Commit | Line | Data |
---|---|---|
95d6976d VG |
1 | /* |
2 | * ARC700 VIPT Cache Management | |
3 | * | |
4 | * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com) | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 as | |
8 | * published by the Free Software Foundation. | |
9 | * | |
10 | * vineetg: May 2011: for Non-aliasing VIPT D-cache following can be NOPs | |
11 | * -flush_cache_dup_mm (fork) | |
12 | * -likewise for flush_cache_mm (exit/execve) | |
13 | * -likewise for flush_cache_range,flush_cache_page (munmap, exit, COW-break) | |
14 | * | |
15 | * vineetg: Apr 2011 | |
16 | * -Now that MMU can support larger pg sz (16K), the determiniation of | |
17 | * aliasing shd not be based on assumption of 8k pg | |
18 | * | |
19 | * vineetg: Mar 2011 | |
20 | * -optimised version of flush_icache_range( ) for making I/D coherent | |
21 | * when vaddr is available (agnostic of num of aliases) | |
22 | * | |
23 | * vineetg: Mar 2011 | |
24 | * -Added documentation about I-cache aliasing on ARC700 and the way it | |
25 | * was handled up until MMU V2. | |
26 | * -Spotted a three year old bug when killing the 4 aliases, which needs | |
27 | * bottom 2 bits, so we need to do paddr | {0x00, 0x01, 0x02, 0x03} | |
28 | * instead of paddr | {0x00, 0x01, 0x10, 0x11} | |
29 | * (Rajesh you owe me one now) | |
30 | * | |
31 | * vineetg: Dec 2010 | |
32 | * -Off-by-one error when computing num_of_lines to flush | |
33 | * This broke signal handling with bionic which uses synthetic sigret stub | |
34 | * | |
35 | * vineetg: Mar 2010 | |
36 | * -GCC can't generate ZOL for core cache flush loops. | |
37 | * Conv them into iterations based as opposed to while (start < end) types | |
38 | * | |
39 | * Vineetg: July 2009 | |
40 | * -In I-cache flush routine we used to chk for aliasing for every line INV. | |
41 | * Instead now we setup routines per cache geometry and invoke them | |
42 | * via function pointers. | |
43 | * | |
44 | * Vineetg: Jan 2009 | |
45 | * -Cache Line flush routines used to flush an extra line beyond end addr | |
46 | * because check was while (end >= start) instead of (end > start) | |
47 | * =Some call sites had to work around by doing -1, -4 etc to end param | |
48 | * =Some callers didnt care. This was spec bad in case of INV routines | |
49 | * which would discard valid data (cause of the horrible ext2 bug | |
50 | * in ARC IDE driver) | |
51 | * | |
52 | * vineetg: June 11th 2008: Fixed flush_icache_range( ) | |
53 | * -Since ARC700 caches are not coherent (I$ doesnt snoop D$) both need | |
54 | * to be flushed, which it was not doing. | |
55 | * -load_module( ) passes vmalloc addr (Kernel Virtual Addr) to the API, | |
56 | * however ARC cache maintenance OPs require PHY addr. Thus need to do | |
57 | * vmalloc_to_phy. | |
58 | * -Also added optimisation there, that for range > PAGE SIZE we flush the | |
59 | * entire cache in one shot rather than line by line. For e.g. a module | |
60 | * with Code sz 600k, old code flushed 600k worth of cache (line-by-line), | |
61 | * while cache is only 16 or 32k. | |
62 | */ | |
63 | ||
64 | #include <linux/module.h> | |
65 | #include <linux/mm.h> | |
66 | #include <linux/sched.h> | |
67 | #include <linux/cache.h> | |
68 | #include <linux/mmu_context.h> | |
69 | #include <linux/syscalls.h> | |
70 | #include <linux/uaccess.h> | |
71 | #include <asm/cacheflush.h> | |
72 | #include <asm/cachectl.h> | |
73 | #include <asm/setup.h> | |
74 | ||
75 | ||
76 | #ifdef CONFIG_ARC_HAS_ICACHE | |
77 | static void __ic_line_inv_no_alias(unsigned long, int); | |
