xref: /linux/arch/arc/mm/cache.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 
119 		/*
120 		 * As for today we don't support both IOC and ZONE_HIGHMEM enabled
121 		 * simultaneously. This happens because as of today IOC aperture covers
122 		 * only ZONE_NORMAL (low mem) and any dma transactions outside this
123 		 * region won't be HW coherent.
124 		 * If we want to use both IOC and ZONE_HIGHMEM we can use
125 		 * bounce_buffer to handle dma transactions to HIGHMEM.
126 		 * Also it is possible to modify dma_direct cache ops or increase IOC
127 		 * aperture size if we are planning to use HIGHMEM without PAE.
128 		 */
129 		if (IS_ENABLED(CONFIG_HIGHMEM) || is_pae40_enabled())
130 			ioc_enable = 0;
131 	} else {
132 		ioc_enable = 0;
133 	}
134 
135 	/* HS 2.0 didn't have AUX_VOL */
136 	if (cpuinfo_arc700[cpu].core.family > 0x51) {
137 		READ_BCR(AUX_VOL, vol);
138 		perip_base = vol.start << 28;
139 		/* HS 3.0 has limit and strict-ordering fields */
140 		if (cpuinfo_arc700[cpu].core.family > 0x52)
141 			perip_end = (vol.limit << 28) - 1;
142 	}
143 }
144 
145 void read_decode_cache_bcr(void)
146 {
147 	struct cpuinfo_arc_cache *p_ic, *p_dc;
148 	unsigned int cpu = smp_processor_id();
149 	struct bcr_cache {
150 #ifdef CONFIG_CPU_BIG_ENDIAN
151 		unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
152 #else
153 		unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
154 #endif
155 	} ibcr, dbcr;
156 
157 	p_ic = &cpuinfo_arc700[cpu].icache;
158 	READ_BCR(ARC_REG_IC_BCR, ibcr);
159 
160 	if (!ibcr.ver)
161 		goto dc_chk;
162 
163 	if (ibcr.ver <= 3) {
164 		BUG_ON(ibcr.config != 3);
165 		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
166 	} else if (ibcr.ver >= 4) {
167 		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
168 	}
169 
170 	p_ic->line_len = 8 << ibcr.line_len;
171 	p_ic->sz_k = 1 << (ibcr.sz - 1);
172 	p_ic->vipt = 1;
173 	p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
174 
175 dc_chk:
176 	p_dc = &cpuinfo_arc700[cpu].dcache;
177 	READ_BCR(ARC_REG_DC_BCR, dbcr);
178 
179 	if (!dbcr.ver)
180 		goto slc_chk;
181 
182 	if (dbcr.ver <= 3) {
183 		BUG_ON(dbcr.config != 2);
184 		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
185 		p_dc->vipt = 1;
186 		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
187 	} else if (dbcr.ver >= 4) {
188 		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
189 		p_dc->vipt = 0;
190 		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
191 	}
192 
193 	p_dc->line_len = 16 << dbcr.line_len;
194 	p_dc->sz_k = 1 << (dbcr.sz - 1);
195 
196 slc_chk:
197 	if (is_isa_arcv2())
198                 read_decode_cache_bcr_arcv2(cpu);
199 }
200 
201 /*
202  * Line Operation on {I,D}-Cache
203  */
204 
205 #define OP_INV		0x1
206 #define OP_FLUSH	0x2
207 #define OP_FLUSH_N_INV	0x3
208 #define OP_INV_IC	0x4
209 
210 /*
211  *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
212  *
213  * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
214  * The orig Cache Management Module "CDU" only required paddr to invalidate a
215  * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
216  * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
217  * the exact same line.
218  *
219  * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
220  * paddr alone could not be used to correctly index the cache.
221  *
222  * ------------------
223  * MMU v1/v2 (Fixed Page Size 8k)
224  * ------------------
225  * The solution was to provide CDU with these additonal vaddr bits. These
226  * would be bits [x:13], x would depend on cache-geometry, 13 comes from
227  * standard page size of 8k.
228  * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
229  * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
230  * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
231  * represent the offset within cache-line. The adv of using this "clumsy"
232  * interface for additional info was no new reg was needed in CDU programming
233  * model.
234  *
235  * 17:13 represented the max num of bits passable, actual bits needed were
236  * fewer, based on the num-of-aliases possible.
237  * -for 2 alias possibility, only bit 13 needed (32K cache)
238  * -for 4 alias possibility, bits 14:13 needed (64K cache)
239  *
240  * ------------------
241  * MMU v3
242  * ------------------
243  * This ver of MMU supports variable page sizes (1k-16k): although Linux will
244  * only support 8k (default), 16k and 4k.
245  * However from hardware perspective, smaller page sizes aggravate aliasing
246  * meaning more vaddr bits needed to disambiguate the cache-line-op ;
247  * the existing scheme of piggybacking won't work for certain configurations.
248  * Two new registers IC_PTAG and DC_PTAG inttoduced.
249  * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
250  */
251 
252 static inline
253 void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
254 			  unsigned long sz, const int op, const int full_page)
255 {
256 	unsigned int aux_cmd;
257 	int num_lines;
258 
259 	if (op == OP_INV_IC) {
260 		aux_cmd = ARC_REG_IC_IVIL;
261 	} else {
262 		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
263 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
264 	}
265 
266 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
267 	 * and have @paddr - aligned to cache line and integral @num_lines.
