xref: /linux/arch/arc/mm/cache.c (revision b68fc09be48edbc47de1a0f3d42ef8adf6c0ac55)
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 #ifdef CONFIG_ISA_ARCV2
25 #define USE_RGN_FLSH	1
26 #endif
27 
28 static int l2_line_sz;
29 static int ioc_exists;
30 int slc_enable = 1, ioc_enable = 1;
31 unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
32 unsigned long perip_end = 0xFFFFFFFF; /* legacy value */
33 
34 void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
35 			       unsigned long sz, const int op, const int full_page);
36 
37 void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
38 void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
39 void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
40 
41 char *arc_cache_mumbojumbo(int c, char *buf, int len)
42 {
43 	int n = 0;
44 	struct cpuinfo_arc_cache *p;
45 
46 #define PR_CACHE(p, cfg, str)						\
47 	if (!(p)->line_len)						\
48 		n += scnprintf(buf + n, len - n, str"\t\t: N/A\n");	\
49 	else								\
50 		n += scnprintf(buf + n, len - n,			\
51 			str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n",	\
52 			(p)->sz_k, (p)->assoc, (p)->line_len,		\
53 			(p)->vipt ? "VIPT" : "PIPT",			\
54 			(p)->alias ? " aliasing" : "",			\
55 			IS_USED_CFG(cfg));
56 
57 	PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
58 	PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
59 
60 	p = &cpuinfo_arc700[c].slc;
61 	if (p->line_len)
62 		n += scnprintf(buf + n, len - n,
63 			       "SLC\t\t: %uK, %uB Line%s\n",
64 			       p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
65 
66 	n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
67 		       perip_base,
68 		       IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency (per-device) "));
69 
70 	return buf;
71 }
72 
73 /*
74  * Read the Cache Build Confuration Registers, Decode them and save into
75  * the cpuinfo structure for later use.
76  * No Validation done here, simply read/convert the BCRs
77  */
78 static void read_decode_cache_bcr_arcv2(int cpu)
79 {
80 	struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
81 	struct bcr_generic sbcr;
82 
83 	struct bcr_slc_cfg {
84 #ifdef CONFIG_CPU_BIG_ENDIAN
85 		unsigned int pad:24, way:2, lsz:2, sz:4;
86 #else
87 		unsigned int sz:4, lsz:2, way:2, pad:24;
88 #endif
89 	} slc_cfg;
90 
91 	struct bcr_clust_cfg {
92 #ifdef CONFIG_CPU_BIG_ENDIAN
93 		unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
94 #else
95 		unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
96 #endif
97 	} cbcr;
98 
99 	struct bcr_volatile {
100 #ifdef CONFIG_CPU_BIG_ENDIAN
101 		unsigned int start:4, limit:4, pad:22, order:1, disable:1;
102 #else
103 		unsigned int disable:1, order:1, pad:22, limit:4, start:4;
104 #endif
105 	} vol;
106 
107 
108 	READ_BCR(ARC_REG_SLC_BCR, sbcr);
109 	if (sbcr.ver) {
110 		READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
111 		p_slc->sz_k = 128 << slc_cfg.sz;
112 		l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
113 	}
114 
115 	READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
116 	if (cbcr.c)
117 		ioc_exists = 1;
118 	else
119 		ioc_enable = 0;
120 
121 	/* HS 2.0 didn't have AUX_VOL */
122 	if (cpuinfo_arc700[cpu].core.family > 0x51) {
123 		READ_BCR(AUX_VOL, vol);
124 		perip_base = vol.start << 28;
125 		/* HS 3.0 has limit and strict-ordering fields */
126 		if (cpuinfo_arc700[cpu].core.family > 0x52)
127 			perip_end = (vol.limit << 28) - 1;
128 	}
129 }
130 
131 void read_decode_cache_bcr(void)
132 {
133 	struct cpuinfo_arc_cache *p_ic, *p_dc;
134 	unsigned int cpu = smp_processor_id();
135 	struct bcr_cache {
136 #ifdef CONFIG_CPU_BIG_ENDIAN
137 		unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
138 #else
139 		unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
140 #endif
141 	} ibcr, dbcr;
142 
143 	p_ic = &cpuinfo_arc700[cpu].icache;
144 	READ_BCR(ARC_REG_IC_BCR, ibcr);
145 
146 	if (!ibcr.ver)
147 		goto dc_chk;
148 
149 	if (ibcr.ver <= 3) {
150 		BUG_ON(ibcr.config != 3);
151 		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
152 	} else if (ibcr.ver >= 4) {
153 		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
154 	}
155 
156 	p_ic->line_len = 8 << ibcr.line_len;
157 	p_ic->sz_k = 1 << (ibcr.sz - 1);
158 	p_ic->vipt = 1;
159 	p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
160 
161 dc_chk:
162 	p_dc = &cpuinfo_arc700[cpu].dcache;
163 	READ_BCR(ARC_REG_DC_BCR, dbcr);
164 
165 	if (!dbcr.ver)
166 		goto slc_chk;
167 
168 	if (dbcr.ver <= 3) {
169 		BUG_ON(dbcr.config != 2);
170 		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
171 		p_dc->vipt = 1;
172 		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
173 	} else if (dbcr.ver >= 4) {
174 		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
175 		p_dc->vipt = 0;
176 		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
177 	}
178 
179 	p_dc->line_len = 16 << dbcr.line_len;
180 	p_dc->sz_k = 1 << (dbcr.sz - 1);
181 
182 slc_chk:
183 	if (is_isa_arcv2())
184                 read_decode_cache_bcr_arcv2(cpu);
185 }
186 
187 /*
188  * Line Operation on {I,D}-Cache
189  */
190 
191 #define OP_INV		0x1
192 #define OP_FLUSH	0x2
193 #define OP_FLUSH_N_INV	0x3
194 #define OP_INV_IC	0x4
195 
196 /*
197  *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
198  *
199  * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
200  * The orig Cache Management Module "CDU" only required paddr to invalidate a
201  * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
202  * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
203  * the exact same line.
