xref: /linux/arch/sh/mm/cache-sh4.c (revision 68550cbc6129159b7a6434796b721e8b66ee12f6)
1 /*
2  * arch/sh/mm/cache-sh4.c
3  *
4  * Copyright (C) 1999, 2000, 2002  Niibe Yutaka
5  * Copyright (C) 2001 - 2009  Paul Mundt
6  * Copyright (C) 2003  Richard Curnow
7  * Copyright (c) 2007 STMicroelectronics (R&D) Ltd.
8  *
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
11  * for more details.
12  */
13 #include <linux/init.h>
14 #include <linux/mm.h>
15 #include <linux/io.h>
16 #include <linux/mutex.h>
17 #include <linux/fs.h>
18 #include <linux/highmem.h>
19 #include <linux/pagemap.h>
20 #include <asm/mmu_context.h>
21 #include <asm/cache_insns.h>
22 #include <asm/cacheflush.h>
23 
24 /*
25  * The maximum number of pages we support up to when doing ranged dcache
26  * flushing. Anything exceeding this will simply flush the dcache in its
27  * entirety.
28  */
29 #define MAX_ICACHE_PAGES	32
30 
31 static void __flush_cache_one(unsigned long addr, unsigned long phys,
32 			       unsigned long exec_offset);
33 
34 /*
35  * Write back the range of D-cache, and purge the I-cache.
36  *
37  * Called from kernel/module.c:sys_init_module and routine for a.out format,
38  * signal handler code and kprobes code
39  */
40 static void sh4_flush_icache_range(void *args)
41 {
42 	struct flusher_data *data = args;
43 	unsigned long start, end;
44 	unsigned long flags, v;
45 	int i;
46 
47 	start = data->addr1;
48 	end = data->addr2;
49 
50 	/* If there are too many pages then just blow away the caches */
51 	if (((end - start) >> PAGE_SHIFT) >= MAX_ICACHE_PAGES) {
52 		local_flush_cache_all(NULL);
53 		return;
54 	}
55 
56 	/*
57 	 * Selectively flush d-cache then invalidate the i-cache.
58 	 * This is inefficient, so only use this for small ranges.
59 	 */
60 	start &= ~(L1_CACHE_BYTES-1);
61 	end += L1_CACHE_BYTES-1;
62 	end &= ~(L1_CACHE_BYTES-1);
63 
64 	local_irq_save(flags);
65 	jump_to_uncached();
66 
67 	for (v = start; v < end; v += L1_CACHE_BYTES) {
68 		unsigned long icacheaddr;
69 		int j, n;
70 
71 		__ocbwb(v);
72 
73 		icacheaddr = CACHE_IC_ADDRESS_ARRAY | (v &
74 				cpu_data->icache.entry_mask);
75 
76 		/* Clear i-cache line valid-bit */
77 		n = boot_cpu_data.icache.n_aliases;
78 		for (i = 0; i < cpu_data->icache.ways; i++) {
79 			for (j = 0; j < n; j++)
80 				__raw_writel(0, icacheaddr + (j * PAGE_SIZE));
81 			icacheaddr += cpu_data->icache.way_incr;
82 		}
83 	}
84 
85 	back_to_cached();
86 	local_irq_restore(flags);
87 }
88 
89 static inline void flush_cache_one(unsigned long start, unsigned long phys)
90 {
91 	unsigned long flags, exec_offset = 0;
92 
93 	/*
94 	 * All types of SH-4 require PC to be uncached to operate on the I-cache.
95 	 * Some types of SH-4 require PC to be uncached to operate on the D-cache.
96 	 */
97 	if ((boot_cpu_data.flags & CPU_HAS_P2_FLUSH_BUG) ||
98 	    (start < CACHE_OC_ADDRESS_ARRAY))
99 		exec_offset = cached_to_uncached;
100 
101 	local_irq_save(flags);
102 	__flush_cache_one(start, phys, exec_offset);
103 	local_irq_restore(flags);
104 }
105 
106 /*
107  * Write back & invalidate the D-cache of the page.