268 	 * This however can be avoided for page sized since:
269 	 *  -@paddr will be cache-line aligned already (being page aligned)
270 	 *  -@sz will be integral multiple of line size (being page sized).
271 	 */
272 	if (!full_page) {
273 		sz += paddr & ~CACHE_LINE_MASK;
274 		paddr &= CACHE_LINE_MASK;
275 		vaddr &= CACHE_LINE_MASK;
276 	}
277 
278 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
279 
280 	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
281 	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
282 
283 	while (num_lines-- > 0) {
284 		write_aux_reg(aux_cmd, paddr);
285 		paddr += L1_CACHE_BYTES;
286 	}
287 }
288 
289 /*
290  * For ARC700 MMUv3 I-cache and D-cache flushes
291  *  - ARC700 programming model requires paddr and vaddr be passed in seperate
292  *    AUX registers (*_IV*L and *_PTAG respectively) irrespective of whether the
293  *    caches actually alias or not.
294  * -  For HS38, only the aliasing I-cache configuration uses the PTAG reg
295  *    (non aliasing I-cache version doesn't; while D-cache can't possibly alias)
296  */
297 static inline
298 void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
299 			  unsigned long sz, const int op, const int full_page)
300 {
301 	unsigned int aux_cmd, aux_tag;
302 	int num_lines;
303 
304 	if (op == OP_INV_IC) {
305 		aux_cmd = ARC_REG_IC_IVIL;
306 		aux_tag = ARC_REG_IC_PTAG;
307 	} else {
308 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
309 		aux_tag = ARC_REG_DC_PTAG;
310 	}
311 
312 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
313 	 * and have @paddr - aligned to cache line and integral @num_lines.
314 	 * This however can be avoided for page sized since:
315 	 *  -@paddr will be cache-line aligned already (being page aligned)
316 	 *  -@sz will be integral multiple of line size (being page sized).
317 	 */
318 	if (!full_page) {
319 		sz += paddr & ~CACHE_LINE_MASK;
320 		paddr &= CACHE_LINE_MASK;
321 		vaddr &= CACHE_LINE_MASK;
322 	}
323 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
324 
325 	/*
326 	 * MMUv3, cache ops require paddr in PTAG reg
327 	 * if V-P const for loop, PTAG can be written once outside loop
328 	 */
329 	if (full_page)
330 		write_aux_reg(aux_tag, paddr);
331 
332 	/*
333 	 * This is technically for MMU v4, using the MMU v3 programming model
334 	 * Special work for HS38 aliasing I-cache configuration with PAE40
335 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
336 	 *   - (and needs to be written before the lower 32 bits)
337 	 * Note that PTAG_HI is hoisted outside the line loop
338 	 */
339 	if (is_pae40_enabled() && op == OP_INV_IC)
340 		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
341 
342 	while (num_lines-- > 0) {
343 		if (!full_page) {
344 			write_aux_reg(aux_tag, paddr);
345 			paddr += L1_CACHE_BYTES;
346 		}
347 
348 		write_aux_reg(aux_cmd, vaddr);
349 		vaddr += L1_CACHE_BYTES;
350 	}
351 }
352 
353 #ifndef USE_RGN_FLSH
354 
355 /*
356  * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
357  * Here's how cache ops are implemented
358  *
359  *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
360  *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
361  *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
362  *    respectively, similar to MMU v3 programming model, hence
363  *    __cache_line_loop_v3() is used)
364  *
365  * If PAE40 is enabled, independent of aliasing considerations, the higher bits
366  * needs to be written into PTAG_HI
367  */
368 static inline
369 void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
370 			  unsigned long sz, const int op, const int full_page)
371 {
372 	unsigned int aux_cmd;
373 	int num_lines;
374 
375 	if (op == OP_INV_IC) {
376 		aux_cmd = ARC_REG_IC_IVIL;
377 	} else {
378 		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
379 		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
380 	}
381 
382 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
383 	 * and have @paddr - aligned to cache line and integral @num_lines.
384 	 * This however can be avoided for page sized since:
385 	 *  -@paddr will be cache-line aligned already (being page aligned)
386 	 *  -@sz will be integral multiple of line size (being page sized).
