1 // SPDX-License-Identifier: GPL-2.0
3 * srmmu.c: SRMMU specific routines for memory management.
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)
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>
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>
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>
54 enum mbus_module srmmu_modtype;
55 static unsigned int hwbug_bitmask;
57 EXPORT_SYMBOL(vac_cache_size);
60 extern struct resource sparc_iomap;
62 extern unsigned long last_valid_pfn;
64 static pgd_t *srmmu_swapper_pg_dir;
66 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
67 EXPORT_SYMBOL(sparc32_cachetlb_ops);
70 const struct sparc32_cachetlb_ops *local_ops;
72 #define FLUSH_BEGIN(mm)
75 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
79 int flush_page_for_dma_global = 1;
83 ctxd_t *srmmu_ctx_table_phys;
84 static ctxd_t *srmmu_context_table;
86 int viking_mxcc_present;
87 static DEFINE_SPINLOCK(srmmu_context_spinlock);
89 static int is_hypersparc;
91 static int srmmu_cache_pagetables;
93 /* these will be initialized in srmmu_nocache_calcsize() */
94 static unsigned long srmmu_nocache_size;
95 static unsigned long srmmu_nocache_end;
97 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
98 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
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)
103 void *srmmu_nocache_pool;
104 static struct bit_map srmmu_nocache_map;
106 static inline int srmmu_pmd_none(pmd_t pmd)
107 { return !(pmd_val(pmd) & 0xFFFFFFF); }
109 /* XXX should we hyper_flush_whole_icache here - Anton */
110 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
114 pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4)));
115 set_pte((pte_t *)ctxp, pte);
119 * Locations of MSI Registers.
121 #define MSI_MBUS_ARBEN 0xe0001008 /* MBus Arbiter Enable register */
124 * Useful bits in the MSI Registers.
126 #define MSI_ASYNC_MODE 0x80000000 /* Operate the MSI asynchronously */
128 static void msi_set_sync(void)
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");
137 void pmd_set(pmd_t *pmdp, pte_t *ptep)
139 unsigned long ptp = __nocache_pa(ptep) >> 4;
140 set_pte((pte_t *)&pmd_val(*pmdp), __pte(SRMMU_ET_PTD | ptp));
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.
148 static void *__srmmu_get_nocache(int size, int align)
150 int offset, minsz = 1 << SRMMU_NOCACHE_BITMAP_SHIFT;
154 printk(KERN_ERR "Size 0x%x too small for nocache request\n",
158 if (size & (minsz - 1)) {
159 printk(KERN_ERR "Size 0x%x unaligned in nocache request\n",
163 BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
165 offset = bit_map_string_get(&srmmu_nocache_map,
166 size >> SRMMU_NOCACHE_BITMAP_SHIFT,
167 align >> SRMMU_NOCACHE_BITMAP_SHIFT);
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);
175 addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
179 void *srmmu_get_nocache(int size, int align)
183 tmp = __srmmu_get_nocache(size, align);
186 memset(tmp, 0, size);
191 void srmmu_free_nocache(void *addr, int size)
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);
202 if (vaddr + size > srmmu_nocache_end) {
203 printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
204 vaddr, srmmu_nocache_end);
207 if (!is_power_of_2(size)) {
208 printk("Size 0x%x is not a power of 2\n", size);
211 if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
212 printk("Size 0x%x is too small\n", size);
215 if (vaddr & (size - 1)) {
216 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
220 offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
221 size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
223 bit_map_clear(&srmmu_nocache_map, offset, size);
226 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
229 /* Return how much physical memory we have. */
230 static unsigned long __init probe_memory(void)
232 unsigned long total = 0;
235 for (i = 0; sp_banks[i].num_bytes; i++)
236 total += sp_banks[i].num_bytes;
242 * Reserve nocache dynamically proportionally to the amount of
243 * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
245 static void __init srmmu_nocache_calcsize(void)
247 unsigned long sysmemavail = probe_memory() / 1024;
248 int srmmu_nocache_npages;
250 srmmu_nocache_npages =
251 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
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;
258 /* anything above 1280 blows up */
259 if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
260 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
262 srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
263 srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
266 static void __init srmmu_nocache_init(void)
268 void *srmmu_nocache_bitmap;
269 unsigned int bitmap_bits;
275 unsigned long paddr, vaddr;
276 unsigned long pteval;
278 bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
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);
287 srmmu_nocache_bitmap =
288 memblock_alloc(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
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);
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;
299 srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
301 paddr = __pa((unsigned long)srmmu_nocache_pool);
302 vaddr = SRMMU_NOCACHE_VADDR;
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);
311 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
313 if (srmmu_cache_pagetables)
314 pteval |= SRMMU_CACHE;
316 set_pte(__nocache_fix(pte), __pte(pteval));
326 pgd_t *get_pgd_fast(void)
330 pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
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));
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.
