Loading...
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 1999-2006 Helge Deller <deller@gmx.de> (07-13-1999)
7 * Copyright (C) 1999 SuSE GmbH Nuernberg
8 * Copyright (C) 2000 Philipp Rumpf (prumpf@tux.org)
9 *
10 * Cache and TLB management
11 *
12 */
13
14#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/mm.h>
17#include <linux/module.h>
18#include <linux/seq_file.h>
19#include <linux/pagemap.h>
20#include <linux/sched.h>
21#include <linux/sched/mm.h>
22#include <asm/pdc.h>
23#include <asm/cache.h>
24#include <asm/cacheflush.h>
25#include <asm/tlbflush.h>
26#include <asm/page.h>
27#include <asm/processor.h>
28#include <asm/sections.h>
29#include <asm/shmparam.h>
30
31int split_tlb __ro_after_init;
32int dcache_stride __ro_after_init;
33int icache_stride __ro_after_init;
34EXPORT_SYMBOL(dcache_stride);
35
36void flush_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
37EXPORT_SYMBOL(flush_dcache_page_asm);
38void purge_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
39void flush_icache_page_asm(unsigned long phys_addr, unsigned long vaddr);
40
41
42/* On some machines (i.e., ones with the Merced bus), there can be
43 * only a single PxTLB broadcast at a time; this must be guaranteed
44 * by software. We need a spinlock around all TLB flushes to ensure
45 * this.
46 */
47DEFINE_SPINLOCK(pa_tlb_flush_lock);
48
49/* Swapper page setup lock. */
50DEFINE_SPINLOCK(pa_swapper_pg_lock);
51
52#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
53int pa_serialize_tlb_flushes __ro_after_init;
54#endif
55
56struct pdc_cache_info cache_info __ro_after_init;
57#ifndef CONFIG_PA20
58static struct pdc_btlb_info btlb_info __ro_after_init;
59#endif
60
61#ifdef CONFIG_SMP
62void
63flush_data_cache(void)
64{
65 on_each_cpu(flush_data_cache_local, NULL, 1);
66}
67void
68flush_instruction_cache(void)
69{
70 on_each_cpu(flush_instruction_cache_local, NULL, 1);
71}
72#endif
73
74void
75flush_cache_all_local(void)
76{
77 flush_instruction_cache_local(NULL);
78 flush_data_cache_local(NULL);
79}
80EXPORT_SYMBOL(flush_cache_all_local);
81
82/* Virtual address of pfn. */
83#define pfn_va(pfn) __va(PFN_PHYS(pfn))
84
85void
86update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
87{
88 unsigned long pfn = pte_pfn(*ptep);
89 struct page *page;
90
91 /* We don't have pte special. As a result, we can be called with
92 an invalid pfn and we don't need to flush the kernel dcache page.
93 This occurs with FireGL card in C8000. */
94 if (!pfn_valid(pfn))
95 return;
96
97 page = pfn_to_page(pfn);
98 if (page_mapping_file(page) &&
99 test_bit(PG_dcache_dirty, &page->flags)) {
100 flush_kernel_dcache_page_addr(pfn_va(pfn));
101 clear_bit(PG_dcache_dirty, &page->flags);
102 } else if (parisc_requires_coherency())
103 flush_kernel_dcache_page_addr(pfn_va(pfn));
104}
105
106void
107show_cache_info(struct seq_file *m)
108{
109 char buf[32];
110
111 seq_printf(m, "I-cache\t\t: %ld KB\n",
112 cache_info.ic_size/1024 );
113 if (cache_info.dc_loop != 1)
114 snprintf(buf, 32, "%lu-way associative", cache_info.dc_loop);
115 seq_printf(m, "D-cache\t\t: %ld KB (%s%s, %s)\n",
116 cache_info.dc_size/1024,
117 (cache_info.dc_conf.cc_wt ? "WT":"WB"),
118 (cache_info.dc_conf.cc_sh ? ", shared I/D":""),
119 ((cache_info.dc_loop == 1) ? "direct mapped" : buf));
120 seq_printf(m, "ITLB entries\t: %ld\n" "DTLB entries\t: %ld%s\n",
121 cache_info.it_size,
122 cache_info.dt_size,
123 cache_info.dt_conf.tc_sh ? " - shared with ITLB":""
124 );
125
126#ifndef CONFIG_PA20
127 /* BTLB - Block TLB */
128 if (btlb_info.max_size==0) {
129 seq_printf(m, "BTLB\t\t: not supported\n" );
130 } else {
131 seq_printf(m,
132 "BTLB fixed\t: max. %d pages, pagesize=%d (%dMB)\n"
133 "BTLB fix-entr.\t: %d instruction, %d data (%d combined)\n"
134 "BTLB var-entr.\t: %d instruction, %d data (%d combined)\n",
135 btlb_info.max_size, (int)4096,
136 btlb_info.max_size>>8,
137 btlb_info.fixed_range_info.num_i,
138 btlb_info.fixed_range_info.num_d,
139 btlb_info.fixed_range_info.num_comb,
140 btlb_info.variable_range_info.num_i,
141 btlb_info.variable_range_info.num_d,
142 btlb_info.variable_range_info.num_comb
143 );
144 }
145#endif
146}
147
148void __init
149parisc_cache_init(void)
150{
151 if (pdc_cache_info(&cache_info) < 0)
152 panic("parisc_cache_init: pdc_cache_info failed");
153
154#if 0
155 printk("ic_size %lx dc_size %lx it_size %lx\n",
156 cache_info.ic_size,
157 cache_info.dc_size,
158 cache_info.it_size);
159
160 printk("DC base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
161 cache_info.dc_base,
162 cache_info.dc_stride,
163 cache_info.dc_count,
164 cache_info.dc_loop);
165
166 printk("dc_conf = 0x%lx alias %d blk %d line %d shift %d\n",
167 *(unsigned long *) (&cache_info.dc_conf),
168 cache_info.dc_conf.cc_alias,
169 cache_info.dc_conf.cc_block,
170 cache_info.dc_conf.cc_line,
171 cache_info.dc_conf.cc_shift);
172 printk(" wt %d sh %d cst %d hv %d\n",
173 cache_info.dc_conf.cc_wt,
174 cache_info.dc_conf.cc_sh,
175 cache_info.dc_conf.cc_cst,
176 cache_info.dc_conf.cc_hv);
177
178 printk("IC base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
179 cache_info.ic_base,
180 cache_info.ic_stride,
181 cache_info.ic_count,
182 cache_info.ic_loop);
183
184 printk("IT base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx off_base 0x%lx off_stride 0x%lx off_count 0x%lx\n",