204  *
205  * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
206  * paddr alone could not be used to correctly index the cache.
207  *
208  * ------------------
209  * MMU v1/v2 (Fixed Page Size 8k)
210  * ------------------
211  * The solution was to provide CDU with these additonal vaddr bits. These
212  * would be bits [x:13], x would depend on cache-geometry, 13 comes from
213  * standard page size of 8k.
214  * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
215  * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
216  * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
217  * represent the offset within cache-line. The adv of using this "clumsy"
218  * interface for additional info was no new reg was needed in CDU programming
219  * model.
220  *
221  * 17:13 represented the max num of bits passable, actual bits needed were
222  * fewer, based on the num-of-aliases possible.
223  * -for 2 alias possibility, only bit 13 needed (32K cache)
224  * -for 4 alias possibility, bits 14:13 needed (64K cache)
225  *
226  * ------------------
227  * MMU v3
228  * ------------------
229  * This ver of MMU supports variable page sizes (1k-16k): although Linux will
230  * only support 8k (default), 16k and 4k.
231  * However from hardware perspective, smaller page sizes aggravate aliasing
232  * meaning more vaddr bits needed to disambiguate the cache-line-op ;
233  * the existing scheme of piggybacking won't work for certain configurations.
234  * Two new registers IC_PTAG and DC_PTAG inttoduced.
235  * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
236  */
237 
238 static inline
239 void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
240 			  unsigned long sz, const int op, const int full_page)
241 {
242 	unsigned int aux_cmd;
243 	int num_lines;
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  *  - ARC700 programming model requires paddr and vaddr be passed in seperate
278  *    AUX registers (*_IV*L and *_PTAG respectively) irrespective of whether the
279  *    caches actually alias or not.
280  * -  For HS38, only the aliasing I-cache configuration uses the PTAG reg
281  *    (non aliasing I-cache version doesn't; while D-cache can't possibly alias)
282  */
283 static inline
284 void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
285 			  unsigned long sz, const int op, const int full_page)
286 {
287 	unsigned int aux_cmd, aux_tag;
288 	int num_lines;
289 
290 	if (op == OP_INV_IC) {
291 		aux_cmd = ARC_REG_IC_IVIL;
292 		aux_tag = ARC_REG_IC_PTAG;
293 	} else {
294 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
295 		aux_tag = ARC_REG_DC_PTAG;
296 	}
297 
298 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
299 	 * and have @paddr - aligned to cache line and integral @num_lines.
300 	 * This however can be avoided for page sized since:
301 	 *  -@paddr will be cache-line aligned already (being page aligned)
302 	 *  -@sz will be integral multiple of line size (being page sized).
303 	 */
304 	if (!full_page) {
305 		sz += paddr & ~CACHE_LINE_MASK;
306 		paddr &= CACHE_LINE_MASK;
307 		vaddr &= CACHE_LINE_MASK;
308 	}
309 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
310 
311 	/*
312 	 * MMUv3, cache ops require paddr in PTAG reg
313 	 * if V-P const for loop, PTAG can be written once outside loop
314 	 */
315 	if (full_page)
316 		write_aux_reg(aux_tag, paddr);
317 
318 	/*
319 	 * This is technically for MMU v4, using the MMU v3 programming model
320 	 * Special work for HS38 aliasing I-cache configuration with PAE40
321 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
322 	 *   - (and needs to be written before the lower 32 bits)
323 	 * Note that PTAG_HI is hoisted outside the line loop
324 	 */
325 	if (is_pae40_enabled() && op == OP_INV_IC)
326 		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
327 
328 	while (num_lines-- > 0) {
329 		if (!full_page) {
330 			write_aux_reg(aux_tag, paddr);
331 			paddr += L1_CACHE_BYTES;
332 		}
333 
334 		write_aux_reg(aux_cmd, vaddr);
335 		vaddr += L1_CACHE_BYTES;
336 	}
337 }
338 
339 #ifndef USE_RGN_FLSH
340 
341 /*
342  * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
343  * Here's how cache ops are implemented
344  *
345  *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
346  *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
347  *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
348  *    respectively, similar to MMU v3 programming model, hence
349  *    __cache_line_loop_v3() is used)
350  *
351  * If PAE40 is enabled, independent of aliasing considerations, the higher bits
352  * needs to be written into PTAG_HI
353  */
354 static inline
355 void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
356 			  unsigned long sz, const int op, const int full_page)
357 {
358 	unsigned int aux_cmd;
359 	int num_lines;
360 
361 	if (op == OP_INV_IC) {
362 		aux_cmd = ARC_REG_IC_IVIL;
363 	} else {
364 		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
365 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
366 	}
367 
368 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
369 	 * and have @paddr - aligned to cache line and integral @num_lines.
