2 * Copyright (C) 1996 David S. Miller (dm@engr.sgi.com)
3 * Copyright (C) 1997, 2001 Ralf Baechle (ralf@gnu.org)
4 * Copyright (C) 2000, 2001, 2002, 2003 Broadcom Corporation
5 * Copyright (C) 2004 Maciej W. Rozycki
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version 2
10 * of the License, or (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
21 #include <linux/init.h>
24 #include <asm/bootinfo.h>
25 #include <asm/cacheops.h>
27 #include <asm/mipsregs.h>
28 #include <asm/mmu_context.h>
29 #include <asm/uaccess.h>
31 extern void sb1_dma_init(void);
33 /* These are probed at ld_mmu time */
34 static unsigned long icache_size;
35 static unsigned long dcache_size;
37 static unsigned short icache_line_size;
38 static unsigned short dcache_line_size;
40 static unsigned int icache_index_mask;
41 static unsigned int dcache_index_mask;
43 static unsigned short icache_assoc;
44 static unsigned short dcache_assoc;
46 static unsigned short icache_sets;
47 static unsigned short dcache_sets;
49 static unsigned int icache_range_cutoff;
50 static unsigned int dcache_range_cutoff;
52 static inline void sb1_on_each_cpu(void (*func) (void *info), void *info,
56 smp_call_function(func, info, retry, wait);
62 * The dcache is fully coherent to the system, with one
63 * big caveat: the instruction stream. In other words,
64 * if we miss in the icache, and have dirty data in the
65 * L1 dcache, then we'll go out to memory (or the L2) and
66 * get the not-as-recent data.
68 * So the only time we have to flush the dcache is when
69 * we're flushing the icache. Since the L2 is fully
70 * coherent to everything, including I/O, we never have
74 #define cache_set_op(op, addr) \
75 __asm__ __volatile__( \
76 " .set noreorder \n" \
77 " .set mips64\n\t \n" \
78 " cache %0, (0<<13)(%1) \n" \
79 " cache %0, (1<<13)(%1) \n" \
80 " cache %0, (2<<13)(%1) \n" \
81 " cache %0, (3<<13)(%1) \n" \
85 : "i" (op), "r" (addr))
89 " .set mips64\n\t \n" \
93 #define mispredict() \
94 __asm__ __volatile__( \
95 " bnezl $0, 1f \n" /* Force mispredict */ \
99 * Writeback and invalidate the entire dcache
101 static inline void __sb1_writeback_inv_dcache_all(void)
103 unsigned long addr = 0;
105 while (addr < dcache_line_size * dcache_sets) {
106 cache_set_op(Index_Writeback_Inv_D, addr);
107 addr += dcache_line_size;
112 * Writeback and invalidate a range of the dcache. The addresses are
113 * virtual, and since we're using index ops and bit 12 is part of both
114 * the virtual frame and physical index, we have to clear both sets
115 * (bit 12 set and cleared).
117 static inline void __sb1_writeback_inv_dcache_range(unsigned long start,
122 start &= ~(dcache_line_size - 1);
123 end = (end + dcache_line_size - 1) & ~(dcache_line_size - 1);
125 while (start != end) {
126 index = start & dcache_index_mask;
127 cache_set_op(Index_Writeback_Inv_D, index);
128 cache_set_op(Index_Writeback_Inv_D, index ^ (1<<12));
129 start += dcache_line_size;
135 * Writeback and invalidate a range of the dcache. With physical
136 * addresseses, we don't have to worry about possible bit 12 aliasing.
137 * XXXKW is it worth turning on KX and using hit ops with xkphys?
139 static inline void __sb1_writeback_inv_dcache_phys_range(unsigned long start,
142 start &= ~(dcache_line_size - 1);
143 end = (end + dcache_line_size - 1) & ~(dcache_line_size - 1);
145 while (start != end) {
146 cache_set_op(Index_Writeback_Inv_D, start & dcache_index_mask);
147 start += dcache_line_size;
154 * Invalidate the entire icache
156 static inline void __sb1_flush_icache_all(void)
158 unsigned long addr = 0;
160 while (addr < icache_line_size * icache_sets) {
161 cache_set_op(Index_Invalidate_I, addr);
162 addr += icache_line_size;
167 * Invalidate a range of the icache. The addresses are virtual, and
168 * the cache is virtually indexed and tagged. However, we don't
169 * necessarily have the right ASID context, so use index ops instead
172 static inline void __sb1_flush_icache_range(unsigned long start,
175 start &= ~(icache_line_size - 1);
176 end = (end + icache_line_size - 1) & ~(icache_line_size - 1);
178 while (start != end) {
179 cache_set_op(Index_Invalidate_I, start & icache_index_mask);
180 start += icache_line_size;
187 * Flush the icache for a given physical page. Need to writeback the
188 * dcache first, then invalidate the icache. If the page isn't
189 * executable, nothing is required.
