Loading...
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 = __nocache_pa(ptep) >> 4;
140 set_pte((pte_t *)&pmd_val(*pmdp), __pte(SRMMU_ET_PTD | ptp));
141}
142
143/*
144 * size: bytes to allocate in the nocache area.
145 * align: bytes, number to align at.
146 * Returns the virtual address of the allocated area.
147 */
148static void *__srmmu_get_nocache(int size, int align)
149{
150 int offset, minsz = 1 << SRMMU_NOCACHE_BITMAP_SHIFT;
151 unsigned long addr;
152
153 if (size < minsz) {
154 printk(KERN_ERR "Size 0x%x too small for nocache request\n",
155 size);
156 size = minsz;
157 }
158 if (size & (minsz - 1)) {
159 printk(KERN_ERR "Size 0x%x unaligned in nocache request\n",
160 size);
161 size += minsz - 1;
162 }
163 BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
164
165 offset = bit_map_string_get(&srmmu_nocache_map,
166 size >> SRMMU_NOCACHE_BITMAP_SHIFT,
167 align >> SRMMU_NOCACHE_BITMAP_SHIFT);
168 if (offset == -1) {
169 printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
170 size, (int) srmmu_nocache_size,
171 srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
172 return NULL;
173 }
174
175 addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
176 return (void *)addr;
177}
178
179void *srmmu_get_nocache(int size, int align)
180{
181 void *tmp;
182
183 tmp = __srmmu_get_nocache(size, align);
184
185 if (tmp)
186 memset(tmp, 0, size);
187
188 return tmp;
189}
190
191void srmmu_free_nocache(void *addr, int size)
192{
193 unsigned long vaddr;
194 int offset;
195
196 vaddr = (unsigned long)addr;
197 if (vaddr < SRMMU_NOCACHE_VADDR) {
198 printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
199 vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
200 BUG();
201 }
202 if (vaddr + size > srmmu_nocache_end) {
203 printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
204 vaddr, srmmu_nocache_end);
205 BUG();
206 }
207 if (!is_power_of_2(size)) {
208 printk("Size 0x%x is not a power of 2\n", size);
209 BUG();
210 }
211 if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
212 printk("Size 0x%x is too small\n", size);
213 BUG();
214 }
215 if (vaddr & (size - 1)) {
216 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
217 BUG();
218 }
219
220 offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
221 size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
222
223 bit_map_clear(&srmmu_nocache_map, offset, size);
224}
225
226static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
227 unsigned long end);
228
229/* Return how much physical memory we have. */
230static unsigned long __init probe_memory(void)
231{
232 unsigned long total = 0;
233 int i;
234
235 for (i = 0; sp_banks[i].num_bytes; i++)
236 total += sp_banks[i].num_bytes;
237
238 return total;
239}
240
241/*
242 * Reserve nocache dynamically proportionally to the amount of
243 * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
244 */
245static void __init srmmu_nocache_calcsize(void)
246{
247 unsigned long sysmemavail = probe_memory() / 1024;
248 int srmmu_nocache_npages;
249
250 srmmu_nocache_npages =
251 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
252
253 /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
254 // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
255 if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
256 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
257
258 /* anything above 1280 blows up */
259 if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
260 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
261
262 srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
263 srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
264}
265
266static void __init srmmu_nocache_init(void)
267{
268 void *srmmu_nocache_bitmap;
269 unsigned int bitmap_bits;
270 pgd_t *pgd;
271 p4d_t *p4d;
272 pud_t *pud;
273 pmd_t *pmd;
274 pte_t *pte;
275 unsigned long paddr, vaddr;
276 unsigned long pteval;
277
278 bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
279
280 srmmu_nocache_pool = memblock_alloc(srmmu_nocache_size,
281 SRMMU_NOCACHE_ALIGN_MAX);
282 if (!srmmu_nocache_pool)
283 panic("%s: Failed to allocate %lu bytes align=0x%x\n",
284 __func__, srmmu_nocache_size, SRMMU_NOCACHE_ALIGN_MAX);
285 memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
286
287 srmmu_nocache_bitmap =
288 memblock_alloc(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
289 SMP_CACHE_BYTES);
290 if (!srmmu_nocache_bitmap)
291 panic("%s: Failed to allocate %zu bytes\n", __func__,
292 BITS_TO_LONGS(bitmap_bits) * sizeof(long));
293 bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
294
295 srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
296 memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
297 init_mm.pgd = srmmu_swapper_pg_dir;
298
299 srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
300
301 paddr = __pa((unsigned long)srmmu_nocache_pool);
302 vaddr = SRMMU_NOCACHE_VADDR;
303
304 while (vaddr < srmmu_nocache_end) {
305 pgd = pgd_offset_k(vaddr);
306 p4d = p4d_offset(pgd, vaddr);
307 pud = pud_offset(p4d, vaddr);
308 pmd = pmd_offset(__nocache_fix(pud), 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)
350{
351 pte_t *ptep;
352 struct page *page;
353
354 if (!(ptep = pte_alloc_one_kernel(mm)))
355 return NULL;
356 page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
357 spin_lock(&mm->page_table_lock);
358 if (page_ref_inc_return(page) == 2 && !pgtable_pte_page_ctor(page)) {
359 page_ref_dec(page);
360 ptep = NULL;
361 }
362 spin_unlock(&mm->page_table_lock);
363
364 return ptep;
365}
366
367void pte_free(struct mm_struct *mm, pgtable_t ptep)
368{
369 struct page *page;
370
371 page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
372 spin_lock(&mm->page_table_lock);
373 if (page_ref_dec_return(page) == 1)
374 pgtable_pte_page_dtor(page);
375 spin_unlock(&mm->page_table_lock);
376
377 srmmu_free_nocache(ptep, SRMMU_PTE_TABLE_SIZE);
378}
379
380/* context handling - a dynamically sized pool is used */
381#define NO_CONTEXT -1
382
383struct ctx_list {
384 struct ctx_list *next;
385 struct ctx_list *prev;
386 unsigned int ctx_number;
387 struct mm_struct *ctx_mm;
388};
389
390static struct ctx_list *ctx_list_pool;
391static struct ctx_list ctx_free;
392static struct ctx_list ctx_used;
393
394/* At boot time we determine the number of contexts */
395static int num_contexts;
396
397static inline void remove_from_ctx_list(struct ctx_list *entry)
398{
399 entry->next->prev = entry->prev;
400 entry->prev->next = entry->next;
401}
402
403static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
404{
405 entry->next = head;
406 (entry->prev = head->prev)->next = entry;
