2 * Copyright (C) 2013-2014 Synopsys, Inc. All rights reserved.
4 * SPDX-License-Identifier: GPL-2.0+
9 #include <linux/compiler.h>
10 #include <linux/kernel.h>
11 #include <linux/log2.h>
12 #include <asm/arcregs.h>
13 #include <asm/cache.h>
17 * Data cache (L1 D$ or SL$) entire invalidate operation or data cache disable
18 * operation may result in unexpected behavior and data loss even if we flush
19 * data cache right before invalidation. That may happens if we store any context
20 * on stack (like we store BLINK register on stack before function call).
21 * BLINK register is the register where return address is automatically saved
22 * when we do function call with instructions like 'bl'.
24 * There is the real example:
25 * We may hang in the next code as we store any BLINK register on stack in
26 * invalidate_dcache_all() function.
28 * void flush_dcache_all() {
29 * __dc_entire_op(OP_FLUSH);
33 * void invalidate_dcache_all() {
34 * __dc_entire_op(OP_INV);
40 * invalidate_dcache_all();
43 * Now let's see what really happens during that code execution:
46 * |->> call flush_dcache_all
47 * [return address is saved to BLINK register]
48 * [push BLINK] (save to stack) ![point 1]
49 * |->> call __dc_entire_op(OP_FLUSH)
50 * [return address is saved to BLINK register]
52 * return [jump to BLINK]
54 * [other flush_dcache_all code]
55 * [pop BLINK] (get from stack)
56 * return [jump to BLINK]
58 * |->> call invalidate_dcache_all
59 * [return address is saved to BLINK register]
60 * [push BLINK] (save to stack) ![point 2]
61 * |->> call __dc_entire_op(OP_FLUSH)
62 * [return address is saved to BLINK register]
63 * [invalidate L1 D$] ![point 3]
65 * // We lose return address from invalidate_dcache_all function:
66 * // we save it to stack and invalidate L1 D$ after that!
67 * return [jump to BLINK]
69 * [other invalidate_dcache_all code]
70 * [pop BLINK] (get from stack)
71 * // we don't have this data in L1 dcache as we invalidated it in [point 3]
72 * // so we get it from next memory level (for example DDR memory)
73 * // but in the memory we have value which we save in [point 1], which
74 * // is return address from flush_dcache_all function (instead of
75 * // address from current invalidate_dcache_all function which we
76 * // saved in [point 2] !)
77 * return [jump to BLINK]
79 * // As BLINK points to invalidate_dcache_all, we call it again and
82 * Fortunately we may fix that by using flush & invalidation of D$ with a single
83 * one instruction (instead of flush and invalidation instructions pair) and
84 * enabling force function inline with '__attribute__((always_inline))' gcc
85 * attribute to avoid any function call (and BLINK store) between cache flush
89 /* Bit values in IC_CTRL */
90 #define IC_CTRL_CACHE_DISABLE BIT(0)
92 /* Bit values in DC_CTRL */
93 #define DC_CTRL_CACHE_DISABLE BIT(0)
94 #define DC_CTRL_INV_MODE_FLUSH BIT(6)
95 #define DC_CTRL_FLUSH_STATUS BIT(8)
96 #define CACHE_VER_NUM_MASK 0xF
99 #define OP_FLUSH BIT(1)
100 #define OP_FLUSH_N_INV (OP_FLUSH | OP_INV)
102 /* Bit val in SLC_CONTROL */
103 #define SLC_CTRL_DIS 0x001
104 #define SLC_CTRL_IM 0x040
105 #define SLC_CTRL_BUSY 0x100
106 #define SLC_CTRL_RGN_OP_INV 0x200
109 * By default that variable will fall into .bss section.
110 * But .bss section is not relocated and so it will be initilized before
111 * relocation but will be used after being zeroed.