185 cache_info.it_sp_base,
186 cache_info.it_sp_stride,
187 cache_info.it_sp_count,
188 cache_info.it_loop,
189 cache_info.it_off_base,
190 cache_info.it_off_stride,
191 cache_info.it_off_count);
192
193 printk("DT base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx off_base 0x%lx off_stride 0x%lx off_count 0x%lx\n",
194 cache_info.dt_sp_base,
195 cache_info.dt_sp_stride,
196 cache_info.dt_sp_count,
197 cache_info.dt_loop,
198 cache_info.dt_off_base,
199 cache_info.dt_off_stride,
200 cache_info.dt_off_count);
201
202 printk("ic_conf = 0x%lx alias %d blk %d line %d shift %d\n",
203 *(unsigned long *) (&cache_info.ic_conf),
204 cache_info.ic_conf.cc_alias,
205 cache_info.ic_conf.cc_block,
206 cache_info.ic_conf.cc_line,
207 cache_info.ic_conf.cc_shift);
208 printk(" wt %d sh %d cst %d hv %d\n",
209 cache_info.ic_conf.cc_wt,
210 cache_info.ic_conf.cc_sh,
211 cache_info.ic_conf.cc_cst,
212 cache_info.ic_conf.cc_hv);
213
214 printk("D-TLB conf: sh %d page %d cst %d aid %d sr %d\n",
215 cache_info.dt_conf.tc_sh,
216 cache_info.dt_conf.tc_page,
217 cache_info.dt_conf.tc_cst,
218 cache_info.dt_conf.tc_aid,
219 cache_info.dt_conf.tc_sr);
220
221 printk("I-TLB conf: sh %d page %d cst %d aid %d sr %d\n",
222 cache_info.it_conf.tc_sh,
223 cache_info.it_conf.tc_page,
224 cache_info.it_conf.tc_cst,
225 cache_info.it_conf.tc_aid,
226 cache_info.it_conf.tc_sr);
227#endif
228
229 split_tlb = 0;
230 if (cache_info.dt_conf.tc_sh == 0 || cache_info.dt_conf.tc_sh == 2) {
231 if (cache_info.dt_conf.tc_sh == 2)
232 printk(KERN_WARNING "Unexpected TLB configuration. "
233 "Will flush I/D separately (could be optimized).\n");
234
235 split_tlb = 1;
236 }
237
238 /* "New and Improved" version from Jim Hull
239 * (1 << (cc_block-1)) * (cc_line << (4 + cnf.cc_shift))
240 * The following CAFL_STRIDE is an optimized version, see
241 * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023625.html
242 * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023671.html
243 */
244#define CAFL_STRIDE(cnf) (cnf.cc_line << (3 + cnf.cc_block + cnf.cc_shift))
245 dcache_stride = CAFL_STRIDE(cache_info.dc_conf);
246 icache_stride = CAFL_STRIDE(cache_info.ic_conf);
247#undef CAFL_STRIDE
248
249#ifndef CONFIG_PA20
250 if (pdc_btlb_info(&btlb_info) < 0) {
251 memset(&btlb_info, 0, sizeof btlb_info);
252 }
253#endif
254
255 if ((boot_cpu_data.pdc.capabilities & PDC_MODEL_NVA_MASK) ==
256 PDC_MODEL_NVA_UNSUPPORTED) {
257 printk(KERN_WARNING "parisc_cache_init: Only equivalent aliasing supported!\n");
258#if 0
259 panic("SMP kernel required to avoid non-equivalent aliasing");
260#endif
261 }
262}
263
264void __init disable_sr_hashing(void)
265{
266 int srhash_type, retval;
267 unsigned long space_bits;
268
269 switch (boot_cpu_data.cpu_type) {
270 case pcx: /* We shouldn't get this far. setup.c should prevent it. */
271 BUG();
272 return;
273
274 case pcxs:
275 case pcxt:
276 case pcxt_:
277 srhash_type = SRHASH_PCXST;
278 break;
279
280 case pcxl:
281 srhash_type = SRHASH_PCXL;
282 break;
283
284 case pcxl2: /* pcxl2 doesn't support space register hashing */
285 return;
286
287 default: /* Currently all PA2.0 machines use the same ins. sequence */
288 srhash_type = SRHASH_PA20;
289 break;
290 }
291
292 disable_sr_hashing_asm(srhash_type);
293
294 retval = pdc_spaceid_bits(&space_bits);
295 /* If this procedure isn't implemented, don't panic. */
296 if (retval < 0 && retval != PDC_BAD_OPTION)
297 panic("pdc_spaceid_bits call failed.\n");
298 if (space_bits != 0)
299 panic("SpaceID hashing is still on!\n");
300}
301
302static inline void
303__flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr,
304 unsigned long physaddr)
305{
306 preempt_disable();
307 flush_dcache_page_asm(physaddr, vmaddr);
308 if (vma->vm_flags & VM_EXEC)
309 flush_icache_page_asm(physaddr, vmaddr);
310 preempt_enable();
311}
312
313static inline void
314__purge_cache_page(struct vm_area_struct *vma, unsigned long vmaddr,
315 unsigned long physaddr)
316{
317 preempt_disable();
318 purge_dcache_page_asm(physaddr, vmaddr);
319 if (vma->vm_flags & VM_EXEC)
320 flush_icache_page_asm(physaddr, vmaddr);
321 preempt_enable();
322}
323
324void flush_dcache_page(struct page *page)
325{
326 struct address_space *mapping = page_mapping_file(page);
327 struct vm_area_struct *mpnt;
328 unsigned long offset;
329 unsigned long addr, old_addr = 0;
330 pgoff_t pgoff;
331
332 if (mapping && !mapping_mapped(mapping)) {
333 set_bit(PG_dcache_dirty, &page->flags);
334 return;
335 }
336
337 flush_kernel_dcache_page(page);
338
339 if (!mapping)
340 return;
341
342 pgoff = page->index;
343
344 /* We have carefully arranged in arch_get_unmapped_area() that
345 * *any* mappings of a file are always congruently mapped (whether
346 * declared as MAP_PRIVATE or MAP_SHARED), so we only need
347 * to flush one address here for them all to become coherent */
348
349 flush_dcache_mmap_lock(mapping);
350 vma_interval_tree_foreach(mpnt, &mapping->i_mmap, pgoff, pgoff) {
351 offset = (pgoff - mpnt->vm_pgoff) << PAGE_SHIFT;
352 addr = mpnt->vm_start + offset;
353
354 /* The TLB is the engine of coherence on parisc: The
355 * CPU is entitled to speculate any page with a TLB
356 * mapping, so here we kill the mapping then flush the
357 * page along a special flush only alias mapping.
