diff options
author | Thiemo Seufer <ths@networkno.de> | 2008-02-18 14:32:49 -0500 |
---|---|---|
committer | Ralf Baechle <ralf@linux-mips.org> | 2008-04-28 12:14:23 -0400 |
commit | fb2a27e743cd565c25cd896911e494482a8b7251 (patch) | |
tree | 791190ead6211e829443185fb1d4b4109fbd9a34 /arch/mips/mm/page.c | |
parent | 064922a805ec7aadfafdd27aa6b4908d737c3c1d (diff) |
[MIPS] Reimplement clear_page/copy_page
Fold the SB-1 specific implementation of clear_page/copy_page in the
generic version, and rewrite that one in tlbex style. The immediate
benefits:
- It converts the compile-time workaround for SB-1 pass 1 prefetches
to a more efficient run-time check.
- It allows adjustment of loop unfolling, which helps to reduce the
number of redundant cdex cache ops.
- It fixes some esoteric cornercases (the cache line length calculations
can go wrong, and support for 64k pages without prefetch instructions
will overflow the addiu immediate).
- Somewhat better guesses of "good" prefetch values.
Signed-off-by: Thiemo Seufer <ths@networkno.de>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Diffstat (limited to 'arch/mips/mm/page.c')
-rw-r--r-- | arch/mips/mm/page.c | 684 |
1 files changed, 684 insertions, 0 deletions
diff --git a/arch/mips/mm/page.c b/arch/mips/mm/page.c new file mode 100644 index 000000000000..d827d6144369 --- /dev/null +++ b/arch/mips/mm/page.c | |||
@@ -0,0 +1,684 @@ | |||
1 | /* | ||
2 | * This file is subject to the terms and conditions of the GNU General Public | ||
3 | * License. See the file "COPYING" in the main directory of this archive | ||
4 | * for more details. | ||
5 | * | ||
6 | * Copyright (C) 2003, 04, 05 Ralf Baechle (ralf@linux-mips.org) | ||
7 | * Copyright (C) 2007 Maciej W. Rozycki | ||
8 | * Copyright (C) 2008 Thiemo Seufer | ||
9 | */ | ||
10 | #include <linux/init.h> | ||
11 | #include <linux/kernel.h> | ||
12 | #include <linux/sched.h> | ||
13 | #include <linux/mm.h> | ||
14 | #include <linux/module.h> | ||
15 | #include <linux/proc_fs.h> | ||
16 | |||
17 | #include <asm/bugs.h> | ||
18 | #include <asm/cacheops.h> | ||
19 | #include <asm/inst.h> | ||
20 | #include <asm/io.h> | ||
21 | #include <asm/page.h> | ||
22 | #include <asm/pgtable.h> | ||
23 | #include <asm/prefetch.h> | ||
24 | #include <asm/system.h> | ||
25 | #include <asm/bootinfo.h> | ||
26 | #include <asm/mipsregs.h> | ||
27 | #include <asm/mmu_context.h> | ||
28 | #include <asm/cpu.h> | ||
29 | #include <asm/war.h> | ||
30 | |||
31 | #ifdef CONFIG_SIBYTE_DMA_PAGEOPS | ||
32 | #include <asm/sibyte/sb1250.h> | ||
33 | #include <asm/sibyte/sb1250_regs.h> | ||
34 | #include <asm/sibyte/sb1250_dma.h> | ||
35 | #endif | ||
36 | |||
37 | #include "uasm.h" | ||
38 | |||
39 | /* Registers used in the assembled routines. */ | ||
40 | #define ZERO 0 | ||
41 | #define AT 2 | ||
42 | #define A0 4 | ||
43 | #define A1 5 | ||
44 | #define A2 6 | ||
45 | #define T0 8 | ||
46 | #define T1 9 | ||
47 | #define T2 10 | ||
48 | #define T3 11 | ||
49 | #define T9 25 | ||
50 | #define RA 31 | ||
51 | |||
52 | /* Handle labels (which must be positive integers). */ | ||
53 | enum label_id { | ||
54 | label_clear_nopref = 1, | ||
55 | label_clear_pref, | ||
56 | label_copy_nopref, | ||
57 | label_copy_pref_both, | ||
58 | label_copy_pref_store, | ||
59 | }; | ||
60 | |||
61 | UASM_L_LA(_clear_nopref) | ||
62 | UASM_L_LA(_clear_pref) | ||
63 | UASM_L_LA(_copy_nopref) | ||
