Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * ARC Cache Management
4 *
5 * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
6 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
7 */
8
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/cache.h>
13#include <linux/mmu_context.h>
14#include <linux/syscalls.h>
15#include <linux/uaccess.h>
16#include <linux/pagemap.h>
17#include <asm/cacheflush.h>
18#include <asm/cachectl.h>
19#include <asm/setup.h>
20
21#ifdef CONFIG_ISA_ARCV2
22#define USE_RGN_FLSH 1
23#endif
24
25static int l2_line_sz;
26static int ioc_exists;
27int slc_enable = 1, ioc_enable = 1;
28unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
29unsigned long perip_end = 0xFFFFFFFF; /* legacy value */
30
31void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
32 unsigned long sz, const int op, const int full_page);
33
34void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
35void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
36void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
37
38char *arc_cache_mumbojumbo(int c, char *buf, int len)
39{
40 int n = 0;
41 struct cpuinfo_arc_cache *p;
42
43#define PR_CACHE(p, cfg, str) \
44 if (!(p)->line_len) \
45 n += scnprintf(buf + n, len - n, str"\t\t: N/A\n"); \
46 else \
47 n += scnprintf(buf + n, len - n, \
48 str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n", \
49 (p)->sz_k, (p)->assoc, (p)->line_len, \
50 (p)->vipt ? "VIPT" : "PIPT", \
51 (p)->alias ? " aliasing" : "", \
52 IS_USED_CFG(cfg));
53
54 PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
55 PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
56
57 p = &cpuinfo_arc700[c].slc;
58 if (p->line_len)
59 n += scnprintf(buf + n, len - n,
60 "SLC\t\t: %uK, %uB Line%s\n",
61 p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
62
63 n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
64 perip_base,
65 IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency (per-device) "));
66
67 return buf;
68}
69
70/*
71 * Read the Cache Build Confuration Registers, Decode them and save into
72 * the cpuinfo structure for later use.
73 * No Validation done here, simply read/convert the BCRs
74 */
75static void read_decode_cache_bcr_arcv2(int cpu)
76{
77 struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
78 struct bcr_generic sbcr;
79
80 struct bcr_slc_cfg {
81#ifdef CONFIG_CPU_BIG_ENDIAN
82 unsigned int pad:24, way:2, lsz:2, sz:4;
83#else
84 unsigned int sz:4, lsz:2, way:2, pad:24;
85#endif
86 } slc_cfg;
87
88 struct bcr_clust_cfg {
89#ifdef CONFIG_CPU_BIG_ENDIAN
90 unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
91#else
92 unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
93#endif
94 } cbcr;
95
96 struct bcr_volatile {
97#ifdef CONFIG_CPU_BIG_ENDIAN
98 unsigned int start:4, limit:4, pad:22, order:1, disable:1;
99#else
100 unsigned int disable:1, order:1, pad:22, limit:4, start:4;
101#endif
102 } vol;
103
104
105 READ_BCR(ARC_REG_SLC_BCR, sbcr);
106 if (sbcr.ver) {
107 READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
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
116 /*
117 * As for today we don't support both IOC and ZONE_HIGHMEM enabled
118 * simultaneously. This happens because as of today IOC aperture covers
119 * only ZONE_NORMAL (low mem) and any dma transactions outside this
120 * region won't be HW coherent.
121 * If we want to use both IOC and ZONE_HIGHMEM we can use
122 * bounce_buffer to handle dma transactions to HIGHMEM.
123 * Also it is possible to modify dma_direct cache ops or increase IOC
124 * aperture size if we are planning to use HIGHMEM without PAE.
125 */
126 if (IS_ENABLED(CONFIG_HIGHMEM) || is_pae40_enabled())
127 ioc_enable = 0;
128 } else {
129 ioc_enable = 0;
130 }
131
132 /* HS 2.0 didn't have AUX_VOL */
133 if (cpuinfo_arc700[cpu].core.family > 0x51) {
134 READ_BCR(AUX_VOL, vol);
135 perip_base = vol.start << 28;
136 /* HS 3.0 has limit and strict-ordering fields */
137 if (cpuinfo_arc700[cpu].core.family > 0x52)
138 perip_end = (vol.limit << 28) - 1;
139 }
140}
141
142void read_decode_cache_bcr(void)
143{
144 struct cpuinfo_arc_cache *p_ic, *p_dc;
145 unsigned int cpu = smp_processor_id();
146 struct bcr_cache {
147#ifdef CONFIG_CPU_BIG_ENDIAN
148 unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
149#else
150 unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
151#endif
152 } ibcr, dbcr;
153
154 p_ic = &cpuinfo_arc700[cpu].icache;
155 READ_BCR(ARC_REG_IC_BCR, ibcr);
156
157 if (!ibcr.ver)
158 goto dc_chk;
159
160 if (ibcr.ver <= 3) {
161 BUG_ON(ibcr.config != 3);
162 p_ic->assoc = 2; /* Fixed to 2w set assoc */
163 } else if (ibcr.ver >= 4) {
164 p_ic->assoc = 1 << ibcr.config; /* 1,2,4,8 */
165 }
166
167 p_ic->line_len = 8 << ibcr.line_len;
168 p_ic->sz_k = 1 << (ibcr.sz - 1);
169 p_ic->vipt = 1;
170 p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
171
172dc_chk:
173 p_dc = &cpuinfo_arc700[cpu].dcache;
174 READ_BCR(ARC_REG_DC_BCR, dbcr);
175
176 if (!dbcr.ver)
177 goto slc_chk;
178
179 if (dbcr.ver <= 3) {
180 BUG_ON(dbcr.config != 2);
181 p_dc->assoc = 4; /* Fixed to 4w set assoc */
182 p_dc->vipt = 1;
183 p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
184 } else if (dbcr.ver >= 4) {
185 p_dc->assoc = 1 << dbcr.config; /* 1,2,4,8 */
186 p_dc->vipt = 0;
187 p_dc->alias = 0; /* PIPT so can't VIPT alias */
188 }
189
190 p_dc->line_len = 16 << dbcr.line_len;
191 p_dc->sz_k = 1 << (dbcr.sz - 1);
192
193slc_chk:
194 if (is_isa_arcv2())
195 read_decode_cache_bcr_arcv2(cpu);
196}
197
198/*
199 * Line Operation on {I,D}-Cache
200 */
201
202#define OP_INV 0x1
203#define OP_FLUSH 0x2
204#define OP_FLUSH_N_INV 0x3
205#define OP_INV_IC 0x4
206
207/*
208 * I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
209 *
210 * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
211 * The orig Cache Management Module "CDU" only required paddr to invalidate a
212 * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
213 * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
214 * the exact same line.
215 *
216 * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
217 * paddr alone could not be used to correctly index the cache.
218 *
219 * ------------------
220 * MMU v1/v2 (Fixed Page Size 8k)
221 * ------------------
222 * The solution was to provide CDU with these additonal vaddr bits. These
223 * would be bits [x:13], x would depend on cache-geometry, 13 comes from
224 * standard page size of 8k.
225 * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
226 * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
227 * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
228 * represent the offset within cache-line. The adv of using this "clumsy"
229 * interface for additional info was no new reg was needed in CDU programming
230 * model.
231 *
232 * 17:13 represented the max num of bits passable, actual bits needed were
233 * fewer, based on the num-of-aliases possible.
234 * -for 2 alias possibility, only bit 13 needed (32K cache)
235 * -for 4 alias possibility, bits 14:13 needed (64K cache)
236 *
237 * ------------------
238 * MMU v3
239 * ------------------
240 * This ver of MMU supports variable page sizes (1k-16k): although Linux will
241 * only support 8k (default), 16k and 4k.
242 * However from hardware perspective, smaller page sizes aggravate aliasing
243 * meaning more vaddr bits needed to disambiguate the cache-line-op ;
244 * the existing scheme of piggybacking won't work for certain configurations.
245 * Two new registers IC_PTAG and DC_PTAG inttoduced.
246 * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
247 */
248
249static inline
250void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
251 unsigned long sz, const int op, const int full_page)
252{
253 unsigned int aux_cmd;
254 int num_lines;
255
256 if (op == OP_INV_IC) {
257 aux_cmd = ARC_REG_IC_IVIL;
258 } else {
259 /* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
260 aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
261 }
262
263 /* Ensure we properly floor/ceil the non-line aligned/sized requests
264 * and have @paddr - aligned to cache line and integral @num_lines.
265 * This however can be avoided for page sized since:
266 * -@paddr will be cache-line aligned already (being page aligned)
267 * -@sz will be integral multiple of line size (being page sized).
268 */
269 if (!full_page) {
270 sz += paddr & ~CACHE_LINE_MASK;
271 paddr &= CACHE_LINE_MASK;
272 vaddr &= CACHE_LINE_MASK;
273 }
274
275 num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
276
277 /* MMUv2 and before: paddr contains stuffed vaddrs bits */
278 paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
279
280 while (num_lines-- > 0) {
281 write_aux_reg(aux_cmd, paddr);
282 paddr += L1_CACHE_BYTES;
283 }
284}
285
286/*
287 * For ARC700 MMUv3 I-cache and D-cache flushes
288 * - ARC700 programming model requires paddr and vaddr be passed in seperate
289 * AUX registers (*_IV*L and *_PTAG respectively) irrespective of whether the
290 * caches actually alias or not.
291 * - For HS38, only the aliasing I-cache configuration uses the PTAG reg
292 * (non aliasing I-cache version doesn't; while D-cache can't possibly alias)
293 */
294static inline
295void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
296 unsigned long sz, const int op, const int full_page)
297{
298 unsigned int aux_cmd, aux_tag;
299 int num_lines;
300
301 if (op == OP_INV_IC) {
302 aux_cmd = ARC_REG_IC_IVIL;
303 aux_tag = ARC_REG_IC_PTAG;
304 } else {
305 aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
306 aux_tag = ARC_REG_DC_PTAG;
307 }
308
309 /* Ensure we properly floor/ceil the non-line aligned/sized requests
310 * and have @paddr - aligned to cache line and integral @num_lines.
