXEC)
1276		turbosparc_flush_icache();
1277	turbosparc_flush_dcache();
1278	FLUSH_END
1279}
1280
1281/* TurboSparc is copy-back, if we turn it on, but this does not work. */
1282static void turbosparc_flush_page_to_ram(unsigned long page)
1283{
1284#ifdef TURBOSPARC_WRITEBACK
1285	volatile unsigned long clear;
1286
1287	if (srmmu_probe(page))
1288		turbosparc_flush_page_cache(page);
1289	clear = srmmu_get_fstatus();
1290#endif
1291}
1292
1293static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1294{
1295}
1296
1297static void turbosparc_flush_page_for_dma(unsigned long page)
1298{
1299	turbosparc_flush_dcache();
1300}
1301
1302static void turbosparc_flush_tlb_all(void)
1303{
1304	srmmu_flush_whole_tlb();
1305}
1306
1307static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1308{
1309	FLUSH_BEGIN(mm)
1310	srmmu_flush_whole_tlb();
1311	FLUSH_END
1312}
1313
1314static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1315{
1316	FLUSH_BEGIN(vma->vm_mm)
1317	srmmu_flush_whole_tlb();
1318	FLUSH_END
1319}
1320
1321static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1322{
1323	FLUSH_BEGIN(vma->vm_mm)
1324	srmmu_flush_whole_tlb();
1325	FLUSH_END
1326}
1327
1328
1329static void poke_turbosparc(void)
1330{
1331	unsigned long mreg = srmmu_get_mmureg();
1332	unsigned long ccreg;
1333
1334	/* Clear any crap from the cache or else... */
1335	turbosparc_flush_cache_all();
1336	/* Temporarily disable I & D caches */
1337	mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1338	mreg &= ~(TURBOSPARC_PCENABLE);		/* Don't check parity */
1339	srmmu_set_mmureg(mreg);
1340
1341	ccreg = turbosparc_get_ccreg();
1342
1343#ifdef TURBOSPARC_WRITEBACK
1344	ccreg |= (TURBOSPARC_SNENABLE);		/* Do DVMA snooping in Dcache */
1345	ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1346			/* Write-back D-cache, emulate VLSI
1347			 * abortion number three, not number one */
1348#else
1349	/* For now let's play safe, optimize later */
1350	ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1351			/* Do DVMA snooping in Dcache, Write-thru D-cache */
1352	ccreg &= ~(TURBOSPARC_uS2);
1353			/* Emulate VLSI abortion number three, not number one */
1354#endif
1355
1356	switch (ccreg & 7) {
1357	case 0: /* No SE cache */
1358	case 7: /* Test mode */
1359		break;
1360	default:
1361		ccreg |= (TURBOSPARC_SCENABLE);
1362	}
1363	turbosparc_set_ccreg(ccreg);
1364
1365	mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1366	mreg |= (TURBOSPARC_ICSNOOP);		/* Icache snooping on */
1367	srmmu_set_mmureg(mreg);
1368}
1369
1370static const struct sparc32_cachetlb_ops turbosparc_ops = {
1371	.cache_all	= turbosparc_flush_cache_all,
1372	.cache_mm	= turbosparc_flush_cache_mm,
1373	.cache_page	= turbosparc_flush_cache_page,
1374	.cache_range	= turbosparc_flush_cache_range,
1375	.tlb_all	= turbosparc_flush_tlb_all,
1376	.tlb_mm		= turbosparc_flush_tlb_mm,
1377	.tlb_page	= turbosparc_flush_tlb_page,
1378	.tlb_range	= turbosparc_flush_tlb_range,
1379	.page_to_ram	= turbosparc_flush_page_to_ram,
1380	.sig_insns	= turbosparc_flush_sig_insns,
1381	.page_for_dma	= turbosparc_flush_page_for_dma,
1382};
1383
1384static void __init init_turbosparc(void)
1385{
1386	srmmu_name = "Fujitsu TurboSparc";