387 	 */
388 	if (!full_page) {
389 		sz += paddr & ~CACHE_LINE_MASK;
390 		paddr &= CACHE_LINE_MASK;
391 	}
392 
393 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
394 
395 	/*
396 	 * For HS38 PAE40 configuration
397 	 *   - upper 8 bits of paddr need to be written into PTAG_HI
398 	 *   - (and needs to be written before the lower 32 bits)
399 	 */
400 	if (is_pae40_enabled()) {
401 		if (op == OP_INV_IC)
402 			/*
403 			 * Non aliasing I-cache in HS38,
404 			 * aliasing I-cache handled in __cache_line_loop_v3()
405 			 */
406 			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
407 		else
408 			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
409 	}
410 
411 	while (num_lines-- > 0) {
412 		write_aux_reg(aux_cmd, paddr);
413 		paddr += L1_CACHE_BYTES;
414 	}
415 }
416 
417 #else
418 
419 /*
420  * optimized flush operation which takes a region as opposed to iterating per line
421  */
422 static inline
423 void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
424 			  unsigned long sz, const int op, const int full_page)
425 {
426 	unsigned int s, e;
427 
428 	/* Only for Non aliasing I-cache in HS38 */
429 	if (op == OP_INV_IC) {
430 		s = ARC_REG_IC_IVIR;
431 		e = ARC_REG_IC_ENDR;
432 	} else {
433 		s = ARC_REG_DC_STARTR;
434 		e = ARC_REG_DC_ENDR;
435 	}
436 
437 	if (!full_page) {
438 		/* for any leading gap between @paddr and start of cache line */
439 		sz += paddr & ~CACHE_LINE_MASK;
440 		paddr &= CACHE_LINE_MASK;
441 
442 		/*
443 		 *  account for any trailing gap to end of cache line
444 		 *  this is equivalent to DIV_ROUND_UP() in line ops above
445 		 */
446 		sz += L1_CACHE_BYTES - 1;
447 	}
448 
449 	if (is_pae40_enabled()) {
450 		/* TBD: check if crossing 4TB boundary */
451 		if (op == OP_INV_IC)
452 			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
453 		else
454 			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
455 	}
456 
457 	/* ENDR needs to be set ahead of START */
458 	write_aux_reg(e, paddr + sz);	/* ENDR is exclusive */
459 	write_aux_reg(s, paddr);
460 
461 	/* caller waits on DC_CTRL.FS */
462 }
463 
464 #endif
465 
466 #if (CONFIG_ARC_MMU_VER < 3)
467 #define __cache_line_loop	__cache_line_loop_v2
468 #elif (CONFIG_ARC_MMU_VER == 3)
469 #define __cache_line_loop	__cache_line_loop_v3
470 #elif (CONFIG_ARC_MMU_VER > 3)
471 #define __cache_line_loop	__cache_line_loop_v4
472 #endif
473 
474 #ifdef CONFIG_ARC_HAS_DCACHE
475 
476 /***************************************************************
477  * Machine specific helpers for Entire D-Cache or Per Line ops
478  */
479 
480 #ifndef USE_RGN_FLSH
481 /*
482  * this version avoids extra read/write of DC_CTRL for flush or invalid ops
483  * in the non region flush regime (such as for ARCompact)
484  */
485 static inline void __before_dc_op(const int op)
486 {
487 	if (op == OP_FLUSH_N_INV) {
488 		/* Dcache provides 2 cmd: FLUSH or INV
489 		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
490 		 * flush-n-inv is achieved by INV cmd but with IM=1
491 		 * So toggle INV sub-mode depending on op request and default
492 		 */
493 		const unsigned int ctl = ARC_REG_DC_CTRL;
494 		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
495 	}
496 }
497 
498 #else
499 
500 static inline void __before_dc_op(const int op)
501 {
502 	const unsigned int ctl = ARC_REG_DC_CTRL;
503 	unsigned int val = read_aux_reg(ctl);
504 
505 	if (op == OP_FLUSH_N_INV) {
506 		val |= DC_CTRL_INV_MODE_FLUSH;
507 	}
508 
509 	if (op != OP_INV_IC) {
510 		/*
511 		 * Flush / Invalidate is provided by DC_CTRL.RNG_OP 0 or 1
512 		 * combined Flush-n-invalidate uses DC_CTRL.IM = 1 set above
513 		 */
514 		val &= ~DC_CTRL_RGN_OP_MSK;
515 		if (op & OP_INV)
516 			val |= DC_CTRL_RGN_OP_INV;
517 	}
518 	write_aux_reg(ctl, val);
519 }
520 
521 #endif
522 
523 
524 static inline void __after_dc_op(const int op)
525 {
526 	if (op & OP_FLUSH) {
527 		const unsigned int ctl = ARC_REG_DC_CTRL;
528 		unsigned int reg;
529 
530 		/* flush / flush-n-inv both wait */
531 		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
532 			;
533 
534 		/* Switch back to default Invalidate mode */
535 		if (op == OP_FLUSH_N_INV)
536 			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
537 	}
538 }
539 
540 /*
541  * Operation on Entire D-Cache
542  * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
543  * Note that constant propagation ensures all the checks are gone
544  * in generated code
545  */
546 static inline void __dc_entire_op(const int op)
547 {
548 	int aux;
549 
550 	__before_dc_op(op);
551 
552 	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
553 		aux = ARC_REG_DC_IVDC;
554 	else
555 		aux = ARC_REG_DC_FLSH;
556 
557 	write_aux_reg(aux, 0x1);
558 
559 	__after_dc_op(op);
560 }
561 
562 static inline void __dc_disable(void)
563 {
564 	const int r = ARC_REG_DC_CTRL;
565 
566 	__dc_entire_op(OP_FLUSH_N_INV);
567 	write_aux_reg(r, read_aux_reg(r) | DC_CTRL_DIS);
568 }
569 
570 static void __dc_enable(void)
571 {
572 	const int r = ARC_REG_DC_CTRL;
573 
574 	write_aux_reg(r, read_aux_reg(r) & ~DC_CTRL_DIS);
575 }
576 
