xref: /openbmc/linux/arch/parisc/kernel/cache.c (revision 8d81cd1a)
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 
45 int split_tlb __ro_after_init;
46 int dcache_stride __ro_after_init;
47 int icache_stride __ro_after_init;
48 EXPORT_SYMBOL(dcache_stride);
49 
50 /* Internal implementation in arch/parisc/kernel/pacache.S */
51 void flush_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
52 EXPORT_SYMBOL(flush_dcache_page_asm);
53 void purge_dcache_page_asm(unsigned long phys_addr, unsigned long vaddr);
54 void flush_icache_page_asm(unsigned long phys_addr, unsigned long vaddr);
55 void flush_data_cache_local(void *);  /* flushes local data-cache only */
56 void flush_instruction_cache_local(void); /* flushes local code-cache only */
57 
58 static 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  */
65 DEFINE_SPINLOCK(pa_tlb_flush_lock);
66 
67 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
68 int pa_serialize_tlb_flushes __ro_after_init;
69 #endif
70 
71 struct pdc_cache_info cache_info __ro_after_init;
72 #ifndef CONFIG_PA20
73 struct pdc_btlb_info btlb_info;
74 #endif
75 
76 DEFINE_STATIC_KEY_TRUE(parisc_has_cache);
77 DEFINE_STATIC_KEY_TRUE(parisc_has_dcache);
78 DEFINE_STATIC_KEY_TRUE(parisc_has_icache);
79 
80 static 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 
88 void flush_cache_all_local(void)
89 {
90 	cache_flush_local_cpu(NULL);
91 }
92 
93 void 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 
99 static 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 
109 void __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 
134 void
135 show_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 
178 void __init
179 parisc_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 	if ((boot_cpu_data.pdc.capabilities & PDC_MODEL_NVA_MASK) ==
280 						PDC_MODEL_NVA_UNSUPPORTED) {
281 		printk(KERN_WARNING "parisc_cache_init: Only equivalent aliasing supported!\n");
282 #if 0
283 		panic("SMP kernel required to avoid non-equivalent aliasing");
284 #endif
285 	}
286 }
287 
288 void disable_sr_hashing(void)
289 {
290 	int srhash_type, retval;
291 	unsigned long space_bits;
292 
293 	switch (boot_cpu_data.cpu_type) {
294 	case pcx: /* We shouldn't get this far.  setup.c should prevent it. */
295 		BUG();
296 		return;
297 
298 	case pcxs:
299 	case pcxt:
300 	case pcxt_:
301 		srhash_type = SRHASH_PCXST;
302 		break;
303 
304 	case pcxl:
305 		srhash_type = SRHASH_PCXL;
306 		break;
307 
308 	case pcxl2: /* pcxl2 doesn't support space register hashing */
309 		return;
310 
311 	default: /* Currently all PA2.0 machines use the same ins. sequence */
312 		srhash_type = SRHASH_PA20;
313 		break;
314 	}
315 
316 	disable_sr_hashing_asm(srhash_type);
317 
318 	retval = pdc_spaceid_bits(&space_bits);
319 	/* If this procedure isn't implemented, don't panic. */
320 	if (retval < 0 && retval != PDC_BAD_OPTION)
321 		panic("pdc_spaceid_bits call failed.\n");
322 	if (space_bits != 0)
323 		panic("SpaceID hashing is still on!\n");
324 }
325 
326 static inline void
327 __flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr,
328 		   unsigned long physaddr)
329 {
330 	if (!static_branch_likely(&parisc_has_cache))
331 		return;
332 
333 	/*
334 	 * The TLB is the engine of coherence on parisc.  The CPU is
335 	 * entitled to speculate any page with a TLB mapping, so here
336 	 * we kill the mapping then flush the page along a special flush
337 	 * only alias mapping. This guarantees that the page is no-longer
338 	 * in the cache for any process and nor may it be speculatively
339 	 * read in (until the user or kernel specifically accesses it,
340 	 * of course).