311 * This however can be avoided for page sized since:
312 * -@paddr will be cache-line aligned already (being page aligned)
313 * -@sz will be integral multiple of line size (being page sized).
314 */
315 if (!full_page) {
316 sz += paddr & ~CACHE_LINE_MASK;
317 paddr &= CACHE_LINE_MASK;
318 vaddr &= CACHE_LINE_MASK;
319 }
320 num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
321
322 /*
323 * MMUv3, cache ops require paddr in PTAG reg
324 * if V-P const for loop, PTAG can be written once outside loop
325 */
326 if (full_page)
327 write_aux_reg(aux_tag, paddr);
328
329 /*
330 * This is technically for MMU v4, using the MMU v3 programming model
331 * Special work for HS38 aliasing I-cache configuration with PAE40
332 * - upper 8 bits of paddr need to be written into PTAG_HI
333 * - (and needs to be written before the lower 32 bits)
334 * Note that PTAG_HI is hoisted outside the line loop
335 */
336 if (is_pae40_enabled() && op == OP_INV_IC)
337 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
338
339 while (num_lines-- > 0) {
340 if (!full_page) {
341 write_aux_reg(aux_tag, paddr);
342 paddr += L1_CACHE_BYTES;
343 }
344
345 write_aux_reg(aux_cmd, vaddr);
346 vaddr += L1_CACHE_BYTES;
347 }
348}
349
350#ifndef USE_RGN_FLSH
351
352/*
353 * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
354 * Here's how cache ops are implemented
355 *
356 * - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
357 * - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
358 * - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
359 * respectively, similar to MMU v3 programming model, hence
360 * __cache_line_loop_v3() is used)
361 *
362 * If PAE40 is enabled, independent of aliasing considerations, the higher bits
363 * needs to be written into PTAG_HI
364 */
365static inline
366void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
367 unsigned long sz, const int op, const int full_page)
368{
369 unsigned int aux_cmd;
370 int num_lines;
371
372 if (op == OP_INV_IC) {
373 aux_cmd = ARC_REG_IC_IVIL;
374 } else {
375 /* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
376 aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
377 }
378
379 /* Ensure we properly floor/ceil the non-line aligned/sized requests
380 * and have @paddr - aligned to cache line and integral @num_lines.
381 * This however can be avoided for page sized since:
382 * -@paddr will be cache-line aligned already (being page aligned)
383 * -@sz will be integral multiple of line size (being page sized).
384 */
385 if (!full_page) {
386 sz += paddr & ~CACHE_LINE_MASK;
387 paddr &= CACHE_LINE_MASK;
388 }
389
390 num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
391
392 /*
393 * For HS38 PAE40 configuration
394 * - upper 8 bits of paddr need to be written into PTAG_HI
395 * - (and needs to be written before the lower 32 bits)
396 */
397 if (is_pae40_enabled()) {
398 if (op == OP_INV_IC)
399 /*
400 * Non aliasing I-cache in HS38,
401 * aliasing I-cache handled in __cache_line_loop_v3()
402 */
403 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
404 else
405 write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
406 }
407
408 while (num_lines-- > 0) {
409 write_aux_reg(aux_cmd, paddr);
410 paddr += L1_CACHE_BYTES;
411 }
412}
413
414#else
415
416/*
417 * optimized flush operation which takes a region as opposed to iterating per line
418 */
419static inline
420void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
421 unsigned long sz, const int op, const int full_page)
422{
423 unsigned int s, e;
424
425 /* Only for Non aliasing I-cache in HS38 */
426 if (op == OP_INV_IC) {
427 s = ARC_REG_IC_IVIR;
428 e = ARC_REG_IC_ENDR;
429 } else {
430 s = ARC_REG_DC_STARTR;
431 e = ARC_REG_DC_ENDR;
432 }
433
434 if (!full_page) {
435 /* for any leading gap between @paddr and start of cache line */
436 sz += paddr & ~CACHE_LINE_MASK;
437 paddr &= CACHE_LINE_MASK;
438
439 /*
440 * account for any trailing gap to end of cache line
441 * this is equivalent to DIV_ROUND_UP() in line ops above
442 */
443 sz += L1_CACHE_BYTES - 1;
444 }
445
446 if (is_pae40_enabled()) {
447 /* TBD: check if crossing 4TB boundary */
448 if (op == OP_INV_IC)
449 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
450 else
451 write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
452 }
453
454 /* ENDR needs to be set ahead of START */
455 write_aux_reg(e, paddr + sz); /* ENDR is exclusive */
456 write_aux_reg(s, paddr);
457
458 /* caller waits on DC_CTRL.FS */
459}
460
461#endif
462
463#if (CONFIG_ARC_MMU_VER < 3)
464#define __cache_line_loop __cache_line_loop_v2
465#elif (CONFIG_ARC_MMU_VER == 3)
466#define __cache_line_loop __cache_line_loop_v3
467#elif (CONFIG_ARC_MMU_VER > 3)
468#define __cache_line_loop __cache_line_loop_v4
469#endif
470
471#ifdef CONFIG_ARC_HAS_DCACHE
472
473/***************************************************************
474 * Machine specific helpers for Entire D-Cache or Per Line ops
475 */
476
477#ifndef USE_RGN_FLSH
478/*
479 * this version avoids extra read/write of DC_CTRL for flush or invalid ops
480 * in the non region flush regime (such as for ARCompact)
481 */
482static inline void __before_dc_op(const int op)
483{
484 if (op == OP_FLUSH_N_INV) {
485 /* Dcache provides 2 cmd: FLUSH or INV
486 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
487 * flush-n-inv is achieved by INV cmd but with IM=1
488 * So toggle INV sub-mode depending on op request and default
489 */
490 const unsigned int ctl = ARC_REG_DC_CTRL;
491 write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
492 }
493}
494
495#else
496
497static inline void __before_dc_op(const int op)
498{
499 const unsigned int ctl = ARC_REG_DC_CTRL;
500 unsigned int val = read_aux_reg(ctl);
501
502 if (op == OP_FLUSH_N_INV) {
503 val |= DC_CTRL_INV_MODE_FLUSH;
504 }
505
506 if (op != OP_INV_IC) {
507 /*
508 * Flush / Invalidate is provided by DC_CTRL.RNG_OP 0 or 1
509 * combined Flush-n-invalidate uses DC_CTRL.IM = 1 set above
510 */
511 val &= ~DC_CTRL_RGN_OP_MSK;
512 if (op & OP_INV)
513 val |= DC_CTRL_RGN_OP_INV;
514 }
515 write_aux_reg(ctl, val);
516}
517
518#endif
519
520
521static inline void __after_dc_op(const int op)
522{
523 if (op & OP_FLUSH) {
524 const unsigned int ctl = ARC_REG_DC_CTRL;
525 unsigned int reg;
526
527 /* flush / flush-n-inv both wait */
528 while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
529 ;
530
531 /* Switch back to default Invalidate mode */
532 if (op == OP_FLUSH_N_INV)
533 write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
534 }
535}
536
537/*
538 * Operation on Entire D-Cache
539 * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
540 * Note that constant propagation ensures all the checks are gone
541 * in generated code
542 */
543static inline void __dc_entire_op(const int op)
544{
545 int aux;
546
547 __before_dc_op(op);
548
549 if (op & OP_INV) /* Inv or flush-n-inv use same cmd reg */
550 aux = ARC_REG_DC_IVDC;
551 else
552 aux = ARC_REG_DC_FLSH;
553
554 write_aux_reg(aux, 0x1);
555
556 __after_dc_op(op);
557}
558
559static inline void __dc_disable(void)
560{
561 const int r = ARC_REG_DC_CTRL;
562
563 __dc_entire_op(OP_FLUSH_N_INV);
564 write_aux_reg(r, read_aux_reg(r) | DC_CTRL_DIS);
565}
566
567static void __dc_enable(void)
568{
569 const int r = ARC_REG_DC_CTRL;
570
571 write_aux_reg(r, read_aux_reg(r) & ~DC_CTRL_DIS);
572}
573
574/* For kernel mappings cache operation: index is same as paddr */
575#define __dc_line_op_k(p, sz, op) __dc_line_op(p, p, sz, op)
576
577/*
578 * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
579 */
580static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
581 unsigned long sz, const int op)
582{
583 const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
584 unsigned long flags;
585
586 local_irq_save(flags);
587
588 __before_dc_op(op);
589
590 __cache_line_loop(paddr, vaddr, sz, op, full_page);
591
592 __after_dc_op(op);
593
594 local_irq_restore(flags);
595}
596
597#else
598
599#define __dc_entire_op(op)
600#define __dc_disable()
601#define __dc_enable()
602#define __dc_line_op(paddr, vaddr, sz, op)
603#define __dc_line_op_k(paddr, sz, op)
604
605#endif /* CONFIG_ARC_HAS_DCACHE */
606
607#ifdef CONFIG_ARC_HAS_ICACHE
608
609static inline void __ic_entire_inv(void)
610{
611 write_aux_reg(ARC_REG_IC_IVIC, 1);
612 read_aux_reg(ARC_REG_IC_CTRL); /* blocks */
613}
614
615static inline void
616__ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
617 unsigned long sz)
618{
619 const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
620 unsigned long flags;
621
622 local_irq_save(flags);
623 (*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC, full_page);
624 local_irq_restore(flags);
625}
626
627#ifndef CONFIG_SMP
628
629#define __ic_line_inv_vaddr(p, v, s) __ic_line_inv_vaddr_local(p, v, s)
630
631#else
632
633struct ic_inv_args {
634 phys_addr_t paddr, vaddr;
635 int sz;
636};
637
638static void __ic_line_inv_vaddr_helper(void *info)
639{
640 struct ic_inv_args *ic_inv = info;
641
642 __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
