xref: /linux/arch/arc/mm/cache.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1 /*
2  * ARC Cache Management
3  *
4  * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
5  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  */
11 
12 #include <linux/module.h>
13 #include <linux/mm.h>
14 #include <linux/sched.h>
15 #include <linux/cache.h>
16 #include <linux/mmu_context.h>
17 #include <linux/syscalls.h>
18 #include <linux/uaccess.h>
19 #include <linux/pagemap.h>
20 #include <asm/cacheflush.h>
21 #include <asm/cachectl.h>
22 #include <asm/setup.h>
23 
24 static int l2_line_sz;
25 static int ioc_exists;
26 int slc_enable = 1, ioc_enable = 1;
27 unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
28 unsigned long perip_end = 0xFFFFFFFF; /* legacy value */
29 
30 void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
31 			       unsigned long sz, const int cacheop);
32 
33 void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
34 void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
35 void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
36 
37 char *arc_cache_mumbojumbo(int c, char *buf, int len)
38 {
39 	int n = 0;
40 	struct cpuinfo_arc_cache *p;
41 
42 #define PR_CACHE(p, cfg, str)						\
43 	if (!(p)->ver)							\
44 		n += scnprintf(buf + n, len - n, str"\t\t: N/A\n");	\
45 	else								\
46 		n += scnprintf(buf + n, len - n,			\
47 			str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n",	\
48 			(p)->sz_k, (p)->assoc, (p)->line_len,		\
49 			(p)->vipt ? "VIPT" : "PIPT",			\
50 			(p)->alias ? " aliasing" : "",			\
51 			IS_USED_CFG(cfg));
52 
53 	PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
54 	PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
55 
56 	p = &cpuinfo_arc700[c].slc;
57 	if (p->ver)
58 		n += scnprintf(buf + n, len - n,
59 			       "SLC\t\t: %uK, %uB Line%s\n",
60 			       p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
61 
62 	n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
63 		       perip_base,
64 		       IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency "));
65 
66 	return buf;
67 }
68 
69 /*
70  * Read the Cache Build Confuration Registers, Decode them and save into
71  * the cpuinfo structure for later use.
72  * No Validation done here, simply read/convert the BCRs
73  */
74 static void read_decode_cache_bcr_arcv2(int cpu)
75 {
76 	struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
77 	struct bcr_generic sbcr;
78 
79 	struct bcr_slc_cfg {
80 #ifdef CONFIG_CPU_BIG_ENDIAN
81 		unsigned int pad:24, way:2, lsz:2, sz:4;
82 #else
83 		unsigned int sz:4, lsz:2, way:2, pad:24;
84 #endif
85 	} slc_cfg;
86 
87 	struct bcr_clust_cfg {
88 #ifdef CONFIG_CPU_BIG_ENDIAN
89 		unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
90 #else
91 		unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
92 #endif
93 	} cbcr;
94 
95 	struct bcr_volatile {
96 #ifdef CONFIG_CPU_BIG_ENDIAN
97 		unsigned int start:4, limit:4, pad:22, order:1, disable:1;
98 #else
99 		unsigned int disable:1, order:1, pad:22, limit:4, start:4;
100 #endif
101 	} vol;
102 
103 
104 	READ_BCR(ARC_REG_SLC_BCR, sbcr);
105 	if (sbcr.ver) {
106 		READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
107 		p_slc->ver = sbcr.ver;
108 		p_slc->sz_k = 128 << slc_cfg.sz;
109 		l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
110 	}
111 
112 	READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
113 	if (cbcr.c)
114 		ioc_exists = 1;
115 	else
116 		ioc_enable = 0;
117 
118 	/* HS 2.0 didn't have AUX_VOL */
119 	if (cpuinfo_arc700[cpu].core.family > 0x51) {
120 		READ_BCR(AUX_VOL, vol);
121 		perip_base = vol.start << 28;
122 		/* HS 3.0 has limit and strict-ordering fields */
123 		if (cpuinfo_arc700[cpu].core.family > 0x52)
124 			perip_end = (vol.limit << 28) - 1;
125 	}
126 }
127 
128 void read_decode_cache_bcr(void)
129 {
130 	struct cpuinfo_arc_cache *p_ic, *p_dc;
131 	unsigned int cpu = smp_processor_id();
132 	struct bcr_cache {
133 #ifdef CONFIG_CPU_BIG_ENDIAN
134 		unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
135 #else