78 | static void __ic_line_inv_2_alias(unsigned long, int); | |
79 | static void __ic_line_inv_4_alias(unsigned long, int); | |
80 | ||
81 | /* Holds the ptr to flush routine, dependign on size due to aliasing issues */ | |
82 | static void (*___flush_icache_rtn) (unsigned long, int); | |
83 | #endif | |
84 | ||
85 | /* | |
86 | * Read the Cache Build Confuration Registers, Decode them and save into | |
87 | * the cpuinfo structure for later use. | |
88 | * No Validation done here, simply read/convert the BCRs | |
89 | */ | |
90 | void __init read_decode_cache_bcr(void) | |
91 | { | |
92 | struct bcr_cache ibcr, dbcr; | |
93 | struct cpuinfo_arc_cache *p_ic, *p_dc; | |
94 | unsigned int cpu = smp_processor_id(); | |
95 | ||
96 | p_ic = &cpuinfo_arc700[cpu].icache; | |
97 | READ_BCR(ARC_REG_IC_BCR, ibcr); | |
98 | ||
99 | if (ibcr.config == 0x3) | |
100 | p_ic->assoc = 2; | |
101 | p_ic->line_len = 8 << ibcr.line_len; | |
102 | p_ic->sz = 0x200 << ibcr.sz; | |
103 | p_ic->ver = ibcr.ver; | |
104 | ||
105 | p_dc = &cpuinfo_arc700[cpu].dcache; | |
106 | READ_BCR(ARC_REG_DC_BCR, dbcr); | |
107 | ||
108 | if (dbcr.config == 0x2) | |
109 | p_dc->assoc = 4; | |
110 | p_dc->line_len = 16 << dbcr.line_len; | |
111 | p_dc->sz = 0x200 << dbcr.sz; | |
112 | p_dc->ver = dbcr.ver; | |
113 | } | |
114 | ||
115 | /* | |
116 | * 1. Validate the Cache Geomtery (compile time config matches hardware) | |
117 | * 2. If I-cache suffers from aliasing, setup work arounds (difft flush rtn) | |
118 | * (aliasing D-cache configurations are not supported YET) | |
119 | * 3. Enable the Caches, setup default flush mode for D-Cache | |
120 | * 3. Calculate the SHMLBA used by user space | |
121 | */ | |
122 | void __init arc_cache_init(void) | |
123 | { | |
124 | unsigned int temp; | |
125 | #ifdef CONFIG_ARC_CACHE | |
126 | unsigned int cpu = smp_processor_id(); | |
127 | #endif | |
128 | #ifdef CONFIG_ARC_HAS_ICACHE | |
129 | struct cpuinfo_arc_cache *ic; | |
130 | #endif | |
131 | #ifdef CONFIG_ARC_HAS_DCACHE | |
132 | struct cpuinfo_arc_cache *dc; | |
133 | #endif | |
134 | int way_pg_ratio = way_pg_ratio; | |
135 | ||
136 | #ifdef CONFIG_ARC_HAS_ICACHE | |
137 | ic = &cpuinfo_arc700[cpu].icache; | |
138 | ||
139 | /* | |
140 | * if Cache way size is <= page size then no aliasing exhibited | |
141 | * otherwise ratio determines num of aliases. | |
142 | * e.g. 32K I$, 2 way set assoc, 8k pg size | |
143 | * way-sz = 32k/2 = 16k | |
144 | * way-pg-ratio = 16k/8k = 2, so 2 aliases possible | |
145 | * (meaning 1 line could be in 2 possible locations). | |
146 | */ | |
147 | way_pg_ratio = ic->sz / ARC_ICACHE_WAYS / PAGE_SIZE; | |
148 | switch (way_pg_ratio) { | |
149 | case 0: | |
150 | case 1: | |
151 | ___flush_icache_rtn = __ic_line_inv_no_alias; | |
152 | break; | |
153 | case 2: | |
154 | ___flush_icache_rtn = __ic_line_inv_2_alias; | |
155 | break; | |
156 | case 4: | |
157 | ___flush_icache_rtn = __ic_line_inv_4_alias; | |
158 | break; | |
159 | default: | |
160 | panic("Unsupported I-Cache Sz\n"); | |
161 | } | |
162 | #endif | |
163 | ||
164 | /* Enable/disable I-Cache */ | |
165 | temp = read_aux_reg(ARC_REG_IC_CTRL); | |
166 | ||
167 | #ifdef CONFIG_ARC_HAS_ICACHE | |
168 | temp &= ~IC_CTRL_CACHE_DISABLE; | |
169 | #else | |
170 | temp |= IC_CTRL_CACHE_DISABLE; | |
171 | #endif | |
172 | ||
173 | write_aux_reg(ARC_REG_IC_CTRL, temp); | |
174 | ||
175 | #ifdef CONFIG_ARC_HAS_DCACHE | |
176 | dc = &cpuinfo_arc700[cpu].dcache; | |
177 | ||
178 | /* check for D-Cache aliasing */ | |
