xref: /linux/arch/sh/mm/cache-sh4.c (revision 273b281fa22c293963ee3e6eec418f5dda2dbc83)
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 <asm/pgtable.h>
20 #include <asm/mmu_context.h>
21 #include <asm/cacheflush.h>
22 
23 /*
24  * The maximum number of pages we support up to when doing ranged dcache
25  * flushing. Anything exceeding this will simply flush the dcache in its
26  * entirety.
27  */
28 #define MAX_ICACHE_PAGES	32
29 
30 static void __flush_cache_one(unsigned long addr, unsigned long phys,
31 			       unsigned long exec_offset);
32 
33 /*
34  * Write back the range of D-cache, and purge the I-cache.
35  *
36  * Called from kernel/module.c:sys_init_module and routine for a.out format,
37  * signal handler code and kprobes code
38  */
39 static void __uses_jump_to_uncached sh4_flush_icache_range(void *args)
40 {
41 	struct flusher_data *data = args;
42 	unsigned long start, end;
43 	unsigned long flags, v;
44 	int i;
45 
46 	start = data->addr1;
47 	end = data->addr2;
48 
49 	/* If there are too many pages then just blow away the caches */
50 	if (((end - start) >> PAGE_SHIFT) >= MAX_ICACHE_PAGES) {
51 		local_flush_cache_all(NULL);
52 		return;
53 	}
54 
55 	/*
56 	 * Selectively flush d-cache then invalidate the i-cache.
57 	 * This is inefficient, so only use this for small ranges.
58 	 */
59 	start &= ~(L1_CACHE_BYTES-1);
60 	end += L1_CACHE_BYTES-1;
61 	end &= ~(L1_CACHE_BYTES-1);
62 
63 	local_irq_save(flags);
64 	jump_to_uncached();
65 
66 	for (v = start; v < end; v += L1_CACHE_BYTES) {
67 		unsigned long icacheaddr;
68 		int j, n;
69 
70 		__ocbwb(v);
71 
72 		icacheaddr = CACHE_IC_ADDRESS_ARRAY | (v &
73 				cpu_data->icache.entry_mask);
74 
75 		/* Clear i-cache line valid-bit */
76 		n = boot_cpu_data.icache.n_aliases;
77 		for (i = 0; i < cpu_data->icache.ways; i++) {
78 			for (j = 0; j < n; j++)
79 				__raw_writel(0, icacheaddr + (j * PAGE_SIZE));
80 			icacheaddr += cpu_data->icache.way_incr;
81 		}
82 	}
83 
84 	back_to_cached();
85 	local_irq_restore(flags);
86 }
87 
88 static inline void flush_cache_one(unsigned long start, unsigned long phys)
89 {
90 	unsigned long flags, exec_offset = 0;
91 
92 	/*
93 	 * All types of SH-4 require PC to be uncached to operate on the I-cache.
94 	 * Some types of SH-4 require PC to be uncached to operate on the D-cache.
95 	 */
96 	if ((boot_cpu_data.flags & CPU_HAS_P2_FLUSH_BUG) ||
97 	    (start < CACHE_OC_ADDRESS_ARRAY))
98 		exec_offset = cached_to_uncached;
99 
100 	local_irq_save(flags);
101 	__flush_cache_one(start, phys, exec_offset);
102 	local_irq_restore(flags);
103 }
104 
105 /*
106  * Write back & invalidate the D-cache of the page.
107  * (To avoid "alias" issues)
108  */
109 static void sh4_flush_dcache_page(void *arg)
110 {
111 	struct page *page = arg;
112 #ifndef CONFIG_SMP
113 	struct address_space *mapping = page_mapping(page);
114 
115 	if (mapping && !mapping_mapped(mapping))
116 		set_bit(PG_dcache_dirty, &page->flags);
117 	else
118 #endif
119 	{
120 		unsigned long phys = page_to_phys(page);
121 		unsigned long addr = CACHE_OC_ADDRESS_ARRAY;
122 		int i, n;
123 
124 		/* Loop all the D-cache */
125 		n = boot_cpu_data.dcache.n_aliases;
126 		for (i = 0; i < n; i++, addr += PAGE_SIZE)
127 			flush_cache_one(addr, phys);
128 	}
129 
130 	wmb();
131 }
132 
133 /* TODO: Selective icache invalidation through IC address array.. */
134 static void __uses_jump_to_uncached flush_icache_all(void)
135 {
136 	unsigned long flags, ccr;
137 
138 	local_irq_save(flags);
139 	jump_to_uncached();
140 
141 	/* Flush I-cache */
142 	ccr = ctrl_inl(CCR);
143 	ccr |= CCR_CACHE_ICI;
144 	ctrl_outl(ccr, CCR);
145 
146 	/*
147 	 * back_to_cached() will take care of the barrier for us, don't add
148 	 * another one!