136 		unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
137 #endif
138 	} ibcr, dbcr;
139 
140 	p_ic = &cpuinfo_arc700[cpu].icache;
141 	READ_BCR(ARC_REG_IC_BCR, ibcr);
142 
143 	if (!ibcr.ver)
144 		goto dc_chk;
145 
146 	if (ibcr.ver <= 3) {
147 		BUG_ON(ibcr.config != 3);
148 		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
149 	} else if (ibcr.ver >= 4) {
150 		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
151 	}
152 
153 	p_ic->line_len = 8 << ibcr.line_len;
154 	p_ic->sz_k = 1 << (ibcr.sz - 1);
155 	p_ic->ver = ibcr.ver;
156 	p_ic->vipt = 1;
157 	p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
158 
159 dc_chk:
160 	p_dc = &cpuinfo_arc700[cpu].dcache;
161 	READ_BCR(ARC_REG_DC_BCR, dbcr);
162 
163 	if (!dbcr.ver)
164 		goto slc_chk;
165 
166 	if (dbcr.ver <= 3) {
167 		BUG_ON(dbcr.config != 2);
168 		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
169 		p_dc->vipt = 1;
170 		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
171 	} else if (dbcr.ver >= 4) {
172 		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
173 		p_dc->vipt = 0;
174 		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
175 	}
176 
177 	p_dc->line_len = 16 << dbcr.line_len;
178 	p_dc->sz_k = 1 << (dbcr.sz - 1);
179 	p_dc->ver = dbcr.ver;
180 
181 slc_chk:
182 	if (is_isa_arcv2())
183                 read_decode_cache_bcr_arcv2(cpu);
184 }
185 
186 /*
187  * Line Operation on {I,D}-Cache
188  */
189 
190 #define OP_INV		0x1
191 #define OP_FLUSH	0x2
192 #define OP_FLUSH_N_INV	0x3
193 #define OP_INV_IC	0x4
194 
195 /*
196  *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
197  *
198  * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
199  * The orig Cache Management Module "CDU" only required paddr to invalidate a
200  * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
201  * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
202  * the exact same line.
203  *
204  * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
205  * paddr alone could not be used to correctly index the cache.
206  *
207  * ------------------
208  * MMU v1/v2 (Fixed Page Size 8k)
209  * ------------------
210  * The solution was to provide CDU with these additonal vaddr bits. These
211  * would be bits [x:13], x would depend on cache-geometry, 13 comes from
212  * standard page size of 8k.
213  * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
214  * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
215  * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
216  * represent the offset within cache-line. The adv of using this "clumsy"
217  * interface for additional info was no new reg was needed in CDU programming
218  * model.
219  *
220  * 17:13 represented the max num of bits passable, actual bits needed were
221  * fewer, based on the num-of-aliases possible.
222  * -for 2 alias possibility, only bit 13 needed (32K cache)
223  * -for 4 alias possibility, bits 14:13 needed (64K cache)
224  *
225  * ------------------
226  * MMU v3
227  * ------------------
228  * This ver of MMU supports variable page sizes (1k-16k): although Linux will
229  * only support 8k (default), 16k and 4k.
230  * However from hardware perspective, smaller page sizes aggravate aliasing
231  * meaning more vaddr bits needed to disambiguate the cache-line-op ;
232  * the existing scheme of piggybacking won't work for certain configurations.
233  * Two new registers IC_PTAG and DC_PTAG inttoduced.
234  * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
235  */
236 
237 static inline
238 void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
239 			  unsigned long sz, const int op)
240 {
241 	unsigned int aux_cmd;
242 	int num_lines;
243 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
244 
245 	if (op == OP_INV_IC) {
246 		aux_cmd = ARC_REG_IC_IVIL;
247 	} else {
248 		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
249 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
250 	}
251 
252 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
253 	 * and have @paddr - aligned to cache line and integral @num_lines.