108  * (To avoid "alias" issues)
109  */
110 static void sh4_flush_dcache_page(void *arg)
111 {
112 	struct page *page = arg;
113 	unsigned long addr = (unsigned long)page_address(page);
114 #ifndef CONFIG_SMP
115 	struct address_space *mapping = page_mapping_file(page);
116 
117 	if (mapping && !mapping_mapped(mapping))
118 		clear_bit(PG_dcache_clean, &page->flags);
119 	else
120 #endif
121 		flush_cache_one(CACHE_OC_ADDRESS_ARRAY |
122 				(addr & shm_align_mask), page_to_phys(page));
123 
124 	wmb();
125 }
126 
127 /* TODO: Selective icache invalidation through IC address array.. */
128 static void flush_icache_all(void)
129 {
130 	unsigned long flags, ccr;
131 
132 	local_irq_save(flags);
133 	jump_to_uncached();
134 
135 	/* Flush I-cache */
136 	ccr = __raw_readl(SH_CCR);
137 	ccr |= CCR_CACHE_ICI;
138 	__raw_writel(ccr, SH_CCR);
139 
140 	/*
141 	 * back_to_cached() will take care of the barrier for us, don't add
142 	 * another one!
143 	 */
144 
145 	back_to_cached();
146 	local_irq_restore(flags);
147 }
148 
149 static void flush_dcache_all(void)
150 {
151 	unsigned long addr, end_addr, entry_offset;
152 
153 	end_addr = CACHE_OC_ADDRESS_ARRAY +
154 		(current_cpu_data.dcache.sets <<
155 		 current_cpu_data.dcache.entry_shift) *
156 			current_cpu_data.dcache.ways;
157 
158 	entry_offset = 1 << current_cpu_data.dcache.entry_shift;
159 
160 	for (addr = CACHE_OC_ADDRESS_ARRAY; addr < end_addr; ) {
161 		__raw_writel(0, addr); addr += entry_offset;
162 		__raw_writel(0, addr); addr += entry_offset;
163 		__raw_writel(0, addr); addr += entry_offset;
164 		__raw_writel(0, addr); addr += entry_offset;
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 	}
170 }
171 
172 static void sh4_flush_cache_all(void *unused)
173 {
174 	flush_dcache_all();
175 	flush_icache_all();
176 }
177 
178 /*
179  * Note : (RPC) since the caches are physically tagged, the only point
180  * of flush_cache_mm for SH-4 is to get rid of aliases from the
181  * D-cache.  The assumption elsewhere, e.g. flush_cache_range, is that
182  * lines can stay resident so long as the virtual address they were
183  * accessed with (hence cache set) is in accord with the physical
184  * address (i.e. tag).  It's no different here.
185  *
186  * Caller takes mm->mmap_lock.
187  */
188 static void sh4_flush_cache_mm(void *arg)
189 {
190 	struct mm_struct *mm = arg;
191 
192 	if (cpu_context(smp_processor_id(), mm) == NO_CONTEXT)
193 		return;
194 
195 	flush_dcache_all();
196 }
197 
198 /*
199  * Write back and invalidate I/D-caches for the page.
200  *
201  * ADDR: Virtual Address (U0 address)
202  * PFN: Physical page number
203  */
204 static void sh4_flush_cache_page(void *args)
205 {
206 	struct flusher_data *data = args;
207 	struct vm_area_struct *vma;
208 	struct page *page;
209 	unsigned long address, pfn, phys;
210 	int map_coherent = 0;
211 	pmd_t *pmd;
212 	pte_t *pte;
213 	void *vaddr;
214 
215 	vma = data->vma;
216 	address = data->addr1 & PAGE_MASK;
217 	pfn = data->addr2;
218 	phys = pfn << PAGE_SHIFT;
219 	page = pfn_to_page(pfn);
220 
221 	if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
222 		return;
223 
224 	pmd = pmd_off(vma->vm_mm, address);
225 	pte = pte_offset_kernel(pmd, address);
226 
227 	/* If the page isn't present, there is nothing to do here. */
228 	if (!(pte_val(*pte) & _PAGE_PRESENT))
229 		return;
230 
231 	if ((vma->vm_mm == current->active_mm))
232 		vaddr = NULL;
233 	else {
234 		/*
235 		 * Use kmap_coherent or kmap_atomic to do flushes for
236 		 * another ASID than the current one.
237 		 */
238 		map_coherent = (current_cpu_data.dcache.n_aliases &&
239 			test_bit(PG_dcache_clean, &page->flags) &&
240 			page_mapcount(page));
241 		if (map_coherent)
242 			vaddr = kmap_coherent(page, address);
243 		else
244 			vaddr = kmap_atomic(page);
245 
246 		address = (unsigned long)vaddr;
247 	}
248 
249 	flush_cache_one(CACHE_OC_ADDRESS_ARRAY |
250 			(address & shm_align_mask), phys);
251 
252 	if (vma->vm_flags & VM_EXEC)
253 		flush_icache_all();
254 
255 	if (vaddr) {
256 		if (map_coherent)
257 			kunmap_coherent(vaddr);
258 		else
259 			kunmap_atomic(vaddr);
260 	}
261 }
262 
263 /*
264  * Write back and invalidate D-caches.