370 	 * This however can be avoided for page sized since:
371 	 *  -@paddr will be cache-line aligned already (being page aligned)
372 	 *  -@sz will be integral multiple of line size (being page sized).
373 	 */
374 	if (!full_page) {
375 		sz += paddr & ~CACHE_LINE_MASK;
376 		paddr &= CACHE_LINE_MASK;
377 	}
378 
379 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
380 
381 	/*
382 	 * For HS38 PAE40 configuration
383 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
384 	 *   - (and needs to be written before the lower 32 bits)
385 	 */
386 	if (is_pae40_enabled()) {
387 		if (op == OP_INV_IC)
388 			/*
389 			 * Non aliasing I-cache in HS38,
390 			 * aliasing I-cache handled in __cache_line_loop_v3()
391 			 */
392 			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
393 		else
394 			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
395 	}
396 
397 	while (num_lines-- > 0) {
398 		write_aux_reg(aux_cmd, paddr);
399 		paddr += L1_CACHE_BYTES;
400 	}
401 }
402 
403 #else
404 
405 /*
406  * optimized flush operation which takes a region as opposed to iterating per line
407  */
408 static inline
409 void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
410 			  unsigned long sz, const int op, const int full_page)
411 {
412 	unsigned int s, e;
413 
414 	/* Only for Non aliasing I-cache in HS38 */
415 	if (op == OP_INV_IC) {
416 		s = ARC_REG_IC_IVIR;
417 		e = ARC_REG_IC_ENDR;
418 	} else {
419 		s = ARC_REG_DC_STARTR;
420 		e = ARC_REG_DC_ENDR;
421 	}
422 
423 	if (!full_page) {
424 		/* for any leading gap between @paddr and start of cache line */
425 		sz += paddr & ~CACHE_LINE_MASK;
426 		paddr &= CACHE_LINE_MASK;
427 
428 		/*
429 		 *  account for any trailing gap to end of cache line
430 		 *  this is equivalent to DIV_ROUND_UP() in line ops above
431 		 */
432 		sz += L1_CACHE_BYTES - 1;
433 	}
434 
435 	if (is_pae40_enabled()) {
436 		/* TBD: check if crossing 4TB boundary */
437 		if (op == OP_INV_IC)
438 			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
439 		else
440 			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
441 	}
442 
443 	/* ENDR needs to be set ahead of START */
444 	write_aux_reg(e, paddr + sz);	/* ENDR is exclusive */
445 	write_aux_reg(s, paddr);
446 
447 	/* caller waits on DC_CTRL.FS */
448 }
449 
450 #endif
451 
452 #if (CONFIG_ARC_MMU_VER < 3)
453 #define __cache_line_loop	__cache_line_loop_v2
454 #elif (CONFIG_ARC_MMU_VER == 3)
455 #define __cache_line_loop	__cache_line_loop_v3
456 #elif (CONFIG_ARC_MMU_VER > 3)
457 #define __cache_line_loop	__cache_line_loop_v4
458 #endif
459 
460 #ifdef CONFIG_ARC_HAS_DCACHE
461 
462 /***************************************************************
463  * Machine specific helpers for Entire D-Cache or Per Line ops
464  */
465 
466 #ifndef USE_RGN_FLSH
467 /*
468  * this version avoids extra read/write of DC_CTRL for flush or invalid ops
469  * in the non region flush regime (such as for ARCompact)
470  */
471 static inline void __before_dc_op(const int op)
472 {
473 	if (op == OP_FLUSH_N_INV) {
474 		/* Dcache provides 2 cmd: FLUSH or INV
475 		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
476 		 * flush-n-inv is achieved by INV cmd but with IM=1
477 		 * So toggle INV sub-mode depending on op request and default
478 		 */
479 		const unsigned int ctl = ARC_REG_DC_CTRL;
480 		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
481 	}
482 }
483 
484 #else
485 
486 static inline void __before_dc_op(const int op)
487 {
488 	const unsigned int ctl = ARC_REG_DC_CTRL;
489 	unsigned int val = read_aux_reg(ctl);
490 
491 	if (op == OP_FLUSH_N_INV) {
492 		val |= DC_CTRL_INV_MODE_FLUSH;
493 	}
494 
495 	if (op != OP_INV_IC) {
496 		/*
497 		 * Flush / Invalidate is provided by DC_CTRL.RNG_OP 0 or 1
498 		 * combined Flush-n-invalidate uses DC_CTRL.IM = 1 set above
499 		 */
500 		val &= ~DC_CTRL_RGN_OP_MSK;
501 		if (op & OP_INV)
502 			val |= DC_CTRL_RGN_OP_INV;
503 	}
504 	write_aux_reg(ctl, val);
505 }
506 
507 #endif
508 
509 
510 static inline void __after_dc_op(const int op)
511 {
512 	if (op & OP_FLUSH) {
513 		const unsigned int ctl = ARC_REG_DC_CTRL;
514 		unsigned int reg;
515 
516 		/* flush / flush-n-inv both wait */
517 		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
518 			;
519 
520 		/* Switch back to default Invalidate mode */
521 		if (op == OP_FLUSH_N_INV)
522 			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
523 	}
524 }
525 
526 /*
527  * Operation on Entire D-Cache
528  * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
529  * Note that constant propagation ensures all the checks are gone
530  * in generated code
531  */
532 static inline void __dc_entire_op(const int op)
533 {
534 	int aux;
535 
536 	__before_dc_op(op);
537 
538 	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
539 		aux = ARC_REG_DC_IVDC;
540 	else
541 		aux = ARC_REG_DC_FLSH;
542 
543 	write_aux_reg(aux, 0x1);
544 
545 	__after_dc_op(op);
546 }
547 
548 static inline void __dc_disable(void)
549 {
550 	const int r = ARC_REG_DC_CTRL;
551 
552 	__dc_entire_op(OP_FLUSH_N_INV);
553 	write_aux_reg(r, read_aux_reg(r) | DC_CTRL_DIS);
554 }
555 
556 static void __dc_enable(void)
