Merge branches 'x86/cleanups', 'x86/fpu' and 'x86/urgent' into x86/core
[linux-2.6] / arch / parisc / kernel / cache.c
1 /* $Id: cache.c,v 1.4 2000/01/25 00:11:38 prumpf Exp $
2  *
3  * This file is subject to the terms and conditions of the GNU General Public
4  * License.  See the file "COPYING" in the main directory of this archive
5  * for more details.
6  *
7  * Copyright (C) 1999-2006 Helge Deller <deller@gmx.de> (07-13-1999)
8  * Copyright (C) 1999 SuSE GmbH Nuernberg
9  * Copyright (C) 2000 Philipp Rumpf (prumpf@tux.org)
10  *
11  * Cache and TLB management
12  *
13  */
14  
15 #include <linux/init.h>
16 #include <linux/kernel.h>
17 #include <linux/mm.h>
18 #include <linux/module.h>
19 #include <linux/seq_file.h>
20 #include <linux/pagemap.h>
21 #include <linux/sched.h>
22 #include <asm/pdc.h>
23 #include <asm/cache.h>
24 #include <asm/cacheflush.h>
25 #include <asm/tlbflush.h>
26 #include <asm/system.h>
27 #include <asm/page.h>
28 #include <asm/pgalloc.h>
29 #include <asm/processor.h>
30 #include <asm/sections.h>
31
32 int split_tlb __read_mostly;
33 int dcache_stride __read_mostly;
34 int icache_stride __read_mostly;
35 EXPORT_SYMBOL(dcache_stride);
36
37
38 /* On some machines (e.g. ones with the Merced bus), there can be
39  * only a single PxTLB broadcast at a time; this must be guaranteed
40  * by software.  We put a spinlock around all TLB flushes  to
41  * ensure this.
42  */
43 DEFINE_SPINLOCK(pa_tlb_lock);
44
45 struct pdc_cache_info cache_info __read_mostly;
46 #ifndef CONFIG_PA20
47 static struct pdc_btlb_info btlb_info __read_mostly;
48 #endif
49
50 #ifdef CONFIG_SMP
51 void
52 flush_data_cache(void)
53 {
54         on_each_cpu(flush_data_cache_local, NULL, 1);
55 }
56 void 
57 flush_instruction_cache(void)
58 {
59         on_each_cpu(flush_instruction_cache_local, NULL, 1);
60 }
61 #endif
62
63 void
64 flush_cache_all_local(void)
65 {
66         flush_instruction_cache_local(NULL);
67         flush_data_cache_local(NULL);
68 }
69 EXPORT_SYMBOL(flush_cache_all_local);
70
71 void
72 update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
73 {
74         struct page *page = pte_page(pte);
75
76         if (pfn_valid(page_to_pfn(page)) && page_mapping(page) &&
77             test_bit(PG_dcache_dirty, &page->flags)) {
78
79                 flush_kernel_dcache_page(page);
80                 clear_bit(PG_dcache_dirty, &page->flags);
81         } else if (parisc_requires_coherency())
82                 flush_kernel_dcache_page(page);
83 }
84
85 void
86 show_cache_info(struct seq_file *m)
87 {
88         char buf[32];
89
90         seq_printf(m, "I-cache\t\t: %ld KB\n", 
91                 cache_info.ic_size/1024 );
92         if (cache_info.dc_loop != 1)
93                 snprintf(buf, 32, "%lu-way associative", cache_info.dc_loop);
94         seq_printf(m, "D-cache\t\t: %ld KB (%s%s, %s)\n",
95                 cache_info.dc_size/1024,
96                 (cache_info.dc_conf.cc_wt ? "WT":"WB"),
97                 (cache_info.dc_conf.cc_sh ? ", shared I/D":""),
98                 ((cache_info.dc_loop == 1) ? "direct mapped" : buf));
99         seq_printf(m, "ITLB entries\t: %ld\n" "DTLB entries\t: %ld%s\n",
100                 cache_info.it_size,
101                 cache_info.dt_size,
102                 cache_info.dt_conf.tc_sh ? " - shared with ITLB":""
103         );
104                 
105 #ifndef CONFIG_PA20
106         /* BTLB - Block TLB */
107         if (btlb_info.max_size==0) {
108                 seq_printf(m, "BTLB\t\t: not supported\n" );
109         } else {
110                 seq_printf(m, 
111                 "BTLB fixed\t: max. %d pages, pagesize=%d (%dMB)\n"
112                 "BTLB fix-entr.\t: %d instruction, %d data (%d combined)\n"
113                 "BTLB var-entr.\t: %d instruction, %d data (%d combined)\n",
114                 btlb_info.max_size, (int)4096,
115                 btlb_info.max_size>>8,
116                 btlb_info.fixed_range_info.num_i,
117                 btlb_info.fixed_range_info.num_d,
118                 btlb_info.fixed_range_info.num_comb, 
119                 btlb_info.variable_range_info.num_i,
120                 btlb_info.variable_range_info.num_d,