346 * Alignments up to the page size are the same for physical and virtual
347 * addresses of the nocache area.
349 pgtable_t pte_alloc_one(struct mm_struct *mm)
354 if (!(ptep = pte_alloc_one_kernel(mm)))
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))) {
363 spin_unlock(&mm->page_table_lock);
368 void pte_free(struct mm_struct *mm, pgtable_t ptep)
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);
378 srmmu_free_nocache(ptep, SRMMU_PTE_TABLE_SIZE);
381 /* context handling - a dynamically sized pool is used */
382 #define NO_CONTEXT -1
385 struct ctx_list *next;
386 struct ctx_list *prev;
387 unsigned int ctx_number;
388 struct mm_struct *ctx_mm;
391 static struct ctx_list *ctx_list_pool;
392 static struct ctx_list ctx_free;
393 static struct ctx_list ctx_used;
395 /* At boot time we determine the number of contexts */
396 static int num_contexts;
398 static inline void remove_from_ctx_list(struct ctx_list *entry)
400 entry->next->prev = entry->prev;
401 entry->prev->next = entry->next;
404 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
407 (entry->prev = head->prev)->next = entry;
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)
414 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
416 struct ctx_list *ctxp;
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;
426 ctxp = ctx_used.next;
427 if (ctxp->ctx_mm == old_mm)
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;
437 mm->context = ctxp->ctx_number;
440 static inline void free_context(int context)
442 struct ctx_list *ctx_old;
444 ctx_old = ctx_list_pool + context;
445 remove_from_ctx_list(ctx_old);
446 add_to_free_ctxlist(ctx_old);
449 static void __init sparc_context_init(int numctx)
454 size = numctx * sizeof(struct ctx_list);
455 ctx_list_pool = memblock_alloc(size, SMP_CACHE_BYTES);
457 panic("%s: Failed to allocate %lu bytes\n", __func__, size);
459 for (ctx = 0; ctx < numctx; ctx++) {
460 struct ctx_list *clist;
462 clist = (ctx_list_pool + ctx);
463 clist->ctx_number = ctx;
464 clist->ctx_mm = NULL;
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);
472 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
473 struct task_struct *tsk)
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);
484 if (sparc_cpu_model == sparc_leon)
488 hyper_flush_whole_icache();
490 srmmu_set_context(mm->context);
493 /* Low level IO area allocation on the SRMMU. */
494 static inline void srmmu_mapioaddr(unsigned long physaddr,
495 unsigned long virt_addr, int bus_type)
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;
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...