64 | UASM_L_LA(_copy_pref_both) | ||
65 | UASM_L_LA(_copy_pref_store) | ||
66 | |||
67 | /* We need one branch and therefore one relocation per target label. */ | ||
68 | static struct uasm_label __cpuinitdata labels[5]; | ||
69 | static struct uasm_reloc __cpuinitdata relocs[5]; | ||
70 | |||
71 | #define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010) | ||
72 | #define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020) | ||
73 | |||
74 | /* | ||
75 | * Maximum sizes: | ||
76 | * | ||
77 | * R4000 128 bytes S-cache: 0x058 bytes | ||
78 | * R4600 v1.7: 0x05c bytes | ||
79 | * R4600 v2.0: 0x060 bytes | ||
80 | * With prefetching, 16 word strides 0x120 bytes | ||
81 | */ | ||
82 | |||
83 | static u32 clear_page_array[0x120 / 4]; | ||
84 | |||
85 | #ifdef CONFIG_SIBYTE_DMA_PAGEOPS | ||
86 | void clear_page_cpu(void *page) __attribute__((alias("clear_page_array"))); | ||
87 | #else | ||
88 | void clear_page(void *page) __attribute__((alias("clear_page_array"))); | ||
89 | #endif | ||
90 | |||
91 | EXPORT_SYMBOL(clear_page); | ||
92 | |||
93 | /* | ||
94 | * Maximum sizes: | ||
95 | * | ||
96 | * R4000 128 bytes S-cache: 0x11c bytes | ||
97 | * R4600 v1.7: 0x080 bytes | ||
98 | * R4600 v2.0: 0x07c bytes | ||
99 | * With prefetching, 16 word strides 0x540 bytes | ||
100 | */ | ||
101 | static u32 copy_page_array[0x540 / 4]; | ||
102 | |||
103 | #ifdef CONFIG_SIBYTE_DMA_PAGEOPS | ||
104 | void | ||
105 | copy_page_cpu(void *to, void *from) __attribute__((alias("copy_page_array"))); | ||
106 | #else | ||
107 | void copy_page(void *to, void *from) __attribute__((alias("copy_page_array"))); | ||
108 | #endif | ||
109 | |||
110 | EXPORT_SYMBOL(copy_page); | ||
111 | |||
112 | |||
113 | static int pref_bias_clear_store __cpuinitdata; | ||
114 | static int pref_bias_copy_load __cpuinitdata; | ||
115 | static int pref_bias_copy_store __cpuinitdata; | ||
116 | |||
117 | static u32 pref_src_mode __cpuinitdata; | ||
118 | static u32 pref_dst_mode __cpuinitdata; | ||
119 | |||
120 | static int clear_word_size __cpuinitdata; | ||
121 | static int copy_word_size __cpuinitdata; | ||
122 | |||
123 | static int half_clear_loop_size __cpuinitdata; | ||
124 | static int half_copy_loop_size __cpuinitdata; | ||
125 | |||
126 | static int cache_line_size __cpuinitdata; | ||
127 | #define cache_line_mask() (cache_line_size - 1) | ||
128 | |||
129 | static inline void __cpuinit | ||
130 | pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off) | ||
131 | { | ||
132 | if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) { | ||
133 | if (off > 0x7fff) { | ||
134 | uasm_i_lui(buf, T9, uasm_rel_hi(off)); | ||
135 | uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off)); | ||
136 | } else | ||
137 | uasm_i_addiu(buf, T9, ZERO, off); | ||
138 | uasm_i_daddu(buf, reg1, reg2, T9); | ||
139 | } else { | ||
140 | if (off > 0x7fff) { | ||
141 | uasm_i_lui(buf, T9, uasm_rel_hi(off)); | ||
142 | uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off)); | ||
143 | UASM_i_ADDU(buf, reg1, reg2, T9); | ||
144 | } else | ||
145 | UASM_i_ADDIU(buf, reg1, reg2, off); | ||
146 | } | ||
147 | } | ||
148 | |||
149 | static void __cpuinit set_prefetch_parameters(void) | ||
150 | { | ||
151 | if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) | ||
152 | clear_word_size = 8; | ||
153 | else | ||
154 | clear_word_size = 4; | ||
155 | |||
156 | if (cpu_has_64bit_gp_regs) | ||