179 | if ((dc->sz / ARC_DCACHE_WAYS) > PAGE_SIZE) | |
180 | panic("D$ aliasing not handled right now\n"); | |
181 | #endif | |
182 | ||
183 | /* Set the default Invalidate Mode to "simpy discard dirty lines" | |
184 | * as this is more frequent then flush before invalidate | |
185 | * Ofcourse we toggle this default behviour when desired | |
186 | */ | |
187 | temp = read_aux_reg(ARC_REG_DC_CTRL); | |
188 | temp &= ~DC_CTRL_INV_MODE_FLUSH; | |
189 | ||
190 | #ifdef CONFIG_ARC_HAS_DCACHE | |
191 | /* Enable D-Cache: Clear Bit 0 */ | |
192 | write_aux_reg(ARC_REG_DC_CTRL, temp & ~IC_CTRL_CACHE_DISABLE); | |
193 | #else | |
194 | /* Flush D cache */ | |
195 | write_aux_reg(ARC_REG_DC_FLSH, 0x1); | |
196 | /* Disable D cache */ | |
197 | write_aux_reg(ARC_REG_DC_CTRL, temp | IC_CTRL_CACHE_DISABLE); | |
198 | #endif | |
199 | ||
200 | return; | |
201 | } | |
202 | ||
203 | #define OP_INV 0x1 | |
204 | #define OP_FLUSH 0x2 | |
205 | #define OP_FLUSH_N_INV 0x3 | |
206 | ||
207 | #ifdef CONFIG_ARC_HAS_DCACHE | |
208 | ||
209 | /*************************************************************** | |
210 | * Machine specific helpers for Entire D-Cache or Per Line ops | |
211 | */ | |
212 | ||
213 | static inline void wait_for_flush(void) | |
214 | { | |
215 | while (read_aux_reg(ARC_REG_DC_CTRL) & DC_CTRL_FLUSH_STATUS) | |
216 | ; | |
217 | } | |
218 | ||
219 | /* | |
220 | * Operation on Entire D-Cache | |
221 | * @cacheop = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV} | |
222 | * Note that constant propagation ensures all the checks are gone | |
223 | * in generated code | |
224 | */ | |
225 | static inline void __dc_entire_op(const int cacheop) | |
226 | { | |
227 | unsigned long flags, tmp = tmp; | |
228 | int aux; | |
229 | ||
230 | local_irq_save(flags); | |
231 | ||
232 | if (cacheop == OP_FLUSH_N_INV) { | |
233 | /* Dcache provides 2 cmd: FLUSH or INV | |
234 | * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE | |
235 | * flush-n-inv is achieved by INV cmd but with IM=1 | |
236 | * Default INV sub-mode is DISCARD, which needs to be toggled | |
237 | */ | |
238 | tmp = read_aux_reg(ARC_REG_DC_CTRL); | |
239 | write_aux_reg(ARC_REG_DC_CTRL, tmp | DC_CTRL_INV_MODE_FLUSH); | |
240 | } | |
241 | ||
242 | if (cacheop & OP_INV) /* Inv or flush-n-inv use same cmd reg */ | |
243 | aux = ARC_REG_DC_IVDC; | |
244 | else | |
245 | aux = ARC_REG_DC_FLSH; | |
246 | ||
247 | write_aux_reg(aux, 0x1); | |
248 | ||
249 | if (cacheop & OP_FLUSH) /* flush / flush-n-inv both wait */ | |
250 | wait_for_flush(); | |
251 | ||
252 | /* Switch back the DISCARD ONLY Invalidate mode */ | |
253 | if (cacheop == OP_FLUSH_N_INV) | |
254 | write_aux_reg(ARC_REG_DC_CTRL, tmp & ~DC_CTRL_INV_MODE_FLUSH); | |
255 | ||
256 | local_irq_restore(flags); | |
257 | } | |
258 | ||
259 | /* | |
260 | * Per Line Operation on D-Cache | |
261 | * Doesn't deal with type-of-op/IRQ-disabling/waiting-for-flush-to-complete | |
262 | * It's sole purpose is to help gcc generate ZOL | |
263 | */ | |
264 | static inline void __dc_line_loop(unsigned long start, unsigned long sz, | |
265 | int aux_reg) | |
266 | { | |
267 | int num_lines, slack; | |
268 | ||
269 | /* Ensure we properly floor/ceil the non-line aligned/sized requests | |
270 | * and have @start - aligned to cache line and integral @num_lines. | |
271 | * This however can be avoided for page sized since: | |
272 | * -@start will be cache-line aligned already (being page aligned) | |
273 | * -@sz will be integral multiple of line size (being page sized). | |