265  *
266  * START, END: Virtual Address (U0 address)
267  *
268  * NOTE: We need to flush the _physical_ page entry.
269  * Flushing the cache lines for U0 only isn't enough.
270  * We need to flush for P1 too, which may contain aliases.
271  */
272 static void sh4_flush_cache_range(void *args)
273 {
274 	struct flusher_data *data = args;
275 	struct vm_area_struct *vma;
276 	unsigned long start, end;
277 
278 	vma = data->vma;
279 	start = data->addr1;
280 	end = data->addr2;
281 
282 	if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
283 		return;
284 
285 	/*
286 	 * If cache is only 4k-per-way, there are never any 'aliases'.  Since
287 	 * the cache is physically tagged, the data can just be left in there.
288 	 */
289 	if (boot_cpu_data.dcache.n_aliases == 0)
290 		return;
291 
292 	flush_dcache_all();
293 
294 	if (vma->vm_flags & VM_EXEC)
295 		flush_icache_all();
296 }
297 
298 /**
299  * __flush_cache_one
300  *
301  * @addr:  address in memory mapped cache array
302  * @phys:  P1 address to flush (has to match tags if addr has 'A' bit
303  *         set i.e. associative write)
304  * @exec_offset: set to 0x20000000 if flush has to be executed from P2
305  *               region else 0x0
306  *
307  * The offset into the cache array implied by 'addr' selects the
308  * 'colour' of the virtual address range that will be flushed.  The
309  * operation (purge/write-back) is selected by the lower 2 bits of
310  * 'phys'.
311  */
312 static void __flush_cache_one(unsigned long addr, unsigned long phys,
313 			       unsigned long exec_offset)
314 {
315 	int way_count;
316 	unsigned long base_addr = addr;
317 	struct cache_info *dcache;
318 	unsigned long way_incr;
319 	unsigned long a, ea, p;
320 	unsigned long temp_pc;
321 
322 	dcache = &boot_cpu_data.dcache;
323 	/* Write this way for better assembly. */
324 	way_count = dcache->ways;
325 	way_incr = dcache->way_incr;
326 
327 	/*
328 	 * Apply exec_offset (i.e. branch to P2 if required.).
329 	 *
330 	 * FIXME:
331 	 *
332 	 *	If I write "=r" for the (temp_pc), it puts this in r6 hence
333 	 *	trashing exec_offset before it's been added on - why?  Hence
334 	 *	"=&r" as a 'workaround'
335 	 */
336 	asm volatile("mov.l 1f, %0\n\t"
337 		     "add   %1, %0\n\t"
338 		     "jmp   @%0\n\t"
339 		     "nop\n\t"
340 		     ".balign 4\n\t"
341 		     "1:  .long 2f\n\t"
342 		     "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
343 
344 	/*
345 	 * We know there will be >=1 iteration, so write as do-while to avoid
346 	 * pointless nead-of-loop check for 0 iterations.
347 	 */
348 	do {
349 		ea = base_addr + PAGE_SIZE;
350 		a = base_addr;
351 		p = phys;
352 
353 		do {
354 			*(volatile unsigned long *)a = p;
355 			/*
356 			 * Next line: intentionally not p+32, saves an add, p
357 			 * will do since only the cache tag bits need to
358 			 * match.
359 			 */
360 			*(volatile unsigned long *)(a+32) = p;
361 			a += 64;
362 			p += 64;
363 		} while (a < ea);
364 
365 		base_addr += way_incr;
366 	} while (--way_count != 0);
367 }
368 
369 extern void __weak sh4__flush_region_init(void);
370 
371 /*
372  * SH-4 has virtually indexed and physically tagged cache.
373  */
374 void __init sh4_cache_init(void)
375 {
376 	printk("PVR=%08x CVR=%08x PRR=%08x\n",
377 		__raw_readl(CCN_PVR),
378 		__raw_readl(CCN_CVR),
379 		__raw_readl(CCN_PRR));
380 
381 	local_flush_icache_range	= sh4_flush_icache_range;
382 	local_flush_dcache_page		= sh4_flush_dcache_page;
383 	local_flush_cache_all		= sh4_flush_cache_all;
384 	local_flush_cache_mm		= sh4_flush_cache_mm;
385 	local_flush_cache_dup_mm	= sh4_flush_cache_mm;
386 	local_flush_cache_page		= sh4_flush_cache_page;
387 	local_flush_cache_range		= sh4_flush_cache_range;
388 
389 	sh4__flush_region_init();
390 }
391