2 * arch/sh/mm/cache-sh4.c
4 * Copyright (C) 1999, 2000, 2002 Niibe Yutaka
5 * Copyright (C) 2001 - 2007 Paul Mundt
6 * Copyright (C) 2003 Richard Curnow
7 * Copyright (c) 2007 STMicroelectronics (R&D) Ltd.
9 * This file is subject to the terms and conditions of the GNU General Public
10 * License. See the file "COPYING" in the main directory of this archive
13 #include <linux/init.h>
16 #include <linux/mutex.h>
17 #include <asm/mmu_context.h>
18 #include <asm/cacheflush.h>
21 * The maximum number of pages we support up to when doing ranged dcache
22 * flushing. Anything exceeding this will simply flush the dcache in its
25 #define MAX_DCACHE_PAGES 64 /* XXX: Tune for ways */
26 #define MAX_ICACHE_PAGES 32
28 static void __flush_dcache_segment_1way(unsigned long start,
29 unsigned long extent);
30 static void __flush_dcache_segment_2way(unsigned long start,
31 unsigned long extent);
32 static void __flush_dcache_segment_4way(unsigned long start,
33 unsigned long extent);
35 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
36 unsigned long exec_offset);
39 * This is initialised here to ensure that it is not placed in the BSS. If
40 * that were to happen, note that cache_init gets called before the BSS is
41 * cleared, so this would get nulled out which would be hopeless.
43 static void (*__flush_dcache_segment_fn)(unsigned long, unsigned long) =
44 (void (*)(unsigned long, unsigned long))0xdeadbeef;
46 static void compute_alias(struct cache_info *c)
48 c->alias_mask = ((c->sets - 1) << c->entry_shift) & ~(PAGE_SIZE - 1);
49 c->n_aliases = c->alias_mask ? (c->alias_mask >> PAGE_SHIFT) + 1 : 0;
52 static void __init emit_cache_params(void)
54 printk("PVR=%08x CVR=%08x PRR=%08x\n",
58 printk("I-cache : n_ways=%d n_sets=%d way_incr=%d\n",
59 boot_cpu_data.icache.ways,
60 boot_cpu_data.icache.sets,
61 boot_cpu_data.icache.way_incr);
62 printk("I-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
63 boot_cpu_data.icache.entry_mask,
64 boot_cpu_data.icache.alias_mask,
65 boot_cpu_data.icache.n_aliases);
66 printk("D-cache : n_ways=%d n_sets=%d way_incr=%d\n",
67 boot_cpu_data.dcache.ways,
68 boot_cpu_data.dcache.sets,
69 boot_cpu_data.dcache.way_incr);
70 printk("D-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
71 boot_cpu_data.dcache.entry_mask,
72 boot_cpu_data.dcache.alias_mask,
73 boot_cpu_data.dcache.n_aliases);
76 * Emit Secondary Cache parameters if the CPU has a probed L2.
78 if (boot_cpu_data.flags & CPU_HAS_L2_CACHE) {
79 printk("S-cache : n_ways=%d n_sets=%d way_incr=%d\n",
80 boot_cpu_data.scache.ways,
81 boot_cpu_data.scache.sets,
82 boot_cpu_data.scache.way_incr);
83 printk("S-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
84 boot_cpu_data.scache.entry_mask,
85 boot_cpu_data.scache.alias_mask,
86 boot_cpu_data.scache.n_aliases);
89 if (!__flush_dcache_segment_fn)
90 panic("unknown number of cache ways\n");
94 * SH-4 has virtually indexed and physically tagged cache.
96 void __init p3_cache_init(void)
98 compute_alias(&boot_cpu_data.icache);
99 compute_alias(&boot_cpu_data.dcache);
100 compute_alias(&boot_cpu_data.scache);
102 switch (boot_cpu_data.dcache.ways) {
104 __flush_dcache_segment_fn = __flush_dcache_segment_1way;
107 __flush_dcache_segment_fn = __flush_dcache_segment_2way;
110 __flush_dcache_segment_fn = __flush_dcache_segment_4way;
113 __flush_dcache_segment_fn = NULL;
121 * Write back the dirty D-caches, but not invalidate them.