121                 btlb_info.variable_range_info.num_comb
122                 );
123         }
124 #endif
125 }
126
127 void __init 
128 parisc_cache_init(void)
129 {
130         if (pdc_cache_info(&cache_info) < 0)
131                 panic("parisc_cache_init: pdc_cache_info failed");
132
133 #if 0
134         printk("ic_size %lx dc_size %lx it_size %lx\n",
135                 cache_info.ic_size,
136                 cache_info.dc_size,
137                 cache_info.it_size);
138
139         printk("DC  base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
140                 cache_info.dc_base,
141                 cache_info.dc_stride,
142                 cache_info.dc_count,
143                 cache_info.dc_loop);
144
145         printk("dc_conf = 0x%lx  alias %d blk %d line %d shift %d\n",
146                 *(unsigned long *) (&cache_info.dc_conf),
147                 cache_info.dc_conf.cc_alias,
148                 cache_info.dc_conf.cc_block,
149                 cache_info.dc_conf.cc_line,
150                 cache_info.dc_conf.cc_shift);
151         printk("        wt %d sh %d cst %d hv %d\n",
152                 cache_info.dc_conf.cc_wt,
153                 cache_info.dc_conf.cc_sh,
154                 cache_info.dc_conf.cc_cst,
155                 cache_info.dc_conf.cc_hv);
156
157         printk("IC  base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
158                 cache_info.ic_base,
159                 cache_info.ic_stride,
160                 cache_info.ic_count,
161                 cache_info.ic_loop);
162
163         printk("ic_conf = 0x%lx  alias %d blk %d line %d shift %d\n",
164                 *(unsigned long *) (&cache_info.ic_conf),
165                 cache_info.ic_conf.cc_alias,
166                 cache_info.ic_conf.cc_block,
167                 cache_info.ic_conf.cc_line,
168                 cache_info.ic_conf.cc_shift);
169         printk("        wt %d sh %d cst %d hv %d\n",
170                 cache_info.ic_conf.cc_wt,
171                 cache_info.ic_conf.cc_sh,
172                 cache_info.ic_conf.cc_cst,
173                 cache_info.ic_conf.cc_hv);
174
175         printk("D-TLB conf: sh %d page %d cst %d aid %d pad1 %d \n",
176                 cache_info.dt_conf.tc_sh,
177                 cache_info.dt_conf.tc_page,
178                 cache_info.dt_conf.tc_cst,
179                 cache_info.dt_conf.tc_aid,
180                 cache_info.dt_conf.tc_pad1);
181
182         printk("I-TLB conf: sh %d page %d cst %d aid %d pad1 %d \n",
183                 cache_info.it_conf.tc_sh,
184                 cache_info.it_conf.tc_page,
185                 cache_info.it_conf.tc_cst,
186                 cache_info.it_conf.tc_aid,
187                 cache_info.it_conf.tc_pad1);
188 #endif
189
190         split_tlb = 0;
191         if (cache_info.dt_conf.tc_sh == 0 || cache_info.dt_conf.tc_sh == 2) {
192                 if (cache_info.dt_conf.tc_sh == 2)
193                         printk(KERN_WARNING "Unexpected TLB configuration. "
194                         "Will flush I/D separately (could be optimized).\n");
195
196                 split_tlb = 1;
197         }
198
199         /* "New and Improved" version from Jim Hull 
200          *      (1 << (cc_block-1)) * (cc_line << (4 + cnf.cc_shift))
201          * The following CAFL_STRIDE is an optimized version, see
202          * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023625.html
203          * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023671.html
204          */
205 #define CAFL_STRIDE(cnf) (cnf.cc_line << (3 + cnf.cc_block + cnf.cc_shift))
206         dcache_stride = CAFL_STRIDE(cache_info.dc_conf);
207         icache_stride = CAFL_STRIDE(cache_info.ic_conf);
208 #undef CAFL_STRIDE
209
210 #ifndef CONFIG_PA20
211         if (pdc_btlb_info(&btlb_info) < 0) {
212                 memset(&btlb_info, 0, sizeof btlb_info);
213         }
214 #endif
215
216         if ((boot_cpu_data.pdc.capabilities & PDC_MODEL_NVA_MASK) ==
217                                                 PDC_MODEL_NVA_UNSUPPORTED) {
218                 printk(KERN_WARNING "parisc_cache_init: Only equivalent aliasing supported!\n");