516 tmp |= (bus_type << 28);
518 __flush_page_to_ram(virt_addr);
519 set_pte(ptep, __pte(tmp));
522 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
523 unsigned long xva, unsigned int len)
527 srmmu_mapioaddr(xpa, xva, bus);
534 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
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);
549 /* No need to flush uncacheable page. */
553 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
557 srmmu_unmapioaddr(virt_addr);
558 virt_addr += PAGE_SIZE;
564 extern void tsunami_flush_cache_all(void);
565 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
566 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
567 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
568 extern void tsunami_flush_page_to_ram(unsigned long page);
569 extern void tsunami_flush_page_for_dma(unsigned long page);
570 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
571 extern void tsunami_flush_tlb_all(void);
572 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
573 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
574 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
575 extern void tsunami_setup_blockops(void);
578 extern void swift_flush_cache_all(void);
579 extern void swift_flush_cache_mm(struct mm_struct *mm);
580 extern void swift_flush_cache_range(struct vm_area_struct *vma,
581 unsigned long start, unsigned long end);
582 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
583 extern void swift_flush_page_to_ram(unsigned long page);
584 extern void swift_flush_page_for_dma(unsigned long page);
585 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
586 extern void swift_flush_tlb_all(void);
587 extern void swift_flush_tlb_mm(struct mm_struct *mm);
588 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
589 unsigned long start, unsigned long end);
590 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
592 #if 0 /* P3: deadwood to debug precise flushes on Swift. */
593 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
598 if ((ctx1 = vma->vm_mm->context) != -1) {
599 cctx = srmmu_get_context();
600 /* Is context # ever different from current context? P3 */
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);
609 /* Rm. prot. bits from virt. c. */
610 /* swift_flush_cache_all(); */
611 /* swift_flush_cache_page(vma, page); */
612 swift_flush_page(page);
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() */
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.
630 extern void viking_flush_cache_all(void);
631 extern void viking_flush_cache_mm(struct mm_struct *mm);
632 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
634 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
635 extern void viking_flush_page_to_ram(unsigned long page);
636 extern void viking_flush_page_for_dma(unsigned long page);
637 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
638 extern void viking_flush_page(unsigned long page);
639 extern void viking_mxcc_flush_page(unsigned long page);
640 extern void viking_flush_tlb_all(void);
641 extern void viking_flush_tlb_mm(struct mm_struct *mm);
642 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
644 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
646 extern void sun4dsmp_flush_tlb_all(void);
647 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
648 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
650 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
654 extern void hypersparc_flush_cache_all(void);
655 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
656 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
657 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
658 extern void hypersparc_flush_page_to_ram(unsigned long page);
659 extern void hypersparc_flush_page_for_dma(unsigned long page);
660 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
661 extern void hypersparc_flush_tlb_all(void);
662 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
663 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
664 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
665 extern void hypersparc_setup_blockops(void);
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.
674 static void __init early_pgtable_allocfail(char *type)
676 prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
680 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
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);
697 early_pgtable_allocfail("pmd");
698 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
699 pud_set(__nocache_fix(pudp), pmdp);
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);
705 early_pgtable_allocfail("pte");
706 memset(__nocache_fix(ptep), 0, PTE_SIZE);
707 pmd_set(__nocache_fix(pmdp), ptep);
709 if (start > (0xffffffffUL - PMD_SIZE))
711 start = (start + PMD_SIZE) & PMD_MASK;
715 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
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);
731 early_pgtable_allocfail("pmd");
732 memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
733 pud_set((pud_t *)pgdp, pmdp);
735 pmdp = pmd_offset(pudp, start);
736 if (srmmu_pmd_none(*pmdp)) {
737 ptep = __srmmu_get_nocache(PTE_SIZE,
740 early_pgtable_allocfail("pte");
741 memset(ptep, 0, PTE_SIZE);
744 if (start > (0xffffffffUL - PMD_SIZE))
746 start = (start + PMD_SIZE) & PMD_MASK;
750 /* These flush types are not available on all chips... */
751 static inline unsigned long srmmu_probe(unsigned long vaddr)
753 unsigned long retval;
755 if (sparc_cpu_model != sparc_leon) {
758 __asm__ __volatile__("lda [%1] %2, %0\n\t" :
760 "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
762 retval = leon_swprobe(vaddr, NULL);
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.