274 | */ | |
275 | if (!(__builtin_constant_p(sz) && sz == PAGE_SIZE)) { | |
276 | slack = start & ~DCACHE_LINE_MASK; | |
277 | sz += slack; | |
278 | start -= slack; | |
279 | } | |
280 | ||
281 | num_lines = DIV_ROUND_UP(sz, ARC_DCACHE_LINE_LEN); | |
282 | ||
283 | while (num_lines-- > 0) { | |
284 | #if (CONFIG_ARC_MMU_VER > 2) | |
285 | /* | |
286 | * Just as for I$, in MMU v3, D$ ops also require | |
287 | * "tag" bits in DC_PTAG, "index" bits in FLDL,IVDL ops | |
288 | * But we pass phy addr for both. This works since Linux | |
289 | * doesn't support aliasing configs for D$, yet. | |
290 | * Thus paddr is enough to provide both tag and index. | |
291 | */ | |
292 | write_aux_reg(ARC_REG_DC_PTAG, start); | |
293 | #endif | |
294 | write_aux_reg(aux_reg, start); | |
295 | start += ARC_DCACHE_LINE_LEN; | |
296 | } | |
297 | } | |
298 | ||
299 | /* | |
300 | * D-Cache : Per Line INV (discard or wback+discard) or FLUSH (wback) | |
301 | */ | |
302 | static inline void __dc_line_op(unsigned long start, unsigned long sz, | |
303 | const int cacheop) | |
304 | { | |
305 | unsigned long flags, tmp = tmp; | |
306 | int aux; | |
307 | ||
308 | local_irq_save(flags); | |
309 | ||
310 | if (cacheop == OP_FLUSH_N_INV) { | |
311 | /* | |
312 | * Dcache provides 2 cmd: FLUSH or INV | |
313 | * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE | |
314 | * flush-n-inv is achieved by INV cmd but with IM=1 | |
315 | * Default INV sub-mode is DISCARD, which needs to be toggled | |
316 | */ | |
317 | tmp = read_aux_reg(ARC_REG_DC_CTRL); | |
318 | write_aux_reg(ARC_REG_DC_CTRL, tmp | DC_CTRL_INV_MODE_FLUSH); | |
319 | } | |
320 | ||
321 | if (cacheop & OP_INV) /* Inv / flush-n-inv use same cmd reg */ | |
322 | aux = ARC_REG_DC_IVDL; | |
323 | else | |
324 | aux = ARC_REG_DC_FLDL; | |
325 | ||
326 | __dc_line_loop(start, sz, aux); | |
327 | ||
328 | if (cacheop & OP_FLUSH) /* flush / flush-n-inv both wait */ | |
329 | wait_for_flush(); | |
330 | ||
331 | /* Switch back the DISCARD ONLY Invalidate mode */ | |
332 | if (cacheop == OP_FLUSH_N_INV) | |
333 | write_aux_reg(ARC_REG_DC_CTRL, tmp & ~DC_CTRL_INV_MODE_FLUSH); | |
334 | ||
335 | local_irq_restore(flags); | |
336 | } | |
337 | ||
338 | #else | |
339 | ||
340 | #define __dc_entire_op(cacheop) | |
341 | #define __dc_line_op(start, sz, cacheop) | |
342 | ||
343 | #endif /* CONFIG_ARC_HAS_DCACHE */ | |
344 | ||
345 | ||
346 | #ifdef CONFIG_ARC_HAS_ICACHE | |
347 | ||
348 | /* | |
349 | * I-Cache Aliasing in ARC700 VIPT caches | |
350 | * | |
351 | * For fetching code from I$, ARC700 uses vaddr (embedded in program code) | |
352 | * to "index" into SET of cache-line and paddr from MMU to match the TAG | |
353 | * in the WAYS of SET. | |
354 | * | |
355 | * However the CDU iterface (to flush/inv) lines from software, only takes | |
356 | * paddr (to have simpler hardware interface). For simpler cases, using paddr | |
357 | * alone suffices. | |
358 | * e.g. 2-way-set-assoc, 16K I$ (8k MMU pg sz, 32b cache line size): | |
359 | * way_sz = cache_sz / num_ways = 16k/2 = 8k | |
360 | * num_sets = way_sz / line_sz = 8k/32 = 256 => 8 bits | |
361 | * Ignoring the bottom 5 bits corresp to the off within a 32b cacheline, | |
362 | * bits req for calc set-index = bits 12:5 (0 based). Since this range fits | |
363 | * inside the bottom 13 bits of paddr, which are same for vaddr and paddr | |
364 | * (with 8k pg sz), paddr alone can be safely used by CDU to unambigously | |
365 | * locate a cache-line. | |
366 | * | |