219 #if 0
220                 panic("SMP kernel required to avoid non-equivalent aliasing");
221 #endif
222         }
223 }
224
225 void disable_sr_hashing(void)
226 {
227         int srhash_type, retval;
228         unsigned long space_bits;
229
230         switch (boot_cpu_data.cpu_type) {
231         case pcx: /* We shouldn't get this far.  setup.c should prevent it. */
232                 BUG();
233                 return;
234
235         case pcxs:
236         case pcxt:
237         case pcxt_:
238                 srhash_type = SRHASH_PCXST;
239                 break;
240
241         case pcxl:
242                 srhash_type = SRHASH_PCXL;
243                 break;
244
245         case pcxl2: /* pcxl2 doesn't support space register hashing */
246                 return;
247
248         default: /* Currently all PA2.0 machines use the same ins. sequence */
249                 srhash_type = SRHASH_PA20;
250                 break;
251         }
252
253         disable_sr_hashing_asm(srhash_type);
254
255         retval = pdc_spaceid_bits(&space_bits);
256         /* If this procedure isn't implemented, don't panic. */
257         if (retval < 0 && retval != PDC_BAD_OPTION)
258                 panic("pdc_spaceid_bits call failed.\n");
259         if (space_bits != 0)
260                 panic("SpaceID hashing is still on!\n");
261 }
262
263 /* Simple function to work out if we have an existing address translation
264  * for a user space vma. */
265 static inline int translation_exists(struct vm_area_struct *vma,
266                                 unsigned long addr, unsigned long pfn)
267 {
268         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
269         pmd_t *pmd;
270         pte_t pte;
271
272         if(pgd_none(*pgd))
273                 return 0;
274
275         pmd = pmd_offset(pgd, addr);
276         if(pmd_none(*pmd) || pmd_bad(*pmd))
277                 return 0;
278
279         /* We cannot take the pte lock here: flush_cache_page is usually
280          * called with pte lock already held.  Whereas flush_dcache_page
281          * takes flush_dcache_mmap_lock, which is lower in the hierarchy:
282          * the vma itself is secure, but the pte might come or go racily.
283          */
284         pte = *pte_offset_map(pmd, addr);
285         /* But pte_unmap() does nothing on this architecture */
286
287         /* Filter out coincidental file entries and swap entries */
288         if (!(pte_val(pte) & (_PAGE_FLUSH|_PAGE_PRESENT)))
289                 return 0;
290
291         return pte_pfn(pte) == pfn;
292 }
293
294 /* Private function to flush a page from the cache of a non-current
295  * process.  cr25 contains the Page Directory of the current user
296  * process; we're going to hijack both it and the user space %sr3 to
297  * temporarily make the non-current process current.  We have to do
298  * this because cache flushing may cause a non-access tlb miss which
299  * the handlers have to fill in from the pgd of the non-current
300  * process. */
301 static inline void
302 flush_user_cache_page_non_current(struct vm_area_struct *vma,
303                                   unsigned long vmaddr)
304 {
305         /* save the current process space and pgd */
306         unsigned long space = mfsp(3), pgd = mfctl(25);
307
308         /* we don't mind taking interrupts since they may not
309          * do anything with user space, but we can't
310          * be preempted here */
311         preempt_disable();
312
313         /* make us current */
314         mtctl(__pa(vma->vm_mm->pgd), 25);
315         mtsp(vma->vm_mm->context, 3);
316
317         flush_user_dcache_page(vmaddr);
318         if(vma->vm_flags & VM_EXEC)
319                 flush_user_icache_page(vmaddr);
320
321         /* put the old current process back */
322         mtsp(space, 3);
323         mtctl(pgd, 25);
324         preempt_enable();
325 }
326
327
328 static inline void
329 __flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr)
330 {
331         if (likely(vma->vm_mm->context == mfsp(3))) {
332                 flush_user_dcache_page(vmaddr);
333                 if (vma->vm_flags & VM_EXEC)