254 	 * This however can be avoided for page sized since:
255 	 *  -@paddr will be cache-line aligned already (being page aligned)
256 	 *  -@sz will be integral multiple of line size (being page sized).
257 	 */
258 	if (!full_page) {
259 		sz += paddr & ~CACHE_LINE_MASK;
260 		paddr &= CACHE_LINE_MASK;
261 		vaddr &= CACHE_LINE_MASK;
262 	}
263 
264 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
265 
266 	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
267 	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
268 
269 	while (num_lines-- > 0) {
270 		write_aux_reg(aux_cmd, paddr);
271 		paddr += L1_CACHE_BYTES;
272 	}
273 }
274 
275 /*
276  * For ARC700 MMUv3 I-cache and D-cache flushes
277  * Also reused for HS38 aliasing I-cache configuration
278  */
279 static inline
280 void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
281 			  unsigned long sz, const int op)
282 {
283 	unsigned int aux_cmd, aux_tag;
284 	int num_lines;
285 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
286 
287 	if (op == OP_INV_IC) {
288 		aux_cmd = ARC_REG_IC_IVIL;
289 		aux_tag = ARC_REG_IC_PTAG;
290 	} else {
291 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
292 		aux_tag = ARC_REG_DC_PTAG;
293 	}
294 
295 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
296 	 * and have @paddr - aligned to cache line and integral @num_lines.
297 	 * This however can be avoided for page sized since:
298 	 *  -@paddr will be cache-line aligned already (being page aligned)
299 	 *  -@sz will be integral multiple of line size (being page sized).
300 	 */
301 	if (!full_page) {
302 		sz += paddr & ~CACHE_LINE_MASK;
303 		paddr &= CACHE_LINE_MASK;
304 		vaddr &= CACHE_LINE_MASK;
305 	}
306 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
307 
308 	/*
309 	 * MMUv3, cache ops require paddr in PTAG reg
310 	 * if V-P const for loop, PTAG can be written once outside loop
311 	 */
312 	if (full_page)
313 		write_aux_reg(aux_tag, paddr);
314 
315 	/*
316 	 * This is technically for MMU v4, using the MMU v3 programming model
317 	 * Special work for HS38 aliasing I-cache configuration with PAE40
318 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
319 	 *   - (and needs to be written before the lower 32 bits)
320 	 * Note that PTAG_HI is hoisted outside the line loop
321 	 */
322 	if (is_pae40_enabled() && op == OP_INV_IC)
323 		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
324 
325 	while (num_lines-- > 0) {
326 		if (!full_page) {
327 			write_aux_reg(aux_tag, paddr);
328 			paddr += L1_CACHE_BYTES;
329 		}
330 
331 		write_aux_reg(aux_cmd, vaddr);
332 		vaddr += L1_CACHE_BYTES;
333 	}
334 }
335 
336 /*
337  * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
338  * Here's how cache ops are implemented
339  *
340  *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
341  *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
342  *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
343  *    respectively, similar to MMU v3 programming model, hence
344  *    __cache_line_loop_v3() is used)
345  *
346  * If PAE40 is enabled, independent of aliasing considerations, the higher bits
347  * needs to be written into PTAG_HI
348  */
349 static inline
350 void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
351 			  unsigned long sz, const int cacheop)
352 {
353 	unsigned int aux_cmd;
354 	int num_lines;
355 	const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
356 
357 	if (cacheop == OP_INV_IC) {
358 		aux_cmd = ARC_REG_IC_IVIL;
359 	} else {
360 		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
361 		aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
362 	}
363 
364 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
365 	 * and have @paddr - aligned to cache line and integral @num_lines.
366 	 * This however can be avoided for page sized since:
367 	 *  -@paddr will be cache-line aligned already (being page aligned)
368 	 *  -@sz will be integral multiple of line size (being page sized).