407 head->prev = entry;
408}
409#define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
410#define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
411
412
413static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
414{
415 struct ctx_list *ctxp;
416
417 ctxp = ctx_free.next;
418 if (ctxp != &ctx_free) {
419 remove_from_ctx_list(ctxp);
420 add_to_used_ctxlist(ctxp);
421 mm->context = ctxp->ctx_number;
422 ctxp->ctx_mm = mm;
423 return;
424 }
425 ctxp = ctx_used.next;
426 if (ctxp->ctx_mm == old_mm)
427 ctxp = ctxp->next;
428 if (ctxp == &ctx_used)
429 panic("out of mmu contexts");
430 flush_cache_mm(ctxp->ctx_mm);
431 flush_tlb_mm(ctxp->ctx_mm);
432 remove_from_ctx_list(ctxp);
433 add_to_used_ctxlist(ctxp);
434 ctxp->ctx_mm->context = NO_CONTEXT;
435 ctxp->ctx_mm = mm;
436 mm->context = ctxp->ctx_number;
437}
438
439static inline void free_context(int context)
440{
441 struct ctx_list *ctx_old;
442
443 ctx_old = ctx_list_pool + context;
444 remove_from_ctx_list(ctx_old);
445 add_to_free_ctxlist(ctx_old);
446}
447
448static void __init sparc_context_init(int numctx)
449{
450 int ctx;
451 unsigned long size;
452
453 size = numctx * sizeof(struct ctx_list);
454 ctx_list_pool = memblock_alloc(size, SMP_CACHE_BYTES);
455 if (!ctx_list_pool)
456 panic("%s: Failed to allocate %lu bytes\n", __func__, size);
457
458 for (ctx = 0; ctx < numctx; ctx++) {
459 struct ctx_list *clist;
460
461 clist = (ctx_list_pool + ctx);
462 clist->ctx_number = ctx;
463 clist->ctx_mm = NULL;
464 }
465 ctx_free.next = ctx_free.prev = &ctx_free;
466 ctx_used.next = ctx_used.prev = &ctx_used;
467 for (ctx = 0; ctx < numctx; ctx++)
468 add_to_free_ctxlist(ctx_list_pool + ctx);
469}
470
471void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
472 struct task_struct *tsk)
473{
474 unsigned long flags;
475
476 if (mm->context == NO_CONTEXT) {
477 spin_lock_irqsave(&srmmu_context_spinlock, flags);
478 alloc_context(old_mm, mm);
479 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
480 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
481 }
482
483 if (sparc_cpu_model == sparc_leon)
484 leon_switch_mm();
485
486 if (is_hypersparc)
487 hyper_flush_whole_icache();
488
489 srmmu_set_context(mm->context);
490}
491
492/* Low level IO area allocation on the SRMMU. */
493static inline void srmmu_mapioaddr(unsigned long physaddr,
494 unsigned long virt_addr, int bus_type)
495{
496 pgd_t *pgdp;
497 p4d_t *p4dp;
498 pud_t *pudp;
499 pmd_t *pmdp;
500 pte_t *ptep;
501 unsigned long tmp;
502
503 physaddr &= PAGE_MASK;
504 pgdp = pgd_offset_k(virt_addr);
505 p4dp = p4d_offset(pgdp, virt_addr);
506 pudp = pud_offset(p4dp, virt_addr);
507 pmdp = pmd_offset(pudp, virt_addr);
508 ptep = pte_offset_kernel(pmdp, virt_addr);
509 tmp = (physaddr >> 4) | SRMMU_ET_PTE;
510
511 /* I need to test whether this is consistent over all
512 * sun4m's. The bus_type represents the upper 4 bits of
513 * 36-bit physical address on the I/O space lines...
514 */
515 tmp |= (bus_type << 28);
516 tmp |= SRMMU_PRIV;
517 __flush_page_to_ram(virt_addr);
518 set_pte(ptep, __pte(tmp));
519}
520
521void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
522 unsigned long xva, unsigned int len)
523{
524 while (len != 0) {
525 len -= PAGE_SIZE;
526 srmmu_mapioaddr(xpa, xva, bus);
527 xva += PAGE_SIZE;
528 xpa += PAGE_SIZE;
529 }
530 flush_tlb_all();
531}
532
533static inline void srmmu_unmapioaddr(unsigned long virt_addr)
534{
535 pgd_t *pgdp;
536 p4d_t *p4dp;
537 pud_t *pudp;
538 pmd_t *pmdp;
539 pte_t *ptep;
540
541
542 pgdp = pgd_offset_k(virt_addr);
543 p4dp = p4d_offset(pgdp, virt_addr);
544 pudp = pud_offset(p4dp, virt_addr);
545 pmdp = pmd_offset(pudp, virt_addr);
546 ptep = pte_offset_kernel(pmdp, virt_addr);
547
548 /* No need to flush uncacheable page. */
549 __pte_clear(ptep);
550}
551
552void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
553{
554 while (len != 0) {
555 len -= PAGE_SIZE;
556 srmmu_unmapioaddr(virt_addr);
557 virt_addr += PAGE_SIZE;
558 }
559 flush_tlb_all();
560}
561
562/* tsunami.S */
563extern void tsunami_flush_cache_all(void);
564extern void tsunami_flush_cache_mm(struct mm_struct *mm);
565extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
566extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
567extern void tsunami_flush_page_to_ram(unsigned long page);
568extern void tsunami_flush_page_for_dma(unsigned long page);
569extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
570extern void tsunami_flush_tlb_all(void);
571extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
572extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
573extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
574extern void tsunami_setup_blockops(void);
575
576/* swift.S */
577extern void swift_flush_cache_all(void);
578extern void swift_flush_cache_mm(struct mm_struct *mm);
579extern void swift_flush_cache_range(struct vm_area_struct *vma,
580 unsigned long start, unsigned long end);
581extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
582extern void swift_flush_page_to_ram(unsigned long page);
583extern void swift_flush_page_for_dma(unsigned long page);
584extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
585extern void swift_flush_tlb_all(void);
586extern void swift_flush_tlb_mm(struct mm_struct *mm);
587extern void swift_flush_tlb_range(struct vm_area_struct *vma,
588 unsigned long start, unsigned long end);
589extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
590
591#if 0 /* P3: deadwood to debug precise flushes on Swift. */
592void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
593{
594 int cctx, ctx1;
595
596 page &= PAGE_MASK;
597 if ((ctx1 = vma->vm_mm->context) != -1) {
598 cctx = srmmu_get_context();
599/* Is context # ever different from current context? P3 */
600 if (cctx != ctx1) {
601 printk("flush ctx %02x curr %02x\n", ctx1, cctx);
602 srmmu_set_context(ctx1);
603 swift_flush_page(page);
604 __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
605 "r" (page), "i" (ASI_M_FLUSH_PROBE));
606 srmmu_set_context(cctx);
607 } else {
608 /* Rm. prot. bits from virt. c. */
609 /* swift_flush_cache_all(); */
610 /* swift_flush_cache_page(vma, page); */
611 swift_flush_page(page);
612
613 __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
614 "r" (page), "i" (ASI_M_FLUSH_PROBE));
615 /* same as above: srmmu_flush_tlb_page() */
616 }
617 }
618}
619#endif
620
621/*
622 * The following are all MBUS based SRMMU modules, and therefore could
623 * be found in a multiprocessor configuration. On the whole, these
624 * chips seems to be much more touchy about DVMA and page tables
625 * with respect to cache coherency.