334                         flush_user_icache_page(vmaddr);
335         } else {
336                 flush_user_cache_page_non_current(vma, vmaddr);
337         }
338 }
339
340 void flush_dcache_page(struct page *page)
341 {
342         struct address_space *mapping = page_mapping(page);
343         struct vm_area_struct *mpnt;
344         struct prio_tree_iter iter;
345         unsigned long offset;
346         unsigned long addr;
347         pgoff_t pgoff;
348         unsigned long pfn = page_to_pfn(page);
349
350
351         if (mapping && !mapping_mapped(mapping)) {
352                 set_bit(PG_dcache_dirty, &page->flags);
353                 return;
354         }
355
356         flush_kernel_dcache_page(page);
357
358         if (!mapping)
359                 return;
360
361         pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
362
363         /* We have carefully arranged in arch_get_unmapped_area() that
364          * *any* mappings of a file are always congruently mapped (whether
365          * declared as MAP_PRIVATE or MAP_SHARED), so we only need
366          * to flush one address here for them all to become coherent */
367
368         flush_dcache_mmap_lock(mapping);
369         vma_prio_tree_foreach(mpnt, &iter, &mapping->i_mmap, pgoff, pgoff) {
370                 offset = (pgoff - mpnt->vm_pgoff) << PAGE_SHIFT;
371                 addr = mpnt->vm_start + offset;
372
373                 /* Flush instructions produce non access tlb misses.
374                  * On PA, we nullify these instructions rather than
375                  * taking a page fault if the pte doesn't exist.
376                  * This is just for speed.  If the page translation
377                  * isn't there, there's no point exciting the
378                  * nadtlb handler into a nullification frenzy.
379                  *
380                  * Make sure we really have this page: the private
381                  * mappings may cover this area but have COW'd this
382                  * particular page.
383                  */
384                 if (translation_exists(mpnt, addr, pfn)) {
385                         __flush_cache_page(mpnt, addr);
386                         break;
387                 }
388         }
389         flush_dcache_mmap_unlock(mapping);
390 }
391 EXPORT_SYMBOL(flush_dcache_page);
392
393 /* Defined in arch/parisc/kernel/pacache.S */
394 EXPORT_SYMBOL(flush_kernel_dcache_range_asm);
395 EXPORT_SYMBOL(flush_kernel_dcache_page_asm);
396 EXPORT_SYMBOL(flush_data_cache_local);
397 EXPORT_SYMBOL(flush_kernel_icache_range_asm);
398
399 void clear_user_page_asm(void *page, unsigned long vaddr)
400 {
401         /* This function is implemented in assembly in pacache.S */
402         extern void __clear_user_page_asm(void *page, unsigned long vaddr);
403
404         purge_tlb_start();
405         __clear_user_page_asm(page, vaddr);
406         purge_tlb_end();
407 }
408
409 #define FLUSH_THRESHOLD 0x80000 /* 0.5MB */
410 int parisc_cache_flush_threshold __read_mostly = FLUSH_THRESHOLD;
411
412 void __init parisc_setup_cache_timing(void)
413 {
414         unsigned long rangetime, alltime;
415         unsigned long size;
416
417         alltime = mfctl(16);
418         flush_data_cache();
419         alltime = mfctl(16) - alltime;
420
421         size = (unsigned long)(_end - _text);
422         rangetime = mfctl(16);
423         flush_kernel_dcache_range((unsigned long)_text, size);
424         rangetime = mfctl(16) - rangetime;
425
426         printk(KERN_DEBUG "Whole cache flush %lu cycles, flushing %lu bytes %lu cycles\n",
427                 alltime, size, rangetime);
428
429         /* Racy, but if we see an intermediate value, it's ok too... */
430         parisc_cache_flush_threshold = size * alltime / rangetime;
431
432         parisc_cache_flush_threshold = (parisc_cache_flush_threshold + L1_CACHE_BYTES - 1) &~ (L1_CACHE_BYTES - 1); 
433         if (!parisc_cache_flush_threshold)
434                 parisc_cache_flush_threshold = FLUSH_THRESHOLD;
435
436         if (parisc_cache_flush_threshold > cache_info.dc_size)
437                 parisc_cache_flush_threshold = cache_info.dc_size;