157 | copy_word_size = 8; | ||
158 | else | ||
159 | copy_word_size = 4; | ||
160 | |||
161 | /* | ||
162 | * The pref's used here are using "streaming" hints, which cause the | ||
163 | * copied data to be kicked out of the cache sooner. A page copy often | ||
164 | * ends up copying a lot more data than is commonly used, so this seems | ||
165 | * to make sense in terms of reducing cache pollution, but I've no real | ||
166 | * performance data to back this up. | ||
167 | */ | ||
168 | if (cpu_has_prefetch) { | ||
169 | /* | ||
170 | * XXX: Most prefetch bias values in here are based on | ||
171 | * guesswork. | ||
172 | */ | ||
173 | cache_line_size = cpu_dcache_line_size(); | ||
174 | switch (current_cpu_type()) { | ||
175 | case CPU_TX49XX: | ||
176 | /* TX49 supports only Pref_Load */ | ||
177 | pref_bias_copy_load = 256; | ||
178 | break; | ||
179 | |||
180 | case CPU_RM9000: | ||
181 | /* | ||
182 | * As a workaround for erratum G105 which make the | ||
183 | * PrepareForStore hint unusable we fall back to | ||
184 | * StoreRetained on the RM9000. Once it is known which | ||
185 | * versions of the RM9000 we'll be able to condition- | ||
186 | * alize this. | ||
187 | */ | ||
188 | |||
189 | case CPU_R10000: | ||
190 | case CPU_R12000: | ||
191 | case CPU_R14000: | ||
192 | /* | ||
193 | * Those values have been experimentally tuned for an | ||
194 | * Origin 200. | ||
195 | */ | ||
196 | pref_bias_clear_store = 512; | ||
197 | pref_bias_copy_load = 256; | ||
198 | pref_bias_copy_store = 256; | ||
199 | pref_src_mode = Pref_LoadStreamed; | ||
200 | pref_dst_mode = Pref_StoreStreamed; | ||
201 | break; | ||
202 | |||
203 | case CPU_SB1: | ||
204 | case CPU_SB1A: | ||
205 | pref_bias_clear_store = 128; | ||
206 | pref_bias_copy_load = 128; | ||
207 | pref_bias_copy_store = 128; | ||
208 | /* | ||
209 | * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed | ||
210 | * hints are broken. | ||
211 | */ | ||
212 | if (current_cpu_type() == CPU_SB1 && | ||
213 | (current_cpu_data.processor_id & 0xff) < 0x02) { | ||
214 | pref_src_mode = Pref_Load; | ||
215 | pref_dst_mode = Pref_Store; | ||
216 | } else { | ||
217 | pref_src_mode = Pref_LoadStreamed; | ||
218 | pref_dst_mode = Pref_StoreStreamed; | ||
219 | } | ||
220 | break; | ||
221 | |||
222 | default: | ||
223 | pref_bias_clear_store = 128; | ||
224 | pref_bias_copy_load = 256; | ||
225 | pref_bias_copy_store = 128; | ||
226 | pref_src_mode = Pref_LoadStreamed; | ||
227 | pref_dst_mode = Pref_PrepareForStore; | ||
228 | break; | ||
229 | } | ||
230 | } else { | ||
231 | if (cpu_has_cache_cdex_s) | ||
232 | cache_line_size = cpu_scache_line_size(); | ||
233 | else if (cpu_has_cache_cdex_p) | ||
234 | cache_line_size = cpu_dcache_line_size(); | ||
235 | } | ||
236 | /* | ||
237 | * Too much unrolling will overflow the available space in | ||
238 | * clear_space_array / copy_page_array. 8 words sounds generous, | ||
239 | * but a R4000 with 128 byte L2 line length can exceed even that. | ||
240 | */ | ||
241 | half_clear_loop_size = min(8 * clear_word_size, | ||
242 | max(cache_line_size >> 1, | ||
243 | 4 * clear_word_size)); | ||
244 | half_copy_loop_size = min(8 * copy_word_size, | ||
245 | max(cache_line_size >> 1, | ||
246 | 4 * copy_word_size)); | ||
247 | } | ||
248 | |||
249 | static void __cpuinit build_clear_store(u32 **buf, int off) | ||
250 | { | ||