341 	 */
342 	flush_tlb_page(vma, vmaddr);
343 
344 	preempt_disable();
345 	flush_dcache_page_asm(physaddr, vmaddr);
346 	if (vma->vm_flags & VM_EXEC)
347 		flush_icache_page_asm(physaddr, vmaddr);
348 	preempt_enable();
349 }
350 
351 static void flush_kernel_dcache_page_addr(const void *addr)
352 {
353 	unsigned long vaddr = (unsigned long)addr;
354 	unsigned long flags;
355 
356 	/* Purge TLB entry to remove translation on all CPUs */
357 	purge_tlb_start(flags);
358 	pdtlb(SR_KERNEL, addr);
359 	purge_tlb_end(flags);
360 
361 	/* Use tmpalias flush to prevent data cache move-in */
362 	preempt_disable();
363 	flush_dcache_page_asm(__pa(vaddr), vaddr);
364 	preempt_enable();
365 }
366 
367 static void flush_kernel_icache_page_addr(const void *addr)
368 {
369 	unsigned long vaddr = (unsigned long)addr;
370 	unsigned long flags;
371 
372 	/* Purge TLB entry to remove translation on all CPUs */
373 	purge_tlb_start(flags);
374 	pdtlb(SR_KERNEL, addr);
375 	purge_tlb_end(flags);
376 
377 	/* Use tmpalias flush to prevent instruction cache move-in */
378 	preempt_disable();
379 	flush_icache_page_asm(__pa(vaddr), vaddr);
380 	preempt_enable();
381 }
382 
383 void kunmap_flush_on_unmap(const void *addr)
384 {
385 	flush_kernel_dcache_page_addr(addr);
386 }
387 EXPORT_SYMBOL(kunmap_flush_on_unmap);
388 
389 void flush_icache_pages(struct vm_area_struct *vma, struct page *page,
390 		unsigned int nr)
391 {
392 	void *kaddr = page_address(page);
393 
394 	for (;;) {
395 		flush_kernel_dcache_page_addr(kaddr);
396 		flush_kernel_icache_page_addr(kaddr);
397 		if (--nr == 0)
398 			break;
399 		kaddr += PAGE_SIZE;
400 	}
401 }
402 
403 /*
404  * Walk page directory for MM to find PTEP pointer for address ADDR.
405  */
406 static inline pte_t *get_ptep(struct mm_struct *mm, unsigned long addr)
407 {
408 	pte_t *ptep = NULL;
409 	pgd_t *pgd = mm->pgd;
410 	p4d_t *p4d;
411 	pud_t *pud;
412 	pmd_t *pmd;
413 
414 	if (!pgd_none(*pgd)) {
415 		p4d = p4d_offset(pgd, addr);
416 		if (!p4d_none(*p4d)) {
417 			pud = pud_offset(p4d, addr);
418 			if (!pud_none(*pud)) {
419 				pmd = pmd_offset(pud, addr);
420 				if (!pmd_none(*pmd))
421 					ptep = pte_offset_map(pmd, addr);
422 			}
423 		}
424 	}
425 	return ptep;
426 }
427 
428 static inline bool pte_needs_flush(pte_t pte)
429 {
430 	return (pte_val(pte) & (_PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_NO_CACHE))
431 		== (_PAGE_PRESENT | _PAGE_ACCESSED);
432 }
433 
434 /*
435  * Return user physical address. Returns 0 if page is not present.