369 	 */
370 	if (!full_page_op) {
371 		sz += paddr & ~CACHE_LINE_MASK;
372 		paddr &= CACHE_LINE_MASK;
373 	}
374 
375 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
376 
377 	/*
378 	 * For HS38 PAE40 configuration
379 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
380 	 *   - (and needs to be written before the lower 32 bits)
381 	 */
382 	if (is_pae40_enabled()) {
383 		if (cacheop == OP_INV_IC)
384 			/*
385 			 * Non aliasing I-cache in HS38,
386 			 * aliasing I-cache handled in __cache_line_loop_v3()
387 			 */
388 			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
389 		else
390 			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
391 	}
392 
393 	while (num_lines-- > 0) {
394 		write_aux_reg(aux_cmd, paddr);
395 		paddr += L1_CACHE_BYTES;
396 	}
397 }
398 
399 #if (CONFIG_ARC_MMU_VER < 3)
400 #define __cache_line_loop	__cache_line_loop_v2
401 #elif (CONFIG_ARC_MMU_VER == 3)
402 #define __cache_line_loop	__cache_line_loop_v3
403 #elif (CONFIG_ARC_MMU_VER > 3)
404 #define __cache_line_loop	__cache_line_loop_v4
405 #endif
406 
407 #ifdef CONFIG_ARC_HAS_DCACHE
408 
409 /***************************************************************
410  * Machine specific helpers for Entire D-Cache or Per Line ops
411  */
412 
413 static inline void __before_dc_op(const int op)
414 {
415 	if (op == OP_FLUSH_N_INV) {
416 		/* Dcache provides 2 cmd: FLUSH or INV
417 		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
418 		 * flush-n-inv is achieved by INV cmd but with IM=1
419 		 * So toggle INV sub-mode depending on op request and default
420 		 */
421 		const unsigned int ctl = ARC_REG_DC_CTRL;
422 		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
423 	}
424 }
425 
426 static inline void __after_dc_op(const int op)
427 {
428 	if (op & OP_FLUSH) {
429 		const unsigned int ctl = ARC_REG_DC_CTRL;
430 		unsigned int reg;
431 
432 		/* flush / flush-n-inv both wait */
433 		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
434 			;
435 
436 		/* Switch back to default Invalidate mode */
437 		if (op == OP_FLUSH_N_INV)
438 			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
439 	}
440 }
441 
442 /*
443  * Operation on Entire D-Cache
444  * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
445  * Note that constant propagation ensures all the checks are gone
446  * in generated code
447  */
448 static inline void __dc_entire_op(const int op)
449 {
450 	int aux;
451 
452 	__before_dc_op(op);
453 
454 	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
455 		aux = ARC_REG_DC_IVDC;
456 	else
457 		aux = ARC_REG_DC_FLSH;
458 
459 	write_aux_reg(aux, 0x1);
460 
461 	__after_dc_op(op);
462 }
463 
464 /* For kernel mappings cache operation: index is same as paddr */
465 #define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
466 
467 /*
468  * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
469  */
470 static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
471 				unsigned long sz, const int op)
472 {
473 	unsigned long flags;
474 
475 	local_irq_save(flags);
476 
477 	__before_dc_op(op);
478 
479 	__cache_line_loop(paddr, vaddr, sz, op);
480 
481 	__after_dc_op(op);
482 
483 	local_irq_restore(flags);
484 }
485 
486 #else
487 
488 #define __dc_entire_op(op)
489 #define __dc_line_op(paddr, vaddr, sz, op)
490 #define __dc_line_op_k(paddr, sz, op)
491 
492 #endif /* CONFIG_ARC_HAS_DCACHE */
493 
494 #ifdef CONFIG_ARC_HAS_ICACHE
495 
496 static inline void __ic_entire_inv(void)
497 {
498 	write_aux_reg(ARC_REG_IC_IVIC, 1);