367 | * However for a difft sized cache, say 32k I$, above math yields need | |
368 | * for 14 bits of vaddr to locate a cache line, which can't be provided by | |
369 | * paddr, since the bit 13 (0 based) might differ between the two. | |
370 | * | |
371 | * This lack of extra bits needed for correct line addressing, defines the | |
372 | * classical problem of Cache aliasing with VIPT architectures | |
373 | * num_aliases = 1 << extra_bits | |
374 | * e.g. 2-way-set-assoc, 32K I$ with 8k MMU pg sz => 2 aliases | |
375 | * 2-way-set-assoc, 64K I$ with 8k MMU pg sz => 4 aliases | |
376 | * 2-way-set-assoc, 16K I$ with 8k MMU pg sz => NO aliases | |
377 | * | |
378 | * ------------------ | |
379 | * MMU v1/v2 (Fixed Page Size 8k) | |
380 | * ------------------ | |
381 | * The solution was to provide CDU with these additonal vaddr bits. These | |
382 | * would be bits [x:13], x would depend on cache-geom. | |
383 | * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits | |
384 | * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the | |
385 | * orig 5 bits of paddr were anyways ignored by CDU line ops, as they | |
386 | * represent the offset within cache-line. The adv of using this "clumsy" | |
387 | * interface for additional info was no new reg was needed in CDU. | |
388 | * | |
389 | * 17:13 represented the max num of bits passable, actual bits needed were | |
390 | * fewer, based on the num-of-aliases possible. | |
391 | * -for 2 alias possibility, only bit 13 needed (32K cache) | |
392 | * -for 4 alias possibility, bits 14:13 needed (64K cache) | |
393 | * | |
394 | * Since vaddr was not available for all instances of I$ flush req by core | |
395 | * kernel, the only safe way (non-optimal though) was to kill all possible | |
396 | * lines which could represent an alias (even if they didnt represent one | |
397 | * in execution). | |
398 | * e.g. for 64K I$, 4 aliases possible, so we did | |
399 | * flush start | |
400 | * flush start | 0x01 | |
401 | * flush start | 0x2 | |
402 | * flush start | 0x3 | |
403 | * | |
404 | * The penalty was invoking the operation itself, since tag match is anyways | |
405 | * paddr based, a line which didn't represent an alias would not match the | |
406 | * paddr, hence wont be killed | |
407 | * | |
408 | * Note that aliasing concerns are independent of line-sz for a given cache | |
409 | * geometry (size + set_assoc) because the extra bits required by line-sz are | |
410 | * reduced from the set calc. | |
411 | * e.g. 2-way-set-assoc, 32K I$ with 8k MMU pg sz and using math above | |
412 | * 32b line-sz: 9 bits set-index-calc, 5 bits offset-in-line => 1 extra bit | |
413 | * 64b line-sz: 8 bits set-index-calc, 6 bits offset-in-line => 1 extra bit | |
414 | * | |
415 | * ------------------ | |
416 | * MMU v3 | |
417 | * ------------------ | |
418 | * This ver of MMU supports var page sizes (1k-16k) - Linux will support | |
419 | * 8k (default), 16k and 4k. | |
420 | * However from hardware perspective, smaller page sizes aggrevate aliasing | |
421 | * meaning more vaddr bits needed to disambiguate the cache-line-op ; | |
422 | * the existing scheme of piggybacking won't work for certain configurations. | |
423 | * Two new registers IC_PTAG and DC_PTAG inttoduced. | |
424 | * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs | |
425 | */ | |
426 | ||
427 | /*********************************************************** | |
428 | * Machine specific helpers for per line I-Cache invalidate. | |
429 | * 3 routines to accpunt for 1, 2, 4 aliases possible | |