557 {
558 	const int r = ARC_REG_DC_CTRL;
559 
560 	write_aux_reg(r, read_aux_reg(r) & ~DC_CTRL_DIS);
561 }
562 
563 /* For kernel mappings cache operation: index is same as paddr */
564 #define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
565 
566 /*
567  * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
568  */
569 static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
570 				unsigned long sz, const int op)
571 {
572 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
573 	unsigned long flags;
574 
575 	local_irq_save(flags);
576 
577 	__before_dc_op(op);
578 
579 	__cache_line_loop(paddr, vaddr, sz, op, full_page);
580 
581 	__after_dc_op(op);
582 
583 	local_irq_restore(flags);
584 }
585 
586 #else
587 
588 #define __dc_entire_op(op)
589 #define __dc_disable()
590 #define __dc_enable()
591 #define __dc_line_op(paddr, vaddr, sz, op)
592 #define __dc_line_op_k(paddr, sz, op)
593 
594 #endif /* CONFIG_ARC_HAS_DCACHE */
595 
596 #ifdef CONFIG_ARC_HAS_ICACHE
597 
598 static inline void __ic_entire_inv(void)
599 {
600 	write_aux_reg(ARC_REG_IC_IVIC, 1);
601 	read_aux_reg(ARC_REG_IC_CTRL);	/* blocks */
602 }
603 
604 static inline void
605 __ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
606 			  unsigned long sz)
607 {
608 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
609 	unsigned long flags;
610 
611 	local_irq_save(flags);
612 	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC, full_page);
613 	local_irq_restore(flags);
614 }
615 
616 #ifndef CONFIG_SMP
617 
618 #define __ic_line_inv_vaddr(p, v, s)	__ic_line_inv_vaddr_local(p, v, s)
619 
620 #else
621 
622 struct ic_inv_args {
623 	phys_addr_t paddr, vaddr;
624 	int sz;
625 };
626 
627 static void __ic_line_inv_vaddr_helper(void *info)
628 {
629         struct ic_inv_args *ic_inv = info;
630 
631         __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
632 }
633 
634 static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
635 				unsigned long sz)
636 {
637 	struct ic_inv_args ic_inv = {
638 		.paddr = paddr,
639 		.vaddr = vaddr,
640 		.sz    = sz
641 	};
642 
643 	on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
644 }
645 
646 #endif	/* CONFIG_SMP */
647 
648 #else	/* !CONFIG_ARC_HAS_ICACHE */
649 
650 #define __ic_entire_inv()
651 #define __ic_line_inv_vaddr(pstart, vstart, sz)
652 
653 #endif /* CONFIG_ARC_HAS_ICACHE */
654 
655 noinline void slc_op_rgn(phys_addr_t paddr, unsigned long sz, const int op)
656 {
657 #ifdef CONFIG_ISA_ARCV2
658 	/*
659 	 * SLC is shared between all cores and concurrent aux operations from
660 	 * multiple cores need to be serialized using a spinlock
661 	 * A concurrent operation can be silently ignored and/or the old/new
662 	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
663 	 * below)
664 	 */
665 	static DEFINE_SPINLOCK(lock);
666 	unsigned long flags;
667 	unsigned int ctrl;
668 	phys_addr_t end;
669 
670 	spin_lock_irqsave(&lock, flags);
671 
672 	/*
673 	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
674 	 *  - b'000 (default) is Flush,
675 	 *  - b'001 is Invalidate if CTRL.IM == 0
676 	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
677 	 */
678 	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
679 
680 	/* Don't rely on default value of IM bit */
681 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
682 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
683 	else
684 		ctrl |= SLC_CTRL_IM;
685 
686 	if (op & OP_INV)
687 		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
688 	else
689 		ctrl &= ~SLC_CTRL_RGN_OP_INV;
690 
691 	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
692 
693 	/*
694 	 * Lower bits are ignored, no need to clip
695 	 * END needs to be setup before START (latter triggers the operation)
696 	 * END can't be same as START, so add (l2_line_sz - 1) to sz
697 	 */
698 	end = paddr + sz + l2_line_sz - 1;
699 	if (is_pae40_enabled())
700 		write_aux_reg(ARC_REG_SLC_RGN_END1, upper_32_bits(end));
701 
702 	write_aux_reg(ARC_REG_SLC_RGN_END, lower_32_bits(end));
703 
704 	if (is_pae40_enabled())
705 		write_aux_reg(ARC_REG_SLC_RGN_START1, upper_32_bits(paddr));
706 
707 	write_aux_reg(ARC_REG_SLC_RGN_START, lower_32_bits(paddr));
708 
709 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
710 	read_aux_reg(ARC_REG_SLC_CTRL);
711 
712 	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
713 
714 	spin_unlock_irqrestore(&lock, flags);
715 #endif
716 }
717 
718 noinline void slc_op_line(phys_addr_t paddr, unsigned long sz, const int op)
719 {
720 #ifdef CONFIG_ISA_ARCV2
721 	/*
722 	 * SLC is shared between all cores and concurrent aux operations from
723 	 * multiple cores need to be serialized using a spinlock
724 	 * A concurrent operation can be silently ignored and/or the old/new
725 	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
726 	 * below)
727 	 */
728 	static DEFINE_SPINLOCK(lock);
729 
730 	const unsigned long SLC_LINE_MASK = ~(l2_line_sz - 1);
731 	unsigned int ctrl, cmd;
732 	unsigned long flags;
733 	int num_lines;
734 
735 	spin_lock_irqsave(&lock, flags);
736 
737 	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
738 
739 	/* Don't rely on default value of IM bit */