499 	read_aux_reg(ARC_REG_IC_CTRL);	/* blocks */
500 }
501 
502 static inline void
503 __ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
504 			  unsigned long sz)
505 {
506 	unsigned long flags;
507 
508 	local_irq_save(flags);
509 	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC);
510 	local_irq_restore(flags);
511 }
512 
513 #ifndef CONFIG_SMP
514 
515 #define __ic_line_inv_vaddr(p, v, s)	__ic_line_inv_vaddr_local(p, v, s)
516 
517 #else
518 
519 struct ic_inv_args {
520 	phys_addr_t paddr, vaddr;
521 	int sz;
522 };
523 
524 static void __ic_line_inv_vaddr_helper(void *info)
525 {
526         struct ic_inv_args *ic_inv = info;
527 
528         __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
529 }
530 
531 static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
532 				unsigned long sz)
533 {
534 	struct ic_inv_args ic_inv = {
535 		.paddr = paddr,
536 		.vaddr = vaddr,
537 		.sz    = sz
538 	};
539 
540 	on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
541 }
542 
543 #endif	/* CONFIG_SMP */
544 
545 #else	/* !CONFIG_ARC_HAS_ICACHE */
546 
547 #define __ic_entire_inv()
548 #define __ic_line_inv_vaddr(pstart, vstart, sz)
549 
550 #endif /* CONFIG_ARC_HAS_ICACHE */
551 
552 noinline void slc_op(phys_addr_t paddr, unsigned long sz, const int op)
553 {
554 #ifdef CONFIG_ISA_ARCV2
555 	/*
556 	 * SLC is shared between all cores and concurrent aux operations from
557 	 * multiple cores need to be serialized using a spinlock
558 	 * A concurrent operation can be silently ignored and/or the old/new
559 	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
560 	 * below)
561 	 */
562 	static DEFINE_SPINLOCK(lock);
563 	unsigned long flags;
564 	unsigned int ctrl;
565 
566 	spin_lock_irqsave(&lock, flags);
567 
568 	/*
569 	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
570 	 *  - b'000 (default) is Flush,
571 	 *  - b'001 is Invalidate if CTRL.IM == 0
572 	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
573 	 */
574 	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
575 
576 	/* Don't rely on default value of IM bit */
577 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
578 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
579 	else
580 		ctrl |= SLC_CTRL_IM;
581 
582 	if (op & OP_INV)
583 		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
584 	else
585 		ctrl &= ~SLC_CTRL_RGN_OP_INV;
586 
587 	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
588 
589 	/*
590 	 * Lower bits are ignored, no need to clip
591 	 * END needs to be setup before START (latter triggers the operation)
592 	 * END can't be same as START, so add (l2_line_sz - 1) to sz
593 	 */
594 	write_aux_reg(ARC_REG_SLC_RGN_END, (paddr + sz + l2_line_sz - 1));
595 	write_aux_reg(ARC_REG_SLC_RGN_START, paddr);
596 
597 	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
598 
599 	spin_unlock_irqrestore(&lock, flags);
600 #endif
601 }
602 
603 /***********************************************************
604  * Exported APIs
605  */
606 
607 /*
608  * Handle cache congruency of kernel and userspace mappings of page when kernel
609  * writes-to/reads-from
610  *
611  * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
612  *  -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
613  *  -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
614  *  -In SMP, if hardware caches are coherent
615  *
616  * There's a corollary case, where kernel READs from a userspace mapped page.
617  * If the U-mapping is not congruent to to K-mapping, former needs flushing.