358 * This guarantees that the page is no-longer in the
359 * cache for any process and nor may it be
360 * speculatively read in (until the user or kernel
361 * specifically accesses it, of course) */
362
363 flush_tlb_page(mpnt, addr);
364 if (old_addr == 0 || (old_addr & (SHM_COLOUR - 1))
365 != (addr & (SHM_COLOUR - 1))) {
366 __flush_cache_page(mpnt, addr, page_to_phys(page));
367 if (parisc_requires_coherency() && old_addr)
368 printk(KERN_ERR "INEQUIVALENT ALIASES 0x%lx and 0x%lx in file %pD\n", old_addr, addr, mpnt->vm_file);
369 old_addr = addr;
370 }
371 }
372 flush_dcache_mmap_unlock(mapping);
373}
374EXPORT_SYMBOL(flush_dcache_page);
375
376/* Defined in arch/parisc/kernel/pacache.S */
377EXPORT_SYMBOL(flush_kernel_dcache_range_asm);
378EXPORT_SYMBOL(flush_kernel_dcache_page_asm);
379EXPORT_SYMBOL(flush_data_cache_local);
380EXPORT_SYMBOL(flush_kernel_icache_range_asm);
381
382#define FLUSH_THRESHOLD 0x80000 /* 0.5MB */
383static unsigned long parisc_cache_flush_threshold __ro_after_init = FLUSH_THRESHOLD;
384
385#define FLUSH_TLB_THRESHOLD (16*1024) /* 16 KiB minimum TLB threshold */
386static unsigned long parisc_tlb_flush_threshold __ro_after_init = FLUSH_TLB_THRESHOLD;
387
388void __init parisc_setup_cache_timing(void)
389{
390 unsigned long rangetime, alltime;
391 unsigned long size, start;
392 unsigned long threshold;
393
394 alltime = mfctl(16);
395 flush_data_cache();
396 alltime = mfctl(16) - alltime;
397
398 size = (unsigned long)(_end - _text);
399 rangetime = mfctl(16);
400 flush_kernel_dcache_range((unsigned long)_text, size);
401 rangetime = mfctl(16) - rangetime;
402
403 printk(KERN_DEBUG "Whole cache flush %lu cycles, flushing %lu bytes %lu cycles\n",
404 alltime, size, rangetime);
405
406 threshold = L1_CACHE_ALIGN(size * alltime / rangetime);
407 if (threshold > cache_info.dc_size)
408 threshold = cache_info.dc_size;
409 if (threshold)
410 parisc_cache_flush_threshold = threshold;
411 printk(KERN_INFO "Cache flush threshold set to %lu KiB\n",
412 parisc_cache_flush_threshold/1024);
413
414 /* calculate TLB flush threshold */
415
416 /* On SMP machines, skip the TLB measure of kernel text which
417 * has been mapped as huge pages. */
418 if (num_online_cpus() > 1 && !parisc_requires_coherency()) {
419 threshold = max(cache_info.it_size, cache_info.dt_size);
420 threshold *= PAGE_SIZE;
421 threshold /= num_online_cpus();
422 goto set_tlb_threshold;
423 }
424
425 size = 0;
426 start = (unsigned long) _text;
427 rangetime = mfctl(16);
428 while (start < (unsigned long) _end) {
429 flush_tlb_kernel_range(start, start + PAGE_SIZE);
430 start += PAGE_SIZE;
431 size += PAGE_SIZE;
432 }
433 rangetime = mfctl(16) - rangetime;
434
435 alltime = mfctl(16);
436 flush_tlb_all();
437 alltime = mfctl(16) - alltime;
438
439 printk(KERN_INFO "Whole TLB flush %lu cycles, Range flush %lu bytes %lu cycles\n",
440 alltime, size, rangetime);
441
442 threshold = PAGE_ALIGN((num_online_cpus() * size * alltime) / rangetime);
443 printk(KERN_INFO "Calculated TLB flush threshold %lu KiB\n",
444 threshold/1024);
445
446set_tlb_threshold:
447 if (threshold > parisc_tlb_flush_threshold)
448 parisc_tlb_flush_threshold = threshold;
449 printk(KERN_INFO "TLB flush threshold set to %lu KiB\n",
450 parisc_tlb_flush_threshold/1024);
451}
452
453extern void purge_kernel_dcache_page_asm(unsigned long);
454extern void clear_user_page_asm(void *, unsigned long);
455extern void copy_user_page_asm(void *, void *, unsigned long);
456
457void flush_kernel_dcache_page_addr(void *addr)
458{
459 unsigned long flags;
460
461 flush_kernel_dcache_page_asm(addr);
462 purge_tlb_start(flags);
463 pdtlb_kernel(addr);
464 purge_tlb_end(flags);
465}
466EXPORT_SYMBOL(flush_kernel_dcache_page_addr);
467
468void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
469 struct page *pg)
470{
471 /* Copy using kernel mapping. No coherency is needed (all in
472 kunmap) for the `to' page. However, the `from' page needs to
473 be flushed through a mapping equivalent to the user mapping
474 before it can be accessed through the kernel mapping. */
475 preempt_disable();
476 flush_dcache_page_asm(__pa(vfrom), vaddr);
477 copy_page_asm(vto, vfrom);
478 preempt_enable();
479}
480EXPORT_SYMBOL(copy_user_page);
481
482/* __flush_tlb_range()
483 *
484 * returns 1 if all TLBs were flushed.
485 */
486int __flush_tlb_range(unsigned long sid, unsigned long start,
487 unsigned long end)
488{
489 unsigned long flags;
490
491 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
492 end - start >= parisc_tlb_flush_threshold) {
493 flush_tlb_all();
494 return 1;
495 }
496
497 /* Purge TLB entries for small ranges using the pdtlb and
498 pitlb instructions. These instructions execute locally
499 but cause a purge request to be broadcast to other TLBs. */
500 while (start < end) {
501 purge_tlb_start(flags);
502 mtsp(sid, 1);
503 pdtlb(start);
504 pitlb(start);
505 purge_tlb_end(flags);
506 start += PAGE_SIZE;
507 }
508 return 0;
509}
510
511static void cacheflush_h_tmp_function(void *dummy)
512{
513 flush_cache_all_local();
514}
515
516void flush_cache_all(void)
517{
518 on_each_cpu(cacheflush_h_tmp_function, NULL, 1);
519}
520
521static inline unsigned long mm_total_size(struct mm_struct *mm)
522{
523 struct vm_area_struct *vma;
524 unsigned long usize = 0;
525
526 for (vma = mm->mmap; vma; vma = vma->vm_next)
527 usize += vma->vm_end - vma->vm_start;
528 return usize;
529}
530
531static inline pte_t *get_ptep(pgd_t *pgd, unsigned long addr)
532{
533 pte_t *ptep = NULL;
534
535 if (!pgd_none(*pgd)) {
536 p4d_t *p4d = p4d_offset(pgd, addr);
537 if (!p4d_none(*p4d)) {
538 pud_t *pud = pud_offset(p4d, addr);
539 if (!pud_none(*pud)) {
540 pmd_t *pmd = pmd_offset(pud, addr);
541 if (!pmd_none(*pmd))
542 ptep = pte_offset_map(pmd, addr);
543 }
544 }
545 }
546 return ptep;
547}
548
549void flush_cache_mm(struct mm_struct *mm)
550{
551 struct vm_area_struct *vma;
552 pgd_t *pgd;
553
554 /* Flushing the whole cache on each cpu takes forever on
555 rp3440, etc. So, avoid it if the mm isn't too big. */
556 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
557 mm_total_size(mm) >= parisc_cache_flush_threshold) {
558 if (mm->context)
559 flush_tlb_all();
560 flush_cache_all();
561 return;
562 }
563
564 if (mm->context == mfsp(3)) {
565 for (vma = mm->mmap; vma; vma = vma->vm_next) {
566 flush_user_dcache_range_asm(vma->vm_start, vma->vm_end);
567 if (vma->vm_flags & VM_EXEC)