191 static void local_sb1_flush_cache_page(struct vm_area_struct *vma, unsigned long addr, unsigned long pfn)
193 int cpu = smp_processor_id();
196 if (!(vma->vm_flags & VM_EXEC))
200 __sb1_writeback_inv_dcache_range(addr, addr + PAGE_SIZE);
203 * Bumping the ASID is probably cheaper than the flush ...
205 if (vma->vm_mm == current->active_mm) {
206 if (cpu_context(cpu, vma->vm_mm) != 0)
207 drop_mmu_context(vma->vm_mm, cpu);
209 __sb1_flush_icache_range(addr, addr + PAGE_SIZE);
213 struct flush_cache_page_args {
214 struct vm_area_struct *vma;
219 static void sb1_flush_cache_page_ipi(void *info)
221 struct flush_cache_page_args *args = info;
223 local_sb1_flush_cache_page(args->vma, args->addr, args->pfn);
226 /* Dirty dcache could be on another CPU, so do the IPIs */
227 static void sb1_flush_cache_page(struct vm_area_struct *vma, unsigned long addr, unsigned long pfn)
229 struct flush_cache_page_args args;
231 if (!(vma->vm_flags & VM_EXEC))
238 sb1_on_each_cpu(sb1_flush_cache_page_ipi, (void *) &args, 1, 1);
241 void sb1_flush_cache_page(struct vm_area_struct *vma, unsigned long addr, unsigned long pfn)
242 __attribute__((alias("local_sb1_flush_cache_page")));
247 * Invalidate all caches on this CPU
249 static void __attribute_used__ local_sb1___flush_cache_all(void)
251 __sb1_writeback_inv_dcache_all();
252 __sb1_flush_icache_all();
256 void sb1___flush_cache_all_ipi(void *ignored)
257 __attribute__((alias("local_sb1___flush_cache_all")));
259 static void sb1___flush_cache_all(void)
261 sb1_on_each_cpu(sb1___flush_cache_all_ipi, 0, 1, 1);
264 void sb1___flush_cache_all(void)
265 __attribute__((alias("local_sb1___flush_cache_all")));
269 * When flushing a range in the icache, we have to first writeback
270 * the dcache for the same range, so new ifetches will see any
271 * data that was dirty in the dcache.
273 * The start/end arguments are Kseg addresses (possibly mapped Kseg).
276 static void local_sb1_flush_icache_range(unsigned long start,
279 /* Just wb-inv the whole dcache if the range is big enough */
280 if ((end - start) > dcache_range_cutoff)
281 __sb1_writeback_inv_dcache_all();
283 __sb1_writeback_inv_dcache_range(start, end);
285 /* Just flush the whole icache if the range is big enough */
286 if ((end - start) > icache_range_cutoff)
287 __sb1_flush_icache_all();
289 __sb1_flush_icache_range(start, end);
293 struct flush_icache_range_args {
298 static void sb1_flush_icache_range_ipi(void *info)
300 struct flush_icache_range_args *args = info;
302 local_sb1_flush_icache_range(args->start, args->end);
305 void sb1_flush_icache_range(unsigned long start, unsigned long end)
307 struct flush_icache_range_args args;
311 sb1_on_each_cpu(sb1_flush_icache_range_ipi, &args, 1, 1);
314 void sb1_flush_icache_range(unsigned long start, unsigned long end)
315 __attribute__((alias("local_sb1_flush_icache_range")));
319 * A signal trampoline must fit into a single cacheline.