251 | if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) { | ||
252 | uasm_i_sd(buf, ZERO, off, A0); | ||
253 | } else { | ||
254 | uasm_i_sw(buf, ZERO, off, A0); | ||
255 | } | ||
256 | } | ||
257 | |||
258 | static inline void __cpuinit build_clear_pref(u32 **buf, int off) | ||
259 | { | ||
260 | if (off & cache_line_mask()) | ||
261 | return; | ||
262 | |||
263 | if (pref_bias_clear_store) { | ||
264 | uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off, | ||
265 | A0); | ||
266 | } else if (cpu_has_cache_cdex_s) { | ||
267 | uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0); | ||
268 | } else if (cpu_has_cache_cdex_p) { | ||
269 | if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) { | ||
270 | uasm_i_nop(buf); | ||
271 | uasm_i_nop(buf); | ||
272 | uasm_i_nop(buf); | ||
273 | uasm_i_nop(buf); | ||
274 | } | ||
275 | |||
276 | if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x()) | ||
277 | uasm_i_lw(buf, ZERO, ZERO, AT); | ||
278 | |||
279 | uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0); | ||
280 | } | ||
281 | } | ||
282 | |||
283 | void __cpuinit build_clear_page(void) | ||
284 | { | ||
285 | int off; | ||
286 | u32 *buf = (u32 *)&clear_page_array; | ||
287 | struct uasm_label *l = labels; | ||
288 | struct uasm_reloc *r = relocs; | ||
289 | int i; | ||
290 | |||
291 | memset(labels, 0, sizeof(labels)); | ||
292 | memset(relocs, 0, sizeof(relocs)); | ||
293 | |||
294 | set_prefetch_parameters(); | ||
295 | |||
296 | /* | ||
297 | * This algorithm makes the following assumptions: | ||
298 | * - The prefetch bias is a multiple of 2 words. | ||
299 | * - The prefetch bias is less than one page. | ||
300 | */ | ||
301 | BUG_ON(pref_bias_clear_store % (2 * clear_word_size)); | ||
302 | BUG_ON(PAGE_SIZE < pref_bias_clear_store); | ||
303 | |||
304 | off = PAGE_SIZE - pref_bias_clear_store; | ||
305 | if (off > 0xffff || !pref_bias_clear_store) | ||
306 | pg_addiu(&buf, A2, A0, off); | ||
307 | else | ||
308 | uasm_i_ori(&buf, A2, A0, off); | ||
309 | |||
310 | if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x()) | ||
311 | uasm_i_lui(&buf, AT, 0xa000); | ||
312 | |||
313 | off = min(8, pref_bias_clear_store / cache_line_size) * | ||
314 | cache_line_size; | ||
315 | while (off) { | ||
316 | build_clear_pref(&buf, -off); | ||
317 | off -= cache_line_size; | ||
318 | } | ||
319 | uasm_l_clear_pref(&l, buf); | ||
320 | do { | ||
321 | build_clear_pref(&buf, off); | ||
322 | build_clear_store(&buf, off); | ||
323 | off += clear_word_size; | ||
324 | } while (off < half_clear_loop_size); | ||
325 | pg_addiu(&buf, A0, A0, 2 * off); | ||
326 | off = -off; | ||
327 | do { | ||
328 | build_clear_pref(&buf, off); | ||
329 | if (off == -clear_word_size) | ||
330 | uasm_il_bne(&buf, &r, A0, A2, label_clear_pref); | ||
331 | build_clear_store(&buf, off); | ||
332 | off += clear_word_size; | ||
333 | } while (off < 0); | ||
334 | |||
335 | if (pref_bias_clear_store) { | ||
336 | pg_addiu(&buf, A2, A0, pref_bias_clear_store); | ||
337 | uasm_l_clear_nopref(&l, buf); | ||
338 | off = 0; | ||
339 | do { | ||
340 | build_clear_store(&buf, off); | ||
341 | off += clear_word_size; | ||
342 | } while (off < half_clear_loop_size); | ||
343 | pg_addiu(&buf, A0, A0, 2 * off); | ||
344 | off = -off; | ||
345 | do { | ||
346 | if (off == -clear_word_size) | ||
347 | uasm_il_bne(&buf, &r, A0, A2, | ||
348 | label_clear_nopref); | ||
