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