568 flush_user_icache_range_asm(vma->vm_start, vma->vm_end);
569 flush_tlb_range(vma, vma->vm_start, vma->vm_end);
570 }
571 return;
572 }
573
574 pgd = mm->pgd;
575 for (vma = mm->mmap; vma; vma = vma->vm_next) {
576 unsigned long addr;
577
578 for (addr = vma->vm_start; addr < vma->vm_end;
579 addr += PAGE_SIZE) {
580 unsigned long pfn;
581 pte_t *ptep = get_ptep(pgd, addr);
582 if (!ptep)
583 continue;
584 pfn = pte_pfn(*ptep);
585 if (!pfn_valid(pfn))
586 continue;
587 if (unlikely(mm->context)) {
588 flush_tlb_page(vma, addr);
589 __flush_cache_page(vma, addr, PFN_PHYS(pfn));
590 } else {
591 __purge_cache_page(vma, addr, PFN_PHYS(pfn));
592 }
593 }
594 }
595}
596
597void flush_cache_range(struct vm_area_struct *vma,
598 unsigned long start, unsigned long end)
599{
600 pgd_t *pgd;
601 unsigned long addr;
602
603 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
604 end - start >= parisc_cache_flush_threshold) {
605 if (vma->vm_mm->context)
606 flush_tlb_range(vma, start, end);
607 flush_cache_all();
608 return;
609 }
610
611 if (vma->vm_mm->context == mfsp(3)) {
612 flush_user_dcache_range_asm(start, end);
613 if (vma->vm_flags & VM_EXEC)
614 flush_user_icache_range_asm(start, end);
615 flush_tlb_range(vma, start, end);
616 return;
617 }
618
619 pgd = vma->vm_mm->pgd;
620 for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
621 unsigned long pfn;
622 pte_t *ptep = get_ptep(pgd, addr);
623 if (!ptep)
624 continue;
625 pfn = pte_pfn(*ptep);
626 if (pfn_valid(pfn)) {
627 if (unlikely(vma->vm_mm->context)) {
628 flush_tlb_page(vma, addr);
629 __flush_cache_page(vma, addr, PFN_PHYS(pfn));
630 } else {
631 __purge_cache_page(vma, addr, PFN_PHYS(pfn));
632 }
633 }
634 }
635}
636
637void
638flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr, unsigned long pfn)
639{
640 if (pfn_valid(pfn)) {
641 if (likely(vma->vm_mm->context)) {
642 flush_tlb_page(vma, vmaddr);
643 __flush_cache_page(vma, vmaddr, PFN_PHYS(pfn));
644 } else {
645 __purge_cache_page(vma, vmaddr, PFN_PHYS(pfn));
646 }
647 }
648}
649
650void flush_kernel_vmap_range(void *vaddr, int size)
651{
652 unsigned long start = (unsigned long)vaddr;
653 unsigned long end = start + size;
654
655 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
656 (unsigned long)size >= parisc_cache_flush_threshold) {
657 flush_tlb_kernel_range(start, end);
658 flush_data_cache();
659 return;
660 }
661
662 flush_kernel_dcache_range_asm(start, end);
663 flush_tlb_kernel_range(start, end);
664}
665EXPORT_SYMBOL(flush_kernel_vmap_range);
666
667void invalidate_kernel_vmap_range(void *vaddr, int size)
668{
669 unsigned long start = (unsigned long)vaddr;
670 unsigned long end = start + size;
671
672 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
673 (unsigned long)size >= parisc_cache_flush_threshold) {
674 flush_tlb_kernel_range(start, end);
675 flush_data_cache();
676 return;
677 }
678
679 purge_kernel_dcache_range_asm(start, end);
680 flush_tlb_kernel_range(start, end);
681}
682EXPORT_SYMBOL(invalidate_kernel_vmap_range);
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 1999-2006 Helge Deller <deller@gmx.de> (07-13-1999)
7 * Copyright (C) 1999 SuSE GmbH Nuernberg
8 * Copyright (C) 2000 Philipp Rumpf (prumpf@tux.org)
9 *
10 * Cache and TLB management
11 *
12 */
13
14#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/mm.h>
17#include <linux/module.h>
18#include <linux/seq_file.h>
19#include <linux/pagemap.h>
20#include <linux/sched.h>
21#include <linux/sched/mm.h>
22#include <linux/syscalls.h>
23#include <linux/vmalloc.h>
24#include <asm/pdc.h>
25#include <asm/cache.h>
26#include <asm/cacheflush.h>
27#include <asm/tlbflush.h>
28#include <asm/page.h>
29#include <asm/processor.h>
30#include <asm/sections.h>
31#include <asm/shmparam.h>
32#include <asm/mmu_context.h>
33#include <asm/cachectl.h>
34
35#define PTR_PAGE_ALIGN_DOWN(addr) PTR_ALIGN_DOWN(addr, PAGE_SIZE)
36
37/*
38 * When nonzero, use _PAGE_ACCESSED bit to try to reduce the number
39 * of page flushes done flush_cache_page_if_present. There are some
40 * pros and cons in using this option. It may increase the risk of
41 * random segmentation faults.
42 */
43#define CONFIG_FLUSH_PAGE_ACCESSED 0
44
45int split_tlb __ro_after_init;
46int dcache_stride __ro_after_init;
47int icache_stride __ro_after_init;
48EXPORT_SYMBOL(dcache_stride);
49
50/* Internal implementation in arch/parisc/kernel/pacache.S */
51void flush_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
52EXPORT_SYMBOL(flush_dcache_page_asm);
53void purge_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
54void flush_icache_page_asm(unsigned long phys_addr, unsigned long vaddr);
55void flush_data_cache_local(void *); /* flushes local data-cache only */
56void flush_instruction_cache_local(void); /* flushes local code-cache only */
57
58static void flush_kernel_dcache_page_addr(const void *addr);
59
60/* On some machines (i.e., ones with the Merced bus), there can be
61 * only a single PxTLB broadcast at a time; this must be guaranteed
62 * by software. We need a spinlock around all TLB flushes to ensure
63 * this.
64 */
65DEFINE_SPINLOCK(pa_tlb_flush_lock);
66
67#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
68int pa_serialize_tlb_flushes __ro_after_init;
69#endif
70
71struct pdc_cache_info cache_info __ro_after_init;
72#ifndef CONFIG_PA20
73struct pdc_btlb_info btlb_info;
74#endif
75
76DEFINE_STATIC_KEY_TRUE(parisc_has_cache);
77DEFINE_STATIC_KEY_TRUE(parisc_has_dcache);
78DEFINE_STATIC_KEY_TRUE(parisc_has_icache);
79
80static void cache_flush_local_cpu(void *dummy)
81{
82 if (static_branch_likely(&parisc_has_icache))
83 flush_instruction_cache_local();
84 if (static_branch_likely(&parisc_has_dcache))
85 flush_data_cache_local(NULL);
86}
87
88void flush_cache_all_local(void)
89{
90 cache_flush_local_cpu(NULL);
91}
92
93void flush_cache_all(void)
94{
95 if (static_branch_likely(&parisc_has_cache))
96 on_each_cpu(cache_flush_local_cpu, NULL, 1);
97}
98
99static inline void flush_data_cache(void)
100{
101 if (static_branch_likely(&parisc_has_dcache))
102 on_each_cpu(flush_data_cache_local, NULL, 1);
103}
104
105
106/* Kernel virtual address of pfn. */
107#define pfn_va(pfn) __va(PFN_PHYS(pfn))
108
109void __update_cache(pte_t pte)
110{
111 unsigned long pfn = pte_pfn(pte);
112 struct folio *folio;
113 unsigned int nr;
114
115 /* We don't have pte special. As a result, we can be called with
116 an invalid pfn and we don't need to flush the kernel dcache page.