349 | build_clear_store(&buf, off); | ||
350 | off += clear_word_size; | ||
351 | } while (off < 0); | ||
352 | } | ||
353 | |||
354 | uasm_i_jr(&buf, RA); | ||
355 | uasm_i_nop(&buf); | ||
356 | |||
357 | BUG_ON(buf > clear_page_array + ARRAY_SIZE(clear_page_array)); | ||
358 | |||
359 | uasm_resolve_relocs(relocs, labels); | ||
360 | |||
361 | pr_debug("Synthesized clear page handler (%u instructions).\n", | ||
362 | (u32)(buf - clear_page_array)); | ||
363 | |||
364 | pr_debug("\t.set push\n"); | ||
365 | pr_debug("\t.set noreorder\n"); | ||
366 | for (i = 0; i < (buf - clear_page_array); i++) | ||
367 | pr_debug("\t.word 0x%08x\n", clear_page_array[i]); | ||
368 | pr_debug("\t.set pop\n"); | ||
369 | } | ||
370 | |||
371 | static void __cpuinit build_copy_load(u32 **buf, int reg, int off) | ||
372 | { | ||
373 | if (cpu_has_64bit_gp_regs) { | ||
374 | uasm_i_ld(buf, reg, off, A1); | ||
375 | } else { | ||
376 | uasm_i_lw(buf, reg, off, A1); | ||
377 | } | ||
378 | } | ||
379 | |||
380 | static void __cpuinit build_copy_store(u32 **buf, int reg, int off) | ||
381 | { | ||
382 | if (cpu_has_64bit_gp_regs) { | ||
383 | uasm_i_sd(buf, reg, off, A0); | ||
384 | } else { | ||
385 | uasm_i_sw(buf, reg, off, A0); | ||
386 | } | ||
387 | } | ||
388 | |||
389 | static inline void build_copy_load_pref(u32 **buf, int off) | ||
390 | { | ||
391 | if (off & cache_line_mask()) | ||
392 | return; | ||
393 | |||
394 | if (pref_bias_copy_load) | ||
395 | uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1); | ||
396 | } | ||
397 | |||
398 | static inline void build_copy_store_pref(u32 **buf, int off) | ||
399 | { | ||
400 | if (off & cache_line_mask()) | ||
401 | return; | ||
402 | |||
403 | if (pref_bias_copy_store) { | ||
404 | uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off, | ||
405 | A0); | ||
406 | } else if (cpu_has_cache_cdex_s) { | ||
407 | uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0); | ||
408 | } else if (cpu_has_cache_cdex_p) { | ||
409 | if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) { | ||
410 | uasm_i_nop(buf); | ||
411 | uasm_i_nop(buf); | ||
412 | uasm_i_nop(buf); | ||
413 | uasm_i_nop(buf); | ||
414 | } | ||
415 | |||
416 | if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x()) | ||
417 | uasm_i_lw(buf, ZERO, ZERO, AT); | ||
418 | |||
419 | uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0); | ||
420 | } | ||
421 | } | ||
422 | |||
423 | void __cpuinit build_copy_page(void) | ||
424 | { | ||
425 | int off; | ||
426 | u32 *buf = (u32 *)©_page_array; | ||
427 | struct uasm_label *l = labels; | ||
428 | struct uasm_reloc *r = relocs; | ||
429 | int i; | ||
430 | |||
431 | memset(labels, 0, sizeof(labels)); | ||
432 | memset(relocs, 0, sizeof(relocs)); | ||
433 | |||
434 | set_prefetch_parameters(); | ||
435 | |||
436 | /* | ||
437 | * This algorithm makes the following assumptions: | ||
438 | * - All prefetch biases are multiples of 8 words. | ||
439 | * - The prefetch biases are less than one page. | ||
440 | * - The store prefetch bias isn't greater than the load | ||
441 | * prefetch bias. | ||
442 | */ | ||
443 | BUG_ON(pref_bias_copy_load % (8 * copy_word_size)); | ||
444 | BUG_ON(pref_bias_copy_store % (8 * copy_word_size)); | ||
445 | BUG_ON(PAGE_SIZE < pref_bias_copy_load); | ||
446 | BUG_ON(pref_bias_copy_store > pref_bias_copy_load); | ||
447 | |||
448 | off = PAGE_SIZE - pref_bias_copy_load; | ||
