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>
18 #include <asm/mmu_context.h>
19 #include <asm/cacheflush.h>
22 * The maximum number of pages we support up to when doing ranged dcache
23 * flushing. Anything exceeding this will simply flush the dcache in its
26 #define MAX_DCACHE_PAGES 64 /* XXX: Tune for ways */
27 #define MAX_ICACHE_PAGES 32
29 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
30 unsigned long exec_offset);
33 * Write back the range of D-cache, and purge the I-cache.
35 * Called from kernel/module.c:sys_init_module and routine for a.out format,
36 * signal handler code and kprobes code
38 static void sh4_flush_icache_range(void *args)
40 struct flusher_data *data = args;
41 unsigned long start, end;
42 unsigned long flags, v;
48 /* If there are too many pages then just blow away the caches */
49 if (((end - start) >> PAGE_SHIFT) >= MAX_ICACHE_PAGES) {
50 local_flush_cache_all(NULL);
55 * Selectively flush d-cache then invalidate the i-cache.
56 * This is inefficient, so only use this for small ranges.
58 start &= ~(L1_CACHE_BYTES-1);
59 end += L1_CACHE_BYTES-1;
60 end &= ~(L1_CACHE_BYTES-1);
62 local_irq_save(flags);
65 for (v = start; v < end; v += L1_CACHE_BYTES) {
66 unsigned long icacheaddr;
70 icacheaddr = CACHE_IC_ADDRESS_ARRAY | (v &
71 cpu_data->icache.entry_mask);
73 /* Clear i-cache line valid-bit */
74 for (i = 0; i < cpu_data->icache.ways; i++) {
75 __raw_writel(0, icacheaddr);
76 icacheaddr += cpu_data->icache.way_incr;
81 local_irq_restore(flags);
84 static inline void flush_cache_4096(unsigned long start,
87 unsigned long flags, exec_offset = 0;
90 * All types of SH-4 require PC to be in P2 to operate on the I-cache.
91 * Some types of SH-4 require PC to be in P2 to operate on the D-cache.
93 if ((boot_cpu_data.flags & CPU_HAS_P2_FLUSH_BUG) ||
94 (start < CACHE_OC_ADDRESS_ARRAY))
95 exec_offset = 0x20000000;
97 local_irq_save(flags);
98 __flush_cache_4096(start | SH_CACHE_ASSOC,
99 P1SEGADDR(phys), exec_offset);
100 local_irq_restore(flags);
104 * Write back & invalidate the D-cache of the page.
105 * (To avoid "alias" issues)
107 static void sh4_flush_dcache_page(void *arg)
109 struct page *page = arg;
111 struct address_space *mapping = page_mapping(page);
113 if (mapping && !mapping_mapped(mapping))
114 set_bit(PG_dcache_dirty, &page->flags);
118 unsigned long phys = page_to_phys(page);
119 unsigned long addr = CACHE_OC_ADDRESS_ARRAY;
122 /* Loop all the D-cache */
123 n = boot_cpu_data.dcache.way_incr >> 12;
124 for (i = 0; i < n; i++, addr += 4096)
125 flush_cache_4096(addr, phys);
131 /* TODO: Selective icache invalidation through IC address array.. */
132 static void __uses_jump_to_uncached flush_icache_all(void)
134 unsigned long flags, ccr;
136 local_irq_save(flags);
141 ccr |= CCR_CACHE_ICI;
145 * back_to_cached() will take care of the barrier for us, don't add
150 local_irq_restore(flags);
153 static void flush_dcache_all(void)
155 unsigned long addr, end_addr, entry_offset;
157 end_addr = CACHE_OC_ADDRESS_ARRAY +
158 (current_cpu_data.dcache.sets <<
159 current_cpu_data.dcache.entry_shift) *
160 current_cpu_data.dcache.ways;
162 entry_offset = 1 << current_cpu_data.dcache.entry_shift;
164 for (addr = CACHE_OC_ADDRESS_ARRAY; addr < end_addr; ) {
165 __raw_writel(0, addr); addr += entry_offset;
166 __raw_writel(0, addr); addr += entry_offset;
167 __raw_writel(0, addr); addr += entry_offset;
168 __raw_writel(0, addr); addr += entry_offset;
169 __raw_writel(0, addr); addr += entry_offset;
170 __raw_writel(0, addr); addr += entry_offset;
171 __raw_writel(0, addr); addr += entry_offset;
172 __raw_writel(0, addr); addr += entry_offset;
176 static void sh4_flush_cache_all(void *unused)
183 * Note : (RPC) since the caches are physically tagged, the only point
184 * of flush_cache_mm for SH-4 is to get rid of aliases from the
185 * D-cache. The assumption elsewhere, e.g. flush_cache_range, is that
186 * lines can stay resident so long as the virtual address they were
187 * accessed with (hence cache set) is in accord with the physical
188 * address (i.e. tag). It's no different here.