643}
644
645static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
646 unsigned long sz)
647{
648 struct ic_inv_args ic_inv = {
649 .paddr = paddr,
650 .vaddr = vaddr,
651 .sz = sz
652 };
653
654 on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
655}
656
657#endif /* CONFIG_SMP */
658
659#else /* !CONFIG_ARC_HAS_ICACHE */
660
661#define __ic_entire_inv()
662#define __ic_line_inv_vaddr(pstart, vstart, sz)
663
664#endif /* CONFIG_ARC_HAS_ICACHE */
665
666noinline void slc_op_rgn(phys_addr_t paddr, unsigned long sz, const int op)
667{
668#ifdef CONFIG_ISA_ARCV2
669 /*
670 * SLC is shared between all cores and concurrent aux operations from
671 * multiple cores need to be serialized using a spinlock
672 * A concurrent operation can be silently ignored and/or the old/new
673 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
674 * below)
675 */
676 static DEFINE_SPINLOCK(lock);
677 unsigned long flags;
678 unsigned int ctrl;
679 phys_addr_t end;
680
681 spin_lock_irqsave(&lock, flags);
682
683 /*
684 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
685 * - b'000 (default) is Flush,
686 * - b'001 is Invalidate if CTRL.IM == 0
687 * - b'001 is Flush-n-Invalidate if CTRL.IM == 1
688 */
689 ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
690
691 /* Don't rely on default value of IM bit */
692 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
693 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
694 else
695 ctrl |= SLC_CTRL_IM;
696
697 if (op & OP_INV)
698 ctrl |= SLC_CTRL_RGN_OP_INV; /* Inv or flush-n-inv */
699 else
700 ctrl &= ~SLC_CTRL_RGN_OP_INV;
701
702 write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
703
704 /*
705 * Lower bits are ignored, no need to clip
706 * END needs to be setup before START (latter triggers the operation)
707 * END can't be same as START, so add (l2_line_sz - 1) to sz
708 */
709 end = paddr + sz + l2_line_sz - 1;
710 if (is_pae40_enabled())
711 write_aux_reg(ARC_REG_SLC_RGN_END1, upper_32_bits(end));
712
713 write_aux_reg(ARC_REG_SLC_RGN_END, lower_32_bits(end));
714
715 if (is_pae40_enabled())
716 write_aux_reg(ARC_REG_SLC_RGN_START1, upper_32_bits(paddr));
717
718 write_aux_reg(ARC_REG_SLC_RGN_START, lower_32_bits(paddr));
719
720 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
721 read_aux_reg(ARC_REG_SLC_CTRL);
722
723 while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
724
725 spin_unlock_irqrestore(&lock, flags);
726#endif
727}
728
729noinline void slc_op_line(phys_addr_t paddr, unsigned long sz, const int op)
730{
731#ifdef CONFIG_ISA_ARCV2
732 /*
733 * SLC is shared between all cores and concurrent aux operations from
734 * multiple cores need to be serialized using a spinlock
735 * A concurrent operation can be silently ignored and/or the old/new
736 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
737 * below)
738 */
739 static DEFINE_SPINLOCK(lock);
740
741 const unsigned long SLC_LINE_MASK = ~(l2_line_sz - 1);
742 unsigned int ctrl, cmd;
743 unsigned long flags;
744 int num_lines;
745
746 spin_lock_irqsave(&lock, flags);
747
748 ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
749
750 /* Don't rely on default value of IM bit */
751 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
752 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
753 else
754 ctrl |= SLC_CTRL_IM;
755
756 write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
757
758 cmd = op & OP_INV ? ARC_AUX_SLC_IVDL : ARC_AUX_SLC_FLDL;
759
760 sz += paddr & ~SLC_LINE_MASK;
761 paddr &= SLC_LINE_MASK;
762
763 num_lines = DIV_ROUND_UP(sz, l2_line_sz);
764
765 while (num_lines-- > 0) {
766 write_aux_reg(cmd, paddr);
767 paddr += l2_line_sz;
768 }
769
770 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
771 read_aux_reg(ARC_REG_SLC_CTRL);
772
773 while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
774
775 spin_unlock_irqrestore(&lock, flags);
776#endif
777}
778
779#define slc_op(paddr, sz, op) slc_op_rgn(paddr, sz, op)
780
781noinline static void slc_entire_op(const int op)
782{
783 unsigned int ctrl, r = ARC_REG_SLC_CTRL;
784
785 ctrl = read_aux_reg(r);
786
787 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
788 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
789 else
790 ctrl |= SLC_CTRL_IM;
791
792 write_aux_reg(r, ctrl);
793
794 if (op & OP_INV) /* Inv or flush-n-inv use same cmd reg */
795 write_aux_reg(ARC_REG_SLC_INVALIDATE, 0x1);
796 else
797 write_aux_reg(ARC_REG_SLC_FLUSH, 0x1);
798
799 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
800 read_aux_reg(r);
801
802 /* Important to wait for flush to complete */
803 while (read_aux_reg(r) & SLC_CTRL_BUSY);
804}
805
806static inline void arc_slc_disable(void)
807{
808 const int r = ARC_REG_SLC_CTRL;
809
810 slc_entire_op(OP_FLUSH_N_INV);
811 write_aux_reg(r, read_aux_reg(r) | SLC_CTRL_DIS);
812}
813
814static inline void arc_slc_enable(void)
815{
816 const int r = ARC_REG_SLC_CTRL;
817
818 write_aux_reg(r, read_aux_reg(r) & ~SLC_CTRL_DIS);
819}
820
821/***********************************************************
822 * Exported APIs
823 */
824
825/*
826 * Handle cache congruency of kernel and userspace mappings of page when kernel
827 * writes-to/reads-from
828 *
829 * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
830 * -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
831 * -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
832 * -In SMP, if hardware caches are coherent
833 *
834 * There's a corollary case, where kernel READs from a userspace mapped page.
835 * If the U-mapping is not congruent to to K-mapping, former needs flushing.
836 */
837void flush_dcache_page(struct page *page)
838{
839 struct address_space *mapping;
840
841 if (!cache_is_vipt_aliasing()) {
842 clear_bit(PG_dc_clean, &page->flags);
843 return;
844 }
845
846 /* don't handle anon pages here */
847 mapping = page_mapping_file(page);
848 if (!mapping)
849 return;
850
851 /*
852 * pagecache page, file not yet mapped to userspace
853 * Make a note that K-mapping is dirty
854 */
855 if (!mapping_mapped(mapping)) {
856 clear_bit(PG_dc_clean, &page->flags);
857 } else if (page_mapcount(page)) {
858
859 /* kernel reading from page with U-mapping */
860 phys_addr_t paddr = (unsigned long)page_address(page);
861 unsigned long vaddr = page->index << PAGE_SHIFT;
862
863 if (addr_not_cache_congruent(paddr, vaddr))
864 __flush_dcache_page(paddr, vaddr);
865 }
866}
867EXPORT_SYMBOL(flush_dcache_page);
868
869/*
870 * DMA ops for systems with L1 cache only
871 * Make memory coherent with L1 cache by flushing/invalidating L1 lines
872 */
873static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
874{
875 __dc_line_op_k(start, sz, OP_FLUSH_N_INV);
876}
877
878static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
879{
880 __dc_line_op_k(start, sz, OP_INV);
881}
882
883static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
884{
885 __dc_line_op_k(start, sz, OP_FLUSH);
886}
887
888/*
889 * DMA ops for systems with both L1 and L2 caches, but without IOC
890 * Both L1 and L2 lines need to be explicitly flushed/invalidated
891 */
892static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
893{
894 __dc_line_op_k(start, sz, OP_FLUSH_N_INV);
895 slc_op(start, sz, OP_FLUSH_N_INV);
896}
897
898static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
899{
900 __dc_line_op_k(start, sz, OP_INV);
901 slc_op(start, sz, OP_INV);
902}
903
904static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
905{
906 __dc_line_op_k(start, sz, OP_FLUSH);
907 slc_op(start, sz, OP_FLUSH);
908}
909
910/*
911 * Exported DMA API
912 */
913void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
914{
915 __dma_cache_wback_inv(start, sz);
916}
917EXPORT_SYMBOL(dma_cache_wback_inv);
918
919void dma_cache_inv(phys_addr_t start, unsigned long sz)
920{
921 __dma_cache_inv(start, sz);
922}
923EXPORT_SYMBOL(dma_cache_inv);
924
925void dma_cache_wback(phys_addr_t start, unsigned long sz)
926{
927 __dma_cache_wback(start, sz);
928}
929EXPORT_SYMBOL(dma_cache_wback);
930
931/*
932 * This is API for making I/D Caches consistent when modifying
933 * kernel code (loadable modules, kprobes, kgdb...)
934 * This is called on insmod, with kernel virtual address for CODE of
935 * the module. ARC cache maintenance ops require PHY address thus we
936 * need to convert vmalloc addr to PHY addr
937 */
938void flush_icache_range(unsigned long kstart, unsigned long kend)
939{
940 unsigned int tot_sz;
941
942 WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
943
944 /* Shortcut for bigger flush ranges.
945 * Here we don't care if this was kernel virtual or phy addr
946 */
947 tot_sz = kend - kstart;
948 if (tot_sz > PAGE_SIZE) {
949 flush_cache_all();
950 return;
951 }
952
953 /* Case: Kernel Phy addr (0x8000_0000 onwards) */
954 if (likely(kstart > PAGE_OFFSET)) {
955 /*
956 * The 2nd arg despite being paddr will be used to index icache
957 * This is OK since no alternate virtual mappings will exist
958 * given the callers for this case: kprobe/kgdb in built-in
959 * kernel code only.