123 * START: Virtual Address (U0, P1, or P3)
124 * SIZE: Size of the region.
126 void __flush_wback_region(void *start, int size)
129 unsigned long begin, end;
131 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
132 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
133 & ~(L1_CACHE_BYTES-1);
134 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
135 asm volatile("ocbwb %0"
142 * Write back the dirty D-caches and invalidate them.
144 * START: Virtual Address (U0, P1, or P3)
145 * SIZE: Size of the region.
147 void __flush_purge_region(void *start, int size)
150 unsigned long begin, end;
152 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
153 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
154 & ~(L1_CACHE_BYTES-1);
155 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
156 asm volatile("ocbp %0"
163 * No write back please
165 void __flush_invalidate_region(void *start, int size)
168 unsigned long begin, end;
170 begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
171 end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
172 & ~(L1_CACHE_BYTES-1);
173 for (v = begin; v < end; v+=L1_CACHE_BYTES) {
174 asm volatile("ocbi %0"
181 * Write back the range of D-cache, and purge the I-cache.
183 * Called from kernel/module.c:sys_init_module and routine for a.out format,
184 * signal handler code and kprobes code
186 void flush_icache_range(unsigned long start, unsigned long end)
189 unsigned long flags, v;
192 /* If there are too many pages then just blow the caches */
193 if (((end - start) >> PAGE_SHIFT) >= MAX_ICACHE_PAGES) {
196 /* selectively flush d-cache then invalidate the i-cache */
197 /* this is inefficient, so only use for small ranges */
198 start &= ~(L1_CACHE_BYTES-1);
199 end += L1_CACHE_BYTES-1;
200 end &= ~(L1_CACHE_BYTES-1);
202 local_irq_save(flags);
205 for (v = start; v < end; v+=L1_CACHE_BYTES) {
206 asm volatile("ocbwb %0"
210 icacheaddr = CACHE_IC_ADDRESS_ARRAY | (
211 v & cpu_data->icache.entry_mask);
213 for (i = 0; i < cpu_data->icache.ways;
214 i++, icacheaddr += cpu_data->icache.way_incr)
215 /* Clear i-cache line valid-bit */
216 ctrl_outl(0, icacheaddr);
220 local_irq_restore(flags);
224 static inline void flush_cache_4096(unsigned long start,
227 unsigned long flags, exec_offset = 0;
230 * All types of SH-4 require PC to be in P2 to operate on the I-cache.
231 * Some types of SH-4 require PC to be in P2 to operate on the D-cache.
233 if ((boot_cpu_data.flags & CPU_HAS_P2_FLUSH_BUG) ||
234 (start < CACHE_OC_ADDRESS_ARRAY))
235 exec_offset = 0x20000000;
237 local_irq_save(flags);
238 __flush_cache_4096(start | SH_CACHE_ASSOC,
239 P1SEGADDR(phys), exec_offset);
240 local_irq_restore(flags);
244 * Write back & invalidate the D-cache of the page.