740 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
741 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
742 	else
743 		ctrl |= SLC_CTRL_IM;
744 
745 	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
746 
747 	cmd = op & OP_INV ? ARC_AUX_SLC_IVDL : ARC_AUX_SLC_FLDL;
748 
749 	sz += paddr & ~SLC_LINE_MASK;
750 	paddr &= SLC_LINE_MASK;
751 
752 	num_lines = DIV_ROUND_UP(sz, l2_line_sz);
753 
754 	while (num_lines-- > 0) {
755 		write_aux_reg(cmd, paddr);
756 		paddr += l2_line_sz;
757 	}
758 
759 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
760 	read_aux_reg(ARC_REG_SLC_CTRL);
761 
762 	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
763 
764 	spin_unlock_irqrestore(&lock, flags);
765 #endif
766 }
767 
768 #define slc_op(paddr, sz, op)	slc_op_rgn(paddr, sz, op)
769 
770 noinline static void slc_entire_op(const int op)
771 {
772 	unsigned int ctrl, r = ARC_REG_SLC_CTRL;
773 
774 	ctrl = read_aux_reg(r);
775 
776 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
777 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
778 	else
779 		ctrl |= SLC_CTRL_IM;
780 
781 	write_aux_reg(r, ctrl);
782 
783 	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
784 		write_aux_reg(ARC_REG_SLC_INVALIDATE, 0x1);
785 	else
786 		write_aux_reg(ARC_REG_SLC_FLUSH, 0x1);
787 
788 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
789 	read_aux_reg(r);
790 
791 	/* Important to wait for flush to complete */
792 	while (read_aux_reg(r) & SLC_CTRL_BUSY);
793 }
794 
795 static inline void arc_slc_disable(void)
796 {
797 	const int r = ARC_REG_SLC_CTRL;
798 
799 	slc_entire_op(OP_FLUSH_N_INV);
800 	write_aux_reg(r, read_aux_reg(r) | SLC_CTRL_DIS);
801 }
802 
803 static inline void arc_slc_enable(void)
804 {
805 	const int r = ARC_REG_SLC_CTRL;
806 
807 	write_aux_reg(r, read_aux_reg(r) & ~SLC_CTRL_DIS);
808 }
809 
810 /***********************************************************
811  * Exported APIs
812  */
813 
814 /*
815  * Handle cache congruency of kernel and userspace mappings of page when kernel
816  * writes-to/reads-from
817  *
818  * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
819  *  -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
820  *  -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
821  *  -In SMP, if hardware caches are coherent
822  *
823  * There's a corollary case, where kernel READs from a userspace mapped page.
824  * If the U-mapping is not congruent to to K-mapping, former needs flushing.
825  */
826 void flush_dcache_page(struct page *page)
827 {
828 	struct address_space *mapping;
829 
830 	if (!cache_is_vipt_aliasing()) {
831 		clear_bit(PG_dc_clean, &page->flags);
832 		return;
833 	}
834 
835 	/* don't handle anon pages here */
836 	mapping = page_mapping_file(page);
837 	if (!mapping)
838 		return;
839 
840 	/*
841 	 * pagecache page, file not yet mapped to userspace
842 	 * Make a note that K-mapping is dirty
843 	 */
844 	if (!mapping_mapped(mapping)) {
845 		clear_bit(PG_dc_clean, &page->flags);
846 	} else if (page_mapcount(page)) {
847 
848 		/* kernel reading from page with U-mapping */
849 		phys_addr_t paddr = (unsigned long)page_address(page);
850 		unsigned long vaddr = page->index << PAGE_SHIFT;
851 
852 		if (addr_not_cache_congruent(paddr, vaddr))
853 			__flush_dcache_page(paddr, vaddr);
854 	}
855 }
856 EXPORT_SYMBOL(flush_dcache_page);
857 
858 /*
859  * DMA ops for systems with L1 cache only
860  * Make memory coherent with L1 cache by flushing/invalidating L1 lines
861  */
862 static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
863 {
864 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
865 }
866 
867 static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
868 {
869 	__dc_line_op_k(start, sz, OP_INV);
870 }
871 
872 static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
873 {
874 	__dc_line_op_k(start, sz, OP_FLUSH);
875 }
876 
877 /*
878  * DMA ops for systems with both L1 and L2 caches, but without IOC
879  * Both L1 and L2 lines need to be explicitly flushed/invalidated
880  */
881 static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
882 {
883 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
884 	slc_op(start, sz, OP_FLUSH_N_INV);
885 }
886 
887 static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
888 {
889 	__dc_line_op_k(start, sz, OP_INV);
890 	slc_op(start, sz, OP_INV);
891 }
892 
893 static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
894 {
895 	__dc_line_op_k(start, sz, OP_FLUSH);
896 	slc_op(start, sz, OP_FLUSH);
897 }
898 
899 /*
900  * Exported DMA API
901  */
902 void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
903 {
904 	__dma_cache_wback_inv(start, sz);
905 }
906 EXPORT_SYMBOL(dma_cache_wback_inv);
907 
908 void dma_cache_inv(phys_addr_t start, unsigned long sz)
909 {
910 	__dma_cache_inv(start, sz);
911 }
912 EXPORT_SYMBOL(dma_cache_inv);
913 
914 void dma_cache_wback(phys_addr_t start, unsigned long sz)
915 {
916 	__dma_cache_wback(start, sz);
917 }
918 EXPORT_SYMBOL(dma_cache_wback);
919 
920 /*
921  * This is API for making I/D Caches consistent when modifying
922  * kernel code (loadable modules, kprobes, kgdb...)