577 /* For kernel mappings cache operation: index is same as paddr */
578 #define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
579 
580 /*
581  * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
582  */
583 static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
584 				unsigned long sz, const int op)
585 {
586 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
587 	unsigned long flags;
588 
589 	local_irq_save(flags);
590 
591 	__before_dc_op(op);
592 
593 	__cache_line_loop(paddr, vaddr, sz, op, full_page);
594 
595 	__after_dc_op(op);
596 
597 	local_irq_restore(flags);
598 }
599 
600 #else
601 
602 #define __dc_entire_op(op)
603 #define __dc_disable()
604 #define __dc_enable()
605 #define __dc_line_op(paddr, vaddr, sz, op)
606 #define __dc_line_op_k(paddr, sz, op)
607 
608 #endif /* CONFIG_ARC_HAS_DCACHE */
609 
610 #ifdef CONFIG_ARC_HAS_ICACHE
611 
612 static inline void __ic_entire_inv(void)
613 {
614 	write_aux_reg(ARC_REG_IC_IVIC, 1);
615 	read_aux_reg(ARC_REG_IC_CTRL);	/* blocks */
616 }
617 
618 static inline void
619 __ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
620 			  unsigned long sz)
621 {
622 	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
623 	unsigned long flags;
624 
625 	local_irq_save(flags);
626 	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC, full_page);
627 	local_irq_restore(flags);
628 }
629 
630 #ifndef CONFIG_SMP
631 
632 #define __ic_line_inv_vaddr(p, v, s)	__ic_line_inv_vaddr_local(p, v, s)
633 
634 #else
635 
636 struct ic_inv_args {
637 	phys_addr_t paddr, vaddr;
638 	int sz;
639 };
640 
641 static void __ic_line_inv_vaddr_helper(void *info)
642 {
643         struct ic_inv_args *ic_inv = info;
644 
645         __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
646 }
647 
648 static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
649 				unsigned long sz)
650 {
651 	struct ic_inv_args ic_inv = {
652 		.paddr = paddr,
653 		.vaddr = vaddr,
654 		.sz    = sz
655 	};
656 
657 	on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
658 }
659 
660 #endif	/* CONFIG_SMP */
661 
662 #else	/* !CONFIG_ARC_HAS_ICACHE */
663 
664 #define __ic_entire_inv()
665 #define __ic_line_inv_vaddr(pstart, vstart, sz)
666 
667 #endif /* CONFIG_ARC_HAS_ICACHE */
668 
669 noinline void slc_op_rgn(phys_addr_t paddr, unsigned long sz, const int op)
670 {
671 #ifdef CONFIG_ISA_ARCV2
672 	/*
673 	 * SLC is shared between all cores and concurrent aux operations from
674 	 * multiple cores need to be serialized using a spinlock
675 	 * A concurrent operation can be silently ignored and/or the old/new
676 	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
677 	 * below)
678 	 */
679 	static DEFINE_SPINLOCK(lock);
680 	unsigned long flags;
681 	unsigned int ctrl;
682 	phys_addr_t end;
683 
684 	spin_lock_irqsave(&lock, flags);
685 
686 	/*
687 	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
688 	 *  - b'000 (default) is Flush,
689 	 *  - b'001 is Invalidate if CTRL.IM == 0
690 	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
691 	 */
692 	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
693 
694 	/* Don't rely on default value of IM bit */
695 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
696 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
697 	else
698 		ctrl |= SLC_CTRL_IM;
699 
700 	if (op & OP_INV)
701 		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
702 	else
703 		ctrl &= ~SLC_CTRL_RGN_OP_INV;
704 
705 	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
706 
707 	/*
708 	 * Lower bits are ignored, no need to clip
709 	 * END needs to be setup before START (latter triggers the operation)
710 	 * END can't be same as START, so add (l2_line_sz - 1) to sz
711 	 */
712 	end = paddr + sz + l2_line_sz - 1;
713 	if (is_pae40_enabled())
714 		write_aux_reg(ARC_REG_SLC_RGN_END1, upper_32_bits(end));
715 
716 	write_aux_reg(ARC_REG_SLC_RGN_END, lower_32_bits(end));
717 
718 	if (is_pae40_enabled())
719 		write_aux_reg(ARC_REG_SLC_RGN_START1, upper_32_bits(paddr));
720 
721 	write_aux_reg(ARC_REG_SLC_RGN_START, lower_32_bits(paddr));
722 
723 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
724 	read_aux_reg(ARC_REG_SLC_CTRL);
725 
726 	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
727 
728 	spin_unlock_irqrestore(&lock, flags);
729 #endif
730 }
731 
732 noinline void slc_op_line(phys_addr_t paddr, unsigned long sz, const int op)
733 {
734 #ifdef CONFIG_ISA_ARCV2
735 	/*
736 	 * SLC is shared between all cores and concurrent aux operations from
737 	 * multiple cores need to be serialized using a spinlock
738 	 * A concurrent operation can be silently ignored and/or the old/new
739 	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
740 	 * below)
741 	 */
742 	static DEFINE_SPINLOCK(lock);
743 
744 	const unsigned long SLC_LINE_MASK = ~(l2_line_sz - 1);
745 	unsigned int ctrl, cmd;
746 	unsigned long flags;
747 	int num_lines;
748 
749 	spin_lock_irqsave(&lock, flags);
750 
751 	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
752 
753 	/* Don't rely on default value of IM bit */
754 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