245 * (To avoid "alias" issues)
247 void flush_dcache_page(struct page *page)
249 if (test_bit(PG_mapped, &page->flags)) {
250 unsigned long phys = PHYSADDR(page_address(page));
251 unsigned long addr = CACHE_OC_ADDRESS_ARRAY;
254 /* Loop all the D-cache */
255 n = boot_cpu_data.dcache.n_aliases;
256 for (i = 0; i < n; i++, addr += 4096)
257 flush_cache_4096(addr, phys);
263 /* TODO: Selective icache invalidation through IC address array.. */
264 static void __uses_jump_to_uncached flush_icache_all(void)
266 unsigned long flags, ccr;
268 local_irq_save(flags);
273 ccr |= CCR_CACHE_ICI;
277 * back_to_cached() will take care of the barrier for us, don't add
282 local_irq_restore(flags);
285 void flush_dcache_all(void)
287 (*__flush_dcache_segment_fn)(0UL, boot_cpu_data.dcache.way_size);
291 void flush_cache_all(void)
297 static void __flush_cache_mm(struct mm_struct *mm, unsigned long start,
300 unsigned long d = 0, p = start & PAGE_MASK;
301 unsigned long alias_mask = boot_cpu_data.dcache.alias_mask;
302 unsigned long n_aliases = boot_cpu_data.dcache.n_aliases;
303 unsigned long select_bit;
304 unsigned long all_aliases_mask;
305 unsigned long addr_offset;
312 dir = pgd_offset(mm, p);
313 pud = pud_offset(dir, p);
314 pmd = pmd_offset(pud, p);
315 end = PAGE_ALIGN(end);
317 all_aliases_mask = (1 << n_aliases) - 1;
320 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) {
328 pte = pte_offset_kernel(pmd, p);
334 if (!(pte_val(entry) & _PAGE_PRESENT)) {
340 phys = pte_val(entry) & PTE_PHYS_MASK;
342 if ((p ^ phys) & alias_mask) {
343 d |= 1 << ((p & alias_mask) >> PAGE_SHIFT);
344 d |= 1 << ((phys & alias_mask) >> PAGE_SHIFT);
346 if (d == all_aliases_mask)
352 } while (p < end && ((unsigned long)pte & ~PAGE_MASK));
360 for (i = 0; i < n_aliases; i++) {
361 if (d & select_bit) {
362 (*__flush_dcache_segment_fn)(addr_offset, PAGE_SIZE);
367 addr_offset += PAGE_SIZE;
372 * Note : (RPC) since the caches are physically tagged, the only point
373 * of flush_cache_mm for SH-4 is to get rid of aliases from the
374 * D-cache. The assumption elsewhere, e.g. flush_cache_range, is that
375 * lines can stay resident so long as the virtual address they were
376 * accessed with (hence cache set) is in accord with the physical
377 * address (i.e. tag). It's no different here. So I reckon we don't
378 * need to flush the I-cache, since aliases don't matter for that. We
381 * Caller takes mm->mmap_sem.
383 void flush_cache_mm(struct mm_struct *mm)
386 * If cache is only 4k-per-way, there are never any 'aliases'. Since
387 * the cache is physically tagged, the data can just be left in there.
389 if (boot_cpu_data.dcache.n_aliases == 0)
393 * Don't bother groveling around the dcache for the VMA ranges
394 * if there are too many PTEs to make it worthwhile.
396 if (mm->nr_ptes >= MAX_DCACHE_PAGES)
399 struct vm_area_struct *vma;
402 * In this case there are reasonably sized ranges to flush,
403 * iterate through the VMA list and take care of any aliases.
405 for (vma = mm->mmap; vma; vma = vma->vm_next)
406 __flush_cache_mm(mm, vma->vm_start, vma->vm_end);
409 /* Only touch the icache if one of the VMAs has VM_EXEC set. */
415 * Write back and invalidate I/D-caches for the page.
417 * ADDR: Virtual Address (U0 address)
418 * PFN: Physical page number
420 void flush_cache_page(struct vm_area_struct *vma, unsigned long address,
423 unsigned long phys = pfn << PAGE_SHIFT;
424 unsigned int alias_mask;
426 alias_mask = boot_cpu_data.dcache.alias_mask;
428 /* We only need to flush D-cache when we have alias */
429 if ((address^phys) & alias_mask) {
430 /* Loop 4K of the D-cache */
432 CACHE_OC_ADDRESS_ARRAY | (address & alias_mask),
434 /* Loop another 4K of the D-cache */
436 CACHE_OC_ADDRESS_ARRAY | (phys & alias_mask),
440 alias_mask = boot_cpu_data.icache.alias_mask;
441 if (vma->vm_flags & VM_EXEC) {
443 * Evict entries from the portion of the cache from which code
444 * may have been executed at this address (virtual). There's
445 * no need to evict from the portion corresponding to the
446 * physical address as for the D-cache, because we know the
447 * kernel has never executed the code through its identity
451 CACHE_IC_ADDRESS_ARRAY | (address & alias_mask),
457 * Write back and invalidate D-caches.