626 */
627
628/* viking.S */
629extern void viking_flush_cache_all(void);
630extern void viking_flush_cache_mm(struct mm_struct *mm);
631extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
632 unsigned long end);
633extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
634extern void viking_flush_page_to_ram(unsigned long page);
635extern void viking_flush_page_for_dma(unsigned long page);
636extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
637extern void viking_flush_page(unsigned long page);
638extern void viking_mxcc_flush_page(unsigned long page);
639extern void viking_flush_tlb_all(void);
640extern void viking_flush_tlb_mm(struct mm_struct *mm);
641extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
642 unsigned long end);
643extern void viking_flush_tlb_page(struct vm_area_struct *vma,
644 unsigned long page);
645extern void sun4dsmp_flush_tlb_all(void);
646extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
647extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
648 unsigned long end);
649extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
650 unsigned long page);
651
652/* hypersparc.S */
653extern void hypersparc_flush_cache_all(void);
654extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
655extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
656extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
657extern void hypersparc_flush_page_to_ram(unsigned long page);
658extern void hypersparc_flush_page_for_dma(unsigned long page);
659extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
660extern void hypersparc_flush_tlb_all(void);
661extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
662extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
663extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
664extern void hypersparc_setup_blockops(void);
665
666/*
667 * NOTE: All of this startup code assumes the low 16mb (approx.) of
668 * kernel mappings are done with one single contiguous chunk of
669 * ram. On small ram machines (classics mainly) we only get
670 * around 8mb mapped for us.
671 */
672
673static void __init early_pgtable_allocfail(char *type)
674{
675 prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
676 prom_halt();
677}
678
679static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
680 unsigned long end)
681{
682 pgd_t *pgdp;
683 p4d_t *p4dp;
684 pud_t *pudp;
685 pmd_t *pmdp;
686 pte_t *ptep;
687
688 while (start < end) {
689 pgdp = pgd_offset_k(start);
690 p4dp = p4d_offset(pgdp, start);
691 pudp = pud_offset(p4dp, start);
692 if (pud_none(*__nocache_fix(pudp))) {
693 pmdp = __srmmu_get_nocache(
694 SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
695 if (pmdp == NULL)
696 early_pgtable_allocfail("pmd");
697 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
698 pud_set(__nocache_fix(pudp), pmdp);
699 }
700 pmdp = pmd_offset(__nocache_fix(pudp), start);
701 if (srmmu_pmd_none(*__nocache_fix(pmdp))) {
702 ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
703 if (ptep == NULL)
704 early_pgtable_allocfail("pte");
705 memset(__nocache_fix(ptep), 0, PTE_SIZE);
706 pmd_set(__nocache_fix(pmdp), ptep);
707 }
708 if (start > (0xffffffffUL - PMD_SIZE))
709 break;
710 start = (start + PMD_SIZE) & PMD_MASK;
711 }
712}
713
714static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
715 unsigned long end)
716{
717 pgd_t *pgdp;
718 p4d_t *p4dp;
719 pud_t *pudp;
720 pmd_t *pmdp;
721 pte_t *ptep;
722
723 while (start < end) {
724 pgdp = pgd_offset_k(start);
725 p4dp = p4d_offset(pgdp, start);
726 pudp = pud_offset(p4dp, start);
727 if (pud_none(*pudp)) {
728 pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
729 if (pmdp == NULL)
730 early_pgtable_allocfail("pmd");
731 memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
732 pud_set((pud_t *)pgdp, pmdp);
733 }
734 pmdp = pmd_offset(pudp, start);
735 if (srmmu_pmd_none(*pmdp)) {
736 ptep = __srmmu_get_nocache(PTE_SIZE,
737 PTE_SIZE);
738 if (ptep == NULL)
739 early_pgtable_allocfail("pte");
740 memset(ptep, 0, PTE_SIZE);
741 pmd_set(pmdp, ptep);
742 }
743 if (start > (0xffffffffUL - PMD_SIZE))
744 break;
745 start = (start + PMD_SIZE) & PMD_MASK;
746 }
747}
748
749/* These flush types are not available on all chips... */
750static inline unsigned long srmmu_probe(unsigned long vaddr)
751{
752 unsigned long retval;
753
754 if (sparc_cpu_model != sparc_leon) {
755
756 vaddr &= PAGE_MASK;
757 __asm__ __volatile__("lda [%1] %2, %0\n\t" :
758 "=r" (retval) :
759 "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
760 } else {
761 retval = leon_swprobe(vaddr, NULL);
762 }
763 return retval;
764}
765
766/*
767 * This is much cleaner than poking around physical address space
768 * looking at the prom's page table directly which is what most
769 * other OS's do. Yuck... this is much better.