618  */
619 void flush_dcache_page(struct page *page)
620 {
621 	struct address_space *mapping;
622 
623 	if (!cache_is_vipt_aliasing()) {
624 		clear_bit(PG_dc_clean, &page->flags);
625 		return;
626 	}
627 
628 	/* don't handle anon pages here */
629 	mapping = page_mapping(page);
630 	if (!mapping)
631 		return;
632 
633 	/*
634 	 * pagecache page, file not yet mapped to userspace
635 	 * Make a note that K-mapping is dirty
636 	 */
637 	if (!mapping_mapped(mapping)) {
638 		clear_bit(PG_dc_clean, &page->flags);
639 	} else if (page_mapcount(page)) {
640 
641 		/* kernel reading from page with U-mapping */
642 		phys_addr_t paddr = (unsigned long)page_address(page);
643 		unsigned long vaddr = page->index << PAGE_SHIFT;
644 
645 		if (addr_not_cache_congruent(paddr, vaddr))
646 			__flush_dcache_page(paddr, vaddr);
647 	}
648 }
649 EXPORT_SYMBOL(flush_dcache_page);
650 
651 /*
652  * DMA ops for systems with L1 cache only
653  * Make memory coherent with L1 cache by flushing/invalidating L1 lines
654  */
655 static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
656 {
657 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
658 }
659 
660 static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
661 {
662 	__dc_line_op_k(start, sz, OP_INV);
663 }
664 
665 static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
666 {
667 	__dc_line_op_k(start, sz, OP_FLUSH);
668 }
669 
670 /*
671  * DMA ops for systems with both L1 and L2 caches, but without IOC
672  * Both L1 and L2 lines need to be explicitly flushed/invalidated
673  */
674 static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
675 {
676 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
677 	slc_op(start, sz, OP_FLUSH_N_INV);
678 }
679 
680 static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
681 {
682 	__dc_line_op_k(start, sz, OP_INV);
683 	slc_op(start, sz, OP_INV);
684 }
685 
686 static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
687 {
688 	__dc_line_op_k(start, sz, OP_FLUSH);
689 	slc_op(start, sz, OP_FLUSH);
690 }
691 
692 /*
693  * DMA ops for systems with IOC
694  * IOC hardware snoops all DMA traffic keeping the caches consistent with
695  * memory - eliding need for any explicit cache maintenance of DMA buffers
696  */
697 static void __dma_cache_wback_inv_ioc(phys_addr_t start, unsigned long sz) {}
698 static void __dma_cache_inv_ioc(phys_addr_t start, unsigned long sz) {}
699 static void __dma_cache_wback_ioc(phys_addr_t start, unsigned long sz) {}
700 
701 /*
702  * Exported DMA API
703  */
704 void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
705 {
706 	__dma_cache_wback_inv(start, sz);
707 }
708 EXPORT_SYMBOL(dma_cache_wback_inv);
709 
710 void dma_cache_inv(phys_addr_t start, unsigned long sz)
711 {
712 	__dma_cache_inv(start, sz);
713 }
714 EXPORT_SYMBOL(dma_cache_inv);
715 
716 void dma_cache_wback(phys_addr_t start, unsigned long sz)
717 {
718 	__dma_cache_wback(start, sz);
719 }
720 EXPORT_SYMBOL(dma_cache_wback);
721 
722 /*
723  * This is API for making I/D Caches consistent when modifying
724  * kernel code (loadable modules, kprobes, kgdb...)
725  * This is called on insmod, with kernel virtual address for CODE of
726  * the module. ARC cache maintenance ops require PHY address thus we
727  * need to convert vmalloc addr to PHY addr
728  */
729 void flush_icache_range(unsigned long kstart, unsigned long kend)
730 {
731 	unsigned int tot_sz;
732 
733 	WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
734 
735 	/* Shortcut for bigger flush ranges.
736 	 * Here we don't care if this was kernel virtual or phy addr
737 	 */
738 	tot_sz = kend - kstart;
739 	if (tot_sz > PAGE_SIZE) {
740 		flush_cache_all();
741 		return;
742 	}
743 
744 	/* Case: Kernel Phy addr (0x8000_0000 onwards) */
745 	if (likely(kstart > PAGE_OFFSET)) {
746 		/*
747 		 * The 2nd arg despite being paddr will be used to index icache
748 		 * This is OK since no alternate virtual mappings will exist
749 		 * given the callers for this case: kprobe/kgdb in built-in
750 		 * kernel code only.
751 		 */
752 		__sync_icache_dcache(kstart, kstart, kend - kstart);
753 		return;
754 	}
755 
756 	/*
757 	 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
758 	 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
759 	 *     handling of kernel vaddr.
760 	 *
761 	 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
762 	 *     it still needs to handle  a 2 page scenario, where the range
763 	 *     straddles across 2 virtual pages and hence need for loop
764 	 */
765 	while (tot_sz > 0) {
766 		unsigned int off, sz;
767 		unsigned long phy, pfn;
768 
769 		off = kstart % PAGE_SIZE;
770 		pfn = vmalloc_to_pfn((void *)kstart);
771 		phy = (pfn << PAGE_SHIFT) + off;
772 		sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
773 		__sync_icache_dcache(phy, kstart, sz);
774 		kstart += sz;
775 		tot_sz -= sz;
776 	}
777 }
778 EXPORT_SYMBOL(flush_icache_range);
779 
780 /*
781  * General purpose helper to make I and D cache lines consistent.