113 int l1_line_sz __section(".data");
114 bool dcache_exists __section(".data") = false;
115 bool icache_exists __section(".data") = false;
117 #define CACHE_LINE_MASK (~(l1_line_sz - 1))
119 #ifdef CONFIG_ISA_ARCV2
120 int slc_line_sz __section(".data");
121 bool slc_exists __section(".data") = false;
122 bool ioc_exists __section(".data") = false;
123 bool pae_exists __section(".data") = false;
125 /* To force enable IOC set ioc_enable to 'true' */
126 bool ioc_enable __section(".data") = false;
128 void read_decode_mmu_bcr(void)
130 /* TODO: should we compare mmu version from BCR and from CONFIG? */
131 #if (CONFIG_ARC_MMU_VER >= 4)
134 tmp = read_aux_reg(ARC_AUX_MMU_BCR);
137 #ifdef CONFIG_CPU_BIG_ENDIAN
138 unsigned int ver:8, sasid:1, sz1:4, sz0:4, res:2, pae:1,
139 n_ways:2, n_entry:2, n_super:2, u_itlb:3, u_dtlb:3;
141 /* DTLB ITLB JES JE JA */
142 unsigned int u_dtlb:3, u_itlb:3, n_super:2, n_entry:2, n_ways:2,
143 pae:1, res:2, sz0:4, sz1:4, sasid:1, ver:8;
144 #endif /* CONFIG_CPU_BIG_ENDIAN */
147 mmu4 = (struct bcr_mmu_4 *)&tmp;
149 pae_exists = !!mmu4->pae;
150 #endif /* (CONFIG_ARC_MMU_VER >= 4) */
153 static void __slc_entire_op(const int op)
157 ctrl = read_aux_reg(ARC_AUX_SLC_CTRL);
159 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
160 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
164 write_aux_reg(ARC_AUX_SLC_CTRL, ctrl);
166 if (op & OP_INV) /* Inv or flush-n-inv use same cmd reg */
167 write_aux_reg(ARC_AUX_SLC_INVALIDATE, 0x1);
169 write_aux_reg(ARC_AUX_SLC_FLUSH, 0x1);
171 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
172 read_aux_reg(ARC_AUX_SLC_CTRL);
174 /* Important to wait for flush to complete */
175 while (read_aux_reg(ARC_AUX_SLC_CTRL) & SLC_CTRL_BUSY);
178 static void slc_upper_region_init(void)
181 * ARC_AUX_SLC_RGN_END1 and ARC_AUX_SLC_RGN_START1 are always == 0
182 * as we don't use PAE40.
184 write_aux_reg(ARC_AUX_SLC_RGN_END1, 0);
185 write_aux_reg(ARC_AUX_SLC_RGN_START1, 0);
188 static void __slc_rgn_op(unsigned long paddr, unsigned long sz, const int op)
194 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
195 * - b'000 (default) is Flush,
196 * - b'001 is Invalidate if CTRL.IM == 0
197 * - b'001 is Flush-n-Invalidate if CTRL.IM == 1
199 ctrl = read_aux_reg(ARC_AUX_SLC_CTRL);
201 /* Don't rely on default value of IM bit */
202 if (!(op & OP_FLUSH)) /* i.e. OP_INV */
203 ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */
208 ctrl |= SLC_CTRL_RGN_OP_INV; /* Inv or flush-n-inv */
210 ctrl &= ~SLC_CTRL_RGN_OP_INV;
212 write_aux_reg(ARC_AUX_SLC_CTRL, ctrl);
215 * Lower bits are ignored, no need to clip
216 * END needs to be setup before START (latter triggers the operation)
217 * END can't be same as START, so add (l2_line_sz - 1) to sz
219 end = paddr + sz + slc_line_sz - 1;
222 * Upper addresses (ARC_AUX_SLC_RGN_END1 and ARC_AUX_SLC_RGN_START1)
223 * are always == 0 as we don't use PAE40, so we only setup lower ones
224 * (ARC_AUX_SLC_RGN_END and ARC_AUX_SLC_RGN_START)
226 write_aux_reg(ARC_AUX_SLC_RGN_END, end);
227 write_aux_reg(ARC_AUX_SLC_RGN_START, paddr);
229 /* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
230 read_aux_reg(ARC_AUX_SLC_CTRL);
232 while (read_aux_reg(ARC_AUX_SLC_CTRL) & SLC_CTRL_BUSY);
235 static void arc_ioc_setup(void)
237 /* IOC Aperture start is equal to DDR start */
238 unsigned int ap_base = CONFIG_SYS_SDRAM_BASE;
239 /* IOC Aperture size is equal to DDR size */
240 long ap_size = CONFIG_SYS_SDRAM_SIZE;
242 flush_n_invalidate_dcache_all();
244 if (!is_power_of_2(ap_size) || ap_size < 4096)
245 panic("IOC Aperture size must be power of 2 and bigger 4Kib");
248 * IOC Aperture size decoded as 2 ^ (SIZE + 2) KB,
249 * so setting 0x11 implies 512M, 0x12 implies 1G...