755 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
756 	else
757 		ctrl |= SLC_CTRL_IM;
758 
759 	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
760 
761 	cmd = op & OP_INV ? ARC_AUX_SLC_IVDL : ARC_AUX_SLC_FLDL;
762 
763 	sz += paddr & ~SLC_LINE_MASK;
764 	paddr &= SLC_LINE_MASK;
765 
766 	num_lines = DIV_ROUND_UP(sz, l2_line_sz);
767 
768 	while (num_lines-- > 0) {
769 		write_aux_reg(cmd, paddr);
770 		paddr += l2_line_sz;
771 	}
772 
773 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
774 	read_aux_reg(ARC_REG_SLC_CTRL);
775 
776 	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
777 
778 	spin_unlock_irqrestore(&lock, flags);
779 #endif
780 }
781 
782 #define slc_op(paddr, sz, op)	slc_op_rgn(paddr, sz, op)
783 
784 noinline static void slc_entire_op(const int op)
785 {
786 	unsigned int ctrl, r = ARC_REG_SLC_CTRL;
787 
788 	ctrl = read_aux_reg(r);
789 
790 	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
791 		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
792 	else
793 		ctrl |= SLC_CTRL_IM;
794 
795 	write_aux_reg(r, ctrl);
796 
797 	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
798 		write_aux_reg(ARC_REG_SLC_INVALIDATE, 0x1);
799 	else
800 		write_aux_reg(ARC_REG_SLC_FLUSH, 0x1);
801 
802 	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
803 	read_aux_reg(r);
804 
805 	/* Important to wait for flush to complete */
806 	while (read_aux_reg(r) & SLC_CTRL_BUSY);
807 }
808 
809 static inline void arc_slc_disable(void)
810 {
811 	const int r = ARC_REG_SLC_CTRL;
812 
813 	slc_entire_op(OP_FLUSH_N_INV);
814 	write_aux_reg(r, read_aux_reg(r) | SLC_CTRL_DIS);
815 }
816 
817 static inline void arc_slc_enable(void)
818 {
819 	const int r = ARC_REG_SLC_CTRL;
820 
821 	write_aux_reg(r, read_aux_reg(r) & ~SLC_CTRL_DIS);
822 }
823 
824 /***********************************************************
825  * Exported APIs
826  */
827 
828 /*
829  * Handle cache congruency of kernel and userspace mappings of page when kernel
830  * writes-to/reads-from
831  *
832  * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
833  *  -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
834  *  -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
835  *  -In SMP, if hardware caches are coherent
836  *
837  * There's a corollary case, where kernel READs from a userspace mapped page.
838  * If the U-mapping is not congruent to to K-mapping, former needs flushing.
839  */
840 void flush_dcache_page(struct page *page)
841 {
842 	struct address_space *mapping;
843 
844 	if (!cache_is_vipt_aliasing()) {
845 		clear_bit(PG_dc_clean, &page->flags);
846 		return;
847 	}
848 
849 	/* don't handle anon pages here */
850 	mapping = page_mapping_file(page);
851 	if (!mapping)
852 		return;
853 
854 	/*
855 	 * pagecache page, file not yet mapped to userspace
856 	 * Make a note that K-mapping is dirty
857 	 */
858 	if (!mapping_mapped(mapping)) {
859 		clear_bit(PG_dc_clean, &page->flags);
860 	} else if (page_mapcount(page)) {
861 
862 		/* kernel reading from page with U-mapping */
863 		phys_addr_t paddr = (unsigned long)page_address(page);
864 		unsigned long vaddr = page->index << PAGE_SHIFT;
865 
866 		if (addr_not_cache_congruent(paddr, vaddr))
867 			__flush_dcache_page(paddr, vaddr);
868 	}
869 }
870 EXPORT_SYMBOL(flush_dcache_page);
871 
872 /*
873  * DMA ops for systems with L1 cache only
874  * Make memory coherent with L1 cache by flushing/invalidating L1 lines
875  */
876 static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
877 {
878 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
879 }
880 
881 static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
882 {
883 	__dc_line_op_k(start, sz, OP_INV);
884 }
885 
886 static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
887 {
888 	__dc_line_op_k(start, sz, OP_FLUSH);
889 }
890 
891 /*
892  * DMA ops for systems with both L1 and L2 caches, but without IOC
893  * Both L1 and L2 lines need to be explicitly flushed/invalidated
894  */
895 static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
896 {
897 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
898 	slc_op(start, sz, OP_FLUSH_N_INV);
899 }
900 
901 static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
902 {
903 	__dc_line_op_k(start, sz, OP_INV);
904 	slc_op(start, sz, OP_INV);
905 }
906 
907 static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
908 {
909 	__dc_line_op_k(start, sz, OP_FLUSH);
910 	slc_op(start, sz, OP_FLUSH);
911 }
912 
913 /*
914  * Exported DMA API
915  */
916 void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
917 {
918 	__dma_cache_wback_inv(start, sz);
919 }
920 EXPORT_SYMBOL(dma_cache_wback_inv);
921 
922 void dma_cache_inv(phys_addr_t start, unsigned long sz)
923 {
924 	__dma_cache_inv(start, sz);
925 }
926 EXPORT_SYMBOL(dma_cache_inv);
927 
928 void dma_cache_wback(phys_addr_t start, unsigned long sz)
929 {
930 	__dma_cache_wback(start, sz);
931 }
932 EXPORT_SYMBOL(dma_cache_wback);
933 
934 /*
935  * This is API for making I/D Caches consistent when modifying
936  * kernel code (loadable modules, kprobes, kgdb...)