772 static void __init srmmu_inherit_prom_mappings(unsigned long start,
775 unsigned long probed;
782 int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
784 while (start <= end) {
786 break; /* probably wrap around */
787 if (start == 0xfef00000)
788 start = KADB_DEBUGGER_BEGVM;
789 probed = srmmu_probe(start);
791 /* continue probing until we find an entry */
796 /* A red snapper, see what it really is. */
798 addr = start - PAGE_SIZE;
800 if (!(start & ~(PMD_MASK))) {
801 if (srmmu_probe(addr + PMD_SIZE) == probed)
805 if (!(start & ~(PGDIR_MASK))) {
806 if (srmmu_probe(addr + PGDIR_SIZE) == probed)
810 pgdp = pgd_offset_k(start);
811 p4dp = p4d_offset(pgdp, start);
812 pudp = pud_offset(p4dp, start);
814 *__nocache_fix(pgdp) = __pgd(probed);
818 if (pud_none(*__nocache_fix(pudp))) {
819 pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
820 SRMMU_PMD_TABLE_SIZE);
822 early_pgtable_allocfail("pmd");
823 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
824 pud_set(__nocache_fix(pudp), pmdp);
826 pmdp = pmd_offset(__nocache_fix(pudp), start);
828 *(pmd_t *)__nocache_fix(pmdp) = __pmd(probed);
832 if (srmmu_pmd_none(*__nocache_fix(pmdp))) {
833 ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
835 early_pgtable_allocfail("pte");
836 memset(__nocache_fix(ptep), 0, PTE_SIZE);
837 pmd_set(__nocache_fix(pmdp), ptep);
839 ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
840 *__nocache_fix(ptep) = __pte(probed);
845 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
847 /* Create a third-level SRMMU 16MB page mapping. */
848 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
850 pgd_t *pgdp = pgd_offset_k(vaddr);
851 unsigned long big_pte;
853 big_pte = KERNEL_PTE(phys_base >> 4);
854 *__nocache_fix(pgdp) = __pgd(big_pte);
857 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
858 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
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;
867 if (vstart < min_vaddr || vstart >= max_vaddr)
870 if (vend > max_vaddr || vend < min_vaddr)
873 while (vstart < vend) {
874 do_large_mapping(vstart, pstart);
875 vstart += PGDIR_SIZE; pstart += PGDIR_SIZE;
880 static void __init map_kernel(void)
885 do_large_mapping(PAGE_OFFSET, phys_base);
888 for (i = 0; sp_banks[i].num_bytes != 0; i++) {
889 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
893 void (*poke_srmmu)(void) = NULL;
895 void __init srmmu_paging_init(void)
905 unsigned long pages_avail;
907 init_mm.context = (unsigned long) NO_CONTEXT;
908 sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
910 if (sparc_cpu_model == sun4d)
911 num_contexts = 65536; /* We know it is Viking */
913 /* Find the number of contexts on the srmmu. */
914 cpunode = prom_getchild(prom_root_node);
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);
922 cpunode = prom_getsibling(cpunode);
927 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
932 last_valid_pfn = bootmem_init(&pages_avail);
934 srmmu_nocache_calcsize();
935 srmmu_nocache_init();
936 srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
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);
943 for (i = 0; i < num_contexts; i++)
944 srmmu_ctxd_set(__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
947 srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
949 /* Stop from hanging here... */
950 local_ops->tlb_all();
956 srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
957 srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
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);
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;
973 sparc_context_init(num_contexts);
976 unsigned long max_zone_pfn[MAX_NR_ZONES] = { 0 };
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;
982 free_area_init(max_zone_pfn);
986 void mmu_info(struct seq_file *m)
991 "nocache total\t: %ld\n"
992 "nocache used\t: %d\n",
996 srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
999 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
1001 mm->context = NO_CONTEXT;
1005 void destroy_context(struct mm_struct *mm)
1007 unsigned long flags;
1009 if (mm->context != NO_CONTEXT) {
1011 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
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;
1020 /* Init various srmmu chip types. */
1021 static void __init srmmu_is_bad(void)
1023 prom_printf("Could not determine SRMMU chip type.\n");
1027 static void __init init_vac_layout(void)
1034 unsigned long max_size = 0;
1035 unsigned long min_line_size = 0x10000000;
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");
1047 cache_lines = prom_getint(nd, "cache-nlines");
1048 if (cache_lines == -1) {
1049 prom_printf("can't determine cache-nlines, halting.\n");
1053 vac_cache_size = cache_lines * vac_line_size;
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!!