960 */
961 __sync_icache_dcache(kstart, kstart, kend - kstart);
962 return;
963 }
964
965 /*
966 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
967 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
968 * handling of kernel vaddr.
969 *
970 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
971 * it still needs to handle a 2 page scenario, where the range
972 * straddles across 2 virtual pages and hence need for loop
973 */
974 while (tot_sz > 0) {
975 unsigned int off, sz;
976 unsigned long phy, pfn;
977
978 off = kstart % PAGE_SIZE;
979 pfn = vmalloc_to_pfn((void *)kstart);
980 phy = (pfn << PAGE_SHIFT) + off;
981 sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
982 __sync_icache_dcache(phy, kstart, sz);
983 kstart += sz;
984 tot_sz -= sz;
985 }
986}
987EXPORT_SYMBOL(flush_icache_range);
988
989/*
990 * General purpose helper to make I and D cache lines consistent.
991 * @paddr is phy addr of region
992 * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
993 * However in one instance, when called by kprobe (for a breakpt in
994 * builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
995 * use a paddr to index the cache (despite VIPT). This is fine since since a
996 * builtin kernel page will not have any virtual mappings.
997 * kprobe on loadable module will be kernel vaddr.
998 */
999void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
1000{
1001 __dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
1002 __ic_line_inv_vaddr(paddr, vaddr, len);
1003}
1004
1005/* wrapper to compile time eliminate alignment checks in flush loop */
1006void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
1007{
1008 __ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
1009}
1010
1011/*
1012 * wrapper to clearout kernel or userspace mappings of a page
1013 * For kernel mappings @vaddr == @paddr
1014 */
1015void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
1016{
1017 __dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
1018}
1019
1020noinline void flush_cache_all(void)
1021{
1022 unsigned long flags;
1023
1024 local_irq_save(flags);
1025
1026 __ic_entire_inv();
1027 __dc_entire_op(OP_FLUSH_N_INV);
1028
1029 local_irq_restore(flags);
1030
1031}
1032
1033#ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
1034
1035void flush_cache_mm(struct mm_struct *mm)
1036{
1037 flush_cache_all();
1038}
1039
1040void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
1041 unsigned long pfn)
1042{
1043 phys_addr_t paddr = pfn << PAGE_SHIFT;
1044
1045 u_vaddr &= PAGE_MASK;
1046
1047 __flush_dcache_page(paddr, u_vaddr);
1048
1049 if (vma->vm_flags & VM_EXEC)
1050 __inv_icache_page(paddr, u_vaddr);
1051}
1052
1053void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
1054 unsigned long end)
1055{
1056 flush_cache_all();
1057}
1058
1059void flush_anon_page(struct vm_area_struct *vma, struct page *page,
1060 unsigned long u_vaddr)
1061{
1062 /* TBD: do we really need to clear the kernel mapping */
1063 __flush_dcache_page((phys_addr_t)page_address(page), u_vaddr);
1064 __flush_dcache_page((phys_addr_t)page_address(page),
1065 (phys_addr_t)page_address(page));
1066
1067}
1068
1069#endif
1070
1071void copy_user_highpage(struct page *to, struct page *from,
1072 unsigned long u_vaddr, struct vm_area_struct *vma)
1073{
1074 void *kfrom = kmap_atomic(from);
1075 void *kto = kmap_atomic(to);
1076 int clean_src_k_mappings = 0;
1077
1078 /*
1079 * If SRC page was already mapped in userspace AND it's U-mapping is
1080 * not congruent with K-mapping, sync former to physical page so that
1081 * K-mapping in memcpy below, sees the right data
1082 *
1083 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
1084 * equally valid for SRC page as well
1085 *
1086 * For !VIPT cache, all of this gets compiled out as
1087 * addr_not_cache_congruent() is 0
1088 */
1089 if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
1090 __flush_dcache_page((unsigned long)kfrom, u_vaddr);
1091 clean_src_k_mappings = 1;
1092 }
1093
1094 copy_page(kto, kfrom);
1095
1096 /*
1097 * Mark DST page K-mapping as dirty for a later finalization by
1098 * update_mmu_cache(). Although the finalization could have been done
1099 * here as well (given that both vaddr/paddr are available).
1100 * But update_mmu_cache() already has code to do that for other
1101 * non copied user pages (e.g. read faults which wire in pagecache page
1102 * directly).
1103 */
1104 clear_bit(PG_dc_clean, &to->flags);
1105
1106 /*
1107 * if SRC was already usermapped and non-congruent to kernel mapping
1108 * sync the kernel mapping back to physical page
1109 */
1110 if (clean_src_k_mappings) {
1111 __flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
1112 set_bit(PG_dc_clean, &from->flags);
1113 } else {
1114 clear_bit(PG_dc_clean, &from->flags);
1115 }
1116
1117 kunmap_atomic(kto);
1118 kunmap_atomic(kfrom);
1119}
1120
1121void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
1122{
1123 clear_page(to);
1124 clear_bit(PG_dc_clean, &page->flags);
1125}
1126
1127
1128/**********************************************************************
1129 * Explicit Cache flush request from user space via syscall
1130 * Needed for JITs which generate code on the fly
1131 */
1132SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
1133{
1134 /* TBD: optimize this */
1135 flush_cache_all();
1136 return 0;
1137}
1138
1139/*
1140 * IO-Coherency (IOC) setup rules:
1141 *
1142 * 1. Needs to be at system level, so only once by Master core
1143 * Non-Masters need not be accessing caches at that time
1144 * - They are either HALT_ON_RESET and kick started much later or
1145 * - if run on reset, need to ensure that arc_platform_smp_wait_to_boot()
1146 * doesn't perturb caches or coherency unit
1147 *
1148 * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC,
1149 * otherwise any straggler data might behave strangely post IOC enabling
1150 *
1151 * 3. All Caches need to be disabled when setting up IOC to elide any in-flight
1152 * Coherency transactions
1153 */
1154noinline void __init arc_ioc_setup(void)
1155{
1156 unsigned int ioc_base, mem_sz;
1157
1158 /*
1159 * If IOC was already enabled (due to bootloader) it technically needs to
1160 * be reconfigured with aperture base,size corresponding to Linux memory map
1161 * which will certainly be different than uboot's. But disabling and
1162 * reenabling IOC when DMA might be potentially active is tricky business.
1163 * To avoid random memory issues later, just panic here and ask user to
1164 * upgrade bootloader to one which doesn't enable IOC
1165 */
1166 if (read_aux_reg(ARC_REG_IO_COH_ENABLE) & ARC_IO_COH_ENABLE_BIT)
1167 panic("IOC already enabled, please upgrade bootloader!\n");
1168
1169 if (!ioc_enable)
1170 return;
1171
1172 /* Flush + invalidate + disable L1 dcache */
1173 __dc_disable();
1174
1175 /* Flush + invalidate SLC */
1176 if (read_aux_reg(ARC_REG_SLC_BCR))
1177 slc_entire_op(OP_FLUSH_N_INV);
1178
1179 /*
1180 * currently IOC Aperture covers entire DDR
1181 * TBD: fix for PGU + 1GB of low mem
1182 * TBD: fix for PAE
1183 */
1184 mem_sz = arc_get_mem_sz();
1185
1186 if (!is_power_of_2(mem_sz) || mem_sz < 4096)
1187 panic("IOC Aperture size must be power of 2 larger than 4KB");
1188
1189 /*
1190 * IOC Aperture size decoded as 2 ^ (SIZE + 2) KB,
1191 * so setting 0x11 implies 512MB, 0x12 implies 1GB...
1192 */
1193 write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, order_base_2(mem_sz >> 10) - 2);
1194
1195 /* for now assume kernel base is start of IOC aperture */
1196 ioc_base = CONFIG_LINUX_RAM_BASE;
1197
1198 if (ioc_base % mem_sz != 0)
1199 panic("IOC Aperture start must be aligned to the size of the aperture");
1200
1201 write_aux_reg(ARC_REG_IO_COH_AP0_BASE, ioc_base >> 12);
1202 write_aux_reg(ARC_REG_IO_COH_PARTIAL, ARC_IO_COH_PARTIAL_BIT);
1203 write_aux_reg(ARC_REG_IO_COH_ENABLE, ARC_IO_COH_ENABLE_BIT);
1204
1205 /* Re-enable L1 dcache */
1206 __dc_enable();
1207}
1208
1209/*
1210 * Cache related boot time checks/setups only needed on master CPU:
1211 * - Geometry checks (kernel build and hardware agree: e.g. L1_CACHE_BYTES)
1212 * Assume SMP only, so all cores will have same cache config. A check on
1213 * one core suffices for all
1214 * - IOC setup / dma callbacks only need to be done once
1215 */
1216void __init arc_cache_init_master(void)
1217{
1218 unsigned int __maybe_unused cpu = smp_processor_id();
1219
1220 if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
1221 struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
1222
1223 if (!ic->line_len)
1224 panic("cache support enabled but non-existent cache\n");
1225
1226 if (ic->line_len != L1_CACHE_BYTES)
1227 panic("ICache line [%d] != kernel Config [%d]",
1228 ic->line_len, L1_CACHE_BYTES);
1229
1230 /*
1231 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
1232 * pair to provide vaddr/paddr respectively, just as in MMU v3
1233 */
1234 if (is_isa_arcv2() && ic->alias)
1235 _cache_line_loop_ic_fn = __cache_line_loop_v3;
1236 else
1237 _cache_line_loop_ic_fn = __cache_line_loop;
1238 }
1239
1240 if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
1241 struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
1242
1243 if (!dc->line_len)
1244 panic("cache support enabled but non-existent cache\n");
1245
1246 if (dc->line_len != L1_CACHE_BYTES)
1247 panic("DCache line [%d] != kernel Config [%d]",
1248 dc->line_len, L1_CACHE_BYTES);
1249
1250 /* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
1251 if (is_isa_arcompact()) {
1252 int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
1253 int num_colors = dc->sz_k/dc->assoc/TO_KB(PAGE_SIZE);
1254
1255 if (dc->alias) {
1256 if (!handled)
1257 panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1258 if (CACHE_COLORS_NUM != num_colors)
1259 panic("CACHE_COLORS_NUM not optimized for config\n");
1260 } else if (!dc->alias && handled) {
1261 panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1262 }
1263 }
1264 }
1265
1266 /*
1267 * Check that SMP_CACHE_BYTES (and hence ARCH_DMA_MINALIGN) is larger
1268 * or equal to any cache line length.