937  * This is called on insmod, with kernel virtual address for CODE of
938  * the module. ARC cache maintenance ops require PHY address thus we
939  * need to convert vmalloc addr to PHY addr
940  */
941 void flush_icache_range(unsigned long kstart, unsigned long kend)
942 {
943 	unsigned int tot_sz;
944 
945 	WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
946 
947 	/* Shortcut for bigger flush ranges.
948 	 * Here we don't care if this was kernel virtual or phy addr
949 	 */
950 	tot_sz = kend - kstart;
951 	if (tot_sz > PAGE_SIZE) {
952 		flush_cache_all();
953 		return;
954 	}
955 
956 	/* Case: Kernel Phy addr (0x8000_0000 onwards) */
957 	if (likely(kstart > PAGE_OFFSET)) {
958 		/*
959 		 * The 2nd arg despite being paddr will be used to index icache
960 		 * This is OK since no alternate virtual mappings will exist
961 		 * given the callers for this case: kprobe/kgdb in built-in
962 		 * kernel code only.
963 		 */
964 		__sync_icache_dcache(kstart, kstart, kend - kstart);
965 		return;
966 	}
967 
968 	/*
969 	 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
970 	 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
971 	 *     handling of kernel vaddr.
972 	 *
973 	 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
974 	 *     it still needs to handle  a 2 page scenario, where the range
975 	 *     straddles across 2 virtual pages and hence need for loop
976 	 */
977 	while (tot_sz > 0) {
978 		unsigned int off, sz;
979 		unsigned long phy, pfn;
980 
981 		off = kstart % PAGE_SIZE;
982 		pfn = vmalloc_to_pfn((void *)kstart);
983 		phy = (pfn << PAGE_SHIFT) + off;
984 		sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
985 		__sync_icache_dcache(phy, kstart, sz);
986 		kstart += sz;
987 		tot_sz -= sz;
988 	}
989 }
990 EXPORT_SYMBOL(flush_icache_range);
991 
992 /*
993  * General purpose helper to make I and D cache lines consistent.
994  * @paddr is phy addr of region
995  * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
996  *    However in one instance, when called by kprobe (for a breakpt in
997  *    builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
998  *    use a paddr to index the cache (despite VIPT). This is fine since since a
999  *    builtin kernel page will not have any virtual mappings.
1000  *    kprobe on loadable module will be kernel vaddr.
1001  */
1002 void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
1003 {
1004 	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
1005 	__ic_line_inv_vaddr(paddr, vaddr, len);
1006 }
1007 
1008 /* wrapper to compile time eliminate alignment checks in flush loop */
1009 void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
1010 {
1011 	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
1012 }
1013 
1014 /*
1015  * wrapper to clearout kernel or userspace mappings of a page
1016  * For kernel mappings @vaddr == @paddr
1017  */
1018 void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
1019 {
1020 	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
1021 }
1022 
1023 noinline void flush_cache_all(void)
1024 {
1025 	unsigned long flags;
1026 
1027 	local_irq_save(flags);
1028 
1029 	__ic_entire_inv();
1030 	__dc_entire_op(OP_FLUSH_N_INV);
1031 
1032 	local_irq_restore(flags);
1033 
1034 }
1035 
1036 #ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
1037 
1038 void flush_cache_mm(struct mm_struct *mm)
1039 {
1040 	flush_cache_all();
1041 }
1042 
1043 void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
1044 		      unsigned long pfn)
1045 {
1046 	phys_addr_t paddr = pfn << PAGE_SHIFT;
1047 
1048 	u_vaddr &= PAGE_MASK;
1049 
1050 	__flush_dcache_page(paddr, u_vaddr);
1051 
1052 	if (vma->vm_flags & VM_EXEC)
1053 		__inv_icache_page(paddr, u_vaddr);
1054 }
1055 
1056 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
1057 		       unsigned long end)
1058 {
1059 	flush_cache_all();
1060 }
1061 
1062 void flush_anon_page(struct vm_area_struct *vma, struct page *page,
1063 		     unsigned long u_vaddr)
1064 {
1065 	/* TBD: do we really need to clear the kernel mapping */
1066 	__flush_dcache_page((phys_addr_t)page_address(page), u_vaddr);
1067 	__flush_dcache_page((phys_addr_t)page_address(page),
1068 			    (phys_addr_t)page_address(page));
1069 
1070 }
1071 
1072 #endif
1073 
1074 void copy_user_highpage(struct page *to, struct page *from,
1075 	unsigned long u_vaddr, struct vm_area_struct *vma)
1076 {
1077 	void *kfrom = kmap_atomic(from);
1078 	void *kto = kmap_atomic(to);
1079 	int clean_src_k_mappings = 0;
1080 
1081 	/*
1082 	 * If SRC page was already mapped in userspace AND it's U-mapping is
1083 	 * not congruent with K-mapping, sync former to physical page so that
1084 	 * K-mapping in memcpy below, sees the right data
1085 	 *
1086 	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
1087 	 * equally valid for SRC page as well
1088 	 *
1089 	 * For !VIPT cache, all of this gets compiled out as
1090 	 * addr_not_cache_congruent() is 0
1091 	 */
1092 	if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
1093 		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
1094 		clean_src_k_mappings = 1;
1095 	}
1096 
1097 	copy_page(kto, kfrom);
1098 
1099 	/*
1100 	 * Mark DST page K-mapping as dirty for a later finalization by
1101 	 * update_mmu_cache(). Although the finalization could have been done
1102 	 * here as well (given that both vaddr/paddr are available).