190 * Caller takes mm->mmap_sem.
192 static void sh4_flush_cache_mm(void *arg)
194 struct mm_struct *mm = arg;
196 if (cpu_context(smp_processor_id(), mm) == NO_CONTEXT)
203 * Write back and invalidate I/D-caches for the page.
205 * ADDR: Virtual Address (U0 address)
206 * PFN: Physical page number
208 static void sh4_flush_cache_page(void *args)
210 struct flusher_data *data = args;
211 struct vm_area_struct *vma;
212 unsigned long address, pfn, phys;
213 unsigned int alias_mask;
216 address = data->addr1;
218 phys = pfn << PAGE_SHIFT;
220 if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
223 alias_mask = boot_cpu_data.dcache.alias_mask;
225 /* We only need to flush D-cache when we have alias */
226 if ((address^phys) & alias_mask) {
227 /* Loop 4K of the D-cache */
229 CACHE_OC_ADDRESS_ARRAY | (address & alias_mask),
231 /* Loop another 4K of the D-cache */
233 CACHE_OC_ADDRESS_ARRAY | (phys & alias_mask),
237 alias_mask = boot_cpu_data.icache.alias_mask;
238 if (vma->vm_flags & VM_EXEC) {
240 * Evict entries from the portion of the cache from which code
241 * may have been executed at this address (virtual). There's
242 * no need to evict from the portion corresponding to the
243 * physical address as for the D-cache, because we know the
244 * kernel has never executed the code through its identity
248 CACHE_IC_ADDRESS_ARRAY | (address & alias_mask),
254 * Write back and invalidate D-caches.
256 * START, END: Virtual Address (U0 address)
258 * NOTE: We need to flush the _physical_ page entry.
259 * Flushing the cache lines for U0 only isn't enough.
260 * We need to flush for P1 too, which may contain aliases.
262 static void sh4_flush_cache_range(void *args)
264 struct flusher_data *data = args;
265 struct vm_area_struct *vma;
266 unsigned long start, end;
272 if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
276 * If cache is only 4k-per-way, there are never any 'aliases'. Since
277 * the cache is physically tagged, the data can just be left in there.
279 if (boot_cpu_data.dcache.n_aliases == 0)
284 if (vma->vm_flags & VM_EXEC)
291 * @addr: address in memory mapped cache array
292 * @phys: P1 address to flush (has to match tags if addr has 'A' bit
293 * set i.e. associative write)
294 * @exec_offset: set to 0x20000000 if flush has to be executed from P2
297 * The offset into the cache array implied by 'addr' selects the
298 * 'colour' of the virtual address range that will be flushed. The
299 * operation (purge/write-back) is selected by the lower 2 bits of
302 static void __flush_cache_4096(unsigned long addr, unsigned long phys,
303 unsigned long exec_offset)
306 unsigned long base_addr = addr;
307 struct cache_info *dcache;
308 unsigned long way_incr;
309 unsigned long a, ea, p;
310 unsigned long temp_pc;
312 dcache = &boot_cpu_data.dcache;
313 /* Write this way for better assembly. */
314 way_count = dcache->ways;
315 way_incr = dcache->way_incr;
318 * Apply exec_offset (i.e. branch to P2 if required.).
322 * If I write "=r" for the (temp_pc), it puts this in r6 hence
323 * trashing exec_offset before it's been added on - why? Hence
324 * "=&r" as a 'workaround'
326 asm volatile("mov.l 1f, %0\n\t"
332 "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
335 * We know there will be >=1 iteration, so write as do-while to avoid
336 * pointless nead-of-loop check for 0 iterations.
339 ea = base_addr + PAGE_SIZE;
344 *(volatile unsigned long *)a = p;
346 * Next line: intentionally not p+32, saves an add, p
347 * will do since only the cache tag bits need to
350 *(volatile unsigned long *)(a+32) = p;
355 base_addr += way_incr;
356 } while (--way_count != 0);
359 extern void __weak sh4__flush_region_init(void);
362 * SH-4 has virtually indexed and physically tagged cache.
364 void __init sh4_cache_init(void)
366 printk("PVR=%08x CVR=%08x PRR=%08x\n",
371 local_flush_icache_range = sh4_flush_icache_range;
372 local_flush_dcache_page = sh4_flush_dcache_page;
373 local_flush_cache_all = sh4_flush_cache_all;
374 local_flush_cache_mm = sh4_flush_cache_mm;
375 local_flush_cache_dup_mm = sh4_flush_cache_mm;
376 local_flush_cache_page = sh4_flush_cache_page;
377 local_flush_cache_range = sh4_flush_cache_range;
379 sh4__flush_region_init();