1269 */
1270 BUILD_BUG_ON_MSG(L1_CACHE_BYTES > SMP_CACHE_BYTES,
1271 "SMP_CACHE_BYTES must be >= any cache line length");
1272 if (is_isa_arcv2() && (l2_line_sz > SMP_CACHE_BYTES))
1273 panic("L2 Cache line [%d] > kernel Config [%d]\n",
1274 l2_line_sz, SMP_CACHE_BYTES);
1275
1276 /* Note that SLC disable not formally supported till HS 3.0 */
1277 if (is_isa_arcv2() && l2_line_sz && !slc_enable)
1278 arc_slc_disable();
1279
1280 if (is_isa_arcv2() && ioc_exists)
1281 arc_ioc_setup();
1282
1283 if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1284 __dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1285 __dma_cache_inv = __dma_cache_inv_slc;
1286 __dma_cache_wback = __dma_cache_wback_slc;
1287 } else {
1288 __dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1289 __dma_cache_inv = __dma_cache_inv_l1;
1290 __dma_cache_wback = __dma_cache_wback_l1;
1291 }
1292 /*
1293 * In case of IOC (say IOC+SLC case), pointers above could still be set
1294 * but end up not being relevant as the first function in chain is not
1295 * called at all for devices using coherent DMA.
1296 * arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
1297 */
1298}
1299
1300void __ref arc_cache_init(void)
1301{
1302 unsigned int __maybe_unused cpu = smp_processor_id();
1303 char str[256];
1304
1305 pr_info("%s", arc_cache_mumbojumbo(0, str, sizeof(str)));
1306
1307 if (!cpu)
1308 arc_cache_init_master();
1309
1310 /*
1311 * In PAE regime, TLB and cache maintenance ops take wider addresses
1312 * And even if PAE is not enabled in kernel, the upper 32-bits still need
1313 * to be zeroed to keep the ops sane.
1314 * As an optimization for more common !PAE enabled case, zero them out
1315 * once at init, rather than checking/setting to 0 for every runtime op
1316 */
1317 if (is_isa_arcv2() && pae40_exist_but_not_enab()) {
1318
1319 if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE))
1320 write_aux_reg(ARC_REG_IC_PTAG_HI, 0);
1321
1322 if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE))
1323 write_aux_reg(ARC_REG_DC_PTAG_HI, 0);
1324
1325 if (l2_line_sz) {
1326 write_aux_reg(ARC_REG_SLC_RGN_END1, 0);
1327 write_aux_reg(ARC_REG_SLC_RGN_START1, 0);
1328 }
1329 }
1330}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * ARC Cache Management
4 *
5 * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
6 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
7 */
8
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/cache.h>
13#include <linux/mmu_context.h>
14#include <linux/syscalls.h>
15#include <linux/uaccess.h>
16#include <linux/pagemap.h>
17#include <asm/cacheflush.h>
18#include <asm/cachectl.h>
19#include <asm/setup.h>
20
21#ifdef CONFIG_ISA_ARCV2
22#define USE_RGN_FLSH 1
23#endif
24
25static int l2_line_sz;
26static int ioc_exists;
27int slc_enable = 1, ioc_enable = 1;
28unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
29unsigned long perip_end = 0xFFFFFFFF; /* legacy value */
30
31static struct cpuinfo_arc_cache {
32 unsigned int sz_k, line_len, colors;
33} ic_info, dc_info, slc_info;
34
35void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
36 unsigned long sz, const int op, const int full_page);
37
38void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
39void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
40void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
41
42static int read_decode_cache_bcr_arcv2(int c, char *buf, int len)
43{
44 struct cpuinfo_arc_cache *p_slc = &slc_info;
45 struct bcr_identity ident;
46 struct bcr_generic sbcr;
47 struct bcr_clust_cfg cbcr;
48 struct bcr_volatile vol;
49 int n = 0;
50
51 READ_BCR(ARC_REG_SLC_BCR, sbcr);
52 if (sbcr.ver) {
53 struct bcr_slc_cfg slc_cfg;
54 READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
55 p_slc->sz_k = 128 << slc_cfg.sz;
56 l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
57 n += scnprintf(buf + n, len - n,
58 "SLC\t\t: %uK, %uB Line%s\n",
59 p_slc->sz_k, p_slc->line_len, IS_USED_RUN(slc_enable));
60 }
61
62 READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
63 if (cbcr.c) {
64 ioc_exists = 1;
65
66 /*
67 * As for today we don't support both IOC and ZONE_HIGHMEM enabled
68 * simultaneously. This happens because as of today IOC aperture covers
69 * only ZONE_NORMAL (low mem) and any dma transactions outside this
70 * region won't be HW coherent.
71 * If we want to use both IOC and ZONE_HIGHMEM we can use
72 * bounce_buffer to handle dma transactions to HIGHMEM.
73 * Also it is possible to modify dma_direct cache ops or increase IOC
74 * aperture size if we are planning to use HIGHMEM without PAE.
75 */
76 if (IS_ENABLED(CONFIG_HIGHMEM) || is_pae40_enabled())
77 ioc_enable = 0;
78 } else {
79 ioc_enable = 0;
80 }
81
82 READ_BCR(AUX_IDENTITY, ident);
83
84 /* HS 2.0 didn't have AUX_VOL */
85 if (ident.family > 0x51) {
86 READ_BCR(AUX_VOL, vol);
87 perip_base = vol.start << 28;
88 /* HS 3.0 has limit and strict-ordering fields */
89 if (ident.family > 0x52)
90 perip_end = (vol.limit << 28) - 1;
91 }
92
93 n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
94 perip_base,
95 IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency (per-device) "));
96
97 return n;
98}
99
100int arc_cache_mumbojumbo(int c, char *buf, int len)
101{
102 struct cpuinfo_arc_cache *p_ic = &ic_info, *p_dc = &dc_info;
103 struct bcr_cache ibcr, dbcr;
104 int vipt, assoc;
105 int n = 0;
106
107 READ_BCR(ARC_REG_IC_BCR, ibcr);
108 if (!ibcr.ver)
109 goto dc_chk;
110
111 if (is_isa_arcompact() && (ibcr.ver <= 3)) {
112 BUG_ON(ibcr.config != 3);
113 assoc = 2; /* Fixed to 2w set assoc */
114 } else if (is_isa_arcv2() && (ibcr.ver >= 4)) {
115 assoc = 1 << ibcr.config; /* 1,2,4,8 */
116 }
117
118 p_ic->line_len = 8 << ibcr.line_len;
119 p_ic->sz_k = 1 << (ibcr.sz - 1);
120 p_ic->colors = p_ic->sz_k/assoc/TO_KB(PAGE_SIZE);
121
122 n += scnprintf(buf + n, len - n,
123 "I-Cache\t\t: %uK, %dway/set, %uB Line, VIPT%s%s\n",
124 p_ic->sz_k, assoc, p_ic->line_len,
125 p_ic->colors > 1 ? " aliasing" : "",
126 IS_USED_CFG(CONFIG_ARC_HAS_ICACHE));
127
128dc_chk:
129 READ_BCR(ARC_REG_DC_BCR, dbcr);
130 if (!dbcr.ver)
131 goto slc_chk;
132
133 if (is_isa_arcompact() && (dbcr.ver <= 3)) {
134 BUG_ON(dbcr.config != 2);
135 vipt = 1;
136 assoc = 4; /* Fixed to 4w set assoc */
137 p_dc->colors = p_dc->sz_k/assoc/TO_KB(PAGE_SIZE);
138 } else if (is_isa_arcv2() && (dbcr.ver >= 4)) {
139 vipt = 0;
140 assoc = 1 << dbcr.config; /* 1,2,4,8 */
141 p_dc->colors = 1; /* PIPT so can't VIPT alias */
142 }
143
144 p_dc->line_len = 16 << dbcr.line_len;
145 p_dc->sz_k = 1 << (dbcr.sz - 1);
146
147 n += scnprintf(buf + n, len - n,
148 "D-Cache\t\t: %uK, %dway/set, %uB Line, %s%s\n",
149 p_dc->sz_k, assoc, p_dc->line_len,
150 vipt ? "VIPT" : "PIPT",
151 IS_USED_CFG(CONFIG_ARC_HAS_DCACHE));
152
153slc_chk:
154 if (is_isa_arcv2())
155 n += read_decode_cache_bcr_arcv2(c, buf + n, len - n);
156
157 return n;
158}
159
160/*
161 * Line Operation on {I,D}-Cache
162 */
163
164#define OP_INV 0x1
165#define OP_FLUSH 0x2
166#define OP_FLUSH_N_INV 0x3
167#define OP_INV_IC 0x4
168
169/*
170 * Cache Flush programming model
171 *
172 * ARC700 MMUv3 I$ and D$ are both VIPT and can potentially alias.
173 * Programming model requires both paddr and vaddr irrespecive of aliasing
174 * considerations:
175 * - vaddr in {I,D}C_IV?L
176 * - paddr in {I,D}C_PTAG
177 *
178 * In HS38x (MMUv4), D$ is PIPT, I$ is VIPT and can still alias.