1103 	 * But update_mmu_cache() already has code to do that for other
1104 	 * non copied user pages (e.g. read faults which wire in pagecache page
1105 	 * directly).
1106 	 */
1107 	clear_bit(PG_dc_clean, &to->flags);
1108 
1109 	/*
1110 	 * if SRC was already usermapped and non-congruent to kernel mapping
1111 	 * sync the kernel mapping back to physical page
1112 	 */
1113 	if (clean_src_k_mappings) {
1114 		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
1115 		set_bit(PG_dc_clean, &from->flags);
1116 	} else {
1117 		clear_bit(PG_dc_clean, &from->flags);
1118 	}
1119 
1120 	kunmap_atomic(kto);
1121 	kunmap_atomic(kfrom);
1122 }
1123 
1124 void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
1125 {
1126 	clear_page(to);
1127 	clear_bit(PG_dc_clean, &page->flags);
1128 }
1129 
1130 
1131 /**********************************************************************
1132  * Explicit Cache flush request from user space via syscall
1133  * Needed for JITs which generate code on the fly
1134  */
1135 SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
1136 {
1137 	/* TBD: optimize this */
1138 	flush_cache_all();
1139 	return 0;
1140 }
1141 
1142 /*
1143  * IO-Coherency (IOC) setup rules:
1144  *
1145  * 1. Needs to be at system level, so only once by Master core
1146  *    Non-Masters need not be accessing caches at that time
1147  *    - They are either HALT_ON_RESET and kick started much later or
1148  *    - if run on reset, need to ensure that arc_platform_smp_wait_to_boot()
1149  *      doesn't perturb caches or coherency unit
1150  *
1151  * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC,
1152  *    otherwise any straggler data might behave strangely post IOC enabling
1153  *
1154  * 3. All Caches need to be disabled when setting up IOC to elide any in-flight
1155  *    Coherency transactions
1156  */
1157 noinline void __init arc_ioc_setup(void)
1158 {
1159 	unsigned int ioc_base, mem_sz;
1160 
1161 	/*
1162 	 * If IOC was already enabled (due to bootloader) it technically needs to
1163 	 * be reconfigured with aperture base,size corresponding to Linux memory map
1164 	 * which will certainly be different than uboot's. But disabling and
1165 	 * reenabling IOC when DMA might be potentially active is tricky business.
1166 	 * To avoid random memory issues later, just panic here and ask user to
1167 	 * upgrade bootloader to one which doesn't enable IOC
1168 	 */
1169 	if (read_aux_reg(ARC_REG_IO_COH_ENABLE) & ARC_IO_COH_ENABLE_BIT)
1170 		panic("IOC already enabled, please upgrade bootloader!\n");
1171 
1172 	if (!ioc_enable)
1173 		return;
1174 
1175 	/* Flush + invalidate + disable L1 dcache */
1176 	__dc_disable();
1177 
1178 	/* Flush + invalidate SLC */
1179 	if (read_aux_reg(ARC_REG_SLC_BCR))
1180 		slc_entire_op(OP_FLUSH_N_INV);
1181 
1182 	/*
1183 	 * currently IOC Aperture covers entire DDR
1184 	 * TBD: fix for PGU + 1GB of low mem
1185 	 * TBD: fix for PAE
1186 	 */
1187 	mem_sz = arc_get_mem_sz();
1188 
1189 	if (!is_power_of_2(mem_sz) || mem_sz < 4096)
1190 		panic("IOC Aperture size must be power of 2 larger than 4KB");
1191 
1192 	/*
1193 	 * IOC Aperture size decoded as 2 ^ (SIZE + 2) KB,
1194 	 * so setting 0x11 implies 512MB, 0x12 implies 1GB...