782  * @paddr is phy addr of region
783  * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
784  *    However in one instance, when called by kprobe (for a breakpt in
785  *    builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
786  *    use a paddr to index the cache (despite VIPT). This is fine since since a
787  *    builtin kernel page will not have any virtual mappings.
788  *    kprobe on loadable module will be kernel vaddr.
789  */
790 void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
791 {
792 	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
793 	__ic_line_inv_vaddr(paddr, vaddr, len);
794 }
795 
796 /* wrapper to compile time eliminate alignment checks in flush loop */
797 void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
798 {
799 	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
800 }
801 
802 /*
803  * wrapper to clearout kernel or userspace mappings of a page
804  * For kernel mappings @vaddr == @paddr
805  */
806 void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
807 {
808 	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
809 }
810 
811 noinline void flush_cache_all(void)
812 {
813 	unsigned long flags;
814 
815 	local_irq_save(flags);
816 
817 	__ic_entire_inv();
818 	__dc_entire_op(OP_FLUSH_N_INV);
819 
820 	local_irq_restore(flags);
821 
822 }
823 
824 #ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
825 
826 void flush_cache_mm(struct mm_struct *mm)
827 {
828 	flush_cache_all();
829 }
830 
831 void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
832 		      unsigned long pfn)
833 {
834 	unsigned int paddr = pfn << PAGE_SHIFT;
835 
836 	u_vaddr &= PAGE_MASK;
837 
838 	__flush_dcache_page(paddr, u_vaddr);
839 
840 	if (vma->vm_flags & VM_EXEC)
841 		__inv_icache_page(paddr, u_vaddr);
842 }
843 
844 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
845 		       unsigned long end)
846 {
847 	flush_cache_all();
848 }
849 
850 void flush_anon_page(struct vm_area_struct *vma, struct page *page,
851 		     unsigned long u_vaddr)
852 {
853 	/* TBD: do we really need to clear the kernel mapping */
854 	__flush_dcache_page(page_address(page), u_vaddr);
855 	__flush_dcache_page(page_address(page), page_address(page));
856 
857 }
858 
859 #endif
860 
861 void copy_user_highpage(struct page *to, struct page *from,
862 	unsigned long u_vaddr, struct vm_area_struct *vma)
863 {
864 	void *kfrom = kmap_atomic(from);
865 	void *kto = kmap_atomic(to);
866 	int clean_src_k_mappings = 0;
867 
868 	/*
869 	 * If SRC page was already mapped in userspace AND it's U-mapping is
870 	 * not congruent with K-mapping, sync former to physical page so that
871 	 * K-mapping in memcpy below, sees the right data
872 	 *
873 	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
874 	 * equally valid for SRC page as well
875 	 *
876 	 * For !VIPT cache, all of this gets compiled out as
877 	 * addr_not_cache_congruent() is 0
878 	 */
879 	if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
880 		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
881 		clean_src_k_mappings = 1;
882 	}
883 
884 	copy_page(kto, kfrom);
885 
886 	/*
887 	 * Mark DST page K-mapping as dirty for a later finalization by
888 	 * update_mmu_cache(). Although the finalization could have been done
889 	 * here as well (given that both vaddr/paddr are available).
890 	 * But update_mmu_cache() already has code to do that for other
891 	 * non copied user pages (e.g. read faults which wire in pagecache page
892 	 * directly).