770 */
771static void __init srmmu_inherit_prom_mappings(unsigned long start,
772 unsigned long end)
773{
774 unsigned long probed;
775 unsigned long addr;
776 pgd_t *pgdp;
777 p4d_t *p4dp;
778 pud_t *pudp;
779 pmd_t *pmdp;
780 pte_t *ptep;
781 int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
782
783 while (start <= end) {
784 if (start == 0)
785 break; /* probably wrap around */
786 if (start == 0xfef00000)
787 start = KADB_DEBUGGER_BEGVM;
788 probed = srmmu_probe(start);
789 if (!probed) {
790 /* continue probing until we find an entry */
791 start += PAGE_SIZE;
792 continue;
793 }
794
795 /* A red snapper, see what it really is. */
796 what = 0;
797 addr = start - PAGE_SIZE;
798
799 if (!(start & ~(PMD_MASK))) {
800 if (srmmu_probe(addr + PMD_SIZE) == probed)
801 what = 1;
802 }
803
804 if (!(start & ~(PGDIR_MASK))) {
805 if (srmmu_probe(addr + PGDIR_SIZE) == probed)
806 what = 2;
807 }
808
809 pgdp = pgd_offset_k(start);
810 p4dp = p4d_offset(pgdp, start);
811 pudp = pud_offset(p4dp, start);
812 if (what == 2) {
813 *__nocache_fix(pgdp) = __pgd(probed);
814 start += PGDIR_SIZE;
815 continue;
816 }
817 if (pud_none(*__nocache_fix(pudp))) {
818 pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
819 SRMMU_PMD_TABLE_SIZE);
820 if (pmdp == NULL)
821 early_pgtable_allocfail("pmd");
822 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
823 pud_set(__nocache_fix(pudp), pmdp);
824 }
825 pmdp = pmd_offset(__nocache_fix(pudp), start);
826 if (what == 1) {
827 *(pmd_t *)__nocache_fix(pmdp) = __pmd(probed);
828 start += PMD_SIZE;
829 continue;
830 }
831 if (srmmu_pmd_none(*__nocache_fix(pmdp))) {
832 ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
833 if (ptep == NULL)
834 early_pgtable_allocfail("pte");
835 memset(__nocache_fix(ptep), 0, PTE_SIZE);
836 pmd_set(__nocache_fix(pmdp), ptep);
837 }
838 ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
839 *__nocache_fix(ptep) = __pte(probed);
840 start += PAGE_SIZE;
841 }
842}
843
844#define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
845
846/* Create a third-level SRMMU 16MB page mapping. */
847static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
848{
849 pgd_t *pgdp = pgd_offset_k(vaddr);
850 unsigned long big_pte;
851
852 big_pte = KERNEL_PTE(phys_base >> 4);
853 *__nocache_fix(pgdp) = __pgd(big_pte);
854}
855
856/* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
857static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
858{
859 unsigned long pstart = (sp_banks[sp_entry].base_addr & PGDIR_MASK);
860 unsigned long vstart = (vbase & PGDIR_MASK);
861 unsigned long vend = PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
862 /* Map "low" memory only */
863 const unsigned long min_vaddr = PAGE_OFFSET;
864 const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
865
866 if (vstart < min_vaddr || vstart >= max_vaddr)
867 return vstart;
868
869 if (vend > max_vaddr || vend < min_vaddr)
870 vend = max_vaddr;
871
872 while (vstart < vend) {
873 do_large_mapping(vstart, pstart);
874 vstart += PGDIR_SIZE; pstart += PGDIR_SIZE;
875 }
876 return vstart;
877}
878
879static void __init map_kernel(void)
880{
881 int i;
882
883 if (phys_base > 0) {
884 do_large_mapping(PAGE_OFFSET, phys_base);
885 }
886
887 for (i = 0; sp_banks[i].num_bytes != 0; i++) {
888 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
889 }
890}
891
892void (*poke_srmmu)(void) = NULL;
893
894void __init srmmu_paging_init(void)
895{
896 int i;
897 phandle cpunode;
898 char node_str[128];
899 pgd_t *pgd;
900 p4d_t *p4d;
901 pud_t *pud;
902 pmd_t *pmd;
903 pte_t *pte;
904 unsigned long pages_avail;
905
906 init_mm.context = (unsigned long) NO_CONTEXT;
907 sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
908
909 if (sparc_cpu_model == sun4d)
910 num_contexts = 65536; /* We know it is Viking */
911 else {
912 /* Find the number of contexts on the srmmu. */
913 cpunode = prom_getchild(prom_root_node);
914 num_contexts = 0;
915 while (cpunode != 0) {
916 prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
917 if (!strcmp(node_str, "cpu")) {
918 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
919 break;
920 }
921 cpunode = prom_getsibling(cpunode);
922 }
923 }
924
925 if (!num_contexts) {
926 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
927 prom_halt();
928 }
929
930 pages_avail = 0;
931 last_valid_pfn = bootmem_init(&pages_avail);
932
933 srmmu_nocache_calcsize();
934 srmmu_nocache_init();
935 srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
936 map_kernel();
937
938 /* ctx table has to be physically aligned to its size */
939 srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
940 srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
941
942 for (i = 0; i < num_contexts; i++)
943 srmmu_ctxd_set(__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
944
945 flush_cache_all();
946 srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
947#ifdef CONFIG_SMP
948 /* Stop from hanging here... */
949 local_ops->tlb_all();
950#else
951 flush_tlb_all();
952#endif
953 poke_srmmu();
954
955 srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
956 srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
957
958 srmmu_allocate_ptable_skeleton(
959 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
960 srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
961
962 pgd = pgd_offset_k(PKMAP_BASE);
963 p4d = p4d_offset(pgd, PKMAP_BASE);
964 pud = pud_offset(p4d, PKMAP_BASE);
965 pmd = pmd_offset(pud, PKMAP_BASE);
966 pte = pte_offset_kernel(pmd, PKMAP_BASE);
967 pkmap_page_table = pte;
968
969 flush_cache_all();
970 flush_tlb_all();
971
972 sparc_context_init(num_contexts);
973
974 {
975 unsigned long max_zone_pfn[MAX_NR_ZONES] = { 0 };
976
977 max_zone_pfn[ZONE_DMA] = max_low_pfn;
978 max_zone_pfn[ZONE_NORMAL] = max_low_pfn;
979 max_zone_pfn[ZONE_HIGHMEM] = highend_pfn;
980
981 free_area_init(max_zone_pfn);
982 }
983}
984
985void mmu_info(struct seq_file *m)
986{
987 seq_printf(m,
988 "MMU type\t: %s\n"
989 "contexts\t: %d\n"
990 "nocache total\t: %ld\n"
991 "nocache used\t: %d\n",
992 srmmu_name,
993 num_contexts,
994 srmmu_nocache_size,
995 srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
996}
997
998int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
999{
1000 mm->context = NO_CONTEXT;
1001 return 0;
1002}
1003
1004void destroy_context(struct mm_struct *mm)
1005{
1006 unsigned long flags;
1007
1008 if (mm->context != NO_CONTEXT) {
1009 flush_cache_mm(mm);
1010 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1011 flush_tlb_mm(mm);
1012 spin_lock_irqsave(&srmmu_context_spinlock, flags);
1013 free_context(mm->context);
1014 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1015 mm->context = NO_CONTEXT;
1016 }
1017}
1018
1019/* Init various srmmu chip types. */
1020static void __init srmmu_is_bad(void)
1021{
1022 prom_printf("Could not determine SRMMU chip type.\n");
1023 prom_halt();
1024}
1025
1026static void __init init_vac_layout(void)
1027{
1028 phandle nd;
1029 int cache_lines;
1030 char node_str[128];
1031#ifdef CONFIG_SMP
1032 int cpu = 0;
1033 unsigned long max_size = 0;
1034 unsigned long min_line_size = 0x10000000;
1035#endif
1036
1037 nd = prom_getchild(prom_root_node);
1038 while ((nd = prom_getsibling(nd)) != 0) {
1039 prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1040 if (!strcmp(node_str, "cpu")) {
1041 vac_line_size = prom_getint(nd, "cache-line-size");
1042 if (vac_line_size == -1) {
1043 prom_printf("can't determine cache-line-size, halting.\n");
1044 prom_halt();
1045 }
1046 cache_lines = prom_getint(nd, "cache-nlines");
1047 if (cache_lines == -1) {
1048 prom_printf("can't determine cache-nlines, halting.\n");
1049 prom_halt();
1050 }
1051
1052 vac_cache_size = cache_lines * vac_line_size;
1053#ifdef CONFIG_SMP
1054 if (vac_cache_size > max_size)
1055 max_size = vac_cache_size;
1056 if (vac_line_size < min_line_size)
1057 min_line_size = vac_line_size;
1058 //FIXME: cpus not contiguous!!
