2 * arch/sh/mm/cache-sh4.c
4 * Copyright (C) 1999, 2000, 2002 Niibe Yutaka
5 * Copyright (C) 2001 - 2006 Paul Mundt
6 * Copyright (C) 2003 Richard Curnow
8 * This file is subject to the terms and conditions of the GNU General Public
9 * License. See the file "COPYING" in the main directory of this archive
12 #include <linux/init.h>
15 #include <linux/mutex.h>
16 #include <asm/mmu_context.h>
17 #include <asm/cacheflush.h>
20 * The maximum number of pages we support up to when doing ranged dcache
21 * flushing. Anything exceeding this will simply flush the dcache in its
24 #define MAX_DCACHE_PAGES 64 /* XXX: Tune for ways */
26 static void __flush_dcache_segment_1way(unsigned long start,
27 unsigned long extent);
28 static void __flush_dcache_segment_2way(unsigned long start,
29 unsigned long extent);
30 static void __flush_dcache_segment_4way(unsigned long start,
31 unsigned long extent);
33 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
34 unsigned long exec_offset);
37 * This is initialised here to ensure that it is not placed in the BSS. If
38 * that were to happen, note that cache_init gets called before the BSS is
39 * cleared, so this would get nulled out which would be hopeless.
41 static void (*__flush_dcache_segment_fn)(unsigned long, unsigned long) =
42 (void (*)(unsigned long, unsigned long))0xdeadbeef;
44 static void compute_alias(struct cache_info *c)
46 c->alias_mask = ((c->sets - 1) << c->entry_shift) & ~(PAGE_SIZE - 1);
47 c->n_aliases = (c->alias_mask >> PAGE_SHIFT) + 1;
50 static void __init emit_cache_params(void)
52 printk("PVR=%08x CVR=%08x PRR=%08x\n",
56 printk("I-cache : n_ways=%d n_sets=%d way_incr=%d\n",
57 current_cpu_data.icache.ways,
58 current_cpu_data.icache.sets,
59 current_cpu_data.icache.way_incr);
60 printk("I-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
61 current_cpu_data.icache.entry_mask,
62 current_cpu_data.icache.alias_mask,
63 current_cpu_data.icache.n_aliases);
64 printk("D-cache : n_ways=%d n_sets=%d way_incr=%d\n",
65 current_cpu_data.dcache.ways,
66 current_cpu_data.dcache.sets,
67 current_cpu_data.dcache.way_incr);
68 printk("D-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
69 current_cpu_data.dcache.entry_mask,
70 current_cpu_data.dcache.alias_mask,
71 current_cpu_data.dcache.n_aliases);
73 if (!__flush_dcache_segment_fn)
74 panic("unknown number of cache ways\n");
78 * SH-4 has virtually indexed and physically tagged cache.
81 /* Worst case assumed to be 64k cache, direct-mapped i.e. 4 synonym bits. */
82 #define MAX_P3_MUTEXES 16
84 struct mutex p3map_mutex[MAX_P3_MUTEXES];
86 void __init p3_cache_init(void)
90 compute_alias(¤t_cpu_data.icache);
91 compute_alias(¤t_cpu_data.dcache);
93 switch (current_cpu_data.dcache.ways) {
95 __flush_dcache_segment_fn = __flush_dcache_segment_1way;
98 __flush_dcache_segment_fn = __flush_dcache_segment_2way;
101 __flush_dcache_segment_fn = __flush_dcache_segment_4way;
104 __flush_dcache_segment_fn = NULL;
110 if (ioremap_page_range(P3SEG, P3SEG + (PAGE_SIZE * 4), 0, PAGE_KERNEL))
111 panic("%s failed.", __FUNCTION__);
113 for (i = 0; i < current_cpu_data.dcache.n_aliases; i++)
114 mutex_init(&p3map_mutex[i]);
118 * Write back the dirty D-caches, but not invalidate them.