1061 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1069 prom_printf("No CPU nodes found, halting.\n");
1073 vac_cache_size = max_size;
1074 vac_line_size = min_line_size;
1076 printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1077 (int)vac_cache_size, (int)vac_line_size);
1080 static void poke_hypersparc(void)
1082 volatile unsigned long clear;
1083 unsigned long mreg = srmmu_get_mmureg();
1085 hyper_flush_unconditional_combined();
1087 mreg &= ~(HYPERSPARC_CWENABLE);
1088 mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1089 mreg |= (HYPERSPARC_CMODE);
1091 srmmu_set_mmureg(mreg);
1093 #if 0 /* XXX I think this is bad news... -DaveM */
1094 hyper_clear_all_tags();
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();
1103 static 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,
1117 static void __init init_hypersparc(void)
1119 srmmu_name = "ROSS HyperSparc";
1120 srmmu_modtype = HyperSparc;
1125 sparc32_cachetlb_ops = &hypersparc_ops;
1127 poke_srmmu = poke_hypersparc;
1129 hypersparc_setup_blockops();
1132 static void poke_swift(void)
1136 /* Clear any crap from the cache or else... */
1137 swift_flush_cache_all();
1139 /* Enable I & D caches */
1140 mreg = srmmu_get_mmureg();
1141 mreg |= (SWIFT_IE | SWIFT_DE);
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.
1151 mreg &= ~(SWIFT_BF);
1152 srmmu_set_mmureg(mreg);
1155 static 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,
1169 #define SWIFT_MASKID_ADDR 0x10003018
1170 static void __init init_swift(void)
1172 unsigned long swift_rev;
1174 __asm__ __volatile__("lda [%1] %2, %0\n\t"
1175 "srl %0, 0x18, %0\n\t" :
1177 "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1178 srmmu_name = "Fujitsu Swift";
1179 switch (swift_rev) {
1184 srmmu_modtype = Swift_lots_o_bugs;
1185 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
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"
1206 srmmu_modtype = Swift_bad_c;
1207 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
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.
1215 srmmu_modtype = Swift_ok;
1219 sparc32_cachetlb_ops = &swift_ops;
1220 flush_page_for_dma_global = 0;
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.
1229 poke_srmmu = poke_swift;
1232 static void turbosparc_flush_cache_all(void)
1234 flush_user_windows();
1235 turbosparc_idflash_clear();
1238 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1241 flush_user_windows();
1242 turbosparc_idflash_clear();
1246 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1248 FLUSH_BEGIN(vma->vm_mm)
1249 flush_user_windows();
1250 turbosparc_idflash_clear();
1254 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
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();
1264 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1265 static void turbosparc_flush_page_to_ram(unsigned long page)
1267 #ifdef TURBOSPARC_WRITEBACK
1268 volatile unsigned long clear;
1270 if (srmmu_probe(page))
1271 turbosparc_flush_page_cache(page);
1272 clear = srmmu_get_fstatus();
1276 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1280 static void turbosparc_flush_page_for_dma(unsigned long page)
1282 turbosparc_flush_dcache();
1285 static void turbosparc_flush_tlb_all(void)
1287 srmmu_flush_whole_tlb();
1290 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1293 srmmu_flush_whole_tlb();
1297 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1299 FLUSH_BEGIN(vma->vm_mm)
1300 srmmu_flush_whole_tlb();
1304 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1306 FLUSH_BEGIN(vma->vm_mm)
1307 srmmu_flush_whole_tlb();
1312 static void poke_turbosparc(void)
1314 unsigned long mreg = srmmu_get_mmureg();
1315 unsigned long ccreg;
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);
1324 ccreg = turbosparc_get_ccreg();
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 */
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 */
1339 switch (ccreg & 7) {
1340 case 0: /* No SE cache */
1341 case 7: /* Test mode */
1344 ccreg |= (TURBOSPARC_SCENABLE);
1346 turbosparc_set_ccreg(ccreg);
1348 mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1349 mreg |= (TURBOSPARC_ICSNOOP); /* Icache snooping on */
1350 srmmu_set_mmureg(mreg);
1353 static 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,
1367 static void __init init_turbosparc(void)
1369 srmmu_name = "Fujitsu TurboSparc";
1370 srmmu_modtype = TurboSparc;
1371 sparc32_cachetlb_ops = &turbosparc_ops;
1372 poke_srmmu = poke_turbosparc;
1375 static void poke_tsunami(void)
1377 unsigned long mreg = srmmu_get_mmureg();
1379 tsunami_flush_icache();
1380 tsunami_flush_dcache();
1381 mreg &= ~TSUNAMI_ITD;
1382 mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1383 srmmu_set_mmureg(mreg);
1386 static 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,
1400 static void __init init_tsunami(void)
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.