179 * Programming model is different for aliasing vs. non-aliasing I$
180 * - D$ / Non-aliasing I$: only paddr in {I,D}C_IV?L
181 * - Aliasing I$: same as ARC700 above (so MMUv3 routine used for MMUv4 I$)
182 *
183 * - If PAE40 is enabled, independent of aliasing considerations, the higher
184 * bits needs to be written into PTAG_HI
185 */
186
187static inline
188void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
189 unsigned long sz, const int op, const int full_page)
190{
191 unsigned int aux_cmd, aux_tag;
192 int num_lines;
193
194 if (op == OP_INV_IC) {
195 aux_cmd = ARC_REG_IC_IVIL;
196 aux_tag = ARC_REG_IC_PTAG;
197 } else {
198 aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
199 aux_tag = ARC_REG_DC_PTAG;
200 }
201
202 /* Ensure we properly floor/ceil the non-line aligned/sized requests
203 * and have @paddr - aligned to cache line and integral @num_lines.
204 * This however can be avoided for page sized since:
205 * -@paddr will be cache-line aligned already (being page aligned)
206 * -@sz will be integral multiple of line size (being page sized).
207 */
208 if (!full_page) {
209 sz += paddr & ~CACHE_LINE_MASK;
210 paddr &= CACHE_LINE_MASK;
211 vaddr &= CACHE_LINE_MASK;
212 }
213 num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
214
215 /*
216 * MMUv3, cache ops require paddr in PTAG reg
217 * if V-P const for loop, PTAG can be written once outside loop
218 */
219 if (full_page)
220 write_aux_reg(aux_tag, paddr);
221
222 /*
223 * This is technically for MMU v4, using the MMU v3 programming model
224 * Special work for HS38 aliasing I-cache configuration with PAE40
225 * - upper 8 bits of paddr need to be written into PTAG_HI
226 * - (and needs to be written before the lower 32 bits)
227 * Note that PTAG_HI is hoisted outside the line loop
228 */
229 if (is_pae40_enabled() && op == OP_INV_IC)
230 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
231
232 while (num_lines-- > 0) {
233 if (!full_page) {
234 write_aux_reg(aux_tag, paddr);
235 paddr += L1_CACHE_BYTES;
236 }
237
238 write_aux_reg(aux_cmd, vaddr);
239 vaddr += L1_CACHE_BYTES;
240 }
241}
242
243#ifndef USE_RGN_FLSH
244
245/*
246 */
247static inline
248void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
249 unsigned long sz, const int op, const int full_page)
250{
251 unsigned int aux_cmd;
252 int num_lines;
253
254 if (op == OP_INV_IC) {
255 aux_cmd = ARC_REG_IC_IVIL;
256 } else {
257 /* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
258 aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
259 }
260
261 /* Ensure we properly floor/ceil the non-line aligned/sized requests
262 * and have @paddr - aligned to cache line and integral @num_lines.
263 * This however can be avoided for page sized since:
264 * -@paddr will be cache-line aligned already (being page aligned)
265 * -@sz will be integral multiple of line size (being page sized).
266 */
267 if (!full_page) {
268 sz += paddr & ~CACHE_LINE_MASK;
269 paddr &= CACHE_LINE_MASK;
270 }
271
272 num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
273
274 /*
275 * For HS38 PAE40 configuration
276 * - upper 8 bits of paddr need to be written into PTAG_HI
277 * - (and needs to be written before the lower 32 bits)
278 */
279 if (is_pae40_enabled()) {
280 if (op == OP_INV_IC)
281 /*
282 * Non aliasing I-cache in HS38,
283 * aliasing I-cache handled in __cache_line_loop_v3()
284 */
285 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
286 else
287 write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
288 }
289
290 while (num_lines-- > 0) {
291 write_aux_reg(aux_cmd, paddr);
292 paddr += L1_CACHE_BYTES;
293 }
294}
295
296#else
297
298/*
299 * optimized flush operation which takes a region as opposed to iterating per line
300 */
301static inline
302void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
303 unsigned long sz, const int op, const int full_page)
304{
305 unsigned int s, e;
306
307 /* Only for Non aliasing I-cache in HS38 */
308 if (op == OP_INV_IC) {
309 s = ARC_REG_IC_IVIR;
310 e = ARC_REG_IC_ENDR;
311 } else {
312 s = ARC_REG_DC_STARTR;
313 e = ARC_REG_DC_ENDR;
314 }
315
316 if (!full_page) {
317 /* for any leading gap between @paddr and start of cache line */
318 sz += paddr & ~CACHE_LINE_MASK;
319 paddr &= CACHE_LINE_MASK;
320
321 /*
322 * account for any trailing gap to end of cache line
323 * this is equivalent to DIV_ROUND_UP() in line ops above
324 */
325 sz += L1_CACHE_BYTES - 1;
326 }
327
328 if (is_pae40_enabled()) {
329 /* TBD: check if crossing 4TB boundary */
330 if (op == OP_INV_IC)
331 write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
332 else
333 write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
334 }
335
336 /* ENDR needs to be set ahead of START */
337 write_aux_reg(e, paddr + sz); /* ENDR is exclusive */
338 write_aux_reg(s, paddr);
339
340 /* caller waits on DC_CTRL.FS */
341}
342
343#endif
344
345#ifdef CONFIG_ARC_MMU_V3
346#define __cache_line_loop __cache_line_loop_v3
347#else
348#define __cache_line_loop __cache_line_loop_v4
349#endif
350
351#ifdef CONFIG_ARC_HAS_DCACHE
352
353/***************************************************************
354 * Machine specific helpers for Entire D-Cache or Per Line ops
355 */
356
357#ifndef USE_RGN_FLSH
358/*
359 * this version avoids extra read/write of DC_CTRL for flush or invalid ops
360 * in the non region flush regime (such as for ARCompact)
361 */
362static inline void __before_dc_op(const int op)
363{
364 if (op == OP_FLUSH_N_INV) {
365 /* Dcache provides 2 cmd: FLUSH or INV
366 * INV in turn has sub-modes: DISCARD or FLUSH-BEFORE
367 * flush-n-inv is achieved by INV cmd but with IM=1
368 * So toggle INV sub-mode depending on op request and default
369 */
370 const unsigned int ctl = ARC_REG_DC_CTRL;
371 write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
372 }
373}
374
375#else
376
377static inline void __before_dc_op(const int op)
378{
379 const unsigned int ctl = ARC_REG_DC_CTRL;
380 unsigned int val = read_aux_reg(ctl);
381
382 if (op == OP_FLUSH_N_INV) {
383 val |= DC_CTRL_INV_MODE_FLUSH;
384 }
385
386 if (op != OP_INV_IC) {
387 /*
388 * Flush / Invalidate is provided by DC_CTRL.RNG_OP 0 or 1
389 * combined Flush-n-invalidate uses DC_CTRL.IM = 1 set above
390 */
391 val &= ~DC_CTRL_RGN_OP_MSK;
392 if (op & OP_INV)
393 val |= DC_CTRL_RGN_OP_INV;
394 }
395 write_aux_reg(ctl, val);
396}
397
398#endif
399
400
401static inline void __after_dc_op(const int op)
402{
403 if (op & OP_FLUSH) {
404 const unsigned int ctl = ARC_REG_DC_CTRL;
405 unsigned int reg;
406
407 /* flush / flush-n-inv both wait */
408 while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
409 ;
410
411 /* Switch back to default Invalidate mode */
412 if (op == OP_FLUSH_N_INV)
413 write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
414 }
415}
416
417/*
418 * Operation on Entire D-Cache
419 * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
420 * Note that constant propagation ensures all the checks are gone
421 * in generated code
422 */
423static inline void __dc_entire_op(const int op)
424{
425 int aux;
426
427 __before_dc_op(op);
428
429 if (op & OP_INV) /* Inv or flush-n-inv use same cmd reg */
430 aux = ARC_REG_DC_IVDC;
431 else
432 aux = ARC_REG_DC_FLSH;
433
434 write_aux_reg(aux, 0x1);
435
436 __after_dc_op(op);
437}
438
439static inline void __dc_disable(void)
440{
441 const int r = ARC_REG_DC_CTRL;
442
443 __dc_entire_op(OP_FLUSH_N_INV);
444 write_aux_reg(r, read_aux_reg(r) | DC_CTRL_DIS);
445}
446
447static void __dc_enable(void)
448{
449 const int r = ARC_REG_DC_CTRL;
450
451 write_aux_reg(r, read_aux_reg(r) & ~DC_CTRL_DIS);
452}
453
454/* For kernel mappings cache operation: index is same as paddr */
455#define __dc_line_op_k(p, sz, op) __dc_line_op(p, p, sz, op)
456
457/*
458 * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
459 */
460static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
461 unsigned long sz, const int op)
462{
463 const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
464 unsigned long flags;
465
466 local_irq_save(flags);
467
468 __before_dc_op(op);
469
470 __cache_line_loop(paddr, vaddr, sz, op, full_page);
471
472 __after_dc_op(op);
473
474 local_irq_restore(flags);
475}
476
477#else
478
479#define __dc_entire_op(op)
480#define __dc_disable()
481#define __dc_enable()
482#define __dc_line_op(paddr, vaddr, sz, op)
483#define __dc_line_op_k(paddr, sz, op)
484
485#endif /* CONFIG_ARC_HAS_DCACHE */
486
487#ifdef CONFIG_ARC_HAS_ICACHE
488
489static inline void __ic_entire_inv(void)
490{
491 write_aux_reg(ARC_REG_IC_IVIC, 1);
492 read_aux_reg(ARC_REG_IC_CTRL); /* blocks */
493}
494
495static inline void
496__ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
497 unsigned long sz)
498{
499 const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
500 unsigned long flags;
501
502 local_irq_save(flags);
503 (*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC, full_page);
504 local_irq_restore(flags);
505}
506
507#ifndef CONFIG_SMP
508
509#define __ic_line_inv_vaddr(p, v, s) __ic_line_inv_vaddr_local(p, v, s)
510
511#else
512
513struct ic_inv_args {