120 * START: Virtual Address (U0, P1, or P3)
121 * SIZE: Size of the region.
123 void __flush_wback_region(void *start, int size)
126 unsigned long begin, end;
128 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
129 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
130 & ~(L1_CACHE_BYTES-1);
131 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
132 asm volatile("ocbwb %0"
139 * Write back the dirty D-caches and invalidate them.
141 * START: Virtual Address (U0, P1, or P3)
142 * SIZE: Size of the region.
144 void __flush_purge_region(void *start, int size)
147 unsigned long begin, end;
149 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
150 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
151 & ~(L1_CACHE_BYTES-1);
152 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
153 asm volatile("ocbp %0"
160 * No write back please
162 void __flush_invalidate_region(void *start, int size)
165 unsigned long begin, end;
167 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
168 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
169 & ~(L1_CACHE_BYTES-1);
170 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
171 asm volatile("ocbi %0"
178 * Write back the range of D-cache, and purge the I-cache.
180 * Called from kernel/module.c:sys_init_module and routine for a.out format.
182 void flush_icache_range(unsigned long start, unsigned long end)
188 * Write back the D-cache and purge the I-cache for signal trampoline.
189 * .. which happens to be the same behavior as flush_icache_range().
190 * So, we simply flush out a line.
192 void flush_cache_sigtramp(unsigned long addr)
194 unsigned long v, index;
198 v = addr & ~(L1_CACHE_BYTES-1);
199 asm volatile("ocbwb %0"
203 index = CACHE_IC_ADDRESS_ARRAY |
204 (v & current_cpu_data.icache.entry_mask);
206 local_irq_save(flags);
209 for (i = 0; i < current_cpu_data.icache.ways;
210 i++, index += current_cpu_data.icache.way_incr)
211 ctrl_outl(0, index); /* Clear out Valid-bit */
215 local_irq_restore(flags);
218 static inline void flush_cache_4096(unsigned long start,
221 unsigned long flags, exec_offset = 0;
224 * All types of SH-4 require PC to be in P2 to operate on the I-cache.
225 * Some types of SH-4 require PC to be in P2 to operate on the D-cache.
227 if ((current_cpu_data.flags & CPU_HAS_P2_FLUSH_BUG) ||
228 (start < CACHE_OC_ADDRESS_ARRAY))
229 exec_offset = 0x20000000;
231 local_irq_save(flags);
232 __flush_cache_4096(start | SH_CACHE_ASSOC,
233 P1SEGADDR(phys), exec_offset);
234 local_irq_restore(flags);
238 * Write back & invalidate the D-cache of the page.
239 * (To avoid "alias" issues)
241 void flush_dcache_page(struct page *page)
243 if (test_bit(PG_mapped, &page->flags)) {
244 unsigned long phys = PHYSADDR(page_address(page));
245 unsigned long addr = CACHE_OC_ADDRESS_ARRAY;
248 /* Loop all the D-cache */
249 n = current_cpu_data.dcache.n_aliases;
250 for (i = 0; i < n; i++, addr += 4096)
251 flush_cache_4096(addr, phys);
257 /* TODO: Selective icache invalidation through IC address array.. */
258 static inline void flush_icache_all(void)
260 unsigned long flags, ccr;
262 local_irq_save(flags);
267 ccr |= CCR_CACHE_ICI;
271 * back_to_P1() will take care of the barrier for us, don't add
276 local_irq_restore(flags);
279 void flush_dcache_all(void)
281 (*__flush_dcache_segment_fn)(0UL, current_cpu_data.dcache.way_size);
285 void flush_cache_all(void)
291 static void __flush_cache_mm(struct mm_struct *mm, unsigned long start,