149 	 */
150 
151 	back_to_cached();
152 	local_irq_restore(flags);
153 }
154 
155 static void flush_dcache_all(void)
156 {
157 	unsigned long addr, end_addr, entry_offset;
158 
159 	end_addr = CACHE_OC_ADDRESS_ARRAY +
160 		(current_cpu_data.dcache.sets <<
161 		 current_cpu_data.dcache.entry_shift) *
162 			current_cpu_data.dcache.ways;
163 
164 	entry_offset = 1 << current_cpu_data.dcache.entry_shift;
165 
166 	for (addr = CACHE_OC_ADDRESS_ARRAY; addr < end_addr; ) {
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;
173 		__raw_writel(0, addr); addr += entry_offset;
174 		__raw_writel(0, addr); addr += entry_offset;
175 	}
176 }
177 
178 static void sh4_flush_cache_all(void *unused)
179 {
180 	flush_dcache_all();
181 	flush_icache_all();
182 }
183 
184 /*
185  * Note : (RPC) since the caches are physically tagged, the only point
186  * of flush_cache_mm for SH-4 is to get rid of aliases from the
187  * D-cache.  The assumption elsewhere, e.g. flush_cache_range, is that
188  * lines can stay resident so long as the virtual address they were
189  * accessed with (hence cache set) is in accord with the physical
190  * address (i.e. tag).  It's no different here.
191  *
192  * Caller takes mm->mmap_sem.
193  */
194 static void sh4_flush_cache_mm(void *arg)
195 {
196 	struct mm_struct *mm = arg;
197 
198 	if (cpu_context(smp_processor_id(), mm) == NO_CONTEXT)
199 		return;
200 
201 	flush_dcache_all();
202 }
203 
204 /*
205  * Write back and invalidate I/D-caches for the page.
206  *
207  * ADDR: Virtual Address (U0 address)
208  * PFN: Physical page number
209  */
210 static void sh4_flush_cache_page(void *args)
211 {
212 	struct flusher_data *data = args;
213 	struct vm_area_struct *vma;
214 	struct page *page;
215 	unsigned long address, pfn, phys;
216 	int map_coherent = 0;
217 	pgd_t *pgd;
218 	pud_t *pud;
219 	pmd_t *pmd;
220 	pte_t *pte;
221 	void *vaddr;
222 
223 	vma = data->vma;
224 	address = data->addr1 & PAGE_MASK;
225 	pfn = data->addr2;
226 	phys = pfn << PAGE_SHIFT;
227 	page = pfn_to_page(pfn);
228 
229 	if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
230 		return;
231 
232 	pgd = pgd_offset(vma->vm_mm, address);
233 	pud = pud_offset(pgd, address);
234 	pmd = pmd_offset(pud, address);
235 	pte = pte_offset_kernel(pmd, address);
236 
237 	/* If the page isn't present, there is nothing to do here. */
238 	if (!(pte_val(*pte) & _PAGE_PRESENT))
239 		return;
240 
241 	if ((vma->vm_mm == current->active_mm))
242 		vaddr = NULL;
243 	else {
244 		/*
245 		 * Use kmap_coherent or kmap_atomic to do flushes for
246 		 * another ASID than the current one.
247 		 */
248 		map_coherent = (current_cpu_data.dcache.n_aliases &&
249 			!test_bit(PG_dcache_dirty, &page->flags) &&
250 			page_mapped(page));
251 		if (map_coherent)
252 			vaddr = kmap_coherent(page, address);
253 		else
254 			vaddr = kmap_atomic(page, KM_USER0);
255 
256 		address = (unsigned long)vaddr;
257 	}
258 
259 	if (pages_do_alias(address, phys))
260 		flush_cache_one(CACHE_OC_ADDRESS_ARRAY |
261 			(address & shm_align_mask), phys);
262 
263 	if (vma->vm_flags & VM_EXEC)
264 		flush_icache_all();
265 
266 	if (vaddr) {
267 		if (map_coherent)
268 			kunmap_coherent(vaddr);
269 		else
270 			kunmap_atomic(vaddr, KM_USER0);
271 	}
272 }
273 
274 /*
275  * Write back and invalidate D-caches.