514 phys_addr_t paddr, vaddr;
515 int sz;
516};
517
518static void __ic_line_inv_vaddr_helper(void *info)
519{
520 struct ic_inv_args *ic_inv = info;
521
522 __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
523}
524
525static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
526 unsigned long sz)
527{
528 struct ic_inv_args ic_inv = {
529 .paddr = paddr,
530 .vaddr = vaddr,
531 .sz = sz
532 };
533
534 on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
535}
536
537#endif /* CONFIG_SMP */
538
539#else /* !CONFIG_ARC_HAS_ICACHE */
540
541#define __ic_entire_inv()
542#define __ic_line_inv_vaddr(pstart, vstart, sz)
543
544#endif /* CONFIG_ARC_HAS_ICACHE */
545
546static noinline void slc_op_rgn(phys_addr_t paddr, unsigned long sz, const int op)
547{
548#ifdef CONFIG_ISA_ARCV2
549 /*
550 * SLC is shared between all cores and concurrent aux operations from
551 * multiple cores need to be serialized using a spinlock
552 * A concurrent operation can be silently ignored and/or the old/new
553 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
554 * below)
555 */
556 static DEFINE_SPINLOCK(lock);
557 unsigned long flags;
558 unsigned int ctrl;
559 phys_addr_t end;
560
561 spin_lock_irqsave(&lock, flags);
562
563 /*
564 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
565 * - b'000 (default) is Flush,
566 * - b'001 is Invalidate if CTRL.IM == 0
567 * - b'001 is Flush-n-Invalidate if CTRL.IM == 1
568 */
569 ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
570
571 /* Don't rely on default value of IM bit */
572 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
573 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
574 else
575 ctrl |= SLC_CTRL_IM;
576
577 if (op & OP_INV)
578 ctrl |= SLC_CTRL_RGN_OP_INV; /* Inv or flush-n-inv */
579 else
580 ctrl &= ~SLC_CTRL_RGN_OP_INV;
581
582 write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
583
584 /*
585 * Lower bits are ignored, no need to clip
586 * END needs to be setup before START (latter triggers the operation)
587 * END can't be same as START, so add (l2_line_sz - 1) to sz
588 */
589 end = paddr + sz + l2_line_sz - 1;
590 if (is_pae40_enabled())
591 write_aux_reg(ARC_REG_SLC_RGN_END1, upper_32_bits(end));
592
593 write_aux_reg(ARC_REG_SLC_RGN_END, lower_32_bits(end));
594
595 if (is_pae40_enabled())
596 write_aux_reg(ARC_REG_SLC_RGN_START1, upper_32_bits(paddr));
597
598 write_aux_reg(ARC_REG_SLC_RGN_START, lower_32_bits(paddr));
599
600 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
601 read_aux_reg(ARC_REG_SLC_CTRL);
602
603 while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
604
605 spin_unlock_irqrestore(&lock, flags);
606#endif
607}
608
609static __maybe_unused noinline void slc_op_line(phys_addr_t paddr, unsigned long sz, const int op)
610{
611#ifdef CONFIG_ISA_ARCV2
612 /*
613 * SLC is shared between all cores and concurrent aux operations from
614 * multiple cores need to be serialized using a spinlock
615 * A concurrent operation can be silently ignored and/or the old/new
616 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
617 * below)
618 */
619 static DEFINE_SPINLOCK(lock);
620
621 const unsigned long SLC_LINE_MASK = ~(l2_line_sz - 1);
622 unsigned int ctrl, cmd;
623 unsigned long flags;
624 int num_lines;
625
626 spin_lock_irqsave(&lock, flags);
627
628 ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
629
630 /* Don't rely on default value of IM bit */
631 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
632 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
633 else
634 ctrl |= SLC_CTRL_IM;
635
636 write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
637
638 cmd = op & OP_INV ? ARC_AUX_SLC_IVDL : ARC_AUX_SLC_FLDL;
639
640 sz += paddr & ~SLC_LINE_MASK;
641 paddr &= SLC_LINE_MASK;
642
643 num_lines = DIV_ROUND_UP(sz, l2_line_sz);
644
645 while (num_lines-- > 0) {
646 write_aux_reg(cmd, paddr);
647 paddr += l2_line_sz;
648 }
649
650 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
651 read_aux_reg(ARC_REG_SLC_CTRL);
652
653 while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
654
655 spin_unlock_irqrestore(&lock, flags);
656#endif
657}
658
659#define slc_op(paddr, sz, op) slc_op_rgn(paddr, sz, op)
660
661noinline static void slc_entire_op(const int op)
662{
663 unsigned int ctrl, r = ARC_REG_SLC_CTRL;
664
665 ctrl = read_aux_reg(r);
666
667 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
668 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
669 else
670 ctrl |= SLC_CTRL_IM;
671
672 write_aux_reg(r, ctrl);
673
674 if (op & OP_INV) /* Inv or flush-n-inv use same cmd reg */
675 write_aux_reg(ARC_REG_SLC_INVALIDATE, 0x1);
676 else
677 write_aux_reg(ARC_REG_SLC_FLUSH, 0x1);
678
679 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
680 read_aux_reg(r);
681
682 /* Important to wait for flush to complete */
683 while (read_aux_reg(r) & SLC_CTRL_BUSY);
684}
685
686static inline void arc_slc_disable(void)
687{
688 const int r = ARC_REG_SLC_CTRL;
689
690 slc_entire_op(OP_FLUSH_N_INV);
691 write_aux_reg(r, read_aux_reg(r) | SLC_CTRL_DIS);
692}
693
694static inline void arc_slc_enable(void)
695{
696 const int r = ARC_REG_SLC_CTRL;
697
698 write_aux_reg(r, read_aux_reg(r) & ~SLC_CTRL_DIS);
699}
700
701/***********************************************************
702 * Exported APIs
703 */
704
705void flush_dcache_folio(struct folio *folio)
706{
707 clear_bit(PG_dc_clean, &folio->flags);
708 return;
709}
710EXPORT_SYMBOL(flush_dcache_folio);
711
712void flush_dcache_page(struct page *page)
713{
714 return flush_dcache_folio(page_folio(page));
715}
716EXPORT_SYMBOL(flush_dcache_page);
717
718/*
719 * DMA ops for systems with L1 cache only
720 * Make memory coherent with L1 cache by flushing/invalidating L1 lines
721 */
722static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
723{
724 __dc_line_op_k(start, sz, OP_FLUSH_N_INV);
725}
726
727static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
728{
729 __dc_line_op_k(start, sz, OP_INV);
730}
731
732static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
733{
734 __dc_line_op_k(start, sz, OP_FLUSH);
735}
736
737/*
738 * DMA ops for systems with both L1 and L2 caches, but without IOC
739 * Both L1 and L2 lines need to be explicitly flushed/invalidated
740 */
741static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
742{
743 __dc_line_op_k(start, sz, OP_FLUSH_N_INV);
744 slc_op(start, sz, OP_FLUSH_N_INV);
745}
746
747static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
748{
749 __dc_line_op_k(start, sz, OP_INV);
750 slc_op(start, sz, OP_INV);
751}
752
753static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
754{
755 __dc_line_op_k(start, sz, OP_FLUSH);
756 slc_op(start, sz, OP_FLUSH);
757}
758
759/*
760 * Exported DMA API
761 */
762void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
763{
764 __dma_cache_wback_inv(start, sz);
765}
766EXPORT_SYMBOL(dma_cache_wback_inv);
767
768void dma_cache_inv(phys_addr_t start, unsigned long sz)
769{
770 __dma_cache_inv(start, sz);
771}
772EXPORT_SYMBOL(dma_cache_inv);
773
774void dma_cache_wback(phys_addr_t start, unsigned long sz)
775{
776 __dma_cache_wback(start, sz);
777}
778EXPORT_SYMBOL(dma_cache_wback);
779
780/*
781 * This is API for making I/D Caches consistent when modifying
782 * kernel code (loadable modules, kprobes, kgdb...)
783 * This is called on insmod, with kernel virtual address for CODE of
784 * the module. ARC cache maintenance ops require PHY address thus we
785 * need to convert vmalloc addr to PHY addr
786 */
787void flush_icache_range(unsigned long kstart, unsigned long kend)
788{
789 unsigned int tot_sz;
790
791 WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
792
793 /* Shortcut for bigger flush ranges.
794 * Here we don't care if this was kernel virtual or phy addr
795 */
796 tot_sz = kend - kstart;
797 if (tot_sz > PAGE_SIZE) {
798 flush_cache_all();
799 return;
800 }
801
802 /* Case: Kernel Phy addr (0x8000_0000 onwards) */
803 if (likely(kstart > PAGE_OFFSET)) {
804 /*
805 * The 2nd arg despite being paddr will be used to index icache
806 * This is OK since no alternate virtual mappings will exist
807 * given the callers for this case: kprobe/kgdb in built-in
808 * kernel code only.
809 */
810 __sync_icache_dcache(kstart, kstart, kend - kstart);
811 return;
812 }
813
814 /*
815 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
816 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
817 * handling of kernel vaddr.
818 *
819 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
820 * it still needs to handle a 2 page scenario, where the range
821 * straddles across 2 virtual pages and hence need for loop
822 */
823 while (tot_sz > 0) {
824 unsigned int off, sz;
825 unsigned long phy, pfn;
826
827 off = kstart % PAGE_SIZE;
828 pfn = vmalloc_to_pfn((void *)kstart);
829 phy = (pfn << PAGE_SHIFT) + off;
830 sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
831 __sync_icache_dcache(phy, kstart, sz);
832 kstart += sz;
833 tot_sz -= sz;
834 }
835}
836EXPORT_SYMBOL(flush_icache_range);
837
838/*
839 * General purpose helper to make I and D cache lines consistent.