1387	srmmu_modtype = TurboSparc;
1388	sparc32_cachetlb_ops = &turbosparc_ops;
1389	poke_srmmu = poke_turbosparc;
1390}
1391
1392static void poke_tsunami(void)
1393{
1394	unsigned long mreg = srmmu_get_mmureg();
1395
1396	tsunami_flush_icache();
1397	tsunami_flush_dcache();
1398	mreg &= ~TSUNAMI_ITD;
1399	mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1400	srmmu_set_mmureg(mreg);
1401}
1402
1403static const struct sparc32_cachetlb_ops tsunami_ops = {
1404	.cache_all	= tsunami_flush_cache_all,
1405	.cache_mm	= tsunami_flush_cache_mm,
1406	.cache_page	= tsunami_flush_cache_page,
1407	.cache_range	= tsunami_flush_cache_range,
1408	.tlb_all	= tsunami_flush_tlb_all,
1409	.tlb_mm		= tsunami_flush_tlb_mm,
1410	.tlb_page	= tsunami_flush_tlb_page,
1411	.tlb_range	= tsunami_flush_tlb_range,
1412	.page_to_ram	= tsunami_flush_page_to_ram,
1413	.sig_insns	= tsunami_flush_sig_insns,
1414	.page_for_dma	= tsunami_flush_page_for_dma,
1415};
1416
1417static void __init init_tsunami(void)
1418{
1419	/*
1420	 * Tsunami's pretty sane, Sun and TI actually got it
1421	 * somewhat right this time.  Fujitsu should have
1422	 * taken some lessons from them.
1423	 */
1424
1425	srmmu_name = "TI Tsunami";
1426	srmmu_modtype = Tsunami;
1427	sparc32_cachetlb_ops = &tsunami_ops;
1428	poke_srmmu = poke_tsunami;
1429
1430	tsunami_setup_blockops();
1431}
1432
1433static void poke_viking(void)
1434{
1435	unsigned long mreg = srmmu_get_mmureg();
1436	static int smp_catch;
1437
1438	if (viking_mxcc_present) {
1439		unsigned long mxcc_control = mxcc_get_creg();
1440
1441		mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1442		mxcc_control &= ~(MXCC_CTL_RRC);
1443		mxcc_set_creg(mxcc_control);
1444
1445		/*
1446		 * We don't need memory parity checks.
1447		 * XXX This is a mess, have to dig out later. ecd.
1448		viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1449		 */
1450
1451		/* We do cache ptables on MXCC. */
1452		mreg |= VIKING_TCENABLE;
1453	} else {
1454		unsigned long bpreg;
1455
1456		mreg &= ~(VIKING_TCENABLE);
1457		if (smp_catch++) {
1458			/* Must disable mixed-cmd mode here for other cpu's. */
1459			bpreg = viking_get_bpreg();
1460			bpreg &= ~(VIKING_ACTION_MIX);
1461			viking_set_bpreg(bpreg);
1462
1463			/* Just in case PROM does something funny. */
1464			msi_set_sync();
1465		}
1466	}
1467
1468	mreg |= VIKING_SPENABLE;
1469	mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1470	mreg |= VIKING_SBENABLE;
1471	mreg &= ~(VIKING_ACENABLE);
1472	srmmu_set_mmureg(mreg);
1473}
1474
1475static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1476	.cache_all	= viking_flush_cache_all,
1477	.cache_mm	= viking_flush_cache_mm,
1478	.cache_page	= viking_flush_cache_page,
1479	.cache_range	= viking_flush_cache_range,
1480	.tlb_all	= viking_flush_tlb_all,
1481	.tlb_mm		= viking_flush_tlb_mm,
1482	.tlb_page	= viking_flush_tlb_page,
1483	.tlb_range	= viking_flush_tlb_range,
1484	.page_to_ram	= viking_flush_page_to_ram,
1485	.sig_insns	= viking_flush_sig_insns,
1486	.page_for_dma	= viking_flush_page_for_dma,
1487};
1488
1489#ifdef CONFIG_SMP
1490/* On sun4d the cpu broadcasts local TLB flushes, so we can just
1491 * perform the local TLB flush and all the other cpus will see it.