117 This occurs with FireGL card in C8000. */
118 if (!pfn_valid(pfn))
119 return;
120
121 folio = page_folio(pfn_to_page(pfn));
122 pfn = folio_pfn(folio);
123 nr = folio_nr_pages(folio);
124 if (folio_flush_mapping(folio) &&
125 test_bit(PG_dcache_dirty, &folio->flags)) {
126 while (nr--)
127 flush_kernel_dcache_page_addr(pfn_va(pfn + nr));
128 clear_bit(PG_dcache_dirty, &folio->flags);
129 } else if (parisc_requires_coherency())
130 while (nr--)
131 flush_kernel_dcache_page_addr(pfn_va(pfn + nr));
132}
133
134void
135show_cache_info(struct seq_file *m)
136{
137 char buf[32];
138
139 seq_printf(m, "I-cache\t\t: %ld KB\n",
140 cache_info.ic_size/1024 );
141 if (cache_info.dc_loop != 1)
142 snprintf(buf, 32, "%lu-way associative", cache_info.dc_loop);
143 seq_printf(m, "D-cache\t\t: %ld KB (%s%s, %s, alias=%d)\n",
144 cache_info.dc_size/1024,
145 (cache_info.dc_conf.cc_wt ? "WT":"WB"),
146 (cache_info.dc_conf.cc_sh ? ", shared I/D":""),
147 ((cache_info.dc_loop == 1) ? "direct mapped" : buf),
148 cache_info.dc_conf.cc_alias
149 );
150 seq_printf(m, "ITLB entries\t: %ld\n" "DTLB entries\t: %ld%s\n",
151 cache_info.it_size,
152 cache_info.dt_size,
153 cache_info.dt_conf.tc_sh ? " - shared with ITLB":""
154 );
155
156#ifndef CONFIG_PA20
157 /* BTLB - Block TLB */
158 if (btlb_info.max_size==0) {
159 seq_printf(m, "BTLB\t\t: not supported\n" );
160 } else {
161 seq_printf(m,
162 "BTLB fixed\t: max. %d pages, pagesize=%d (%dMB)\n"
163 "BTLB fix-entr.\t: %d instruction, %d data (%d combined)\n"
164 "BTLB var-entr.\t: %d instruction, %d data (%d combined)\n",
165 btlb_info.max_size, (int)4096,
166 btlb_info.max_size>>8,
167 btlb_info.fixed_range_info.num_i,
168 btlb_info.fixed_range_info.num_d,
169 btlb_info.fixed_range_info.num_comb,
170 btlb_info.variable_range_info.num_i,
171 btlb_info.variable_range_info.num_d,
172 btlb_info.variable_range_info.num_comb
173 );
174 }
175#endif
176}
177
178void __init
179parisc_cache_init(void)
180{
181 if (pdc_cache_info(&cache_info) < 0)
182 panic("parisc_cache_init: pdc_cache_info failed");
183
184#if 0
185 printk("ic_size %lx dc_size %lx it_size %lx\n",
186 cache_info.ic_size,
187 cache_info.dc_size,
188 cache_info.it_size);
189
190 printk("DC base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
191 cache_info.dc_base,
192 cache_info.dc_stride,
193 cache_info.dc_count,
194 cache_info.dc_loop);
195
196 printk("dc_conf = 0x%lx alias %d blk %d line %d shift %d\n",
197 *(unsigned long *) (&cache_info.dc_conf),
198 cache_info.dc_conf.cc_alias,
199 cache_info.dc_conf.cc_block,
200 cache_info.dc_conf.cc_line,
201 cache_info.dc_conf.cc_shift);
202 printk(" wt %d sh %d cst %d hv %d\n",
203 cache_info.dc_conf.cc_wt,
204 cache_info.dc_conf.cc_sh,
205 cache_info.dc_conf.cc_cst,
206 cache_info.dc_conf.cc_hv);
207
208 printk("IC base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
209 cache_info.ic_base,
210 cache_info.ic_stride,
211 cache_info.ic_count,
212 cache_info.ic_loop);
213
214 printk("IT base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx off_base 0x%lx off_stride 0x%lx off_count 0x%lx\n",
215 cache_info.it_sp_base,
216 cache_info.it_sp_stride,
217 cache_info.it_sp_count,
218 cache_info.it_loop,
219 cache_info.it_off_base,
220 cache_info.it_off_stride,
221 cache_info.it_off_count);
222
223 printk("DT base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx off_base 0x%lx off_stride 0x%lx off_count 0x%lx\n",
224 cache_info.dt_sp_base,
225 cache_info.dt_sp_stride,
226 cache_info.dt_sp_count,
227 cache_info.dt_loop,
228 cache_info.dt_off_base,
229 cache_info.dt_off_stride,
230 cache_info.dt_off_count);
231
232 printk("ic_conf = 0x%lx alias %d blk %d line %d shift %d\n",
233 *(unsigned long *) (&cache_info.ic_conf),
234 cache_info.ic_conf.cc_alias,
235 cache_info.ic_conf.cc_block,
236 cache_info.ic_conf.cc_line,
237 cache_info.ic_conf.cc_shift);
238 printk(" wt %d sh %d cst %d hv %d\n",
239 cache_info.ic_conf.cc_wt,
240 cache_info.ic_conf.cc_sh,
241 cache_info.ic_conf.cc_cst,
242 cache_info.ic_conf.cc_hv);
243
244 printk("D-TLB conf: sh %d page %d cst %d aid %d sr %d\n",
245 cache_info.dt_conf.tc_sh,
246 cache_info.dt_conf.tc_page,
247 cache_info.dt_conf.tc_cst,
248 cache_info.dt_conf.tc_aid,
249 cache_info.dt_conf.tc_sr);
250
251 printk("I-TLB conf: sh %d page %d cst %d aid %d sr %d\n",
252 cache_info.it_conf.tc_sh,
253 cache_info.it_conf.tc_page,
254 cache_info.it_conf.tc_cst,
255 cache_info.it_conf.tc_aid,
256 cache_info.it_conf.tc_sr);
257#endif
258
259 split_tlb = 0;
260 if (cache_info.dt_conf.tc_sh == 0 || cache_info.dt_conf.tc_sh == 2) {
261 if (cache_info.dt_conf.tc_sh == 2)
262 printk(KERN_WARNING "Unexpected TLB configuration. "
263 "Will flush I/D separately (could be optimized).\n");
264
265 split_tlb = 1;
266 }
267
268 /* "New and Improved" version from Jim Hull
269 * (1 << (cc_block-1)) * (cc_line << (4 + cnf.cc_shift))
270 * The following CAFL_STRIDE is an optimized version, see
271 * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023625.html
272 * http://lists.parisc-linux.org/pipermail/parisc-linux/2004-June/023671.html
273 */
274#define CAFL_STRIDE(cnf) (cnf.cc_line << (3 + cnf.cc_block + cnf.cc_shift))
275 dcache_stride = CAFL_STRIDE(cache_info.dc_conf);
276 icache_stride = CAFL_STRIDE(cache_info.ic_conf);
277#undef CAFL_STRIDE
278
279 /* stride needs to be non-zero, otherwise cache flushes will not work */
280 WARN_ON(cache_info.dc_size && dcache_stride == 0);
281 WARN_ON(cache_info.ic_size && icache_stride == 0);
282
283 if ((boot_cpu_data.pdc.capabilities & PDC_MODEL_NVA_MASK) ==
284 PDC_MODEL_NVA_UNSUPPORTED) {
285 printk(KERN_WARNING "parisc_cache_init: Only equivalent aliasing supported!\n");
286#if 0
287 panic("SMP kernel required to avoid non-equivalent aliasing");
288#endif
289 }
290}
291
292void disable_sr_hashing(void)
293{
294 int srhash_type, retval;
295 unsigned long space_bits;
296
297 switch (boot_cpu_data.cpu_type) {
298 case pcx: /* We shouldn't get this far. setup.c should prevent it. */
299 BUG();
300 return;
301
302 case pcxs:
303 case pcxt:
304 case pcxt_:
305 srhash_type = SRHASH_PCXST;
306 break;
307
308 case pcxl:
309 srhash_type = SRHASH_PCXL;
310 break;
311
312 case pcxl2: /* pcxl2 doesn't support space register hashing */
313 return;
314
315 default: /* Currently all PA2.0 machines use the same ins. sequence */
316 srhash_type = SRHASH_PA20;
317 break;
318 }
319
320 disable_sr_hashing_asm(srhash_type);
321
322 retval = pdc_spaceid_bits(&space_bits);
323 /* If this procedure isn't implemented, don't panic. */
324 if (retval < 0 && retval != PDC_BAD_OPTION)
325 panic("pdc_spaceid_bits call failed.\n");
326 if (space_bits != 0)
327 panic("SpaceID hashing is still on!\n");
328}
329
330static inline void
331__flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr,
332 unsigned long physaddr)
333{
334 if (!static_branch_likely(&parisc_has_cache))
335 return;
336
337 /*
338 * The TLB is the engine of coherence on parisc. The CPU is
339 * entitled to speculate any page with a TLB mapping, so here
340 * we kill the mapping then flush the page along a special flush
341 * only alias mapping. This guarantees that the page is no-longer
342 * in the cache for any process and nor may it be speculatively
343 * read in (until the user or kernel specifically accesses it,
344 * of course).