1059 cpu++;
1060 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1061 break;
1062#else
1063 break;
1064#endif
1065 }
1066 }
1067 if (nd == 0) {
1068 prom_printf("No CPU nodes found, halting.\n");
1069 prom_halt();
1070 }
1071#ifdef CONFIG_SMP
1072 vac_cache_size = max_size;
1073 vac_line_size = min_line_size;
1074#endif
1075 printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1076 (int)vac_cache_size, (int)vac_line_size);
1077}
1078
1079static void poke_hypersparc(void)
1080{
1081 volatile unsigned long clear;
1082 unsigned long mreg = srmmu_get_mmureg();
1083
1084 hyper_flush_unconditional_combined();
1085
1086 mreg &= ~(HYPERSPARC_CWENABLE);
1087 mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1088 mreg |= (HYPERSPARC_CMODE);
1089
1090 srmmu_set_mmureg(mreg);
1091
1092#if 0 /* XXX I think this is bad news... -DaveM */
1093 hyper_clear_all_tags();
1094#endif
1095
1096 put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1097 hyper_flush_whole_icache();
1098 clear = srmmu_get_faddr();
1099 clear = srmmu_get_fstatus();
1100}
1101
1102static const struct sparc32_cachetlb_ops hypersparc_ops = {
1103 .cache_all = hypersparc_flush_cache_all,
1104 .cache_mm = hypersparc_flush_cache_mm,
1105 .cache_page = hypersparc_flush_cache_page,
1106 .cache_range = hypersparc_flush_cache_range,
1107 .tlb_all = hypersparc_flush_tlb_all,
1108 .tlb_mm = hypersparc_flush_tlb_mm,
1109 .tlb_page = hypersparc_flush_tlb_page,
1110 .tlb_range = hypersparc_flush_tlb_range,
1111 .page_to_ram = hypersparc_flush_page_to_ram,
1112 .sig_insns = hypersparc_flush_sig_insns,
1113 .page_for_dma = hypersparc_flush_page_for_dma,
1114};
1115
1116static void __init init_hypersparc(void)
1117{
1118 srmmu_name = "ROSS HyperSparc";
1119 srmmu_modtype = HyperSparc;
1120
1121 init_vac_layout();
1122
1123 is_hypersparc = 1;
1124 sparc32_cachetlb_ops = &hypersparc_ops;
1125
1126 poke_srmmu = poke_hypersparc;
1127
1128 hypersparc_setup_blockops();
1129}
1130
1131static void poke_swift(void)
1132{
1133 unsigned long mreg;
1134
1135 /* Clear any crap from the cache or else... */
1136 swift_flush_cache_all();
1137
1138 /* Enable I & D caches */
1139 mreg = srmmu_get_mmureg();
1140 mreg |= (SWIFT_IE | SWIFT_DE);
1141 /*
1142 * The Swift branch folding logic is completely broken. At
1143 * trap time, if things are just right, if can mistakenly
1144 * think that a trap is coming from kernel mode when in fact
1145 * it is coming from user mode (it mis-executes the branch in
1146 * the trap code). So you see things like crashme completely
1147 * hosing your machine which is completely unacceptable. Turn
1148 * this shit off... nice job Fujitsu.
1149 */
1150 mreg &= ~(SWIFT_BF);
1151 srmmu_set_mmureg(mreg);
1152}
1153
1154static const struct sparc32_cachetlb_ops swift_ops = {
1155 .cache_all = swift_flush_cache_all,
1156 .cache_mm = swift_flush_cache_mm,
1157 .cache_page = swift_flush_cache_page,
1158 .cache_range = swift_flush_cache_range,
1159 .tlb_all = swift_flush_tlb_all,
1160 .tlb_mm = swift_flush_tlb_mm,
1161 .tlb_page = swift_flush_tlb_page,
1162 .tlb_range = swift_flush_tlb_range,
1163 .page_to_ram = swift_flush_page_to_ram,
1164 .sig_insns = swift_flush_sig_insns,
1165 .page_for_dma = swift_flush_page_for_dma,
1166};
1167
1168#define SWIFT_MASKID_ADDR 0x10003018
1169static void __init init_swift(void)
1170{
1171 unsigned long swift_rev;
1172
1173 __asm__ __volatile__("lda [%1] %2, %0\n\t"
1174 "srl %0, 0x18, %0\n\t" :
1175 "=r" (swift_rev) :
1176 "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1177 srmmu_name = "Fujitsu Swift";
1178 switch (swift_rev) {
1179 case 0x11:
1180 case 0x20:
1181 case 0x23:
1182 case 0x30:
1183 srmmu_modtype = Swift_lots_o_bugs;
1184 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1185 /*
1186 * Gee george, I wonder why Sun is so hush hush about
1187 * this hardware bug... really braindamage stuff going
1188 * on here. However I think we can find a way to avoid
1189 * all of the workaround overhead under Linux. Basically,
1190 * any page fault can cause kernel pages to become user
1191 * accessible (the mmu gets confused and clears some of
1192 * the ACC bits in kernel ptes). Aha, sounds pretty
1193 * horrible eh? But wait, after extensive testing it appears
1194 * that if you use pgd_t level large kernel pte's (like the
1195 * 4MB pages on the Pentium) the bug does not get tripped
1196 * at all. This avoids almost all of the major overhead.
1197 * Welcome to a world where your vendor tells you to,
1198 * "apply this kernel patch" instead of "sorry for the
1199 * broken hardware, send it back and we'll give you
1200 * properly functioning parts"
1201 */
1202 break;
1203 case 0x25:
1204 case 0x31:
1205 srmmu_modtype = Swift_bad_c;
1206 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1207 /*
1208 * You see Sun allude to this hardware bug but never
1209 * admit things directly, they'll say things like,
1210 * "the Swift chip cache problems" or similar.
1211 */
1212 break;
1213 default:
1214 srmmu_modtype = Swift_ok;
1215 break;
1216 }
1217
1218 sparc32_cachetlb_ops = &swift_ops;
1219 flush_page_for_dma_global = 0;
1220
1221 /*
1222 * Are you now convinced that the Swift is one of the
1223 * biggest VLSI abortions of all time? Bravo Fujitsu!