251 write_aux_reg(ARC_AUX_IO_COH_AP0_SIZE,
252 order_base_2(ap_size / 1024) - 2);
254 /* IOC Aperture start must be aligned to the size of the aperture */
255 if (ap_base % ap_size != 0)
256 panic("IOC Aperture start must be aligned to the size of the aperture");
258 write_aux_reg(ARC_AUX_IO_COH_AP0_BASE, ap_base >> 12);
259 write_aux_reg(ARC_AUX_IO_COH_PARTIAL, 1);
260 write_aux_reg(ARC_AUX_IO_COH_ENABLE, 1);
262 #endif /* CONFIG_ISA_ARCV2 */
264 #ifdef CONFIG_ISA_ARCV2
265 static void read_decode_cache_bcr_arcv2(void)
269 #ifdef CONFIG_CPU_BIG_ENDIAN
270 unsigned int pad:24, way:2, lsz:2, sz:4;
272 unsigned int sz:4, lsz:2, way:2, pad:24;
280 #ifdef CONFIG_CPU_BIG_ENDIAN
281 unsigned int pad:24, ver:8;
283 unsigned int ver:8, pad:24;
289 sbcr.word = read_aux_reg(ARC_BCR_SLC);
290 if (sbcr.fields.ver) {
291 slc_cfg.word = read_aux_reg(ARC_AUX_SLC_CONFIG);
293 slc_line_sz = (slc_cfg.fields.lsz == 0) ? 128 : 64;
297 struct bcr_clust_cfg {
298 #ifdef CONFIG_CPU_BIG_ENDIAN
299 unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
301 unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
307 cbcr.word = read_aux_reg(ARC_BCR_CLUSTER);
308 if (cbcr.fields.c && ioc_enable)
313 void read_decode_cache_bcr(void)
315 int dc_line_sz = 0, ic_line_sz = 0;
319 #ifdef CONFIG_CPU_BIG_ENDIAN
320 unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
322 unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
328 ibcr.word = read_aux_reg(ARC_BCR_IC_BUILD);
329 if (ibcr.fields.ver) {
330 icache_exists = true;
331 l1_line_sz = ic_line_sz = 8 << ibcr.fields.line_len;
333 panic("Instruction exists but line length is 0\n");
336 dbcr.word = read_aux_reg(ARC_BCR_DC_BUILD);
337 if (dbcr.fields.ver) {
338 dcache_exists = true;
339 l1_line_sz = dc_line_sz = 16 << dbcr.fields.line_len;
341 panic("Data cache exists but line length is 0\n");
344 if (ic_line_sz && dc_line_sz && (ic_line_sz != dc_line_sz))
345 panic("Instruction and data cache line lengths differ\n");
348 void cache_init(void)
350 read_decode_cache_bcr();
352 #ifdef CONFIG_ISA_ARCV2
353 read_decode_cache_bcr_arcv2();
358 read_decode_mmu_bcr();
361 * ARC_AUX_SLC_RGN_START1 and ARC_AUX_SLC_RGN_END1 register exist
362 * only if PAE exists in current HW. So we had to check pae_exist
365 if (slc_exists && pae_exists)
366 slc_upper_region_init();
367 #endif /* CONFIG_ISA_ARCV2 */
370 int icache_status(void)
375 if (read_aux_reg(ARC_AUX_IC_CTRL) & IC_CTRL_CACHE_DISABLE)
381 void icache_enable(void)
384 write_aux_reg(ARC_AUX_IC_CTRL, read_aux_reg(ARC_AUX_IC_CTRL) &
385 ~IC_CTRL_CACHE_DISABLE);
388 void icache_disable(void)
391 write_aux_reg(ARC_AUX_IC_CTRL, read_aux_reg(ARC_AUX_IC_CTRL) |
392 IC_CTRL_CACHE_DISABLE);
395 /* IC supports only invalidation */
396 static inline void __ic_entire_invalidate(void)
398 if (!icache_status())
401 /* Any write to IC_IVIC register triggers invalidation of entire I$ */
402 write_aux_reg(ARC_AUX_IC_IVIC, 1);
404 * As per ARC HS databook (see chapter 5.3.3.2)
405 * it is required to add 3 NOPs after each write to IC_IVIC.