345 */
346 flush_tlb_page(vma, vmaddr);
347
348 preempt_disable();
349 flush_dcache_page_asm(physaddr, vmaddr);
350 if (vma->vm_flags & VM_EXEC)
351 flush_icache_page_asm(physaddr, vmaddr);
352 preempt_enable();
353}
354
355static void flush_kernel_dcache_page_addr(const void *addr)
356{
357 unsigned long vaddr = (unsigned long)addr;
358 unsigned long flags;
359
360 /* Purge TLB entry to remove translation on all CPUs */
361 purge_tlb_start(flags);
362 pdtlb(SR_KERNEL, addr);
363 purge_tlb_end(flags);
364
365 /* Use tmpalias flush to prevent data cache move-in */
366 preempt_disable();
367 flush_dcache_page_asm(__pa(vaddr), vaddr);
368 preempt_enable();
369}
370
371static void flush_kernel_icache_page_addr(const void *addr)
372{
373 unsigned long vaddr = (unsigned long)addr;
374 unsigned long flags;
375
376 /* Purge TLB entry to remove translation on all CPUs */
377 purge_tlb_start(flags);
378 pdtlb(SR_KERNEL, addr);
379 purge_tlb_end(flags);
380
381 /* Use tmpalias flush to prevent instruction cache move-in */
382 preempt_disable();
383 flush_icache_page_asm(__pa(vaddr), vaddr);
384 preempt_enable();
385}
386
387void kunmap_flush_on_unmap(const void *addr)
388{
389 flush_kernel_dcache_page_addr(addr);
390}
391EXPORT_SYMBOL(kunmap_flush_on_unmap);
392
393void flush_icache_pages(struct vm_area_struct *vma, struct page *page,
394 unsigned int nr)
395{
396 void *kaddr = page_address(page);
397
398 for (;;) {
399 flush_kernel_dcache_page_addr(kaddr);
400 flush_kernel_icache_page_addr(kaddr);
401 if (--nr == 0)
402 break;
403 kaddr += PAGE_SIZE;
404 }
405}
406
407/*
408 * Walk page directory for MM to find PTEP pointer for address ADDR.
409 */
410static inline pte_t *get_ptep(struct mm_struct *mm, unsigned long addr)
411{
412 pte_t *ptep = NULL;
413 pgd_t *pgd = mm->pgd;
414 p4d_t *p4d;
415 pud_t *pud;
416 pmd_t *pmd;
417
418 if (!pgd_none(*pgd)) {
419 p4d = p4d_offset(pgd, addr);
420 if (!p4d_none(*p4d)) {
421 pud = pud_offset(p4d, addr);
422 if (!pud_none(*pud)) {
423 pmd = pmd_offset(pud, addr);
424 if (!pmd_none(*pmd))
425 ptep = pte_offset_map(pmd, addr);
426 }
427 }
428 }
429 return ptep;
430}
431
432static inline bool pte_needs_flush(pte_t pte)
433{
434 return (pte_val(pte) & (_PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_NO_CACHE))
435 == (_PAGE_PRESENT | _PAGE_ACCESSED);
436}
437
438/*
439 * Return user physical address. Returns 0 if page is not present.
440 */
441static inline unsigned long get_upa(struct mm_struct *mm, unsigned long addr)
442{
443 unsigned long flags, space, pgd, prot, pa;
444#ifdef CONFIG_TLB_PTLOCK
445 unsigned long pgd_lock;
446#endif
447
448 /* Save context */
449 local_irq_save(flags);
450 prot = mfctl(8);
451 space = mfsp(SR_USER);
452 pgd = mfctl(25);
453#ifdef CONFIG_TLB_PTLOCK
454 pgd_lock = mfctl(28);
455#endif
456
457 /* Set context for lpa_user */
458 switch_mm_irqs_off(NULL, mm, NULL);
459 pa = lpa_user(addr);
460
461 /* Restore previous context */
462#ifdef CONFIG_TLB_PTLOCK
463 mtctl(pgd_lock, 28);
464#endif
465 mtctl(pgd, 25);
466 mtsp(space, SR_USER);
467 mtctl(prot, 8);
468 local_irq_restore(flags);
469
470 return pa;
471}
472
473void flush_dcache_folio(struct folio *folio)
474{
475 struct address_space *mapping = folio_flush_mapping(folio);
476 struct vm_area_struct *vma;
477 unsigned long addr, old_addr = 0;
478 void *kaddr;
479 unsigned long count = 0;
480 unsigned long i, nr, flags;
481 pgoff_t pgoff;
482
483 if (mapping && !mapping_mapped(mapping)) {
484 set_bit(PG_dcache_dirty, &folio->flags);
485 return;
486 }
487
488 nr = folio_nr_pages(folio);
489 kaddr = folio_address(folio);
490 for (i = 0; i < nr; i++)
491 flush_kernel_dcache_page_addr(kaddr + i * PAGE_SIZE);
492
493 if (!mapping)
494 return;
495
496 pgoff = folio->index;
497
498 /*
499 * We have carefully arranged in arch_get_unmapped_area() that
500 * *any* mappings of a file are always congruently mapped (whether
501 * declared as MAP_PRIVATE or MAP_SHARED), so we only need
502 * to flush one address here for them all to become coherent
503 * on machines that support equivalent aliasing
504 */
505 flush_dcache_mmap_lock_irqsave(mapping, flags);
506 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff + nr - 1) {
507 unsigned long offset = pgoff - vma->vm_pgoff;
508 unsigned long pfn = folio_pfn(folio);
509
510 addr = vma->vm_start;
511 nr = folio_nr_pages(folio);
512 if (offset > -nr) {
513 pfn -= offset;
514 nr += offset;
515 } else {
516 addr += offset * PAGE_SIZE;
517 }
518 if (addr + nr * PAGE_SIZE > vma->vm_end)
519 nr = (vma->vm_end - addr) / PAGE_SIZE;
520
521 if (old_addr == 0 || (old_addr & (SHM_COLOUR - 1))
522 != (addr & (SHM_COLOUR - 1))) {
523 for (i = 0; i < nr; i++)
524 __flush_cache_page(vma,
525 addr + i * PAGE_SIZE,
526 (pfn + i) * PAGE_SIZE);
527 /*
528 * Software is allowed to have any number
529 * of private mappings to a page.
530 */
531 if (!(vma->vm_flags & VM_SHARED))
532 continue;
533 if (old_addr)
534 pr_err("INEQUIVALENT ALIASES 0x%lx and 0x%lx in file %pD\n",
535 old_addr, addr, vma->vm_file);
536 if (nr == folio_nr_pages(folio))
537 old_addr = addr;
538 }
539 WARN_ON(++count == 4096);
540 }
541 flush_dcache_mmap_unlock_irqrestore(mapping, flags);
542}
543EXPORT_SYMBOL(flush_dcache_folio);
544
545/* Defined in arch/parisc/kernel/pacache.S */
546EXPORT_SYMBOL(flush_kernel_dcache_range_asm);
547EXPORT_SYMBOL(flush_kernel_icache_range_asm);
548
549#define FLUSH_THRESHOLD 0x80000 /* 0.5MB */
550static unsigned long parisc_cache_flush_threshold __ro_after_init = FLUSH_THRESHOLD;
551
552#define FLUSH_TLB_THRESHOLD (16*1024) /* 16 KiB minimum TLB threshold */
553static unsigned long parisc_tlb_flush_threshold __ro_after_init = ~0UL;
554
555void __init parisc_setup_cache_timing(void)
556{
557 unsigned long rangetime, alltime;
558 unsigned long size;
559 unsigned long threshold, threshold2;
560
561 alltime = mfctl(16);
562 flush_data_cache();
563 alltime = mfctl(16) - alltime;
564
565 size = (unsigned long)(_end - _text);
566 rangetime = mfctl(16);
567 flush_kernel_dcache_range((unsigned long)_text, size);
568 rangetime = mfctl(16) - rangetime;
569
570 printk(KERN_DEBUG "Whole cache flush %lu cycles, flushing %lu bytes %lu cycles\n",
571 alltime, size, rangetime);
572
573 threshold = L1_CACHE_ALIGN((unsigned long)((uint64_t)size * alltime / rangetime));
574 pr_info("Calculated flush threshold is %lu KiB\n",
575 threshold/1024);
576
577 /*
578 * The threshold computed above isn't very reliable. The following
579 * heuristic works reasonably well on c8000/rp3440.