1408 srmmu_name = "TI Tsunami";
1409 srmmu_modtype = Tsunami;
1410 sparc32_cachetlb_ops = &tsunami_ops;
1411 poke_srmmu = poke_tsunami;
1413 tsunami_setup_blockops();
1416 static void poke_viking(void)
1418 unsigned long mreg = srmmu_get_mmureg();
1419 static int smp_catch;
1421 if (viking_mxcc_present) {
1422 unsigned long mxcc_control = mxcc_get_creg();
1424 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1425 mxcc_control &= ~(MXCC_CTL_RRC);
1426 mxcc_set_creg(mxcc_control);
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);
1434 /* We do cache ptables on MXCC. */
1435 mreg |= VIKING_TCENABLE;
1437 unsigned long bpreg;
1439 mreg &= ~(VIKING_TCENABLE);
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);
1446 /* Just in case PROM does something funny. */
1451 mreg |= VIKING_SPENABLE;
1452 mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1453 mreg |= VIKING_SBENABLE;
1454 mreg &= ~(VIKING_ACENABLE);
1455 srmmu_set_mmureg(mreg);
1458 static 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,
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.
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.
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.
1489 static 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,
1504 static void __init init_viking(void)
1506 unsigned long mreg = srmmu_get_mmureg();
1508 /* Ahhh, the viking. SRMMU VLSI abortion number two... */
1509 if (mreg & VIKING_MMODE) {
1510 srmmu_name = "TI Viking";
1511 viking_mxcc_present = 0;
1515 * We need this to make sure old viking takes no hits
1516 * on it's 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.
1521 viking_ops.page_for_dma = viking_flush_page;
1523 viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1525 flush_page_for_dma_global = 0;
1527 srmmu_name = "TI Viking/MXCC";
1528 viking_mxcc_present = 1;
1529 srmmu_cache_pagetables = 1;
1532 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1535 if (sparc_cpu_model == sun4d)
1536 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1537 &viking_sun4d_smp_ops;
1540 poke_srmmu = poke_viking;
1543 /* Probe for the srmmu chip version. */
1544 static void __init get_srmmu_type(void)
1546 unsigned long mreg, psr;
1547 unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1549 srmmu_modtype = SRMMU_INVAL_MOD;
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;
1558 /* First, check for sparc-leon. */
1559 if (sparc_cpu_model == sparc_leon) {
1564 /* Second, check for HyperSparc or Cypress. */
1568 /* UP or MP Hypersparc */
1580 prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1587 /* Now Fujitsu TurboSparc. It might happen that it is
1588 * in Swift emulation mode, so we will check later...