1224 * Fujitsu, the !#?!%$'d up processor people. I bet if
1225 * you examined the microcode of the Swift you'd find
1226 * XXX's all over the place.
1227 */
1228 poke_srmmu = poke_swift;
1229}
1230
1231static void turbosparc_flush_cache_all(void)
1232{
1233 flush_user_windows();
1234 turbosparc_idflash_clear();
1235}
1236
1237static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1238{
1239 FLUSH_BEGIN(mm)
1240 flush_user_windows();
1241 turbosparc_idflash_clear();
1242 FLUSH_END
1243}
1244
1245static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1246{
1247 FLUSH_BEGIN(vma->vm_mm)
1248 flush_user_windows();
1249 turbosparc_idflash_clear();
1250 FLUSH_END
1251}
1252
1253static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1254{
1255 FLUSH_BEGIN(vma->vm_mm)
1256 flush_user_windows();
1257 if (vma->vm_flags & VM_EXEC)
1258 turbosparc_flush_icache();
1259 turbosparc_flush_dcache();
1260 FLUSH_END
1261}
1262
1263/* TurboSparc is copy-back, if we turn it on, but this does not work. */
1264static void turbosparc_flush_page_to_ram(unsigned long page)
1265{
1266#ifdef TURBOSPARC_WRITEBACK
1267 volatile unsigned long clear;
1268
1269 if (srmmu_probe(page))
1270 turbosparc_flush_page_cache(page);
1271 clear = srmmu_get_fstatus();
1272#endif
1273}
1274
1275static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1276{
1277}
1278
1279static void turbosparc_flush_page_for_dma(unsigned long page)
1280{
1281 turbosparc_flush_dcache();
1282}
1283
1284static void turbosparc_flush_tlb_all(void)
1285{
1286 srmmu_flush_whole_tlb();
1287}
1288
1289static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1290{
1291 FLUSH_BEGIN(mm)
1292 srmmu_flush_whole_tlb();
1293 FLUSH_END
1294}
1295
1296static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1297{
1298 FLUSH_BEGIN(vma->vm_mm)
1299 srmmu_flush_whole_tlb();
1300 FLUSH_END
1301}
1302
1303static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1304{
1305 FLUSH_BEGIN(vma->vm_mm)
1306 srmmu_flush_whole_tlb();
1307 FLUSH_END
1308}
1309
1310
1311static void poke_turbosparc(void)
1312{
1313 unsigned long mreg = srmmu_get_mmureg();
1314 unsigned long ccreg;
1315
1316 /* Clear any crap from the cache or else... */
1317 turbosparc_flush_cache_all();
1318 /* Temporarily disable I & D caches */
1319 mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1320 mreg &= ~(TURBOSPARC_PCENABLE); /* Don't check parity */
1321 srmmu_set_mmureg(mreg);
1322
1323 ccreg = turbosparc_get_ccreg();
1324
1325#ifdef TURBOSPARC_WRITEBACK
1326 ccreg |= (TURBOSPARC_SNENABLE); /* Do DVMA snooping in Dcache */
1327 ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1328 /* Write-back D-cache, emulate VLSI
1329 * abortion number three, not number one */
1330#else
1331 /* For now let's play safe, optimize later */
1332 ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1333 /* Do DVMA snooping in Dcache, Write-thru D-cache */
1334 ccreg &= ~(TURBOSPARC_uS2);
1335 /* Emulate VLSI abortion number three, not number one */
1336#endif
1337
1338 switch (ccreg & 7) {
1339 case 0: /* No SE cache */
1340 case 7: /* Test mode */
1341 break;
1342 default:
1343 ccreg |= (TURBOSPARC_SCENABLE);
1344 }
1345 turbosparc_set_ccreg(ccreg);
1346
1347 mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1348 mreg |= (TURBOSPARC_ICSNOOP); /* Icache snooping on */
1349 srmmu_set_mmureg(mreg);
1350}
1351
1352static const struct sparc32_cachetlb_ops turbosparc_ops = {
1353 .cache_all = turbosparc_flush_cache_all,
1354 .cache_mm = turbosparc_flush_cache_mm,
1355 .cache_page = turbosparc_flush_cache_page,
1356 .cache_range = turbosparc_flush_cache_range,
1357 .tlb_all = turbosparc_flush_tlb_all,
1358 .tlb_mm = turbosparc_flush_tlb_mm,
1359 .tlb_page = turbosparc_flush_tlb_page,
1360 .tlb_range = turbosparc_flush_tlb_range,
1361 .page_to_ram = turbosparc_flush_page_to_ram,
1362 .sig_insns = turbosparc_flush_sig_insns,
1363 .page_for_dma = turbosparc_flush_page_for_dma,
1364};
1365
1366static void __init init_turbosparc(void)
1367{
1368 srmmu_name = "Fujitsu TurboSparc";
1369 srmmu_modtype = TurboSparc;
1370 sparc32_cachetlb_ops = &turbosparc_ops;
1371 poke_srmmu = poke_turbosparc;
1372}
1373
1374static void poke_tsunami(void)
1375{
1376 unsigned long mreg = srmmu_get_mmureg();
1377
1378 tsunami_flush_icache();
1379 tsunami_flush_dcache();
1380 mreg &= ~TSUNAMI_ITD;
1381 mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1382 srmmu_set_mmureg(mreg);
1383}
1384
1385static const struct sparc32_cachetlb_ops tsunami_ops = {
1386 .cache_all = tsunami_flush_cache_all,
1387 .cache_mm = tsunami_flush_cache_mm,
1388 .cache_page = tsunami_flush_cache_page,
1389 .cache_range = tsunami_flush_cache_range,
1390 .tlb_all = tsunami_flush_tlb_all,
1391 .tlb_mm = tsunami_flush_tlb_mm,
1392 .tlb_page = tsunami_flush_tlb_page,
1393 .tlb_range = tsunami_flush_tlb_range,
1394 .page_to_ram = tsunami_flush_page_to_ram,
1395 .sig_insns = tsunami_flush_sig_insns,
1396 .page_for_dma = tsunami_flush_page_for_dma,
1397};
1398
1399static void __init init_tsunami(void)
1400{
1401 /*
1402 * Tsunami's pretty sane, Sun and TI actually got it
1403 * somewhat right this time. Fujitsu should have
1404 * taken some lessons from them.
1405 */
1406
1407 srmmu_name = "TI Tsunami";
1408 srmmu_modtype = Tsunami;
1409 sparc32_cachetlb_ops = &tsunami_ops;
1410 poke_srmmu = poke_tsunami;
1411
1412 tsunami_setup_blockops();
1413}
1414
1415static void poke_viking(void)
1416{
1417 unsigned long mreg = srmmu_get_mmureg();
1418 static int smp_catch;
1419
1420 if (viking_mxcc_present) {
1421 unsigned long mxcc_control = mxcc_get_creg();
1422
1423 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1424 mxcc_control &= ~(MXCC_CTL_RRC);
1425 mxcc_set_creg(mxcc_control);
1426
1427 /*
1428 * We don't need memory parity checks.