436  */
437 static inline unsigned long get_upa(struct mm_struct *mm, unsigned long addr)
438 {
439 	unsigned long flags, space, pgd, prot, pa;
440 #ifdef CONFIG_TLB_PTLOCK
441 	unsigned long pgd_lock;
442 #endif
443 
444 	/* Save context */
445 	local_irq_save(flags);
446 	prot = mfctl(8);
447 	space = mfsp(SR_USER);
448 	pgd = mfctl(25);
449 #ifdef CONFIG_TLB_PTLOCK
450 	pgd_lock = mfctl(28);
451 #endif
452 
453 	/* Set context for lpa_user */
454 	switch_mm_irqs_off(NULL, mm, NULL);
455 	pa = lpa_user(addr);
456 
457 	/* Restore previous context */
458 #ifdef CONFIG_TLB_PTLOCK
459 	mtctl(pgd_lock, 28);
460 #endif
461 	mtctl(pgd, 25);
462 	mtsp(space, SR_USER);
463 	mtctl(prot, 8);
464 	local_irq_restore(flags);
465 
466 	return pa;
467 }
468 
469 void flush_dcache_folio(struct folio *folio)
470 {
471 	struct address_space *mapping = folio_flush_mapping(folio);
472 	struct vm_area_struct *vma;
473 	unsigned long addr, old_addr = 0;
474 	void *kaddr;
475 	unsigned long count = 0;
476 	unsigned long i, nr, flags;
477 	pgoff_t pgoff;
478 
479 	if (mapping && !mapping_mapped(mapping)) {
480 		set_bit(PG_dcache_dirty, &folio->flags);
481 		return;
482 	}
483 
484 	nr = folio_nr_pages(folio);
485 	kaddr = folio_address(folio);
486 	for (i = 0; i < nr; i++)
487 		flush_kernel_dcache_page_addr(kaddr + i * PAGE_SIZE);
488 
489 	if (!mapping)
490 		return;
491 
492 	pgoff = folio->index;
493 
494 	/*
495 	 * We have carefully arranged in arch_get_unmapped_area() that
496 	 * *any* mappings of a file are always congruently mapped (whether
497 	 * declared as MAP_PRIVATE or MAP_SHARED), so we only need
498 	 * to flush one address here for them all to become coherent
499 	 * on machines that support equivalent aliasing
500 	 */
501 	flush_dcache_mmap_lock_irqsave(mapping, flags);
502 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff + nr - 1) {
503 		unsigned long offset = pgoff - vma->vm_pgoff;
504 		unsigned long pfn = folio_pfn(folio);
505 
506 		addr = vma->vm_start;
507 		nr = folio_nr_pages(folio);
508 		if (offset > -nr) {
509 			pfn -= offset;
510 			nr += offset;
511 		} else {
512 			addr += offset * PAGE_SIZE;
513 		}
514 		if (addr + nr * PAGE_SIZE > vma->vm_end)
515 			nr = (vma->vm_end - addr) / PAGE_SIZE;
516 
517 		if (old_addr == 0 || (old_addr & (SHM_COLOUR - 1))
518 					!= (addr & (SHM_COLOUR - 1))) {
519 			for (i = 0; i < nr; i++)
520 				__flush_cache_page(vma,
521 					addr + i * PAGE_SIZE,
522 					(pfn + i) * PAGE_SIZE);
523 			/*
524 			 * Software is allowed to have any number
525 			 * of private mappings to a page.