923  * This is called on insmod, with kernel virtual address for CODE of
924  * the module. ARC cache maintenance ops require PHY address thus we
925  * need to convert vmalloc addr to PHY addr
926  */
927 void flush_icache_range(unsigned long kstart, unsigned long kend)
928 {
929 	unsigned int tot_sz;
930 
931 	WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
932 
933 	/* Shortcut for bigger flush ranges.
934 	 * Here we don't care if this was kernel virtual or phy addr
935 	 */
936 	tot_sz = kend - kstart;
937 	if (tot_sz > PAGE_SIZE) {
938 		flush_cache_all();
939 		return;
940 	}
941 
942 	/* Case: Kernel Phy addr (0x8000_0000 onwards) */
943 	if (likely(kstart > PAGE_OFFSET)) {
944 		/*
945 		 * The 2nd arg despite being paddr will be used to index icache
946 		 * This is OK since no alternate virtual mappings will exist
947 		 * given the callers for this case: kprobe/kgdb in built-in
948 		 * kernel code only.
949 		 */
950 		__sync_icache_dcache(kstart, kstart, kend - kstart);
951 		return;
952 	}
953 
954 	/*
955 	 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
956 	 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
957 	 *     handling of kernel vaddr.
958 	 *
959 	 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
960 	 *     it still needs to handle  a 2 page scenario, where the range
961 	 *     straddles across 2 virtual pages and hence need for loop
962 	 */
963 	while (tot_sz > 0) {
964 		unsigned int off, sz;
965 		unsigned long phy, pfn;
966 
967 		off = kstart % PAGE_SIZE;
968 		pfn = vmalloc_to_pfn((void *)kstart);
969 		phy = (pfn << PAGE_SHIFT) + off;
970 		sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
971 		__sync_icache_dcache(phy, kstart, sz);
972 		kstart += sz;
973 		tot_sz -= sz;
974 	}
975 }
976 EXPORT_SYMBOL(flush_icache_range);
977 
978 /*
979  * General purpose helper to make I and D cache lines consistent.
980  * @paddr is phy addr of region
981  * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
982  *    However in one instance, when called by kprobe (for a breakpt in
983  *    builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
984  *    use a paddr to index the cache (despite VIPT). This is fine since since a
985  *    builtin kernel page will not have any virtual mappings.
986  *    kprobe on loadable module will be kernel vaddr.
987  */
988 void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
989 {
990 	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
991 	__ic_line_inv_vaddr(paddr, vaddr, len);
992 }
993 
994 /* wrapper to compile time eliminate alignment checks in flush loop */
995 void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
996 {
997 	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
998 }
999 
1000 /*
1001  * wrapper to clearout kernel or userspace mappings of a page
1002  * For kernel mappings @vaddr == @paddr
1003  */
1004 void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
1005 {
1006 	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
1007 }
1008 
1009 noinline void flush_cache_all(void)
1010 {
1011 	unsigned long flags;
1012 
1013 	local_irq_save(flags);
1014 
1015 	__ic_entire_inv();
1016 	__dc_entire_op(OP_FLUSH_N_INV);
1017 
1018 	local_irq_restore(flags);
1019 
1020 }
1021 
1022 #ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
1023 
1024 void flush_cache_mm(struct mm_struct *mm)
1025 {
1026 	flush_cache_all();
1027 }
1028 
1029 void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
1030 		      unsigned long pfn)
1031 {
1032 	phys_addr_t paddr = pfn << PAGE_SHIFT;
1033 
1034 	u_vaddr &= PAGE_MASK;
1035 
1036 	__flush_dcache_page(paddr, u_vaddr);
1037 
1038 	if (vma->vm_flags & VM_EXEC)
1039 		__inv_icache_page(paddr, u_vaddr);
1040 }
1041 
1042 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
1043 		       unsigned long end)
1044 {
1045 	flush_cache_all();
1046 }
1047 
1048 void flush_anon_page(struct vm_area_struct *vma, struct page *page,
1049 		     unsigned long u_vaddr)
1050 {
1051 	/* TBD: do we really need to clear the kernel mapping */
1052 	__flush_dcache_page((phys_addr_t)page_address(page), u_vaddr);
1053 	__flush_dcache_page((phys_addr_t)page_address(page),
1054 			    (phys_addr_t)page_address(page));
1055 
1056 }
1057 
1058 #endif
1059 
1060 void copy_user_highpage(struct page *to, struct page *from,
1061 	unsigned long u_vaddr, struct vm_area_struct *vma)
1062 {
1063 	void *kfrom = kmap_atomic(from);
1064 	void *kto = kmap_atomic(to);
1065 	int clean_src_k_mappings = 0;
1066 
1067 	/*
1068 	 * If SRC page was already mapped in userspace AND it's U-mapping is
1069 	 * not congruent with K-mapping, sync former to physical page so that
1070 	 * K-mapping in memcpy below, sees the right data
1071 	 *
1072 	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
1073 	 * equally valid for SRC page as well
1074 	 *
1075 	 * For !VIPT cache, all of this gets compiled out as
1076 	 * addr_not_cache_congruent() is 0
1077 	 */
1078 	if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
1079 		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
1080 		clean_src_k_mappings = 1;
1081 	}
1082 
1083 	copy_page(kto, kfrom);
1084 
1085 	/*
1086 	 * Mark DST page K-mapping as dirty for a later finalization by
1087 	 * update_mmu_cache(). Although the finalization could have been done
1088 	 * here as well (given that both vaddr/paddr are available).