410 read_aux_reg(ARC_AUX_IC_CTRL); /* blocks */
413 void invalidate_icache_all(void)
415 __ic_entire_invalidate();
417 #ifdef CONFIG_ISA_ARCV2
419 __slc_entire_op(OP_INV);
423 int dcache_status(void)
428 if (read_aux_reg(ARC_AUX_DC_CTRL) & DC_CTRL_CACHE_DISABLE)
434 void dcache_enable(void)
439 write_aux_reg(ARC_AUX_DC_CTRL, read_aux_reg(ARC_AUX_DC_CTRL) &
440 ~(DC_CTRL_INV_MODE_FLUSH | DC_CTRL_CACHE_DISABLE));
443 void dcache_disable(void)
448 write_aux_reg(ARC_AUX_DC_CTRL, read_aux_reg(ARC_AUX_DC_CTRL) |
449 DC_CTRL_CACHE_DISABLE);
452 #ifndef CONFIG_SYS_DCACHE_OFF
453 /* Common Helper for Line Operations on D-cache */
454 static inline void __dcache_line_loop(unsigned long paddr, unsigned long sz,
457 unsigned int aux_cmd;
460 /* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
461 aux_cmd = cacheop & OP_INV ? ARC_AUX_DC_IVDL : ARC_AUX_DC_FLDL;
463 sz += paddr & ~CACHE_LINE_MASK;
464 paddr &= CACHE_LINE_MASK;
466 num_lines = DIV_ROUND_UP(sz, l1_line_sz);
468 while (num_lines-- > 0) {
469 #if (CONFIG_ARC_MMU_VER == 3)
470 write_aux_reg(ARC_AUX_DC_PTAG, paddr);
472 write_aux_reg(aux_cmd, paddr);
477 static void __before_dc_op(const int op)
481 ctrl = read_aux_reg(ARC_AUX_DC_CTRL);
483 /* IM bit implies flush-n-inv, instead of vanilla inv */
485 ctrl &= ~DC_CTRL_INV_MODE_FLUSH;
487 ctrl |= DC_CTRL_INV_MODE_FLUSH;
489 write_aux_reg(ARC_AUX_DC_CTRL, ctrl);
492 static void __after_dc_op(const int op)
494 if (op & OP_FLUSH) /* flush / flush-n-inv both wait */
495 while (read_aux_reg(ARC_AUX_DC_CTRL) & DC_CTRL_FLUSH_STATUS);
498 static inline void __dc_entire_op(const int cacheop)
502 __before_dc_op(cacheop);
504 if (cacheop & OP_INV) /* Inv or flush-n-inv use same cmd reg */
505 aux = ARC_AUX_DC_IVDC;
507 aux = ARC_AUX_DC_FLSH;
509 write_aux_reg(aux, 0x1);
511 __after_dc_op(cacheop);
514 static inline void __dc_line_op(unsigned long paddr, unsigned long sz,
517 __before_dc_op(cacheop);
518 __dcache_line_loop(paddr, sz, cacheop);
519 __after_dc_op(cacheop);
522 #define __dc_entire_op(cacheop)
523 #define __dc_line_op(paddr, sz, cacheop)
524 #endif /* !CONFIG_SYS_DCACHE_OFF */
526 void invalidate_dcache_range(unsigned long start, unsigned long end)
531 #ifdef CONFIG_ISA_ARCV2
534 __dc_line_op(start, end - start, OP_INV);
536 #ifdef CONFIG_ISA_ARCV2
537 if (slc_exists && !ioc_exists)
538 __slc_rgn_op(start, end - start, OP_INV);
542 void flush_dcache_range(unsigned long start, unsigned long end)
547 #ifdef CONFIG_ISA_ARCV2
550 __dc_line_op(start, end - start, OP_FLUSH);
552 #ifdef CONFIG_ISA_ARCV2
553 if (slc_exists && !ioc_exists)
554 __slc_rgn_op(start, end - start, OP_FLUSH);
558 void flush_cache(unsigned long start, unsigned long size)
560 flush_dcache_range(start, start + size);
564 * As invalidate_dcache_all() is not used in generic U-Boot code and as we
565 * don't need it in arch/arc code alone (invalidate without flush) we implement
566 * flush_n_invalidate_dcache_all (flush and invalidate in 1 operation) because
567 * it's much safer. See [ NOTE 1 ] for more details.
569 void flush_n_invalidate_dcache_all(void)
571 __dc_entire_op(OP_FLUSH_N_INV);
573 #ifdef CONFIG_ISA_ARCV2
575 __slc_entire_op(OP_FLUSH_N_INV);
579 void flush_dcache_all(void)
581 __dc_entire_op(OP_FLUSH);
583 #ifdef CONFIG_ISA_ARCV2
585 __slc_entire_op(OP_FLUSH);