840 * @paddr is phy addr of region
841 * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
842 * However in one instance, when called by kprobe (for a breakpt in
843 * builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
844 * use a paddr to index the cache (despite VIPT). This is fine since a
845 * builtin kernel page will not have any virtual mappings.
846 * kprobe on loadable module will be kernel vaddr.
847 */
848void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
849{
850 __dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
851 __ic_line_inv_vaddr(paddr, vaddr, len);
852}
853
854/* wrapper to compile time eliminate alignment checks in flush loop */
855void __inv_icache_pages(phys_addr_t paddr, unsigned long vaddr, unsigned nr)
856{
857 __ic_line_inv_vaddr(paddr, vaddr, nr * PAGE_SIZE);
858}
859
860/*
861 * wrapper to clearout kernel or userspace mappings of a page
862 * For kernel mappings @vaddr == @paddr
863 */
864void __flush_dcache_pages(phys_addr_t paddr, unsigned long vaddr, unsigned nr)
865{
866 __dc_line_op(paddr, vaddr & PAGE_MASK, nr * PAGE_SIZE, OP_FLUSH_N_INV);
867}
868
869noinline void flush_cache_all(void)
870{
871 unsigned long flags;
872
873 local_irq_save(flags);
874
875 __ic_entire_inv();
876 __dc_entire_op(OP_FLUSH_N_INV);
877
878 local_irq_restore(flags);
879
880}
881
882void copy_user_highpage(struct page *to, struct page *from,
883 unsigned long u_vaddr, struct vm_area_struct *vma)
884{
885 struct folio *src = page_folio(from);
886 struct folio *dst = page_folio(to);
887 void *kfrom = kmap_atomic(from);
888 void *kto = kmap_atomic(to);
889
890 copy_page(kto, kfrom);
891
892 clear_bit(PG_dc_clean, &dst->flags);
893 clear_bit(PG_dc_clean, &src->flags);
894
895 kunmap_atomic(kto);
896 kunmap_atomic(kfrom);
897}
898
899void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
900{
901 struct folio *folio = page_folio(page);
902 clear_page(to);
903 clear_bit(PG_dc_clean, &folio->flags);
904}
905EXPORT_SYMBOL(clear_user_page);
906
907/**********************************************************************
908 * Explicit Cache flush request from user space via syscall
909 * Needed for JITs which generate code on the fly
910 */
911SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
912{
913 /* TBD: optimize this */
914 flush_cache_all();
915 return 0;
916}
917
918/*
919 * IO-Coherency (IOC) setup rules:
920 *
921 * 1. Needs to be at system level, so only once by Master core
922 * Non-Masters need not be accessing caches at that time
923 * - They are either HALT_ON_RESET and kick started much later or
924 * - if run on reset, need to ensure that arc_platform_smp_wait_to_boot()
925 * doesn't perturb caches or coherency unit
926 *
927 * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC,
928 * otherwise any straggler data might behave strangely post IOC enabling
929 *
930 * 3. All Caches need to be disabled when setting up IOC to elide any in-flight
931 * Coherency transactions
932 */
933static noinline void __init arc_ioc_setup(void)
934{
935 unsigned int ioc_base, mem_sz;
936
937 /*
938 * If IOC was already enabled (due to bootloader) it technically needs to
939 * be reconfigured with aperture base,size corresponding to Linux memory map
940 * which will certainly be different than uboot's. But disabling and
941 * reenabling IOC when DMA might be potentially active is tricky business.
942 * To avoid random memory issues later, just panic here and ask user to
943 * upgrade bootloader to one which doesn't enable IOC
944 */
945 if (read_aux_reg(ARC_REG_IO_COH_ENABLE) & ARC_IO_COH_ENABLE_BIT)
946 panic("IOC already enabled, please upgrade bootloader!\n");
947
948 if (!ioc_enable)
949 return;
950
951 /* Flush + invalidate + disable L1 dcache */
952 __dc_disable();
953
954 /* Flush + invalidate SLC */
955 if (read_aux_reg(ARC_REG_SLC_BCR))
956 slc_entire_op(OP_FLUSH_N_INV);
957
958 /*
959 * currently IOC Aperture covers entire DDR
960 * TBD: fix for PGU + 1GB of low mem
961 * TBD: fix for PAE
962 */
963 mem_sz = arc_get_mem_sz();
964
965 if (!is_power_of_2(mem_sz) || mem_sz < 4096)
966 panic("IOC Aperture size must be power of 2 larger than 4KB");
967
968 /*
969 * IOC Aperture size decoded as 2 ^ (SIZE + 2) KB,
970 * so setting 0x11 implies 512MB, 0x12 implies 1GB...
971 */
972 write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, order_base_2(mem_sz >> 10) - 2);
973
974 /* for now assume kernel base is start of IOC aperture */
975 ioc_base = CONFIG_LINUX_RAM_BASE;
976
977 if (ioc_base % mem_sz != 0)
978 panic("IOC Aperture start must be aligned to the size of the aperture");
979
980 write_aux_reg(ARC_REG_IO_COH_AP0_BASE, ioc_base >> 12);
981 write_aux_reg(ARC_REG_IO_COH_PARTIAL, ARC_IO_COH_PARTIAL_BIT);
982 write_aux_reg(ARC_REG_IO_COH_ENABLE, ARC_IO_COH_ENABLE_BIT);
983
984 /* Re-enable L1 dcache */
985 __dc_enable();
986}
987
988/*
989 * Cache related boot time checks/setups only needed on master CPU:
990 * - Geometry checks (kernel build and hardware agree: e.g. L1_CACHE_BYTES)
991 * Assume SMP only, so all cores will have same cache config. A check on
992 * one core suffices for all
993 * - IOC setup / dma callbacks only need to be done once
994 */
995static noinline void __init arc_cache_init_master(void)
996{
997 if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
998 struct cpuinfo_arc_cache *ic = &ic_info;
999
1000 if (!ic->line_len)
1001 panic("cache support enabled but non-existent cache\n");
1002
1003 if (ic->line_len != L1_CACHE_BYTES)
1004 panic("ICache line [%d] != kernel Config [%d]",
1005 ic->line_len, L1_CACHE_BYTES);
1006
1007 /*
1008 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
1009 * pair to provide vaddr/paddr respectively, just as in MMU v3
1010 */
1011 if (is_isa_arcv2() && ic->colors > 1)
1012 _cache_line_loop_ic_fn = __cache_line_loop_v3;
1013 else
1014 _cache_line_loop_ic_fn = __cache_line_loop;
1015 }
1016
1017 if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
1018 struct cpuinfo_arc_cache *dc = &dc_info;
1019
1020 if (!dc->line_len)
1021 panic("cache support enabled but non-existent cache\n");
1022
1023 if (dc->line_len != L1_CACHE_BYTES)
1024 panic("DCache line [%d] != kernel Config [%d]",
1025 dc->line_len, L1_CACHE_BYTES);
1026
1027 /* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
1028 if (is_isa_arcompact() && dc->colors > 1) {
1029 panic("Aliasing VIPT cache not supported\n");
1030 }
1031 }
1032
1033 /*
1034 * Check that SMP_CACHE_BYTES (and hence ARCH_DMA_MINALIGN) is larger
1035 * or equal to any cache line length.
1036 */
1037 BUILD_BUG_ON_MSG(L1_CACHE_BYTES > SMP_CACHE_BYTES,
1038 "SMP_CACHE_BYTES must be >= any cache line length");
1039 if (is_isa_arcv2() && (l2_line_sz > SMP_CACHE_BYTES))
1040 panic("L2 Cache line [%d] > kernel Config [%d]\n",
1041 l2_line_sz, SMP_CACHE_BYTES);
1042
1043 /* Note that SLC disable not formally supported till HS 3.0 */
1044 if (is_isa_arcv2() && l2_line_sz && !slc_enable)
1045 arc_slc_disable();
1046
1047 if (is_isa_arcv2() && ioc_exists)
1048 arc_ioc_setup();
1049
1050 if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1051 __dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1052 __dma_cache_inv = __dma_cache_inv_slc;
1053 __dma_cache_wback = __dma_cache_wback_slc;
1054 } else {
1055 __dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1056 __dma_cache_inv = __dma_cache_inv_l1;
1057 __dma_cache_wback = __dma_cache_wback_l1;
1058 }
1059 /*
1060 * In case of IOC (say IOC+SLC case), pointers above could still be set
1061 * but end up not being relevant as the first function in chain is not
1062 * called at all for devices using coherent DMA.
1063 * arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
1064 */
1065}
1066
1067void __ref arc_cache_init(void)
1068{
1069 unsigned int __maybe_unused cpu = smp_processor_id();
1070
1071 if (!cpu)
1072 arc_cache_init_master();
1073
1074 /*
1075 * In PAE regime, TLB and cache maintenance ops take wider addresses
1076 * And even if PAE is not enabled in kernel, the upper 32-bits still need
1077 * to be zeroed to keep the ops sane.
1078 * As an optimization for more common !PAE enabled case, zero them out
1079 * once at init, rather than checking/setting to 0 for every runtime op
1080 */
1081 if (is_isa_arcv2() && pae40_exist_but_not_enab()) {
1082
1083 if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE))
1084 write_aux_reg(ARC_REG_IC_PTAG_HI, 0);
1085
1086 if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE))
1087 write_aux_reg(ARC_REG_DC_PTAG_HI, 0);
1088
1089 if (l2_line_sz) {
1090 write_aux_reg(ARC_REG_SLC_RGN_END1, 0);
1091 write_aux_reg(ARC_REG_SLC_RGN_START1, 0);
1092 }
1093 }
1094}