1492 * But, unfortunately, there is a bug in the sun4d XBUS backplane
1493 * that requires that we add some synchronization to these flushes.
1494 *
1495 * The bug is that the fifo which keeps track of all the pending TLB
1496 * broadcasts in the system is an entry or two too small, so if we
1497 * have too many going at once we'll overflow that fifo and lose a TLB
1498 * flush resulting in corruption.
1499 *
1500 * Our workaround is to take a global spinlock around the TLB flushes,
1501 * which guarentees we won't ever have too many pending.  It's a big
1502 * hammer, but a semaphore like system to make sure we only have N TLB
1503 * flushes going at once will require SMP locking anyways so there's
1504 * no real value in trying any harder than this.
1505 */
1506static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1507	.cache_all	= viking_flush_cache_all,
1508	.cache_mm	= viking_flush_cache_mm,
1509	.cache_page	= viking_flush_cache_page,
1510	.cache_range	= viking_flush_cache_range,
1511	.tlb_all	= sun4dsmp_flush_tlb_all,
1512	.tlb_mm		= sun4dsmp_flush_tlb_mm,
1513	.tlb_page	= sun4dsmp_flush_tlb_page,
1514	.tlb_range	= sun4dsmp_flush_tlb_range,
1515	.page_to_ram	= viking_flush_page_to_ram,
1516	.sig_insns	= viking_flush_sig_insns,
1517	.page_for_dma	= viking_flush_page_for_dma,
1518};
1519#endif
1520
1521static void __init init_viking(void)
1522{
1523	unsigned long mreg = srmmu_get_mmureg();
1524
1525	/* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1526	if (mreg & VIKING_MMODE) {
1527		srmmu_name = "TI Viking";
1528		viking_mxcc_present = 0;
1529		msi_set_sync();
1530
1531		/*
1532		 * We need this to make sure old viking takes no hits
1533		 * on it's cache for dma snoops to workaround the
1534		 * "load from non-cacheable memory" interrupt bug.
1535		 * This is only necessary because of the new way in
1536		 * which we use the IOMMU.
1537		 */
1538		viking_ops.page_for_dma = viking_flush_page;
1539#ifdef CONFIG_SMP
1540		viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1541#endif
1542		flush_page_for_dma_global = 0;
1543	} else {
1544		srmmu_name = "TI Viking/MXCC";
1545		viking_mxcc_present = 1;
1546		srmmu_cache_pagetables = 1;
1547	}
1548
1549	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1550		&viking_ops;
1551#ifdef CONFIG_SMP
1552	if (sparc_cpu_model == sun4d)
1553		sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1554			&viking_sun4d_smp_ops;
1555#endif
1556
1557	poke_srmmu = poke_viking;
1558}
1559
1560/* Probe for the srmmu chip version. */
1561static void __init get_srmmu_type(void)
1562{
1563	unsigned long mreg, psr;
1564	unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1565
1566	srmmu_modtype = SRMMU_INVAL_MOD;
1567	hwbug_bitmask = 0;
1568
1569	mreg = srmmu_get_mmureg(); psr = get_psr();
1570	mod_typ = (mreg & 0xf0000000) >> 28;
1571	mod_rev = (mreg & 0x0f000000) >> 24;
1572	psr_typ = (psr >> 28) & 0xf;
1573	psr_vers = (psr >> 24) & 0xf;
1574
1575	/* First, check for sparc-leon. */
1576	if (sparc_cpu_model == sparc_leon) {
1577		init_leon();
1578		return;
1579	}
1580
1581	/* Second, check for HyperSparc or Cypress. */
1582	if (mod_typ == 1) {
1583		switch (mod_rev) {
1584		case 7:
1585			/* UP or MP Hypersparc */
1586			init_hypersparc();
1587			break;
1588		case 0:
1589		case 2:
1590		case 10:
1591		case 11:
1592		case 12:
1593		case 13:
1594		case 14:
1595		case 15:
1596		default:
1597			prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1598			prom_halt();
1599			break;
1600		}
1601		return;
1602	}
1603
1604	/* Now Fujitsu TurboSparc. It might happen that it is
1605	 * in Swift emulation mode, so we will check later...