430 | */ | |
431 | ||
432 | static void __ic_line_inv_no_alias(unsigned long start, int num_lines) | |
433 | { | |
434 | while (num_lines-- > 0) { | |
435 | #if (CONFIG_ARC_MMU_VER > 2) | |
436 | write_aux_reg(ARC_REG_IC_PTAG, start); | |
437 | #endif | |
438 | write_aux_reg(ARC_REG_IC_IVIL, start); | |
439 | start += ARC_ICACHE_LINE_LEN; | |
440 | } | |
441 | } | |
442 | ||
443 | static void __ic_line_inv_2_alias(unsigned long start, int num_lines) | |
444 | { | |
445 | while (num_lines-- > 0) { | |
446 | ||
447 | #if (CONFIG_ARC_MMU_VER > 2) | |
448 | /* | |
449 | * MMU v3, CDU prog model (for line ops) now uses a new IC_PTAG | |
450 | * reg to pass the "tag" bits and existing IVIL reg only looks | |
451 | * at bits relevant for "index" (details above) | |
452 | * Programming Notes: | |
453 | * -when writing tag to PTAG reg, bit chopping can be avoided, | |
454 | * CDU ignores non-tag bits. | |
455 | * -Ideally "index" must be computed from vaddr, but it is not | |
456 | * avail in these rtns. So to be safe, we kill the lines in all | |
457 | * possible indexes corresp to num of aliases possible for | |
458 | * given cache config. | |
459 | */ | |
460 | write_aux_reg(ARC_REG_IC_PTAG, start); | |
461 | write_aux_reg(ARC_REG_IC_IVIL, | |
462 | start & ~(0x1 << PAGE_SHIFT)); | |
463 | write_aux_reg(ARC_REG_IC_IVIL, start | (0x1 << PAGE_SHIFT)); | |
464 | #else | |
465 | write_aux_reg(ARC_REG_IC_IVIL, start); | |
466 | write_aux_reg(ARC_REG_IC_IVIL, start | 0x01); | |
467 | #endif | |
468 | start += ARC_ICACHE_LINE_LEN; | |
469 | } | |
470 | } | |
471 | ||
472 | static void __ic_line_inv_4_alias(unsigned long start, int num_lines) | |
473 | { | |
474 | while (num_lines-- > 0) { | |
475 | ||
476 | #if (CONFIG_ARC_MMU_VER > 2) | |
477 | write_aux_reg(ARC_REG_IC_PTAG, start); | |
478 | ||
479 | write_aux_reg(ARC_REG_IC_IVIL, | |
480 | start & ~(0x3 << PAGE_SHIFT)); | |
481 | write_aux_reg(ARC_REG_IC_IVIL, | |
482 | start & ~(0x2 << PAGE_SHIFT)); | |
483 | write_aux_reg(ARC_REG_IC_IVIL, | |
484 | start & ~(0x1 << PAGE_SHIFT)); | |
485 | write_aux_reg(ARC_REG_IC_IVIL, start | (0x3 << PAGE_SHIFT)); | |
486 | #else | |
487 | write_aux_reg(ARC_REG_IC_IVIL, start); | |
488 | write_aux_reg(ARC_REG_IC_IVIL, start | 0x01); | |
489 | write_aux_reg(ARC_REG_IC_IVIL, start | 0x02); | |
490 | write_aux_reg(ARC_REG_IC_IVIL, start | 0x03); | |
491 | #endif | |
492 | start += ARC_ICACHE_LINE_LEN; | |
493 | } | |
494 | } | |
495 | ||
496 | static void __ic_line_inv(unsigned long start, unsigned long sz) | |
497 | { | |
498 | unsigned long flags; | |
499 | int num_lines, slack; | |
500 | ||
501 | /* | |
502 | * Ensure we properly floor/ceil the non-line aligned/sized requests | |
503 | * and have @start - aligned to cache line, and integral @num_lines | |
504 | * However page sized flushes can be compile time optimised. | |
505 | * -@start will be cache-line aligned already (being page aligned) | |
506 | * -@sz will be integral multiple of line size (being page sized). | |
507 | */ | |
508 | if (!(__builtin_constant_p(sz) && sz == PAGE_SIZE)) { | |
509 | slack = start & ~ICACHE_LINE_MASK; | |
510 | sz += slack; | |
511 | start -= slack; | |
512 | } | |
513 | ||
514 | num_lines = DIV_ROUND_UP(sz, ARC_ICACHE_LINE_LEN); | |
515 | ||
516 | local_irq_save(flags); | |
517 | (*___flush_icache_rtn) (start, num_lines); | |
518 | local_irq_restore(flags); | |
519 | } | |
520 | ||
521 | /* Unlike routines above, having vaddr for flush op (along with paddr), | |
522 | * prevents the need to speculatively kill the lines in multiple sets | |