893 	 */
894 	clear_bit(PG_dc_clean, &to->flags);
895 
896 	/*
897 	 * if SRC was already usermapped and non-congruent to kernel mapping
898 	 * sync the kernel mapping back to physical page
899 	 */
900 	if (clean_src_k_mappings) {
901 		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
902 		set_bit(PG_dc_clean, &from->flags);
903 	} else {
904 		clear_bit(PG_dc_clean, &from->flags);
905 	}
906 
907 	kunmap_atomic(kto);
908 	kunmap_atomic(kfrom);
909 }
910 
911 void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
912 {
913 	clear_page(to);
914 	clear_bit(PG_dc_clean, &page->flags);
915 }
916 
917 
918 /**********************************************************************
919  * Explicit Cache flush request from user space via syscall
920  * Needed for JITs which generate code on the fly
921  */
922 SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
923 {
924 	/* TBD: optimize this */
925 	flush_cache_all();
926 	return 0;
927 }
928 
929 void arc_cache_init(void)
930 {
931 	unsigned int __maybe_unused cpu = smp_processor_id();
932 	char str[256];
933 
934 	printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
935 
936 	/*
937 	 * Only master CPU needs to execute rest of function:
938 	 *  - Assume SMP so all cores will have same cache config so
939 	 *    any geomtry checks will be same for all
940 	 *  - IOC setup / dma callbacks only need to be setup once
941 	 */
942 	if (cpu)
943 		return;
944 
945 	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
946 		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
947 
948 		if (!ic->ver)
949 			panic("cache support enabled but non-existent cache\n");
950 
951 		if (ic->line_len != L1_CACHE_BYTES)
952 			panic("ICache line [%d] != kernel Config [%d]",
953 			      ic->line_len, L1_CACHE_BYTES);
954 
955 		if (ic->ver != CONFIG_ARC_MMU_VER)
956 			panic("Cache ver [%d] doesn't match MMU ver [%d]\n",
957 			      ic->ver, CONFIG_ARC_MMU_VER);
958 
959 		/*
960 		 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
961 		 * pair to provide vaddr/paddr respectively, just as in MMU v3
962 		 */
963 		if (is_isa_arcv2() && ic->alias)
964 			_cache_line_loop_ic_fn = __cache_line_loop_v3;
965 		else
966 			_cache_line_loop_ic_fn = __cache_line_loop;
967 	}
968 
969 	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
970 		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
971 
972 		if (!dc->ver)
973 			panic("cache support enabled but non-existent cache\n");
974 
975 		if (dc->line_len != L1_CACHE_BYTES)
976 			panic("DCache line [%d] != kernel Config [%d]",
977 			      dc->line_len, L1_CACHE_BYTES);
978 
979 		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
980 		if (is_isa_arcompact()) {
981 			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
982 
983 			if (dc->alias && !handled)
984 				panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
985 			else if (!dc->alias && handled)
986 				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
987 		}
988 	}
989 
990 	if (is_isa_arcv2() && l2_line_sz && !slc_enable) {
991 
992 		/* IM set : flush before invalidate */
993 		write_aux_reg(ARC_REG_SLC_CTRL,
994 			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_IM);
995 
996 		write_aux_reg(ARC_REG_SLC_INVALIDATE, 1);
997 
998 		/* Important to wait for flush to complete */
999 		while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
1000 		write_aux_reg(ARC_REG_SLC_CTRL,
1001 			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_DISABLE);
1002 	}
1003 
1004 	if (is_isa_arcv2() && ioc_enable) {
1005 		/* IO coherency base - 0x8z */
1006 		write_aux_reg(ARC_REG_IO_COH_AP0_BASE, 0x80000);
1007 		/* IO coherency aperture size - 512Mb: 0x8z-0xAz */
1008 		write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, 0x11);
1009 		/* Enable partial writes */
1010 		write_aux_reg(ARC_REG_IO_COH_PARTIAL, 1);
1011 		/* Enable IO coherency */
1012 		write_aux_reg(ARC_REG_IO_COH_ENABLE, 1);
1013 
1014 		__dma_cache_wback_inv = __dma_cache_wback_inv_ioc;
1015 		__dma_cache_inv = __dma_cache_inv_ioc;
1016 		__dma_cache_wback = __dma_cache_wback_ioc;
1017 	} else if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1018 		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1019 		__dma_cache_inv = __dma_cache_inv_slc;
1020 		__dma_cache_wback = __dma_cache_wback_slc;
1021 	} else {
1022 		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1023 		__dma_cache_inv = __dma_cache_inv_l1;
1024 		__dma_cache_wback = __dma_cache_wback_l1;
1025 	}
1026 }
1027