294 unsigned long d = 0, p = start & PAGE_MASK;
295 unsigned long alias_mask = current_cpu_data.dcache.alias_mask;
296 unsigned long n_aliases = current_cpu_data.dcache.n_aliases;
297 unsigned long select_bit;
298 unsigned long all_aliases_mask;
299 unsigned long addr_offset;
306 dir = pgd_offset(mm, p);
307 pud = pud_offset(dir, p);
308 pmd = pmd_offset(pud, p);
309 end = PAGE_ALIGN(end);
311 all_aliases_mask = (1 << n_aliases) - 1;
314 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) {
322 pte = pte_offset_kernel(pmd, p);
328 if (!(pte_val(entry) & _PAGE_PRESENT)) {
334 phys = pte_val(entry) & PTE_PHYS_MASK;
336 if ((p ^ phys) & alias_mask) {
337 d |= 1 << ((p & alias_mask) >> PAGE_SHIFT);
338 d |= 1 << ((phys & alias_mask) >> PAGE_SHIFT);
340 if (d == all_aliases_mask)
346 } while (p < end && ((unsigned long)pte & ~PAGE_MASK));
354 for (i = 0; i < n_aliases; i++) {
355 if (d & select_bit) {
356 (*__flush_dcache_segment_fn)(addr_offset, PAGE_SIZE);
361 addr_offset += PAGE_SIZE;
366 * Note : (RPC) since the caches are physically tagged, the only point
367 * of flush_cache_mm for SH-4 is to get rid of aliases from the
368 * D-cache. The assumption elsewhere, e.g. flush_cache_range, is that
369 * lines can stay resident so long as the virtual address they were
370 * accessed with (hence cache set) is in accord with the physical
371 * address (i.e. tag). It's no different here. So I reckon we don't
372 * need to flush the I-cache, since aliases don't matter for that. We
375 * Caller takes mm->mmap_sem.
377 void flush_cache_mm(struct mm_struct *mm)
380 * If cache is only 4k-per-way, there are never any 'aliases'. Since
381 * the cache is physically tagged, the data can just be left in there.
383 if (current_cpu_data.dcache.n_aliases == 0)
387 * Don't bother groveling around the dcache for the VMA ranges
388 * if there are too many PTEs to make it worthwhile.
390 if (mm->nr_ptes >= MAX_DCACHE_PAGES)
393 struct vm_area_struct *vma;
396 * In this case there are reasonably sized ranges to flush,
397 * iterate through the VMA list and take care of any aliases.
399 for (vma = mm->mmap; vma; vma = vma->vm_next)
400 __flush_cache_mm(mm, vma->vm_start, vma->vm_end);
403 /* Only touch the icache if one of the VMAs has VM_EXEC set. */
409 * Write back and invalidate I/D-caches for the page.
411 * ADDR: Virtual Address (U0 address)
412 * PFN: Physical page number
414 void flush_cache_page(struct vm_area_struct *vma, unsigned long address,
417 unsigned long phys = pfn << PAGE_SHIFT;
418 unsigned int alias_mask;
420 alias_mask = current_cpu_data.dcache.alias_mask;
422 /* We only need to flush D-cache when we have alias */
423 if ((address^phys) & alias_mask) {
424 /* Loop 4K of the D-cache */
426 CACHE_OC_ADDRESS_ARRAY | (address & alias_mask),
428 /* Loop another 4K of the D-cache */
430 CACHE_OC_ADDRESS_ARRAY | (phys & alias_mask),
434 alias_mask = current_cpu_data.icache.alias_mask;
435 if (vma->vm_flags & VM_EXEC) {
437 * Evict entries from the portion of the cache from which code
438 * may have been executed at this address (virtual). There's
439 * no need to evict from the portion corresponding to the
440 * physical address as for the D-cache, because we know the
441 * kernel has never executed the code through its identity
445 CACHE_IC_ADDRESS_ARRAY | (address & alias_mask),
451 * Write back and invalidate D-caches.
453 * START, END: Virtual Address (U0 address)
455 * NOTE: We need to flush the _physical_ page entry.