1089 	 * But update_mmu_cache() already has code to do that for other
1090 	 * non copied user pages (e.g. read faults which wire in pagecache page
1091 	 * directly).
1092 	 */
1093 	clear_bit(PG_dc_clean, &to->flags);
1094 
1095 	/*
1096 	 * if SRC was already usermapped and non-congruent to kernel mapping
1097 	 * sync the kernel mapping back to physical page
1098 	 */
1099 	if (clean_src_k_mappings) {
1100 		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
1101 		set_bit(PG_dc_clean, &from->flags);
1102 	} else {
1103 		clear_bit(PG_dc_clean, &from->flags);
1104 	}
1105 
1106 	kunmap_atomic(kto);
1107 	kunmap_atomic(kfrom);
1108 }
1109 
1110 void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
1111 {
1112 	clear_page(to);
1113 	clear_bit(PG_dc_clean, &page->flags);
1114 }
1115 
1116 
1117 /**********************************************************************
1118  * Explicit Cache flush request from user space via syscall
1119  * Needed for JITs which generate code on the fly
1120  */
1121 SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
1122 {
1123 	/* TBD: optimize this */
1124 	flush_cache_all();
1125 	return 0;
1126 }
1127 
1128 /*
1129  * IO-Coherency (IOC) setup rules:
1130  *
1131  * 1. Needs to be at system level, so only once by Master core
1132  *    Non-Masters need not be accessing caches at that time
1133  *    - They are either HALT_ON_RESET and kick started much later or
1134  *    - if run on reset, need to ensure that arc_platform_smp_wait_to_boot()
1135  *      doesn't perturb caches or coherency unit
1136  *
1137  * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC,
1138  *    otherwise any straggler data might behave strangely post IOC enabling
1139  *
1140  * 3. All Caches need to be disabled when setting up IOC to elide any in-flight
1141  *    Coherency transactions
1142  */
1143 noinline void __init arc_ioc_setup(void)
1144 {
1145 	unsigned int ioc_base, mem_sz;
1146 
1147 	/*
1148 	 * As for today we don't support both IOC and ZONE_HIGHMEM enabled
1149 	 * simultaneously. This happens because as of today IOC aperture covers
1150 	 * only ZONE_NORMAL (low mem) and any dma transactions outside this
1151 	 * region won't be HW coherent.
1152 	 * If we want to use both IOC and ZONE_HIGHMEM we can use
1153 	 * bounce_buffer to handle dma transactions to HIGHMEM.
1154 	 * Also it is possible to modify dma_direct cache ops or increase IOC
1155 	 * aperture size if we are planning to use HIGHMEM without PAE.
1156 	 */
1157 	if (IS_ENABLED(CONFIG_HIGHMEM))
1158 		panic("IOC and HIGHMEM can't be used simultaneously");
1159 
1160 	/* Flush + invalidate + disable L1 dcache */
1161 	__dc_disable();
1162 
1163 	/* Flush + invalidate SLC */
1164 	if (read_aux_reg(ARC_REG_SLC_BCR))
1165 		slc_entire_op(OP_FLUSH_N_INV);
1166 
1167 	/*
1168 	 * currently IOC Aperture covers entire DDR
1169 	 * TBD: fix for PGU + 1GB of low mem
1170 	 * TBD: fix for PAE
1171 	 */
1172 	mem_sz = arc_get_mem_sz();
1173 
1174 	if (!is_power_of_2(mem_sz) || mem_sz < 4096)
1175 		panic("IOC Aperture size must be power of 2 larger than 4KB");
1176 
1177 	/*
1178 	 * IOC Aperture size decoded as 2 ^ (SIZE + 2) KB,
1179 	 * so setting 0x11 implies 512MB, 0x12 implies 1GB...