1429 * XXX This is a mess, have to dig out later. ecd.
1430 viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1431 */
1432
1433 /* We do cache ptables on MXCC. */
1434 mreg |= VIKING_TCENABLE;
1435 } else {
1436 unsigned long bpreg;
1437
1438 mreg &= ~(VIKING_TCENABLE);
1439 if (smp_catch++) {
1440 /* Must disable mixed-cmd mode here for other cpu's. */
1441 bpreg = viking_get_bpreg();
1442 bpreg &= ~(VIKING_ACTION_MIX);
1443 viking_set_bpreg(bpreg);
1444
1445 /* Just in case PROM does something funny. */
1446 msi_set_sync();
1447 }
1448 }
1449
1450 mreg |= VIKING_SPENABLE;
1451 mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1452 mreg |= VIKING_SBENABLE;
1453 mreg &= ~(VIKING_ACENABLE);
1454 srmmu_set_mmureg(mreg);
1455}
1456
1457static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1458 .cache_all = viking_flush_cache_all,
1459 .cache_mm = viking_flush_cache_mm,
1460 .cache_page = viking_flush_cache_page,
1461 .cache_range = viking_flush_cache_range,
1462 .tlb_all = viking_flush_tlb_all,
1463 .tlb_mm = viking_flush_tlb_mm,
1464 .tlb_page = viking_flush_tlb_page,
1465 .tlb_range = viking_flush_tlb_range,
1466 .page_to_ram = viking_flush_page_to_ram,
1467 .sig_insns = viking_flush_sig_insns,
1468 .page_for_dma = viking_flush_page_for_dma,
1469};
1470
1471#ifdef CONFIG_SMP
1472/* On sun4d the cpu broadcasts local TLB flushes, so we can just
1473 * perform the local TLB flush and all the other cpus will see it.
1474 * But, unfortunately, there is a bug in the sun4d XBUS backplane
1475 * that requires that we add some synchronization to these flushes.
1476 *
1477 * The bug is that the fifo which keeps track of all the pending TLB
1478 * broadcasts in the system is an entry or two too small, so if we
1479 * have too many going at once we'll overflow that fifo and lose a TLB
1480 * flush resulting in corruption.
1481 *
1482 * Our workaround is to take a global spinlock around the TLB flushes,
1483 * which guarentees we won't ever have too many pending. It's a big
1484 * hammer, but a semaphore like system to make sure we only have N TLB
1485 * flushes going at once will require SMP locking anyways so there's
1486 * no real value in trying any harder than this.
1487 */
1488static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1489 .cache_all = viking_flush_cache_all,
1490 .cache_mm = viking_flush_cache_mm,
1491 .cache_page = viking_flush_cache_page,
1492 .cache_range = viking_flush_cache_range,
1493 .tlb_all = sun4dsmp_flush_tlb_all,
1494 .tlb_mm = sun4dsmp_flush_tlb_mm,
1495 .tlb_page = sun4dsmp_flush_tlb_page,
1496 .tlb_range = sun4dsmp_flush_tlb_range,
1497 .page_to_ram = viking_flush_page_to_ram,
1498 .sig_insns = viking_flush_sig_insns,
1499 .page_for_dma = viking_flush_page_for_dma,
1500};
1501#endif
1502
1503static void __init init_viking(void)
1504{
1505 unsigned long mreg = srmmu_get_mmureg();
1506
1507 /* Ahhh, the viking. SRMMU VLSI abortion number two... */
1508 if (mreg & VIKING_MMODE) {
1509 srmmu_name = "TI Viking";
1510 viking_mxcc_present = 0;
1511 msi_set_sync();
1512
1513 /*
1514 * We need this to make sure old viking takes no hits
1515 * on it's cache for dma snoops to workaround the
1516 * "load from non-cacheable memory" interrupt bug.
1517 * This is only necessary because of the new way in
1518 * which we use the IOMMU.
1519 */
1520 viking_ops.page_for_dma = viking_flush_page;
1521#ifdef CONFIG_SMP
1522 viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1523#endif
1524 flush_page_for_dma_global = 0;
1525 } else {
1526 srmmu_name = "TI Viking/MXCC";
1527 viking_mxcc_present = 1;
1528 srmmu_cache_pagetables = 1;
1529 }
1530
1531 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1532 &viking_ops;
1533#ifdef CONFIG_SMP
1534 if (sparc_cpu_model == sun4d)
1535 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1536 &viking_sun4d_smp_ops;
1537#endif
1538
1539 poke_srmmu = poke_viking;
1540}
1541
1542/* Probe for the srmmu chip version. */
1543static void __init get_srmmu_type(void)
1544{
1545 unsigned long mreg, psr;
1546 unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1547
1548 srmmu_modtype = SRMMU_INVAL_MOD;
1549 hwbug_bitmask = 0;
1550
1551 mreg = srmmu_get_mmureg(); psr = get_psr();
1552 mod_typ = (mreg & 0xf0000000) >> 28;
1553 mod_rev = (mreg & 0x0f000000) >> 24;
1554 psr_typ = (psr >> 28) & 0xf;
1555 psr_vers = (psr >> 24) & 0xf;
1556
1557 /* First, check for sparc-leon. */
1558 if (sparc_cpu_model == sparc_leon) {
1559 init_leon();
1560 return;
1561 }
1562
1563 /* Second, check for HyperSparc or Cypress. */
1564 if (mod_typ == 1) {
1565 switch (mod_rev) {
1566 case 7:
1567 /* UP or MP Hypersparc */
1568 init_hypersparc();
1569 break;
1570 case 0:
1571 case 2:
1572 case 10:
1573 case 11:
1574 case 12:
1575 case 13:
1576 case 14:
1577 case 15:
1578 default:
1579 prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1580 prom_halt();
1581 break;
1582 }
1583 return;
1584 }
1585
1586 /* Now Fujitsu TurboSparc. It might happen that it is
1587 * in Swift emulation mode, so we will check later...