1606	 */
1607	if (psr_typ == 0 && psr_vers == 5) {
1608		init_turbosparc();
1609		return;
1610	}
1611
1612	/* Next check for Fujitsu Swift. */
1613	if (psr_typ == 0 && psr_vers == 4) {
1614		phandle cpunode;
1615		char node_str[128];
1616
1617		/* Look if it is not a TurboSparc emulating Swift... */
1618		cpunode = prom_getchild(prom_root_node);
1619		while ((cpunode = prom_getsibling(cpunode)) != 0) {
1620			prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1621			if (!strcmp(node_str, "cpu")) {
1622				if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1623				    prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1624					init_turbosparc();
1625					return;
1626				}
1627				break;
1628			}
1629		}
1630
1631		init_swift();
1632		return;
1633	}
1634
1635	/* Now the Viking family of srmmu. */
1636	if (psr_typ == 4 &&
1637	   ((psr_vers == 0) ||
1638	    ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1639		init_viking();
1640		return;
1641	}
1642
1643	/* Finally the Tsunami. */
1644	if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1645		init_tsunami();
1646		return;
1647	}
1648
1649	/* Oh well */
1650	srmmu_is_bad();
1651}
1652
1653#ifdef CONFIG_SMP
1654/* Local cross-calls. */
1655static void smp_flush_page_for_dma(unsigned long page)
1656{
1657	xc1((smpfunc_t) local_ops->page_for_dma, page);
1658	local_ops->page_for_dma(page);
1659}
1660
1661static void smp_flush_cache_all(void)
1662{
1663	xc0((smpfunc_t) local_ops->cache_all);
1664	local_ops->cache_all();
1665}
1666
1667static void smp_flush_tlb_all(void)
1668{
1669	xc0((smpfunc_t) local_ops->tlb_all);
1670	local_ops->tlb_all();
1671}
1672
1673static void smp_flush_cache_mm(struct mm_struct *mm)
1674{
1675	if (mm->context != NO_CONTEXT) {
1676		cpumask_t cpu_mask;
1677		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1678		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1679		if (!cpumask_empty(&cpu_mask))
1680			xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1681		local_ops->cache_mm(mm);
1682	}
1683}
1684
1685static void smp_flush_tlb_mm(struct mm_struct *mm)
1686{
1687	if (mm->context != NO_CONTEXT) {
1688		cpumask_t cpu_mask;
1689		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1690		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1691		if (!cpumask_empty(&cpu_mask)) {
1692			xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1693			if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1694				cpumask_copy(mm_cpumask(mm),
1695					     cpumask_of(smp_processor_id()));
1696		}
1697		local_ops->tlb_mm(mm);
1698	}
1699}
1700
1701static void smp_flush_cache_range(struct vm_area_struct *vma,
1702				  unsigned long start,
1703				  unsigned long end)
1704{
1705	struct mm_struct *mm = vma->vm_mm;
1706
1707	if (mm->context != NO_CONTEXT) {
1708		cpumask_t cpu_mask;
1709		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1710		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1711		if (!cpumask_empty(&cpu_mask))
1712			xc3((smpfunc_t) local_ops->cache_range,
1713			    (unsigned long) vma, start, end);
1714		local_ops->cache_range(vma, start, end);
1715	}
1716}
1717
1718static void smp_flush_tlb_range(struct vm_area_struct *vma,
1719				unsigned long start,
1720				unsigned long end)
1721{
1722	struct mm_struct *mm = vma->vm_mm;
1723
1724	if (mm->context != NO_CONTEXT) {
1725		cpumask_t cpu_mask;
1726		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1727		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1728		if (!cpumask_empty(&cpu_mask))
1729			xc3((smpfunc_t) local_ops->tlb_range,
1730			    (unsigned long) vma, start, end);
1731		local_ops->tlb_range(vma, start, end);
1732	}
1733}
1734
1735static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1736{