449 | if (off > 0xffff || !pref_bias_copy_load) | ||
450 | pg_addiu(&buf, A2, A0, off); | ||
451 | else | ||
452 | uasm_i_ori(&buf, A2, A0, off); | ||
453 | |||
454 | if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x()) | ||
455 | uasm_i_lui(&buf, AT, 0xa000); | ||
456 | |||
457 | off = min(8, pref_bias_copy_load / cache_line_size) * cache_line_size; | ||
458 | while (off) { | ||
459 | build_copy_load_pref(&buf, -off); | ||
460 | off -= cache_line_size; | ||
461 | } | ||
462 | off = min(8, pref_bias_copy_store / cache_line_size) * cache_line_size; | ||
463 | while (off) { | ||
464 | build_copy_store_pref(&buf, -off); | ||
465 | off -= cache_line_size; | ||
466 | } | ||
467 | uasm_l_copy_pref_both(&l, buf); | ||
468 | do { | ||
469 | build_copy_load_pref(&buf, off); | ||
470 | build_copy_load(&buf, T0, off); | ||
471 | build_copy_load_pref(&buf, off + copy_word_size); | ||
472 | build_copy_load(&buf, T1, off + copy_word_size); | ||
473 | build_copy_load_pref(&buf, off + 2 * copy_word_size); | ||
474 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
475 | build_copy_load_pref(&buf, off + 3 * copy_word_size); | ||
476 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
477 | build_copy_store_pref(&buf, off); | ||
478 | build_copy_store(&buf, T0, off); | ||
479 | build_copy_store_pref(&buf, off + copy_word_size); | ||
480 | build_copy_store(&buf, T1, off + copy_word_size); | ||
481 | build_copy_store_pref(&buf, off + 2 * copy_word_size); | ||
482 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
483 | build_copy_store_pref(&buf, off + 3 * copy_word_size); | ||
484 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
485 | off += 4 * copy_word_size; | ||
486 | } while (off < half_copy_loop_size); | ||
487 | pg_addiu(&buf, A1, A1, 2 * off); | ||
488 | pg_addiu(&buf, A0, A0, 2 * off); | ||
489 | off = -off; | ||
490 | do { | ||
491 | build_copy_load_pref(&buf, off); | ||
492 | build_copy_load(&buf, T0, off); | ||
493 | build_copy_load_pref(&buf, off + copy_word_size); | ||
494 | build_copy_load(&buf, T1, off + copy_word_size); | ||
495 | build_copy_load_pref(&buf, off + 2 * copy_word_size); | ||
496 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
497 | build_copy_load_pref(&buf, off + 3 * copy_word_size); | ||
498 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
499 | build_copy_store_pref(&buf, off); | ||
500 | build_copy_store(&buf, T0, off); | ||
501 | build_copy_store_pref(&buf, off + copy_word_size); | ||
502 | build_copy_store(&buf, T1, off + copy_word_size); | ||
503 | build_copy_store_pref(&buf, off + 2 * copy_word_size); | ||
504 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
505 | build_copy_store_pref(&buf, off + 3 * copy_word_size); | ||
506 | if (off == -(4 * copy_word_size)) | ||
507 | uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both); | ||
508 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
509 | off += 4 * copy_word_size; | ||
510 | } while (off < 0); | ||
511 | |||
512 | if (pref_bias_copy_load - pref_bias_copy_store) { | ||
513 | pg_addiu(&buf, A2, A0, | ||
514 | pref_bias_copy_load - pref_bias_copy_store); | ||
515 | uasm_l_copy_pref_store(&l, buf); | ||
516 | off = 0; | ||
517 | do { | ||
518 | build_copy_load(&buf, T0, off); | ||
519 | build_copy_load(&buf, T1, off + copy_word_size); | ||
520 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