580 */
581 threshold2 = cache_info.dc_size * num_online_cpus();
582 parisc_cache_flush_threshold = threshold2;
583 printk(KERN_INFO "Cache flush threshold set to %lu KiB\n",
584 parisc_cache_flush_threshold/1024);
585
586 /* calculate TLB flush threshold */
587
588 /* On SMP machines, skip the TLB measure of kernel text which
589 * has been mapped as huge pages. */
590 if (num_online_cpus() > 1 && !parisc_requires_coherency()) {
591 threshold = max(cache_info.it_size, cache_info.dt_size);
592 threshold *= PAGE_SIZE;
593 threshold /= num_online_cpus();
594 goto set_tlb_threshold;
595 }
596
597 size = (unsigned long)_end - (unsigned long)_text;
598 rangetime = mfctl(16);
599 flush_tlb_kernel_range((unsigned long)_text, (unsigned long)_end);
600 rangetime = mfctl(16) - rangetime;
601
602 alltime = mfctl(16);
603 flush_tlb_all();
604 alltime = mfctl(16) - alltime;
605
606 printk(KERN_INFO "Whole TLB flush %lu cycles, Range flush %lu bytes %lu cycles\n",
607 alltime, size, rangetime);
608
609 threshold = PAGE_ALIGN((num_online_cpus() * size * alltime) / rangetime);
610 printk(KERN_INFO "Calculated TLB flush threshold %lu KiB\n",
611 threshold/1024);
612
613set_tlb_threshold:
614 parisc_tlb_flush_threshold = max(threshold, FLUSH_TLB_THRESHOLD);
615 printk(KERN_INFO "TLB flush threshold set to %lu KiB\n",
616 parisc_tlb_flush_threshold/1024);
617}
618
619extern void purge_kernel_dcache_page_asm(unsigned long);
620extern void clear_user_page_asm(void *, unsigned long);
621extern void copy_user_page_asm(void *, void *, unsigned long);
622
623static void flush_cache_page_if_present(struct vm_area_struct *vma,
624 unsigned long vmaddr)
625{
626#if CONFIG_FLUSH_PAGE_ACCESSED
627 bool needs_flush = false;
628 pte_t *ptep, pte;
629
630 ptep = get_ptep(vma->vm_mm, vmaddr);
631 if (ptep) {
632 pte = ptep_get(ptep);
633 needs_flush = pte_needs_flush(pte);
634 pte_unmap(ptep);
635 }
636 if (needs_flush)
637 __flush_cache_page(vma, vmaddr, PFN_PHYS(pte_pfn(pte)));
638#else
639 struct mm_struct *mm = vma->vm_mm;
640 unsigned long physaddr = get_upa(mm, vmaddr);
641
642 if (physaddr)
643 __flush_cache_page(vma, vmaddr, PAGE_ALIGN_DOWN(physaddr));
644#endif
645}
646
647void copy_user_highpage(struct page *to, struct page *from,
648 unsigned long vaddr, struct vm_area_struct *vma)
649{
650 void *kto, *kfrom;
651
652 kfrom = kmap_local_page(from);
653 kto = kmap_local_page(to);
654 __flush_cache_page(vma, vaddr, PFN_PHYS(page_to_pfn(from)));
655 copy_page_asm(kto, kfrom);
656 kunmap_local(kto);
657 kunmap_local(kfrom);
658}
659
660void copy_to_user_page(struct vm_area_struct *vma, struct page *page,
661 unsigned long user_vaddr, void *dst, void *src, int len)
662{
663 __flush_cache_page(vma, user_vaddr, PFN_PHYS(page_to_pfn(page)));
664 memcpy(dst, src, len);
665 flush_kernel_dcache_page_addr(PTR_PAGE_ALIGN_DOWN(dst));
666}
667
668void copy_from_user_page(struct vm_area_struct *vma, struct page *page,
669 unsigned long user_vaddr, void *dst, void *src, int len)
670{
671 __flush_cache_page(vma, user_vaddr, PFN_PHYS(page_to_pfn(page)));
672 memcpy(dst, src, len);
673 flush_kernel_dcache_page_addr(PTR_PAGE_ALIGN_DOWN(src));
674}
675
676/* __flush_tlb_range()
677 *
678 * returns 1 if all TLBs were flushed.
679 */
680int __flush_tlb_range(unsigned long sid, unsigned long start,
681 unsigned long end)
682{
683 unsigned long flags;
684
685 if ((!IS_ENABLED(CONFIG_SMP) || !arch_irqs_disabled()) &&
686 end - start >= parisc_tlb_flush_threshold) {
687 flush_tlb_all();
688 return 1;
689 }
690
691 /* Purge TLB entries for small ranges using the pdtlb and
692 pitlb instructions. These instructions execute locally
693 but cause a purge request to be broadcast to other TLBs. */
694 while (start < end) {
695 purge_tlb_start(flags);
696 mtsp(sid, SR_TEMP1);
697 pdtlb(SR_TEMP1, start);
698 pitlb(SR_TEMP1, start);
699 purge_tlb_end(flags);
700 start += PAGE_SIZE;
701 }
702 return 0;
703}
704
705static void flush_cache_pages(struct vm_area_struct *vma, unsigned long start, unsigned long end)
706{
707 unsigned long addr;
708
709 for (addr = start; addr < end; addr += PAGE_SIZE)
710 flush_cache_page_if_present(vma, addr);
711}
712
713static inline unsigned long mm_total_size(struct mm_struct *mm)
714{
715 struct vm_area_struct *vma;
716 unsigned long usize = 0;
717 VMA_ITERATOR(vmi, mm, 0);
718
719 for_each_vma(vmi, vma) {
720 if (usize >= parisc_cache_flush_threshold)
721 break;
722 usize += vma->vm_end - vma->vm_start;
723 }
724 return usize;
725}
726
727void flush_cache_mm(struct mm_struct *mm)
728{
729 struct vm_area_struct *vma;
730 VMA_ITERATOR(vmi, mm, 0);
731
732 /*
733 * Flushing the whole cache on each cpu takes forever on
734 * rp3440, etc. So, avoid it if the mm isn't too big.
735 *
736 * Note that we must flush the entire cache on machines
737 * with aliasing caches to prevent random segmentation
738 * faults.