456 * Flushing the cache lines for U0 only isn't enough.
457 * We need to flush for P1 too, which may contain aliases.
459 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
463 * If cache is only 4k-per-way, there are never any 'aliases'. Since
464 * the cache is physically tagged, the data can just be left in there.
466 if (current_cpu_data.dcache.n_aliases == 0)
470 * Don't bother with the lookup and alias check if we have a
471 * wide range to cover, just blow away the dcache in its
472 * entirety instead. -- PFM.
474 if (((end - start) >> PAGE_SHIFT) >= MAX_DCACHE_PAGES)
477 __flush_cache_mm(vma->vm_mm, start, end);
479 if (vma->vm_flags & VM_EXEC) {
481 * TODO: Is this required??? Need to look at how I-cache
482 * coherency is assured when new programs are loaded to see if
490 * flush_icache_user_range
491 * @vma: VMA of the process
494 * @len: length of the range (< page size)
496 void flush_icache_user_range(struct vm_area_struct *vma,
497 struct page *page, unsigned long addr, int len)
499 flush_cache_page(vma, addr, page_to_pfn(page));
506 * @addr: address in memory mapped cache array
507 * @phys: P1 address to flush (has to match tags if addr has 'A' bit
508 * set i.e. associative write)
509 * @exec_offset: set to 0x20000000 if flush has to be executed from P2
512 * The offset into the cache array implied by 'addr' selects the
513 * 'colour' of the virtual address range that will be flushed. The
514 * operation (purge/write-back) is selected by the lower 2 bits of
517 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
518 unsigned long exec_offset)
521 unsigned long base_addr = addr;
522 struct cache_info *dcache;
523 unsigned long way_incr;
524 unsigned long a, ea, p;
525 unsigned long temp_pc;
527 dcache = ¤t_cpu_data.dcache;
528 /* Write this way for better assembly. */
529 way_count = dcache->ways;
530 way_incr = dcache->way_incr;
533 * Apply exec_offset (i.e. branch to P2 if required.).
537 * If I write "=r" for the (temp_pc), it puts this in r6 hence
538 * trashing exec_offset before it's been added on - why? Hence
539 * "=&r" as a 'workaround'
541 asm volatile("mov.l 1f, %0\n\t"
547 "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
550 * We know there will be >=1 iteration, so write as do-while to avoid
551 * pointless nead-of-loop check for 0 iterations.
554 ea = base_addr + PAGE_SIZE;
559 *(volatile unsigned long *)a = p;
561 * Next line: intentionally not p+32, saves an add, p
562 * will do since only the cache tag bits need to
565 *(volatile unsigned long *)(a+32) = p;
570 base_addr += way_incr;
571 } while (--way_count != 0);
575 * Break the 1, 2 and 4 way variants of this out into separate functions to
576 * avoid nearly all the overhead of having the conditional stuff in the function
577 * bodies (+ the 1 and 2 way cases avoid saving any registers too).