523 | * based on ratio of way_sz : pg_sz | |
524 | */ | |
525 | static void __ic_line_inv_vaddr(unsigned long phy_start, | |
526 | unsigned long vaddr, unsigned long sz) | |
527 | { | |
528 | unsigned long flags; | |
529 | int num_lines, slack; | |
530 | unsigned int addr; | |
531 | ||
532 | slack = phy_start & ~ICACHE_LINE_MASK; | |
533 | sz += slack; | |
534 | phy_start -= slack; | |
535 | num_lines = DIV_ROUND_UP(sz, ARC_ICACHE_LINE_LEN); | |
536 | ||
537 | #if (CONFIG_ARC_MMU_VER > 2) | |
538 | vaddr &= ~ICACHE_LINE_MASK; | |
539 | addr = phy_start; | |
540 | #else | |
541 | /* bits 17:13 of vaddr go as bits 4:0 of paddr */ | |
542 | addr = phy_start | ((vaddr >> 13) & 0x1F); | |
543 | #endif | |
544 | ||
545 | local_irq_save(flags); | |
546 | while (num_lines-- > 0) { | |
547 | #if (CONFIG_ARC_MMU_VER > 2) | |
548 | /* tag comes from phy addr */ | |
549 | write_aux_reg(ARC_REG_IC_PTAG, addr); | |
550 | ||
551 | /* index bits come from vaddr */ | |
552 | write_aux_reg(ARC_REG_IC_IVIL, vaddr); | |
553 | vaddr += ARC_ICACHE_LINE_LEN; | |
554 | #else | |
555 | /* this paddr contains vaddrs bits as needed */ | |
556 | write_aux_reg(ARC_REG_IC_IVIL, addr); | |
557 | #endif | |
558 | addr += ARC_ICACHE_LINE_LEN; | |
559 | } | |
560 | local_irq_restore(flags); | |
561 | } | |
562 | ||
563 | #else | |
564 | ||
565 | #define __ic_line_inv(start, sz) | |
566 | #define __ic_line_inv_vaddr(pstart, vstart, sz) | |
567 | ||
568 | #endif /* CONFIG_ARC_HAS_ICACHE */ | |
569 | ||
570 | ||
571 | /*********************************************************** | |
572 | * Exported APIs | |
573 | */ | |
574 | ||
575 | /* TBD: use pg_arch_1 to optimize this */ | |
576 | void flush_dcache_page(struct page *page) | |
577 | { | |
578 | __dc_line_op((unsigned long)page_address(page), PAGE_SIZE, OP_FLUSH); | |
579 | } | |
580 | EXPORT_SYMBOL(flush_dcache_page); | |
581 | ||
582 | ||
583 | void dma_cache_wback_inv(unsigned long start, unsigned long sz) | |
584 | { | |
585 | __dc_line_op(start, sz, OP_FLUSH_N_INV); | |
586 | } | |
587 | EXPORT_SYMBOL(dma_cache_wback_inv); | |
588 | ||
589 | void dma_cache_inv(unsigned long start, unsigned long sz) | |
590 | { | |
591 | __dc_line_op(start, sz, OP_INV); | |
592 | } | |
593 | EXPORT_SYMBOL(dma_cache_inv); | |
594 | ||
595 | void dma_cache_wback(unsigned long start, unsigned long sz) | |
596 | { | |
597 | __dc_line_op(start, sz, OP_FLUSH); | |
598 | } | |
599 | EXPORT_SYMBOL(dma_cache_wback); | |
600 | ||
601 | /* | |
602 | * This is API for making I/D Caches consistent when modifying code | |
603 | * (loadable modules, kprobes, etc) | |
604 | * This is called on insmod, with kernel virtual address for CODE of | |
605 | * the module. ARC cache maintenance ops require PHY address thus we | |
606 | * need to convert vmalloc addr to PHY addr | |
607 | */ | |
608 | void flush_icache_range(unsigned long kstart, unsigned long kend) | |
609 | { | |
610 | unsigned int tot_sz, off, sz; | |
611 | unsigned long phy, pfn; | |
612 | unsigned long flags; | |
613 | ||
614 | /* printk("Kernel Cache Cohenercy: %lx to %lx\n",kstart, kend); */ | |
615 | ||
616 | /* This is not the right API for user virtual address */ | |
617 | if (kstart < TASK_SIZE) { | |
618 | BUG_ON("Flush icache range for user virtual addr space"); | |
619 | return; | |
620 | } | |
621 | ||
622 | /* Shortcut for bigger flush ranges. | |
623 | * Here we don't care if this was kernel virtual or phy addr | |
624 | */ | |
625 | tot_sz = kend - kstart; | |
626 | if (tot_sz > PAGE_SIZE) { | |