1588 */
1589 if (psr_typ == 0 && psr_vers == 5) {
1590 init_turbosparc();
1591 return;
1592 }
1593
1594 /* Next check for Fujitsu Swift. */
1595 if (psr_typ == 0 && psr_vers == 4) {
1596 phandle cpunode;
1597 char node_str[128];
1598
1599 /* Look if it is not a TurboSparc emulating Swift... */
1600 cpunode = prom_getchild(prom_root_node);
1601 while ((cpunode = prom_getsibling(cpunode)) != 0) {
1602 prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1603 if (!strcmp(node_str, "cpu")) {
1604 if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1605 prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1606 init_turbosparc();
1607 return;
1608 }
1609 break;
1610 }
1611 }
1612
1613 init_swift();
1614 return;
1615 }
1616
1617 /* Now the Viking family of srmmu. */
1618 if (psr_typ == 4 &&
1619 ((psr_vers == 0) ||
1620 ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1621 init_viking();
1622 return;
1623 }
1624
1625 /* Finally the Tsunami. */
1626 if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1627 init_tsunami();
1628 return;
1629 }
1630
1631 /* Oh well */
1632 srmmu_is_bad();
1633}
1634
1635#ifdef CONFIG_SMP
1636/* Local cross-calls. */
1637static void smp_flush_page_for_dma(unsigned long page)
1638{
1639 xc1(local_ops->page_for_dma, page);
1640 local_ops->page_for_dma(page);
1641}
1642
1643static void smp_flush_cache_all(void)
1644{
1645 xc0(local_ops->cache_all);
1646 local_ops->cache_all();
1647}
1648
1649static void smp_flush_tlb_all(void)
1650{
1651 xc0(local_ops->tlb_all);
1652 local_ops->tlb_all();
1653}
1654
1655static void smp_flush_cache_mm(struct mm_struct *mm)
1656{
1657 if (mm->context != NO_CONTEXT) {
1658 cpumask_t cpu_mask;
1659 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1660 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1661 if (!cpumask_empty(&cpu_mask))
1662 xc1(local_ops->cache_mm, (unsigned long)mm);
1663 local_ops->cache_mm(mm);
1664 }
1665}
1666
1667static void smp_flush_tlb_mm(struct mm_struct *mm)
1668{
1669 if (mm->context != NO_CONTEXT) {
1670 cpumask_t cpu_mask;
1671 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1672 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1673 if (!cpumask_empty(&cpu_mask)) {
1674 xc1(local_ops->tlb_mm, (unsigned long)mm);
1675 if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1676 cpumask_copy(mm_cpumask(mm),
1677 cpumask_of(smp_processor_id()));
1678 }
1679 local_ops->tlb_mm(mm);
1680 }
1681}
1682
1683static void smp_flush_cache_range(struct vm_area_struct *vma,
1684 unsigned long start,
1685 unsigned long end)
1686{
1687 struct mm_struct *mm = vma->vm_mm;
1688
1689 if (mm->context != NO_CONTEXT) {
1690 cpumask_t cpu_mask;
1691 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1692 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1693 if (!cpumask_empty(&cpu_mask))
1694 xc3(local_ops->cache_range, (unsigned long)vma, start,
1695 end);
1696 local_ops->cache_range(vma, start, end);
1697 }
1698}
1699
1700static void smp_flush_tlb_range(struct vm_area_struct *vma,
1701 unsigned long start,
1702 unsigned long end)
1703{
1704 struct mm_struct *mm = vma->vm_mm;
1705
1706 if (mm->context != NO_CONTEXT) {
1707 cpumask_t cpu_mask;
1708 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1709 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1710 if (!cpumask_empty(&cpu_mask))
1711 xc3(local_ops->tlb_range, (unsigned long)vma, start,
1712 end);
1713 local_ops->tlb_range(vma, start, end);
1714 }
1715}
1716
1717static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1718{
1719 struct mm_struct *mm = vma->vm_mm;
1720
1721 if (mm->context != NO_CONTEXT) {
1722 cpumask_t cpu_mask;
1723 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1724 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1725 if (!cpumask_empty(&cpu_mask))
1726 xc2(local_ops->cache_page, (unsigned long)vma, page);
1727 local_ops->cache_page(vma, page);
1728 }
1729}
1730
1731static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1732{
1733 struct mm_struct *mm = vma->vm_mm;
1734
1735 if (mm->context != NO_CONTEXT) {
1736 cpumask_t cpu_mask;
1737 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1738 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1739 if (!cpumask_empty(&cpu_mask))
1740 xc2(local_ops->tlb_page, (unsigned long)vma, page);
1741 local_ops->tlb_page(vma, page);
1742 }
1743}
1744
1745static void smp_flush_page_to_ram(unsigned long page)
1746{
1747 /* Current theory is that those who call this are the one's
1748 * who have just dirtied their cache with the pages contents
1749 * in kernel space, therefore we only run this on local cpu.
1750 *
1751 * XXX This experiment failed, research further... -DaveM
1752 */
1753#if 1
1754 xc1(local_ops->page_to_ram, page);
1755#endif
1756 local_ops->page_to_ram(page);
1757}
1758
1759static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1760{
1761 cpumask_t cpu_mask;
1762 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1763 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1764 if (!cpumask_empty(&cpu_mask))
1765 xc2(local_ops->sig_insns, (unsigned long)mm, insn_addr);
1766 local_ops->sig_insns(mm, insn_addr);
1767}
1768
1769static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1770 .cache_all = smp_flush_cache_all,
1771 .cache_mm = smp_flush_cache_mm,
1772 .cache_page = smp_flush_cache_page,
1773 .cache_range = smp_flush_cache_range,
1774 .tlb_all = smp_flush_tlb_all,
1775 .tlb_mm = smp_flush_tlb_mm,
1776 .tlb_page = smp_flush_tlb_page,
1777 .tlb_range = smp_flush_tlb_range,
1778 .page_to_ram = smp_flush_page_to_ram,
1779 .sig_insns = smp_flush_sig_insns,
1780 .page_for_dma = smp_flush_page_for_dma,
1781};
1782#endif
1783
1784/* Load up routines and constants for sun4m and sun4d mmu */
1785void __init load_mmu(void)
1786{
1787 /* Functions */
1788 get_srmmu_type();
1789
1790#ifdef CONFIG_SMP
1791 /* El switcheroo... */
1792 local_ops = sparc32_cachetlb_ops;
1793
1794 if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1795 smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1796 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1797 smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1798 smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1799 }
1800
1801 if (poke_srmmu == poke_viking) {
1802 /* Avoid unnecessary cross calls. */
1803 smp_cachetlb_ops.cache_all = local_ops->cache_all;
1804 smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1805 smp_cachetlb_ops.cache_range = local_ops->cache_range;
1806 smp_cachetlb_ops.cache_page = local_ops->cache_page;
1807
1808 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1809 smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1810 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1811 }
1812
1813 /* It really is const after this point. */
1814 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1815 &smp_cachetlb_ops;
1816#endif
1817
1818 if (sparc_cpu_model != sun4d)
1819 ld_mmu_iommu();
1820#ifdef CONFIG_SMP
1821 if (sparc_cpu_model == sun4d)
1822 sun4d_init_smp();
1823 else if (sparc_cpu_model == sparc_leon)
1824 leon_init_smp();
1825 else
1826 sun4m_init_smp();
1827#endif
1828}
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}