1737	struct mm_struct *mm = vma->vm_mm;
1738
1739	if (mm->context != NO_CONTEXT) {
1740		cpumask_t cpu_mask;
1741		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1742		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1743		if (!cpumask_empty(&cpu_mask))
1744			xc2((smpfunc_t) local_ops->cache_page,
1745			    (unsigned long) vma, page);
1746		local_ops->cache_page(vma, page);
1747	}
1748}
1749
1750static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1751{
1752	struct mm_struct *mm = vma->vm_mm;
1753
1754	if (mm->context != NO_CONTEXT) {
1755		cpumask_t cpu_mask;
1756		cpumask_copy(&cpu_mask, mm_cpumask(mm));
1757		cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1758		if (!cpumask_empty(&cpu_mask))
1759			xc2((smpfunc_t) local_ops->tlb_page,
1760			    (unsigned long) vma, page);
1761		local_ops->tlb_page(vma, page);
1762	}
1763}
1764
1765static void smp_flush_page_to_ram(unsigned long page)
1766{
1767	/* Current theory is that those who call this are the one's
1768	 * who have just dirtied their cache with the pages contents
1769	 * in kernel space, therefore we only run this on local cpu.
1770	 *
1771	 * XXX This experiment failed, research further... -DaveM
1772	 */
1773#if 1
1774	xc1((smpfunc_t) local_ops->page_to_ram, page);
1775#endif
1776	local_ops->page_to_ram(page);
1777}
1778
1779static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1780{
1781	cpumask_t cpu_mask;
1782	cpumask_copy(&cpu_mask, mm_cpumask(mm));
1783	cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1784	if (!cpumask_empty(&cpu_mask))
1785		xc2((smpfunc_t) local_ops->sig_insns,
1786		    (unsigned long) mm, insn_addr);
1787	local_ops->sig_insns(mm, insn_addr);
1788}
1789
1790static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1791	.cache_all	= smp_flush_cache_all,
1792	.cache_mm	= smp_flush_cache_mm,
1793	.cache_page	= smp_flush_cache_page,
1794	.cache_range	= smp_flush_cache_range,
1795	.tlb_all	= smp_flush_tlb_all,
1796	.tlb_mm		= smp_flush_tlb_mm,
1797	.tlb_page	= smp_flush_tlb_page,
1798	.tlb_range	= smp_flush_tlb_range,
1799	.page_to_ram	= smp_flush_page_to_ram,
1800	.sig_insns	= smp_flush_sig_insns,
1801	.page_for_dma	= smp_flush_page_for_dma,
1802};
1803#endif
1804
1805/* Load up routines and constants for sun4m and sun4d mmu */
1806void __init load_mmu(void)
1807{
1808	/* Functions */
1809	get_srmmu_type();
1810
1811#ifdef CONFIG_SMP
1812	/* El switcheroo... */
1813	local_ops = sparc32_cachetlb_ops;
1814
1815	if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1816		smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1817		smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1818		smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1819		smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1820	}
1821
1822	if (poke_srmmu == poke_viking) {
1823		/* Avoid unnecessary cross calls. */
1824		smp_cachetlb_ops.cache_all = local_ops->cache_all;
1825		smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1826		smp_cachetlb_ops.cache_range = local_ops->cache_range;
1827		smp_cachetlb_ops.cache_page = local_ops->cache_page;
1828
1829		smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1830		smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1831		smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1832	}
1833
1834	/* It really is const after this point. */
1835	sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1836		&smp_cachetlb_ops;
1837#endif
1838
1839	if (sparc_cpu_model == sun4d)
1840		ld_mmu_iounit();
1841	else
1842		ld_mmu_iommu();
1843#ifdef CONFIG_SMP
1844	if (sparc_cpu_model == sun4d)
1845		sun4d_init_smp();
1846	else if (sparc_cpu_model == sparc_leon)
1847		leon_init_smp();
1848	else
1849		sun4m_init_smp();
1850#endif
1851}