579 static void __flush_dcache_segment_1way(unsigned long start,
580 unsigned long extent_per_way)
582 unsigned long orig_sr, sr_with_bl;
583 unsigned long base_addr;
584 unsigned long way_incr, linesz, way_size;
585 struct cache_info *dcache;
586 register unsigned long a0, a0e;
588 asm volatile("stc sr, %0" : "=r" (orig_sr));
589 sr_with_bl = orig_sr | (1<<28);
590 base_addr = ((unsigned long)&empty_zero_page[0]);
593 * The previous code aligned base_addr to 16k, i.e. the way_size of all
594 * existing SH-4 D-caches. Whilst I don't see a need to have this
595 * aligned to any better than the cache line size (which it will be
596 * anyway by construction), let's align it to at least the way_size of
597 * any existing or conceivable SH-4 D-cache. -- RPC
599 base_addr = ((base_addr >> 16) << 16);
602 dcache = ¤t_cpu_data.dcache;
603 linesz = dcache->linesz;
604 way_incr = dcache->way_incr;
605 way_size = dcache->way_size;
608 a0e = base_addr + extent_per_way;
610 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
611 asm volatile("movca.l r0, @%0\n\t"
612 "ocbi @%0" : : "r" (a0));
614 asm volatile("movca.l r0, @%0\n\t"
615 "ocbi @%0" : : "r" (a0));
617 asm volatile("movca.l r0, @%0\n\t"
618 "ocbi @%0" : : "r" (a0));
620 asm volatile("movca.l r0, @%0\n\t"
621 "ocbi @%0" : : "r" (a0));
622 asm volatile("ldc %0, sr" : : "r" (orig_sr));
627 static void __flush_dcache_segment_2way(unsigned long start,
628 unsigned long extent_per_way)
630 unsigned long orig_sr, sr_with_bl;
631 unsigned long base_addr;
632 unsigned long way_incr, linesz, way_size;
633 struct cache_info *dcache;
634 register unsigned long a0, a1, a0e;
636 asm volatile("stc sr, %0" : "=r" (orig_sr));
637 sr_with_bl = orig_sr | (1<<28);
638 base_addr = ((unsigned long)&empty_zero_page[0]);
640 /* See comment under 1-way above */
641 base_addr = ((base_addr >> 16) << 16);
644 dcache = ¤t_cpu_data.dcache;
645 linesz = dcache->linesz;
646 way_incr = dcache->way_incr;
647 way_size = dcache->way_size;
651 a0e = base_addr + extent_per_way;
653 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
654 asm volatile("movca.l r0, @%0\n\t"
655 "movca.l r0, @%1\n\t"
661 asm volatile("movca.l r0, @%0\n\t"
662 "movca.l r0, @%1\n\t"
668 asm volatile("movca.l r0, @%0\n\t"
669 "movca.l r0, @%1\n\t"
675 asm volatile("movca.l r0, @%0\n\t"
676 "movca.l r0, @%1\n\t"
680 asm volatile("ldc %0, sr" : : "r" (orig_sr));
686 static void __flush_dcache_segment_4way(unsigned long start,
687 unsigned long extent_per_way)
689 unsigned long orig_sr, sr_with_bl;
690 unsigned long base_addr;
691 unsigned long way_incr, linesz, way_size;
692 struct cache_info *dcache;
693 register unsigned long a0, a1, a2, a3, a0e;
695 asm volatile("stc sr, %0" : "=r" (orig_sr));
696 sr_with_bl = orig_sr | (1<<28);
697 base_addr = ((unsigned long)&empty_zero_page[0]);
699 /* See comment under 1-way above */
700 base_addr = ((base_addr >> 16) << 16);
703 dcache = ¤t_cpu_data.dcache;
704 linesz = dcache->linesz;
705 way_incr = dcache->way_incr;
706 way_size = dcache->way_size;
712 a0e = base_addr + extent_per_way;
714 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
715 asm volatile("movca.l r0, @%0\n\t"
716 "movca.l r0, @%1\n\t"
717 "movca.l r0, @%2\n\t"
718 "movca.l r0, @%3\n\t"
723 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
728 asm volatile("movca.l r0, @%0\n\t"
729 "movca.l r0, @%1\n\t"
730 "movca.l r0, @%2\n\t"
731 "movca.l r0, @%3\n\t"
736 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
741 asm volatile("movca.l r0, @%0\n\t"
742 "movca.l r0, @%1\n\t"
743 "movca.l r0, @%2\n\t"
744 "movca.l r0, @%3\n\t"
749 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
754 asm volatile("movca.l r0, @%0\n\t"
755 "movca.l r0, @%1\n\t"
756 "movca.l r0, @%2\n\t"
757 "movca.l r0, @%3\n\t"
762 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
763 asm volatile("ldc %0, sr" : : "r" (orig_sr));