521 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
522 | build_copy_store_pref(&buf, off); | ||
523 | build_copy_store(&buf, T0, off); | ||
524 | build_copy_store_pref(&buf, off + copy_word_size); | ||
525 | build_copy_store(&buf, T1, off + copy_word_size); | ||
526 | build_copy_store_pref(&buf, off + 2 * copy_word_size); | ||
527 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
528 | build_copy_store_pref(&buf, off + 3 * copy_word_size); | ||
529 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
530 | off += 4 * copy_word_size; | ||
531 | } while (off < half_copy_loop_size); | ||
532 | pg_addiu(&buf, A1, A1, 2 * off); | ||
533 | pg_addiu(&buf, A0, A0, 2 * off); | ||
534 | off = -off; | ||
535 | do { | ||
536 | build_copy_load(&buf, T0, off); | ||
537 | build_copy_load(&buf, T1, off + copy_word_size); | ||
538 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
539 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
540 | build_copy_store_pref(&buf, off); | ||
541 | build_copy_store(&buf, T0, off); | ||
542 | build_copy_store_pref(&buf, off + copy_word_size); | ||
543 | build_copy_store(&buf, T1, off + copy_word_size); | ||
544 | build_copy_store_pref(&buf, off + 2 * copy_word_size); | ||
545 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
546 | build_copy_store_pref(&buf, off + 3 * copy_word_size); | ||
547 | if (off == -(4 * copy_word_size)) | ||
548 | uasm_il_bne(&buf, &r, A2, A0, | ||
549 | label_copy_pref_store); | ||
550 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
551 | off += 4 * copy_word_size; | ||
552 | } while (off < 0); | ||
553 | } | ||
554 | |||
555 | if (pref_bias_copy_store) { | ||
556 | pg_addiu(&buf, A2, A0, pref_bias_copy_store); | ||
557 | uasm_l_copy_nopref(&l, buf); | ||
558 | off = 0; | ||
559 | do { | ||
560 | build_copy_load(&buf, T0, off); | ||
561 | build_copy_load(&buf, T1, off + copy_word_size); | ||
562 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
563 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
564 | build_copy_store(&buf, T0, off); | ||
565 | build_copy_store(&buf, T1, off + copy_word_size); | ||
566 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
567 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
568 | off += 4 * copy_word_size; | ||
569 | } while (off < half_copy_loop_size); | ||
570 | pg_addiu(&buf, A1, A1, 2 * off); | ||
571 | pg_addiu(&buf, A0, A0, 2 * off); | ||
572 | off = -off; | ||
573 | do { | ||
574 | build_copy_load(&buf, T0, off); | ||
575 | build_copy_load(&buf, T1, off + copy_word_size); | ||
576 | build_copy_load(&buf, T2, off + 2 * copy_word_size); | ||
577 | build_copy_load(&buf, T3, off + 3 * copy_word_size); | ||
578 | build_copy_store(&buf, T0, off); | ||
579 | build_copy_store(&buf, T1, off + copy_word_size); | ||
580 | build_copy_store(&buf, T2, off + 2 * copy_word_size); | ||
581 | if (off == -(4 * copy_word_size)) | ||
582 | uasm_il_bne(&buf, &r, A2, A0, | ||
583 | label_copy_nopref); | ||
584 | build_copy_store(&buf, T3, off + 3 * copy_word_size); | ||
585 | off += 4 * copy_word_size; | ||
586 | } while (off < 0); | ||
587 | } | ||
588 | |||
589 | uasm_i_jr(&buf, RA); | ||
590 | uasm_i_nop(&buf); | ||
591 | |||
592 | BUG_ON(buf > copy_page_array + ARRAY_SIZE(copy_page_array)); | ||
593 | |||
594 | uasm_resolve_relocs(relocs, labels); | ||
595 | |||
596 | pr_debug("Synthesized copy page handler (%u instructions).\n", | ||