459 * START, END: Virtual Address (U0 address)
461 * NOTE: We need to flush the _physical_ page entry.
462 * Flushing the cache lines for U0 only isn't enough.
463 * We need to flush for P1 too, which may contain aliases.
465 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
469 * If cache is only 4k-per-way, there are never any 'aliases'. Since
470 * the cache is physically tagged, the data can just be left in there.
472 if (boot_cpu_data.dcache.n_aliases == 0)
476 * Don't bother with the lookup and alias check if we have a
477 * wide range to cover, just blow away the dcache in its
478 * entirety instead. -- PFM.
480 if (((end - start) >> PAGE_SHIFT) >= MAX_DCACHE_PAGES)
483 __flush_cache_mm(vma->vm_mm, start, end);
485 if (vma->vm_flags & VM_EXEC) {
487 * TODO: Is this required??? Need to look at how I-cache
488 * coherency is assured when new programs are loaded to see if
496 * flush_icache_user_range
497 * @vma: VMA of the process
500 * @len: length of the range (< page size)
502 void flush_icache_user_range(struct vm_area_struct *vma,
503 struct page *page, unsigned long addr, int len)
505 flush_cache_page(vma, addr, page_to_pfn(page));
512 * @addr: address in memory mapped cache array
513 * @phys: P1 address to flush (has to match tags if addr has 'A' bit
514 * set i.e. associative write)
515 * @exec_offset: set to 0x20000000 if flush has to be executed from P2
518 * The offset into the cache array implied by 'addr' selects the
519 * 'colour' of the virtual address range that will be flushed. The
520 * operation (purge/write-back) is selected by the lower 2 bits of
523 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
524 unsigned long exec_offset)
527 unsigned long base_addr = addr;
528 struct cache_info *dcache;
529 unsigned long way_incr;
530 unsigned long a, ea, p;
531 unsigned long temp_pc;
533 dcache = &boot_cpu_data.dcache;
534 /* Write this way for better assembly. */
535 way_count = dcache->ways;
536 way_incr = dcache->way_incr;
539 * Apply exec_offset (i.e. branch to P2 if required.).
543 * If I write "=r" for the (temp_pc), it puts this in r6 hence
544 * trashing exec_offset before it's been added on - why? Hence
545 * "=&r" as a 'workaround'
547 asm volatile("mov.l 1f, %0\n\t"
553 "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
556 * We know there will be >=1 iteration, so write as do-while to avoid
557 * pointless nead-of-loop check for 0 iterations.
560 ea = base_addr + PAGE_SIZE;
565 *(volatile unsigned long *)a = p;
567 * Next line: intentionally not p+32, saves an add, p
568 * will do since only the cache tag bits need to
571 *(volatile unsigned long *)(a+32) = p;
576 base_addr += way_incr;
577 } while (--way_count != 0);
581 * Break the 1, 2 and 4 way variants of this out into separate functions to
582 * avoid nearly all the overhead of having the conditional stuff in the function
583 * bodies (+ the 1 and 2 way cases avoid saving any registers too).
585 static void __flush_dcache_segment_1way(unsigned long start,
586 unsigned long extent_per_way)
588 unsigned long orig_sr, sr_with_bl;
589 unsigned long base_addr;
590 unsigned long way_incr, linesz, way_size;
591 struct cache_info *dcache;
592 register unsigned long a0, a0e;
594 asm volatile("stc sr, %0" : "=r" (orig_sr));
595 sr_with_bl = orig_sr | (1<<28);
596 base_addr = ((unsigned long)&empty_zero_page[0]);
599 * The previous code aligned base_addr to 16k, i.e. the way_size of all
600 * existing SH-4 D-caches. Whilst I don't see a need to have this
601 * aligned to any better than the cache line size (which it will be
602 * anyway by construction), let's align it to at least the way_size of
603 * any existing or conceivable SH-4 D-cache. -- RPC
605 base_addr = ((base_addr >> 16) << 16);
608 dcache = &boot_cpu_data.dcache;
609 linesz = dcache->linesz;
610 way_incr = dcache->way_incr;
611 way_size = dcache->way_size;
614 a0e = base_addr + extent_per_way;
616 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
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));
623 asm volatile("movca.l r0, @%0\n\t"
624 "ocbi @%0" : : "r" (a0));
626 asm volatile("movca.l r0, @%0\n\t"
627 "ocbi @%0" : : "r" (a0));
628 asm volatile("ldc %0, sr" : : "r" (orig_sr));
633 static void __flush_dcache_segment_2way(unsigned long start,
634 unsigned long extent_per_way)
636 unsigned long orig_sr, sr_with_bl;
637 unsigned long base_addr;
638 unsigned long way_incr, linesz, way_size;
639 struct cache_info *dcache;
640 register unsigned long a0, a1, a0e;
642 asm volatile("stc sr, %0" : "=r" (orig_sr));
643 sr_with_bl = orig_sr | (1<<28);
644 base_addr = ((unsigned long)&empty_zero_page[0]);
646 /* See comment under 1-way above */
647 base_addr = ((base_addr >> 16) << 16);
650 dcache = &boot_cpu_data.dcache;
651 linesz = dcache->linesz;
652 way_incr = dcache->way_incr;
653 way_size = dcache->way_size;
657 a0e = base_addr + extent_per_way;
659 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
660 asm volatile("movca.l r0, @%0\n\t"
661 "movca.l r0, @%1\n\t"
667 asm volatile("movca.l r0, @%0\n\t"
668 "movca.l r0, @%1\n\t"
674 asm volatile("movca.l r0, @%0\n\t"
675 "movca.l r0, @%1\n\t"
681 asm volatile("movca.l r0, @%0\n\t"
682 "movca.l r0, @%1\n\t"
686 asm volatile("ldc %0, sr" : : "r" (orig_sr));
692 static void __flush_dcache_segment_4way(unsigned long start,
693 unsigned long extent_per_way)
695 unsigned long orig_sr, sr_with_bl;
696 unsigned long base_addr;
697 unsigned long way_incr, linesz, way_size;
698 struct cache_info *dcache;
699 register unsigned long a0, a1, a2, a3, a0e;
701 asm volatile("stc sr, %0" : "=r" (orig_sr));
702 sr_with_bl = orig_sr | (1<<28);
703 base_addr = ((unsigned long)&empty_zero_page[0]);
705 /* See comment under 1-way above */
706 base_addr = ((base_addr >> 16) << 16);
709 dcache = &boot_cpu_data.dcache;
710 linesz = dcache->linesz;
711 way_incr = dcache->way_incr;
712 way_size = dcache->way_size;
718 a0e = base_addr + extent_per_way;
720 asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
721 asm volatile("movca.l r0, @%0\n\t"
722 "movca.l r0, @%1\n\t"
723 "movca.l r0, @%2\n\t"
724 "movca.l r0, @%3\n\t"
729 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
734 asm volatile("movca.l r0, @%0\n\t"
735 "movca.l r0, @%1\n\t"
736 "movca.l r0, @%2\n\t"
737 "movca.l r0, @%3\n\t"
742 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
747 asm volatile("movca.l r0, @%0\n\t"
748 "movca.l r0, @%1\n\t"
749 "movca.l r0, @%2\n\t"
750 "movca.l r0, @%3\n\t"
755 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
760 asm volatile("movca.l r0, @%0\n\t"
761 "movca.l r0, @%1\n\t"
762 "movca.l r0, @%2\n\t"
763 "movca.l r0, @%3\n\t"
768 "r" (a0), "r" (a1), "r" (a2), "r" (a3));
769 asm volatile("ldc %0, sr" : : "r" (orig_sr));