276  *
277  * START, END: Virtual Address (U0 address)
278  *
279  * NOTE: We need to flush the _physical_ page entry.
280  * Flushing the cache lines for U0 only isn't enough.
281  * We need to flush for P1 too, which may contain aliases.
282  */
283 static void sh4_flush_cache_range(void *args)
284 {
285 	struct flusher_data *data = args;
286 	struct vm_area_struct *vma;
287 	unsigned long start, end;
288 
289 	vma = data->vma;
290 	start = data->addr1;
291 	end = data->addr2;
292 
293 	if (cpu_context(smp_processor_id(), vma->vm_mm) == NO_CONTEXT)
294 		return;
295 
296 	/*
297 	 * If cache is only 4k-per-way, there are never any 'aliases'.  Since
298 	 * the cache is physically tagged, the data can just be left in there.
299 	 */
300 	if (boot_cpu_data.dcache.n_aliases == 0)
301 		return;
302 
303 	flush_dcache_all();
304 
305 	if (vma->vm_flags & VM_EXEC)
306 		flush_icache_all();
307 }
308 
309 /**
310  * __flush_cache_one
311  *
312  * @addr:  address in memory mapped cache array
313  * @phys:  P1 address to flush (has to match tags if addr has 'A' bit
314  *         set i.e. associative write)
315  * @exec_offset: set to 0x20000000 if flush has to be executed from P2
316  *               region else 0x0
317  *
318  * The offset into the cache array implied by 'addr' selects the
319  * 'colour' of the virtual address range that will be flushed.  The
320  * operation (purge/write-back) is selected by the lower 2 bits of
321  * 'phys'.
322  */
323 static void __flush_cache_one(unsigned long addr, unsigned long phys,
324 			       unsigned long exec_offset)
325 {
326 	int way_count;
327 	unsigned long base_addr = addr;
328 	struct cache_info *dcache;
329 	unsigned long way_incr;
330 	unsigned long a, ea, p;
331 	unsigned long temp_pc;
332 
333 	dcache = &boot_cpu_data.dcache;
334 	/* Write this way for better assembly. */
335 	way_count = dcache->ways;
336 	way_incr = dcache->way_incr;
337 
338 	/*
339 	 * Apply exec_offset (i.e. branch to P2 if required.).
340 	 *
341 	 * FIXME:
342 	 *
343 	 *	If I write "=r" for the (temp_pc), it puts this in r6 hence
344 	 *	trashing exec_offset before it's been added on - why?  Hence
345 	 *	"=&r" as a 'workaround'
346 	 */
347 	asm volatile("mov.l 1f, %0\n\t"
348 		     "add   %1, %0\n\t"
349 		     "jmp   @%0\n\t"
350 		     "nop\n\t"
351 		     ".balign 4\n\t"
352 		     "1:  .long 2f\n\t"
353 		     "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
354 
355 	/*
356 	 * We know there will be >=1 iteration, so write as do-while to avoid
357 	 * pointless nead-of-loop check for 0 iterations.
358 	 */
359 	do {
360 		ea = base_addr + PAGE_SIZE;
361 		a = base_addr;
362 		p = phys;
363 
364 		do {
365 			*(volatile unsigned long *)a = p;
366 			/*
367 			 * Next line: intentionally not p+32, saves an add, p
368 			 * will do since only the cache tag bits need to
369 			 * match.
370 			 */
371 			*(volatile unsigned long *)(a+32) = p;
372 			a += 64;
373 			p += 64;
374 		} while (a < ea);
375 
376 		base_addr += way_incr;
377 	} while (--way_count != 0);
378 }
379 
380 extern void __weak sh4__flush_region_init(void);
381 
382 /*
383  * SH-4 has virtually indexed and physically tagged cache.
384  */
385 void __init sh4_cache_init(void)
386 {
387 	printk("PVR=%08x CVR=%08x PRR=%08x\n",
388 		ctrl_inl(CCN_PVR),
389 		ctrl_inl(CCN_CVR),
390 		ctrl_inl(CCN_PRR));
391 
392 	local_flush_icache_range	= sh4_flush_icache_range;
393 	local_flush_dcache_page		= sh4_flush_dcache_page;
394 	local_flush_cache_all		= sh4_flush_cache_all;
395 	local_flush_cache_mm		= sh4_flush_cache_mm;
396 	local_flush_cache_dup_mm	= sh4_flush_cache_mm;
397 	local_flush_cache_page		= sh4_flush_cache_page;
398 	local_flush_cache_range		= sh4_flush_cache_range;
399 
400 	sh4__flush_region_init();
401 }
402