526 			 */
527 			if (!(vma->vm_flags & VM_SHARED))
528 				continue;
529 			if (old_addr)
530 				pr_err("INEQUIVALENT ALIASES 0x%lx and 0x%lx in file %pD\n",
531 					old_addr, addr, vma->vm_file);
532 			if (nr == folio_nr_pages(folio))
533 				old_addr = addr;
534 		}
535 		WARN_ON(++count == 4096);
536 	}
537 	flush_dcache_mmap_unlock_irqrestore(mapping, flags);
538 }
539 EXPORT_SYMBOL(flush_dcache_folio);
540 
541 /* Defined in arch/parisc/kernel/pacache.S */
542 EXPORT_SYMBOL(flush_kernel_dcache_range_asm);
543 EXPORT_SYMBOL(flush_kernel_icache_range_asm);
544 
545 #define FLUSH_THRESHOLD 0x80000 /* 0.5MB */
546 static unsigned long parisc_cache_flush_threshold __ro_after_init = FLUSH_THRESHOLD;
547 
548 #define FLUSH_TLB_THRESHOLD (16*1024) /* 16 KiB minimum TLB threshold */
549 static unsigned long parisc_tlb_flush_threshold __ro_after_init = ~0UL;
550 
551 void __init parisc_setup_cache_timing(void)
552 {
553 	unsigned long rangetime, alltime;
554 	unsigned long size;
555 	unsigned long threshold, threshold2;
556 
557 	alltime = mfctl(16);
558 	flush_data_cache();
559 	alltime = mfctl(16) - alltime;
560 
561 	size = (unsigned long)(_end - _text);
562 	rangetime = mfctl(16);
563 	flush_kernel_dcache_range((unsigned long)_text, size);
564 	rangetime = mfctl(16) - rangetime;
565 
566 	printk(KERN_DEBUG "Whole cache flush %lu cycles, flushing %lu bytes %lu cycles\n",
567 		alltime, size, rangetime);
568 
569 	threshold = L1_CACHE_ALIGN((unsigned long)((uint64_t)size * alltime / rangetime));
570 	pr_info("Calculated flush threshold is %lu KiB\n",
571 		threshold/1024);
572 
573 	/*
574 	 * The threshold computed above isn't very reliable. The following
575 	 * heuristic works reasonably well on c8000/rp3440.
576 	 */
577 	threshold2 = cache_info.dc_size * num_online_cpus();
578 	parisc_cache_flush_threshold = threshold2;
579 	printk(KERN_INFO "Cache flush threshold set to %lu KiB\n",
580 		parisc_cache_flush_threshold/1024);
581 
582 	/* calculate TLB flush threshold */
583 
584 	/* On SMP machines, skip the TLB measure of kernel text which
585 	 * has been mapped as huge pages. */
586 	if (num_online_cpus() > 1 && !parisc_requires_coherency()) {
587 		threshold = max(cache_info.it_size, cache_info.dt_size);
588 		threshold *= PAGE_SIZE;
589 		threshold /= num_online_cpus();
590 		goto set_tlb_threshold;
591 	}
592 
593 	size = (unsigned long)_end - (unsigned long)_text;
594 	rangetime = mfctl(16);
595 	flush_tlb_kernel_range((unsigned long)_text, (unsigned long)_end);
596 	rangetime = mfctl(16) - rangetime;
597 
598 	alltime = mfctl(16);
599 	flush_tlb_all();
600 	alltime = mfctl(16) - alltime;
601 
602 	printk(KERN_INFO "Whole TLB flush %lu cycles, Range flush %lu bytes %lu cycles\n",
603 		alltime, size, rangetime);
604 
605 	threshold = PAGE_ALIGN((num_online_cpus() * size * alltime) / rangetime);
606 	printk(KERN_INFO "Calculated TLB flush threshold %lu KiB\n",
607 		threshold/1024);
608 
609 set_tlb_threshold:
610 	if (threshold > FLUSH_TLB_THRESHOLD)
611 		parisc_tlb_flush_threshold = threshold;
612 	else
613 		parisc_tlb_flush_threshold = FLUSH_TLB_THRESHOLD;
614 
615 	printk(KERN_INFO "TLB flush threshold set to %lu KiB\n",
616 		parisc_tlb_flush_threshold/1024);
617 }
618 
619 extern void purge_kernel_dcache_page_asm(unsigned long);
620 extern void clear_user_page_asm(void *, unsigned long);
621 extern void copy_user_page_asm(void *, void *, unsigned long);
622 
623 static 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 
647 void 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 
660 void 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 
668 void 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  */
680 int __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 
705 static 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 
713 static 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 
727 void 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 
754 void 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 
771 void 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 
776 void 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 
784 int 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  */
804 pte_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  */
825 void 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 }
864 EXPORT_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  */
872 void 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 }
878 EXPORT_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  */
891 void 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 }
909 EXPORT_SYMBOL(flush_kernel_vmap_range);
910 
911 void 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 }
932 EXPORT_SYMBOL(invalidate_kernel_vmap_range);
933 
934 
935 SYSCALL_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 }
980