1590 if (psr_typ == 0 && psr_vers == 5) {
1595 /* Next check for Fujitsu Swift. */
1596 if (psr_typ == 0 && psr_vers == 4) {
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) {
1618 /* Now the Viking family of srmmu. */
1621 ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1626 /* Finally the Tsunami. */
1627 if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1637 /* Local cross-calls. */
1638 static void smp_flush_page_for_dma(unsigned long page)
1640 xc1(local_ops->page_for_dma, page);
1641 local_ops->page_for_dma(page);
1644 static void smp_flush_cache_all(void)
1646 xc0(local_ops->cache_all);
1647 local_ops->cache_all();
1650 static void smp_flush_tlb_all(void)
1652 xc0(local_ops->tlb_all);
1653 local_ops->tlb_all();
1656 static void smp_flush_cache_mm(struct mm_struct *mm)
1658 if (mm->context != NO_CONTEXT) {
1660 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1661 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1662 if (!cpumask_empty(&cpu_mask))
1663 xc1(local_ops->cache_mm, (unsigned long)mm);
1664 local_ops->cache_mm(mm);
1668 static void smp_flush_tlb_mm(struct mm_struct *mm)
1670 if (mm->context != NO_CONTEXT) {
1672 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1673 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1674 if (!cpumask_empty(&cpu_mask)) {
1675 xc1(local_ops->tlb_mm, (unsigned long)mm);
1676 if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1677 cpumask_copy(mm_cpumask(mm),
1678 cpumask_of(smp_processor_id()));
1680 local_ops->tlb_mm(mm);
1684 static void smp_flush_cache_range(struct vm_area_struct *vma,
1685 unsigned long start,
1688 struct mm_struct *mm = vma->vm_mm;
1690 if (mm->context != NO_CONTEXT) {
1692 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1693 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1694 if (!cpumask_empty(&cpu_mask))
1695 xc3(local_ops->cache_range, (unsigned long)vma, start,
1697 local_ops->cache_range(vma, start, end);
1701 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1702 unsigned long start,
1705 struct mm_struct *mm = vma->vm_mm;
1707 if (mm->context != NO_CONTEXT) {
1709 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1710 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1711 if (!cpumask_empty(&cpu_mask))
1712 xc3(local_ops->tlb_range, (unsigned long)vma, start,
1714 local_ops->tlb_range(vma, start, end);
1718 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1720 struct mm_struct *mm = vma->vm_mm;
1722 if (mm->context != NO_CONTEXT) {
1724 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1725 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1726 if (!cpumask_empty(&cpu_mask))
1727 xc2(local_ops->cache_page, (unsigned long)vma, page);
1728 local_ops->cache_page(vma, page);
1732 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1734 struct mm_struct *mm = vma->vm_mm;
1736 if (mm->context != NO_CONTEXT) {
1738 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1739 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1740 if (!cpumask_empty(&cpu_mask))
1741 xc2(local_ops->tlb_page, (unsigned long)vma, page);
1742 local_ops->tlb_page(vma, page);
1746 static void smp_flush_page_to_ram(unsigned long page)
1748 /* Current theory is that those who call this are the one's
1749 * who have just dirtied their cache with the pages contents
1750 * in kernel space, therefore we only run this on local cpu.
1752 * XXX This experiment failed, research further... -DaveM
1755 xc1(local_ops->page_to_ram, page);
1757 local_ops->page_to_ram(page);
1760 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1763 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1764 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1765 if (!cpumask_empty(&cpu_mask))
1766 xc2(local_ops->sig_insns, (unsigned long)mm, insn_addr);
1767 local_ops->sig_insns(mm, insn_addr);
1770 static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1771 .cache_all = smp_flush_cache_all,
1772 .cache_mm = smp_flush_cache_mm,
1773 .cache_page = smp_flush_cache_page,
1774 .cache_range = smp_flush_cache_range,
1775 .tlb_all = smp_flush_tlb_all,
1776 .tlb_mm = smp_flush_tlb_mm,
1777 .tlb_page = smp_flush_tlb_page,
1778 .tlb_range = smp_flush_tlb_range,
1779 .page_to_ram = smp_flush_page_to_ram,
1780 .sig_insns = smp_flush_sig_insns,
1781 .page_for_dma = smp_flush_page_for_dma,
1785 /* Load up routines and constants for sun4m and sun4d mmu */
1786 void __init load_mmu(void)
1792 /* El switcheroo... */
1793 local_ops = sparc32_cachetlb_ops;
1795 if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1796 smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1797 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1798 smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1799 smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1802 if (poke_srmmu == poke_viking) {
1803 /* Avoid unnecessary cross calls. */
1804 smp_cachetlb_ops.cache_all = local_ops->cache_all;
1805 smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1806 smp_cachetlb_ops.cache_range = local_ops->cache_range;
1807 smp_cachetlb_ops.cache_page = local_ops->cache_page;
1809 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1810 smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1811 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1814 /* It really is const after this point. */
1815 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1819 if (sparc_cpu_model != sun4d)
1822 if (sparc_cpu_model == sun4d)
1824 else if (sparc_cpu_model == sparc_leon)