1180 	 */
1181 	write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, order_base_2(mem_sz >> 10) - 2);
1182 
1183 	/* for now assume kernel base is start of IOC aperture */
1184 	ioc_base = CONFIG_LINUX_RAM_BASE;
1185 
1186 	if (ioc_base % mem_sz != 0)
1187 		panic("IOC Aperture start must be aligned to the size of the aperture");
1188 
1189 	write_aux_reg(ARC_REG_IO_COH_AP0_BASE, ioc_base >> 12);
1190 	write_aux_reg(ARC_REG_IO_COH_PARTIAL, 1);
1191 	write_aux_reg(ARC_REG_IO_COH_ENABLE, 1);
1192 
1193 	/* Re-enable L1 dcache */
1194 	__dc_enable();
1195 }
1196 
1197 /*
1198  * Cache related boot time checks/setups only needed on master CPU:
1199  *  - Geometry checks (kernel build and hardware agree: e.g. L1_CACHE_BYTES)
1200  *    Assume SMP only, so all cores will have same cache config. A check on
1201  *    one core suffices for all
1202  *  - IOC setup / dma callbacks only need to be done once
1203  */
1204 void __init arc_cache_init_master(void)
1205 {
1206 	unsigned int __maybe_unused cpu = smp_processor_id();
1207 
1208 	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
1209 		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
1210 
1211 		if (!ic->line_len)
1212 			panic("cache support enabled but non-existent cache\n");
1213 
1214 		if (ic->line_len != L1_CACHE_BYTES)
1215 			panic("ICache line [%d] != kernel Config [%d]",
1216 			      ic->line_len, L1_CACHE_BYTES);
1217 
1218 		/*
1219 		 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
1220 		 * pair to provide vaddr/paddr respectively, just as in MMU v3
1221 		 */
1222 		if (is_isa_arcv2() && ic->alias)
1223 			_cache_line_loop_ic_fn = __cache_line_loop_v3;
1224 		else
1225 			_cache_line_loop_ic_fn = __cache_line_loop;
1226 	}
1227 
1228 	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
1229 		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
1230 
1231 		if (!dc->line_len)
1232 			panic("cache support enabled but non-existent cache\n");
1233 
1234 		if (dc->line_len != L1_CACHE_BYTES)
1235 			panic("DCache line [%d] != kernel Config [%d]",
1236 			      dc->line_len, L1_CACHE_BYTES);
1237 
1238 		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
1239 		if (is_isa_arcompact()) {
1240 			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
1241 			int num_colors = dc->sz_k/dc->assoc/TO_KB(PAGE_SIZE);
1242 
1243 			if (dc->alias) {
1244 				if (!handled)
1245 					panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1246 				if (CACHE_COLORS_NUM != num_colors)
1247 					panic("CACHE_COLORS_NUM not optimized for config\n");
1248 			} else if (!dc->alias && handled) {
1249 				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1250 			}
1251 		}
1252 	}
1253 
1254 	/*
1255 	 * Check that SMP_CACHE_BYTES (and hence ARCH_DMA_MINALIGN) is larger
1256 	 * or equal to any cache line length.
1257 	 */
1258 	BUILD_BUG_ON_MSG(L1_CACHE_BYTES > SMP_CACHE_BYTES,
1259 			 "SMP_CACHE_BYTES must be >= any cache line length");
1260 	if (is_isa_arcv2() && (l2_line_sz > SMP_CACHE_BYTES))
1261 		panic("L2 Cache line [%d] > kernel Config [%d]\n",
1262 		      l2_line_sz, SMP_CACHE_BYTES);
1263 
1264 	/* Note that SLC disable not formally supported till HS 3.0 */
1265 	if (is_isa_arcv2() && l2_line_sz && !slc_enable)
1266 		arc_slc_disable();
1267 
1268 	if (is_isa_arcv2() && ioc_enable)
1269 		arc_ioc_setup();
1270 
1271 	if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1272 		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1273 		__dma_cache_inv = __dma_cache_inv_slc;
1274 		__dma_cache_wback = __dma_cache_wback_slc;
1275 	} else {
1276 		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1277 		__dma_cache_inv = __dma_cache_inv_l1;
1278 		__dma_cache_wback = __dma_cache_wback_l1;
1279 	}
1280 	/*
1281 	 * In case of IOC (say IOC+SLC case), pointers above could still be set
1282 	 * but end up not being relevant as the first function in chain is not
1283 	 * called at all for @dma_direct_ops
1284 	 *     arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
1285 	 */
1286 }
1287 
1288 void __ref arc_cache_init(void)
1289 {
1290 	unsigned int __maybe_unused cpu = smp_processor_id();
1291 	char str[256];
1292 
1293 	pr_info("%s", arc_cache_mumbojumbo(0, str, sizeof(str)));
1294 
1295 	if (!cpu)
1296 		arc_cache_init_master();
1297 
1298 	/*
1299 	 * In PAE regime, TLB and cache maintenance ops take wider addresses
1300 	 * And even if PAE is not enabled in kernel, the upper 32-bits still need
1301 	 * to be zeroed to keep the ops sane.
1302 	 * As an optimization for more common !PAE enabled case, zero them out
1303 	 * once at init, rather than checking/setting to 0 for every runtime op
1304 	 */
1305 	if (is_isa_arcv2() && pae40_exist_but_not_enab()) {
1306 
1307 		if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE))
1308 			write_aux_reg(ARC_REG_IC_PTAG_HI, 0);
1309 
1310 		if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE))
1311 			write_aux_reg(ARC_REG_DC_PTAG_HI, 0);
1312 
1313 		if (l2_line_sz) {
1314 			write_aux_reg(ARC_REG_SLC_RGN_END1, 0);
1315 			write_aux_reg(ARC_REG_SLC_RGN_START1, 0);
1316 		}
1317 	}
1318 }
1319