321 static void local_sb1_flush_cache_sigtramp(unsigned long addr)
323 cache_set_op(Index_Writeback_Inv_D, addr & dcache_index_mask);
324 cache_set_op(Index_Writeback_Inv_D, (addr ^ (1<<12)) & dcache_index_mask);
325 cache_set_op(Index_Invalidate_I, addr & icache_index_mask);
330 static void sb1_flush_cache_sigtramp_ipi(void *info)
332 unsigned long iaddr = (unsigned long) info;
333 local_sb1_flush_cache_sigtramp(iaddr);
336 static void sb1_flush_cache_sigtramp(unsigned long addr)
338 sb1_on_each_cpu(sb1_flush_cache_sigtramp_ipi, (void *) addr, 1, 1);
341 void sb1_flush_cache_sigtramp(unsigned long addr)
342 __attribute__((alias("local_sb1_flush_cache_sigtramp")));
347 * Anything that just flushes dcache state can be ignored, as we're always
348 * coherent in dcache space. This is just a dummy function that all the
349 * nop'ed routines point to
351 static void sb1_nop(void)
356 * Cache set values (from the mips64 spec)
367 static unsigned int decode_cache_sets(unsigned int config_field)
369 if (config_field == 7) {
370 /* JDCXXX - Find a graceful way to abort. */
373 return (1<<(config_field + 6));
377 * Cache line size values (from the mips64 spec)
378 * 0 - No cache present.
388 static unsigned int decode_cache_line_size(unsigned int config_field)
390 if (config_field == 0) {
392 } else if (config_field == 7) {
393 /* JDCXXX - Find a graceful way to abort. */
396 return (1<<(config_field + 1));
400 * Relevant bits of the config1 register format (from the MIPS32/MIPS64 specs)
402 * 24:22 Icache sets per way
403 * 21:19 Icache line size
404 * 18:16 Icache Associativity
405 * 15:13 Dcache sets per way
406 * 12:10 Dcache line size
407 * 9:7 Dcache Associativity
410 static char *way_string[] = {
411 "direct mapped", "2-way", "3-way", "4-way",
412 "5-way", "6-way", "7-way", "8-way",
415 static __init void probe_cache_sizes(void)
419 config1 = read_c0_config1();
420 icache_line_size = decode_cache_line_size((config1 >> 19) & 0x7);
421 dcache_line_size = decode_cache_line_size((config1 >> 10) & 0x7);
422 icache_sets = decode_cache_sets((config1 >> 22) & 0x7);
423 dcache_sets = decode_cache_sets((config1 >> 13) & 0x7);
424 icache_assoc = ((config1 >> 16) & 0x7) + 1;
425 dcache_assoc = ((config1 >> 7) & 0x7) + 1;
426 icache_size = icache_line_size * icache_sets * icache_assoc;
427 dcache_size = dcache_line_size * dcache_sets * dcache_assoc;
428 /* Need to remove non-index bits for index ops */
429 icache_index_mask = (icache_sets - 1) * icache_line_size;
430 dcache_index_mask = (dcache_sets - 1) * dcache_line_size;
432 * These are for choosing range (index ops) versus all.
433 * icache flushes all ways for each set, so drop icache_assoc.
434 * dcache flushes all ways and each setting of bit 12 for each
435 * index, so drop dcache_assoc and halve the dcache_sets.
437 icache_range_cutoff = icache_sets * icache_line_size;
438 dcache_range_cutoff = (dcache_sets / 2) * icache_line_size;
440 printk("Primary instruction cache %ldkB, %s, linesize %d bytes.\n",
441 icache_size >> 10, way_string[icache_assoc - 1],
443 printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
444 dcache_size >> 10, way_string[dcache_assoc - 1],
449 * This is called from cache.c. We have to set up all the
450 * memory management function pointers, as well as initialize
451 * the caches and tlbs
453 void sb1_cache_init(void)
455 extern char except_vec2_sb1;
457 /* Special cache error handler for SB1 */
458 set_uncached_handler (0x100, &except_vec2_sb1, 0x80);
462 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
467 * None of these are needed for the SB1 - the Dcache is
468 * physically indexed and tagged, so no virtual aliasing can
471 flush_cache_range = (void *) sb1_nop;
472 flush_cache_mm = (void (*)(struct mm_struct *))sb1_nop;
473 flush_cache_all = sb1_nop;
475 /* These routines are for Icache coherence with the Dcache */
476 flush_icache_range = sb1_flush_icache_range;
477 flush_icache_all = __sb1_flush_icache_all; /* local only */
479 /* This implies an Icache flush too, so can't be nop'ed */
480 flush_cache_page = sb1_flush_cache_page;
482 flush_cache_sigtramp = sb1_flush_cache_sigtramp;
483 local_flush_data_cache_page = (void *) sb1_nop;
484 flush_data_cache_page = (void *) sb1_nop;
487 __flush_cache_all = sb1___flush_cache_all;
489 change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
492 * This is the only way to force the update of K0 to complete
493 * before subsequent instruction fetch.
495 __asm__ __volatile__(
500 " " STR(PTR_LA) " $1, 1f \n"
501 " " STR(MTC0) " $1, $14 \n"
508 local_sb1___flush_cache_all();