438
439         printk(KERN_INFO "Setting cache flush threshold to %x (%d CPUs online)\n", parisc_cache_flush_threshold, num_online_cpus());
440 }
441
442 extern void purge_kernel_dcache_page(unsigned long);
443 extern void clear_user_page_asm(void *page, unsigned long vaddr);
444
445 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
446 {
447         purge_kernel_dcache_page((unsigned long)page);
448         purge_tlb_start();
449         pdtlb_kernel(page);
450         purge_tlb_end();
451         clear_user_page_asm(page, vaddr);
452 }
453 EXPORT_SYMBOL(clear_user_page);
454
455 void flush_kernel_dcache_page_addr(void *addr)
456 {
457         flush_kernel_dcache_page_asm(addr);
458         purge_tlb_start();
459         pdtlb_kernel(addr);
460         purge_tlb_end();
461 }
462 EXPORT_SYMBOL(flush_kernel_dcache_page_addr);
463
464 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
465                     struct page *pg)
466 {
467         /* no coherency needed (all in kmap/kunmap) */
468         copy_user_page_asm(vto, vfrom);
469         if (!parisc_requires_coherency())
470                 flush_kernel_dcache_page_asm(vto);
471 }
472 EXPORT_SYMBOL(copy_user_page);
473
474 #ifdef CONFIG_PA8X00
475
476 void kunmap_parisc(void *addr)
477 {
478         if (parisc_requires_coherency())
479                 flush_kernel_dcache_page_addr(addr);
480 }
481 EXPORT_SYMBOL(kunmap_parisc);
482 #endif
483
484 void __flush_tlb_range(unsigned long sid, unsigned long start,
485                        unsigned long end)
486 {
487         unsigned long npages;
488
489         npages = ((end - (start & PAGE_MASK)) + (PAGE_SIZE - 1)) >> PAGE_SHIFT;
490         if (npages >= 512)  /* 2MB of space: arbitrary, should be tuned */
491                 flush_tlb_all();
492         else {
493                 mtsp(sid, 1);
494                 purge_tlb_start();
495                 if (split_tlb) {
496                         while (npages--) {
497                                 pdtlb(start);
498                                 pitlb(start);
499                                 start += PAGE_SIZE;
500                         }
501                 } else {
502                         while (npages--) {
503                                 pdtlb(start);
504                                 start += PAGE_SIZE;
505                         }
506                 }
507                 purge_tlb_end();
508         }
509 }
510
511 static void cacheflush_h_tmp_function(void *dummy)
512 {
513         flush_cache_all_local();
514 }
515
516 void flush_cache_all(void)
517 {
518         on_each_cpu(cacheflush_h_tmp_function, NULL, 1);
519 }
520
521 void flush_cache_mm(struct mm_struct *mm)
522 {
523 #ifdef CONFIG_SMP
524         flush_cache_all();
525 #else
526         flush_cache_all_local();
527 #endif
528 }
529
530 void
531 flush_user_dcache_range(unsigned long start, unsigned long end)
532 {
533         if ((end - start) < parisc_cache_flush_threshold)
534                 flush_user_dcache_range_asm(start,end);
535         else
536                 flush_data_cache();
537 }
538
539 void
540 flush_user_icache_range(unsigned long start, unsigned long end)
541 {
542         if ((end - start) < parisc_cache_flush_threshold)
543                 flush_user_icache_range_asm(start,end);
544         else
545                 flush_instruction_cache();
546 }
547
548
549 void flush_cache_range(struct vm_area_struct *vma,
550                 unsigned long start, unsigned long end)
551 {
552         int sr3;
553
554         if (!vma->vm_mm->context) {
555                 BUG();
556                 return;
557         }
558
559         sr3 = mfsp(3);
560         if (vma->vm_mm->context == sr3) {
561                 flush_user_dcache_range(start,end);
562                 flush_user_icache_range(start,end);
563         } else {
564                 flush_cache_all();
565         }
566 }
567
568 void
569 flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr, unsigned long pfn)
570 {
571         BUG_ON(!vma->vm_mm->context);
572
573         if (likely(translation_exists(vma, vmaddr, pfn)))
574                 __flush_cache_page(vma, vmaddr);
575
576 }