627 | flush_cache_all(); | |
628 | return; | |
629 | } | |
630 | ||
631 | /* Case: Kernel Phy addr (0x8000_0000 onwards) */ | |
632 | if (likely(kstart > PAGE_OFFSET)) { | |
633 | __ic_line_inv(kstart, kend - kstart); | |
634 | __dc_line_op(kstart, kend - kstart, OP_FLUSH); | |
635 | return; | |
636 | } | |
637 | ||
638 | /* | |
639 | * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff) | |
640 | * (1) ARC Cache Maintenance ops only take Phy addr, hence special | |
641 | * handling of kernel vaddr. | |
642 | * | |
643 | * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already), | |
644 | * it still needs to handle a 2 page scenario, where the range | |
645 | * straddles across 2 virtual pages and hence need for loop | |
646 | */ | |
647 | while (tot_sz > 0) { | |
648 | off = kstart % PAGE_SIZE; | |
649 | pfn = vmalloc_to_pfn((void *)kstart); | |
650 | phy = (pfn << PAGE_SHIFT) + off; | |
651 | sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off); | |
652 | local_irq_save(flags); | |
653 | __dc_line_op(phy, sz, OP_FLUSH); | |
654 | __ic_line_inv(phy, sz); | |
655 | local_irq_restore(flags); | |
656 | kstart += sz; | |
657 | tot_sz -= sz; | |
658 | } | |
659 | } | |
660 | ||
661 | /* | |
662 | * Optimised ver of flush_icache_range() with spec callers: ptrace/signals | |
663 | * where vaddr is also available. This allows passing both vaddr and paddr | |
664 | * bits to CDU for cache flush, short-circuting the current pessimistic algo | |
665 | * which kills all possible aliases. | |
666 | * An added adv of knowing that vaddr is user-vaddr avoids various checks | |
667 | * and handling for k-vaddr, k-paddr as done in orig ver above | |
668 | */ | |
669 | void flush_icache_range_vaddr(unsigned long paddr, unsigned long u_vaddr, | |
670 | int len) | |
671 | { | |
672 | __ic_line_inv_vaddr(paddr, u_vaddr, len); | |
673 | __dc_line_op(paddr, len, OP_FLUSH); | |
674 | } | |
675 | ||
676 | /* | |
677 | * XXX: This also needs to be optim using pg_arch_1 | |
678 | * This is called when a page-cache page is about to be mapped into a | |
679 | * user process' address space. It offers an opportunity for a | |
680 | * port to ensure d-cache/i-cache coherency if necessary. | |
681 | */ | |
682 | void flush_icache_page(struct vm_area_struct *vma, struct page *page) | |
683 | { | |
684 | if (!(vma->vm_flags & VM_EXEC)) | |
685 | return; | |
686 | ||
687 | __ic_line_inv((unsigned long)page_address(page), PAGE_SIZE); | |
688 | } | |
689 | ||
690 | void flush_icache_all(void) | |
691 | { | |
692 | unsigned long flags; | |
693 | ||
694 | local_irq_save(flags); | |
695 | ||
696 | write_aux_reg(ARC_REG_IC_IVIC, 1); | |
697 | ||
698 | /* lr will not complete till the icache inv operation is not over */ | |
699 | read_aux_reg(ARC_REG_IC_CTRL); | |
700 | local_irq_restore(flags); | |
701 | } | |
702 | ||
703 | noinline void flush_cache_all(void) | |
704 | { | |
705 | unsigned long flags; | |
706 | ||
707 | local_irq_save(flags); | |
708 | ||
709 | flush_icache_all(); | |
710 | __dc_entire_op(OP_FLUSH_N_INV); | |
711 | ||
712 | local_irq_restore(flags); | |
713 | ||
714 | } | |
715 | ||
716 | /********************************************************************** | |
717 | * Explicit Cache flush request from user space via syscall | |
718 | * Needed for JITs which generate code on the fly | |
719 | */ | |
720 | SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags) | |
721 | { | |
722 | /* TBD: optimize this */ | |
723 | flush_cache_all(); | |
724 | return 0; | |
725 | } |