739 */
740 if (!parisc_requires_coherency()
741 || mm_total_size(mm) >= parisc_cache_flush_threshold) {
742 if (WARN_ON(IS_ENABLED(CONFIG_SMP) && arch_irqs_disabled()))
743 return;
744 flush_tlb_all();
745 flush_cache_all();
746 return;
747 }
748
749 /* Flush mm */
750 for_each_vma(vmi, vma)
751 flush_cache_pages(vma, vma->vm_start, vma->vm_end);
752}
753
754void flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
755{
756 if (!parisc_requires_coherency()
757 || end - start >= parisc_cache_flush_threshold) {
758 if (WARN_ON(IS_ENABLED(CONFIG_SMP) && arch_irqs_disabled()))
759 return;
760 flush_tlb_range(vma, start, end);
761 if (vma->vm_flags & VM_EXEC)
762 flush_cache_all();
763 else
764 flush_data_cache();
765 return;
766 }
767
768 flush_cache_pages(vma, start & PAGE_MASK, end);
769}
770
771void flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr, unsigned long pfn)
772{
773 __flush_cache_page(vma, vmaddr, PFN_PHYS(pfn));
774}
775
776void flush_anon_page(struct vm_area_struct *vma, struct page *page, unsigned long vmaddr)
777{
778 if (!PageAnon(page))
779 return;
780
781 __flush_cache_page(vma, vmaddr, PFN_PHYS(page_to_pfn(page)));
782}
783
784int ptep_clear_flush_young(struct vm_area_struct *vma, unsigned long addr,
785 pte_t *ptep)
786{
787 pte_t pte = ptep_get(ptep);
788
789 if (!pte_young(pte))
790 return 0;
791 set_pte(ptep, pte_mkold(pte));
792#if CONFIG_FLUSH_PAGE_ACCESSED
793 __flush_cache_page(vma, addr, PFN_PHYS(pte_pfn(pte)));
794#endif
795 return 1;
796}
797
798/*
799 * After a PTE is cleared, we have no way to flush the cache for
800 * the physical page. On PA8800 and PA8900 processors, these lines
801 * can cause random cache corruption. Thus, we must flush the cache
802 * as well as the TLB when clearing a PTE that's valid.
803 */
804pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long addr,
805 pte_t *ptep)
806{
807 struct mm_struct *mm = (vma)->vm_mm;
808 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
809 unsigned long pfn = pte_pfn(pte);
810
811 if (pfn_valid(pfn))
812 __flush_cache_page(vma, addr, PFN_PHYS(pfn));
813 else if (pte_accessible(mm, pte))
814 flush_tlb_page(vma, addr);
815
816 return pte;
817}
818
819/*
820 * The physical address for pages in the ioremap case can be obtained
821 * from the vm_struct struct. I wasn't able to successfully handle the
822 * vmalloc and vmap cases. We have an array of struct page pointers in
823 * the uninitialized vmalloc case but the flush failed using page_to_pfn.
824 */
825void flush_cache_vmap(unsigned long start, unsigned long end)
826{
827 unsigned long addr, physaddr;
828 struct vm_struct *vm;
829
830 /* Prevent cache move-in */
831 flush_tlb_kernel_range(start, end);
832
833 if (end - start >= parisc_cache_flush_threshold) {
834 flush_cache_all();
835 return;
836 }
837
838 if (WARN_ON_ONCE(!is_vmalloc_addr((void *)start))) {
839 flush_cache_all();
840 return;
841 }
842
843 vm = find_vm_area((void *)start);
844 if (WARN_ON_ONCE(!vm)) {
845 flush_cache_all();
846 return;
847 }
848
849 /* The physical addresses of IOREMAP regions are contiguous */
850 if (vm->flags & VM_IOREMAP) {
851 physaddr = vm->phys_addr;
852 for (addr = start; addr < end; addr += PAGE_SIZE) {
853 preempt_disable();
854 flush_dcache_page_asm(physaddr, start);
855 flush_icache_page_asm(physaddr, start);
856 preempt_enable();
857 physaddr += PAGE_SIZE;
858 }
859 return;
860 }
861
862 flush_cache_all();
863}
864EXPORT_SYMBOL(flush_cache_vmap);
865
866/*
867 * The vm_struct has been retired and the page table is set up. The
868 * last page in the range is a guard page. Its physical address can't
869 * be determined using lpa, so there is no way to flush the range
870 * using flush_dcache_page_asm.
871 */
872void flush_cache_vunmap(unsigned long start, unsigned long end)
873{
874 /* Prevent cache move-in */
875 flush_tlb_kernel_range(start, end);
876 flush_data_cache();
877}
878EXPORT_SYMBOL(flush_cache_vunmap);
879
880/*
881 * On systems with PA8800/PA8900 processors, there is no way to flush
882 * a vmap range other than using the architected loop to flush the
883 * entire cache. The page directory is not set up, so we can't use
884 * fdc, etc. FDCE/FICE don't work to flush a portion of the cache.
885 * L2 is physically indexed but FDCE/FICE instructions in virtual
886 * mode output their virtual address on the core bus, not their
887 * real address. As a result, the L2 cache index formed from the
888 * virtual address will most likely not be the same as the L2 index
889 * formed from the real address.
890 */
891void flush_kernel_vmap_range(void *vaddr, int size)
892{
893 unsigned long start = (unsigned long)vaddr;
894 unsigned long end = start + size;
895
896 flush_tlb_kernel_range(start, end);
897
898 if (!static_branch_likely(&parisc_has_dcache))
899 return;
900
901 /* If interrupts are disabled, we can only do local flush */
902 if (WARN_ON(IS_ENABLED(CONFIG_SMP) && arch_irqs_disabled())) {
903 flush_data_cache_local(NULL);
904 return;
905 }
906
907 flush_data_cache();
908}
909EXPORT_SYMBOL(flush_kernel_vmap_range);
910
911void invalidate_kernel_vmap_range(void *vaddr, int size)
912{
913 unsigned long start = (unsigned long)vaddr;
914 unsigned long end = start + size;
915
916 /* Ensure DMA is complete */
917 asm_syncdma();
918
919 flush_tlb_kernel_range(start, end);
920
921 if (!static_branch_likely(&parisc_has_dcache))
922 return;
923
924 /* If interrupts are disabled, we can only do local flush */
925 if (WARN_ON(IS_ENABLED(CONFIG_SMP) && arch_irqs_disabled())) {
926 flush_data_cache_local(NULL);
927 return;
928 }
929
930 flush_data_cache();
931}
932EXPORT_SYMBOL(invalidate_kernel_vmap_range);
933
934
935SYSCALL_DEFINE3(cacheflush, unsigned long, addr, unsigned long, bytes,
936 unsigned int, cache)
937{
938 unsigned long start, end;
939 ASM_EXCEPTIONTABLE_VAR(error);
940
941 if (bytes == 0)
942 return 0;
943 if (!access_ok((void __user *) addr, bytes))
944 return -EFAULT;
945
946 end = addr + bytes;
947
948 if (cache & DCACHE) {
949 start = addr;
950 __asm__ __volatile__ (
951#ifdef CONFIG_64BIT
952 "1: cmpb,*<<,n %0,%2,1b\n"
953#else
954 "1: cmpb,<<,n %0,%2,1b\n"
955#endif
956 " fic,m %3(%4,%0)\n"
957 "2: sync\n"
958 ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 2b, "%1")
959 : "+r" (start), "+r" (error)
960 : "r" (end), "r" (dcache_stride), "i" (SR_USER));
961 }
962
963 if (cache & ICACHE && error == 0) {
964 start = addr;
965 __asm__ __volatile__ (
966#ifdef CONFIG_64BIT
967 "1: cmpb,*<<,n %0,%2,1b\n"
968#else
969 "1: cmpb,<<,n %0,%2,1b\n"
970#endif
971 " fdc,m %3(%4,%0)\n"
972 "2: sync\n"
973 ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 2b, "%1")
974 : "+r" (start), "+r" (error)
975 : "r" (end), "r" (icache_stride), "i" (SR_USER));
976 }
977
978 return error;
979}