1195 	 */
1196 	write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, order_base_2(mem_sz >> 10) - 2);
1197 
1198 	/* for now assume kernel base is start of IOC aperture */
1199 	ioc_base = CONFIG_LINUX_RAM_BASE;
1200 
1201 	if (ioc_base % mem_sz != 0)
1202 		panic("IOC Aperture start must be aligned to the size of the aperture");
1203 
1204 	write_aux_reg(ARC_REG_IO_COH_AP0_BASE, ioc_base >> 12);
1205 	write_aux_reg(ARC_REG_IO_COH_PARTIAL, ARC_IO_COH_PARTIAL_BIT);
1206 	write_aux_reg(ARC_REG_IO_COH_ENABLE, ARC_IO_COH_ENABLE_BIT);
1207 
1208 	/* Re-enable L1 dcache */
1209 	__dc_enable();
1210 }
1211 
1212 /*
1213  * Cache related boot time checks/setups only needed on master CPU:
1214  *  - Geometry checks (kernel build and hardware agree: e.g. L1_CACHE_BYTES)
1215  *    Assume SMP only, so all cores will have same cache config. A check on
1216  *    one core suffices for all
1217  *  - IOC setup / dma callbacks only need to be done once
1218  */
1219 void __init arc_cache_init_master(void)
1220 {
1221 	unsigned int __maybe_unused cpu = smp_processor_id();
1222 
1223 	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
1224 		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
1225 
1226 		if (!ic->line_len)
1227 			panic("cache support enabled but non-existent cache\n");
1228 
1229 		if (ic->line_len != L1_CACHE_BYTES)
1230 			panic("ICache line [%d] != kernel Config [%d]",
1231 			      ic->line_len, L1_CACHE_BYTES);
1232 
1233 		/*
1234 		 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
1235 		 * pair to provide vaddr/paddr respectively, just as in MMU v3
1236 		 */
1237 		if (is_isa_arcv2() && ic->alias)
1238 			_cache_line_loop_ic_fn = __cache_line_loop_v3;
1239 		else
1240 			_cache_line_loop_ic_fn = __cache_line_loop;
1241 	}
1242 
1243 	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
1244 		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
1245 
1246 		if (!dc->line_len)
1247 			panic("cache support enabled but non-existent cache\n");
1248 
1249 		if (dc->line_len != L1_CACHE_BYTES)
1250 			panic("DCache line [%d] != kernel Config [%d]",
1251 			      dc->line_len, L1_CACHE_BYTES);
1252 
1253 		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
1254 		if (is_isa_arcompact()) {
1255 			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
1256 			int num_colors = dc->sz_k/dc->assoc/TO_KB(PAGE_SIZE);
1257 
1258 			if (dc->alias) {
1259 				if (!handled)
1260 					panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1261 				if (CACHE_COLORS_NUM != num_colors)
1262 					panic("CACHE_COLORS_NUM not optimized for config\n");
1263 			} else if (!dc->alias && handled) {
1264 				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1265 			}
1266 		}
1267 	}
1268 
1269 	/*
1270 	 * Check that SMP_CACHE_BYTES (and hence ARCH_DMA_MINALIGN) is larger
1271 	 * or equal to any cache line length.
1272 	 */
1273 	BUILD_BUG_ON_MSG(L1_CACHE_BYTES > SMP_CACHE_BYTES,
1274 			 "SMP_CACHE_BYTES must be >= any cache line length");
1275 	if (is_isa_arcv2() && (l2_line_sz > SMP_CACHE_BYTES))
1276 		panic("L2 Cache line [%d] > kernel Config [%d]\n",
1277 		      l2_line_sz, SMP_CACHE_BYTES);
1278 
1279 	/* Note that SLC disable not formally supported till HS 3.0 */
1280 	if (is_isa_arcv2() && l2_line_sz && !slc_enable)
1281 		arc_slc_disable();
1282 
1283 	if (is_isa_arcv2() && ioc_exists)
1284 		arc_ioc_setup();
1285 
1286 	if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1287 		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1288 		__dma_cache_inv = __dma_cache_inv_slc;
1289 		__dma_cache_wback = __dma_cache_wback_slc;
1290 	} else {
1291 		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1292 		__dma_cache_inv = __dma_cache_inv_l1;
1293 		__dma_cache_wback = __dma_cache_wback_l1;
1294 	}
1295 	/*
1296 	 * In case of IOC (say IOC+SLC case), pointers above could still be set
1297 	 * but end up not being relevant as the first function in chain is not
1298 	 * called at all for devices using coherent DMA.
1299 	 *     arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
1300 	 */
1301 }
1302 
1303 void __ref arc_cache_init(void)
1304 {
1305 	unsigned int __maybe_unused cpu = smp_processor_id();
1306 	char str[256];
1307 
1308 	pr_info("%s", arc_cache_mumbojumbo(0, str, sizeof(str)));
1309 
1310 	if (!cpu)
1311 		arc_cache_init_master();
1312 
1313 	/*
1314 	 * In PAE regime, TLB and cache maintenance ops take wider addresses
1315 	 * And even if PAE is not enabled in kernel, the upper 32-bits still need
1316 	 * to be zeroed to keep the ops sane.
1317 	 * As an optimization for more common !PAE enabled case, zero them out
1318 	 * once at init, rather than checking/setting to 0 for every runtime op
1319 	 */
1320 	if (is_isa_arcv2() && pae40_exist_but_not_enab()) {
1321 
1322 		if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE))
1323 			write_aux_reg(ARC_REG_IC_PTAG_HI, 0);
1324 
1325 		if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE))
1326 			write_aux_reg(ARC_REG_DC_PTAG_HI, 0);
1327 
1328 		if (l2_line_sz) {
1329 			write_aux_reg(ARC_REG_SLC_RGN_END1, 0);
1330 			write_aux_reg(ARC_REG_SLC_RGN_START1, 0);
1331 		}
1332 	}
1333 }
1334