597 | (u32)(buf - copy_page_array)); | ||
598 | |||
599 | pr_debug("\t.set push\n"); | ||
600 | pr_debug("\t.set noreorder\n"); | ||
601 | for (i = 0; i < (buf - copy_page_array); i++) | ||
602 | pr_debug("\t.word 0x%08x\n", copy_page_array[i]); | ||
603 | pr_debug("\t.set pop\n"); | ||
604 | } | ||
605 | |||
606 | #ifdef CONFIG_SIBYTE_DMA_PAGEOPS | ||
607 | |||
608 | /* | ||
609 | * Pad descriptors to cacheline, since each is exclusively owned by a | ||
610 | * particular CPU. | ||
611 | */ | ||
612 | struct dmadscr { | ||
613 | u64 dscr_a; | ||
614 | u64 dscr_b; | ||
615 | u64 pad_a; | ||
616 | u64 pad_b; | ||
617 | } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS]; | ||
618 | |||
619 | void sb1_dma_init(void) | ||
620 | { | ||
621 | int i; | ||
622 | |||
623 | for (i = 0; i < DM_NUM_CHANNELS; i++) { | ||
624 | const u64 base_val = CPHYSADDR((unsigned long)&page_descr[i]) | | ||
625 | V_DM_DSCR_BASE_RINGSZ(1); | ||
626 | void *base_reg = IOADDR(A_DM_REGISTER(i, R_DM_DSCR_BASE)); | ||
627 | |||
628 | __raw_writeq(base_val, base_reg); | ||
629 | __raw_writeq(base_val | M_DM_DSCR_BASE_RESET, base_reg); | ||
630 | __raw_writeq(base_val | M_DM_DSCR_BASE_ENABL, base_reg); | ||
631 | } | ||
632 | } | ||
633 | |||
634 | void clear_page(void *page) | ||
635 | { | ||
636 | u64 to_phys = CPHYSADDR((unsigned long)page); | ||
637 | unsigned int cpu = smp_processor_id(); | ||
638 | |||
639 | /* if the page is not in KSEG0, use old way */ | ||
640 | if ((long)KSEGX((unsigned long)page) != (long)CKSEG0) | ||
641 | return clear_page_cpu(page); | ||
642 | |||
643 | page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM | | ||
644 | M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT; | ||
645 | page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE); | ||
646 | __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT))); | ||
647 | |||
648 | /* | ||
649 | * Don't really want to do it this way, but there's no | ||
650 | * reliable way to delay completion detection. | ||
651 | */ | ||
652 | while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG))) | ||
653 | & M_DM_DSCR_BASE_INTERRUPT)) | ||
654 | ; | ||
655 | __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE))); | ||
656 | } | ||
657 | |||
658 | void copy_page(void *to, void *from) | ||
659 | { | ||
660 | u64 from_phys = CPHYSADDR((unsigned long)from); | ||
661 | u64 to_phys = CPHYSADDR((unsigned long)to); | ||
662 | unsigned int cpu = smp_processor_id(); | ||
663 | |||
664 | /* if any page is not in KSEG0, use old way */ | ||
665 | if ((long)KSEGX((unsigned long)to) != (long)CKSEG0 | ||
666 | || (long)KSEGX((unsigned long)from) != (long)CKSEG0) | ||
667 | return copy_page_cpu(to, from); | ||
668 | |||
669 | page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST | | ||
670 | M_DM_DSCRA_INTERRUPT; | ||
671 | page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE); | ||
672 | __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT))); | ||
673 | |||
674 | /* | ||
675 | * Don't really want to do it this way, but there's no | ||
676 | * reliable way to delay completion detection. | ||
677 | */ | ||
678 | while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG))) | ||
679 | & M_DM_DSCR_BASE_INTERRUPT)) | ||
680 | ; | ||
681 | __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE))); | ||
682 | } | ||
683 | |||
684 | #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */ | ||