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/smp.h>
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/proc_fs.h>
17 
18 #include <asm/bugs.h>
19 #include <asm/cacheops.h>
20 #include <asm/inst.h>
21 #include <asm/io.h>
22 #include <asm/page.h>
23 #include <asm/pgtable.h>
24 #include <asm/prefetch.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 <asm/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
pg_addiu(u32 ** buf,unsigned int reg1,unsigned int reg2,unsigned int off)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 
set_prefetch_parameters(void)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_R5500:
176 		case CPU_TX49XX:
177 			/* These processors only support the Pref_Load. */
178 			pref_bias_copy_load = 256;
179 			break;
180 
181 		case CPU_RM9000:
182 			/*
183 			 * As a workaround for erratum G105 which make the
184 			 * PrepareForStore hint unusable we fall back to
185 			 * StoreRetained on the RM9000.  Once it is known which
186 			 * versions of the RM9000 we'll be able to condition-
187 			 * alize this.
188 			 */
189 
190 		case CPU_R10000:
191 		case CPU_R12000:
192 		case CPU_R14000:
193 			/*
194 			 * Those values have been experimentally tuned for an
195 			 * Origin 200.
196 			 */
197 			pref_bias_clear_store = 512;
198 			pref_bias_copy_load = 256;
199 			pref_bias_copy_store = 256;
200 			pref_src_mode = Pref_LoadStreamed;
201 			pref_dst_mode = Pref_StoreStreamed;
202 			break;
203 
204 		case CPU_SB1:
205 		case CPU_SB1A:
206 			pref_bias_clear_store = 128;
207 			pref_bias_copy_load = 128;
208 			pref_bias_copy_store = 128;
209 			/*
210 			 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
211 			 * hints are broken.
212 			 */
213 			if (current_cpu_type() == CPU_SB1 &&
214 			    (current_cpu_data.processor_id & 0xff) < 0x02) {
215 				pref_src_mode = Pref_Load;
216 				pref_dst_mode = Pref_Store;
217 			} else {
218 				pref_src_mode = Pref_LoadStreamed;
219 				pref_dst_mode = Pref_StoreStreamed;
220 			}
221 			break;
222 
223 		default:
224 			pref_bias_clear_store = 128;
225 			pref_bias_copy_load = 256;
226 			pref_bias_copy_store = 128;
227 			pref_src_mode = Pref_LoadStreamed;
228 			pref_dst_mode = Pref_PrepareForStore;
229 			break;
230 		}
231 	} else {
232 		if (cpu_has_cache_cdex_s)
233 			cache_line_size = cpu_scache_line_size();
234 		else if (cpu_has_cache_cdex_p)
235 			cache_line_size = cpu_dcache_line_size();
236 	}
237 	/*
238 	 * Too much unrolling will overflow the available space in
239 	 * clear_space_array / copy_page_array.
240 	 */
241 	half_clear_loop_size = min(16 * clear_word_size,
242 				   max(cache_line_size >> 1,
243 				       4 * clear_word_size));
244 	half_copy_loop_size = min(16 * copy_word_size,
245 				  max(cache_line_size >> 1,
246 				      4 * copy_word_size));
247 }
248 
build_clear_store(u32 ** buf,int off)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 
build_clear_pref(u32 ** buf,int off)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 (cache_line_size == (half_clear_loop_size << 1)) {
267 		if (cpu_has_cache_cdex_s) {
268 			uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
269 		} else if (cpu_has_cache_cdex_p) {
270 			if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
271 				uasm_i_nop(buf);
272 				uasm_i_nop(buf);
273 				uasm_i_nop(buf);
274 				uasm_i_nop(buf);
275 			}
276 
277 			if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
278 				uasm_i_lw(buf, ZERO, ZERO, AT);
279 
280 			uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
281 		}
282 		}
283 }
284 
build_clear_page(void)285 void __cpuinit build_clear_page(void)
286 {
287 	int off;
288 	u32 *buf = (u32 *)&clear_page_array;
289 	struct uasm_label *l = labels;
290 	struct uasm_reloc *r = relocs;
291 	int i;
292 
293 	memset(labels, 0, sizeof(labels));
294 	memset(relocs, 0, sizeof(relocs));
295 
296 	set_prefetch_parameters();
297 
298 	/*
299 	 * This algorithm makes the following assumptions:
300 	 *   - The prefetch bias is a multiple of 2 words.
301 	 *   - The prefetch bias is less than one page.
302 	 */
303 	BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
304 	BUG_ON(PAGE_SIZE < pref_bias_clear_store);
305 
306 	off = PAGE_SIZE - pref_bias_clear_store;
307 	if (off > 0xffff || !pref_bias_clear_store)
308 		pg_addiu(&buf, A2, A0, off);
309 	else
310 		uasm_i_ori(&buf, A2, A0, off);
311 
312 	if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
313 		uasm_i_lui(&buf, AT, 0xa000);
314 
315 	off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
316 	                        * cache_line_size : 0;
317 	while (off) {
318 		build_clear_pref(&buf, -off);
319 		off -= cache_line_size;
320 	}
321 	uasm_l_clear_pref(&l, buf);
322 	do {
323 		build_clear_pref(&buf, off);
324 		build_clear_store(&buf, off);
325 		off += clear_word_size;
326 	} while (off < half_clear_loop_size);
327 	pg_addiu(&buf, A0, A0, 2 * off);
328 	off = -off;
329 	do {
330 		build_clear_pref(&buf, off);
331 		if (off == -clear_word_size)
332 			uasm_il_bne(&buf, &r, A0, A2, label_clear_pref);
333 		build_clear_store(&buf, off);
334 		off += clear_word_size;
335 	} while (off < 0);
336 
337 	if (pref_bias_clear_store) {
338 		pg_addiu(&buf, A2, A0, pref_bias_clear_store);
339 		uasm_l_clear_nopref(&l, buf);
340 		off = 0;
341 		do {
342 			build_clear_store(&buf, off);
343 			off += clear_word_size;
344 		} while (off < half_clear_loop_size);
345 		pg_addiu(&buf, A0, A0, 2 * off);
346 		off = -off;
347 		do {
348 			if (off == -clear_word_size)
349 				uasm_il_bne(&buf, &r, A0, A2,
350 					    label_clear_nopref);
351 			build_clear_store(&buf, off);
352 			off += clear_word_size;
353 		} while (off < 0);
354 	}
355 
356 	uasm_i_jr(&buf, RA);
357 	uasm_i_nop(&buf);
358 
359 	BUG_ON(buf > clear_page_array + ARRAY_SIZE(clear_page_array));
360 
361 	uasm_resolve_relocs(relocs, labels);
362 
363 	pr_debug("Synthesized clear page handler (%u instructions).\n",
364 		 (u32)(buf - clear_page_array));
365 
366 	pr_debug("\t.set push\n");
367 	pr_debug("\t.set noreorder\n");
368 	for (i = 0; i < (buf - clear_page_array); i++)
369 		pr_debug("\t.word 0x%08x\n", clear_page_array[i]);
370 	pr_debug("\t.set pop\n");
371 }
372 
build_copy_load(u32 ** buf,int reg,int off)373 static void __cpuinit build_copy_load(u32 **buf, int reg, int off)
374 {
375 	if (cpu_has_64bit_gp_regs) {
376 		uasm_i_ld(buf, reg, off, A1);
377 	} else {
378 		uasm_i_lw(buf, reg, off, A1);
379 	}
380 }
381 
build_copy_store(u32 ** buf,int reg,int off)382 static void __cpuinit build_copy_store(u32 **buf, int reg, int off)
383 {
384 	if (cpu_has_64bit_gp_regs) {
385 		uasm_i_sd(buf, reg, off, A0);
386 	} else {
387 		uasm_i_sw(buf, reg, off, A0);
388 	}
389 }
390 
build_copy_load_pref(u32 ** buf,int off)391 static inline void build_copy_load_pref(u32 **buf, int off)
392 {
393 	if (off & cache_line_mask())
394 		return;
395 
396 	if (pref_bias_copy_load)
397 		uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1);
398 }
399 
build_copy_store_pref(u32 ** buf,int off)400 static inline void build_copy_store_pref(u32 **buf, int off)
401 {
402 	if (off & cache_line_mask())
403 		return;
404 
405 	if (pref_bias_copy_store) {
406 		uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
407 			    A0);
408 	} else if (cache_line_size == (half_copy_loop_size << 1)) {
409 		if (cpu_has_cache_cdex_s) {
410 			uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
411 		} else if (cpu_has_cache_cdex_p) {
412 			if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
413 				uasm_i_nop(buf);
414 				uasm_i_nop(buf);
415 				uasm_i_nop(buf);
416 				uasm_i_nop(buf);
417 			}
418 
419 			if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
420 				uasm_i_lw(buf, ZERO, ZERO, AT);
421 
422 			uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
423 		}
424 	}
425 }
426 
build_copy_page(void)427 void __cpuinit build_copy_page(void)
428 {
429 	int off;
430 	u32 *buf = (u32 *)&copy_page_array;
431 	struct uasm_label *l = labels;
432 	struct uasm_reloc *r = relocs;
433 	int i;
434 
435 	memset(labels, 0, sizeof(labels));
436 	memset(relocs, 0, sizeof(relocs));
437 
438 	set_prefetch_parameters();
439 
440 	/*
441 	 * This algorithm makes the following assumptions:
442 	 *   - All prefetch biases are multiples of 8 words.
443 	 *   - The prefetch biases are less than one page.
444 	 *   - The store prefetch bias isn't greater than the load
445 	 *     prefetch bias.
446 	 */
447 	BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
448 	BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
449 	BUG_ON(PAGE_SIZE < pref_bias_copy_load);
450 	BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
451 
452 	off = PAGE_SIZE - pref_bias_copy_load;
453 	if (off > 0xffff || !pref_bias_copy_load)
454 		pg_addiu(&buf, A2, A0, off);
455 	else
456 		uasm_i_ori(&buf, A2, A0, off);
457 
458 	if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
459 		uasm_i_lui(&buf, AT, 0xa000);
460 
461 	off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
462 	                        cache_line_size : 0;
463 	while (off) {
464 		build_copy_load_pref(&buf, -off);
465 		off -= cache_line_size;
466 	}
467 	off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
468 	                        cache_line_size : 0;
469 	while (off) {
470 		build_copy_store_pref(&buf, -off);
471 		off -= cache_line_size;
472 	}
473 	uasm_l_copy_pref_both(&l, buf);
474 	do {
475 		build_copy_load_pref(&buf, off);
476 		build_copy_load(&buf, T0, off);
477 		build_copy_load_pref(&buf, off + copy_word_size);
478 		build_copy_load(&buf, T1, off + copy_word_size);
479 		build_copy_load_pref(&buf, off + 2 * copy_word_size);
480 		build_copy_load(&buf, T2, off + 2 * copy_word_size);
481 		build_copy_load_pref(&buf, off + 3 * copy_word_size);
482 		build_copy_load(&buf, T3, off + 3 * copy_word_size);
483 		build_copy_store_pref(&buf, off);
484 		build_copy_store(&buf, T0, off);
485 		build_copy_store_pref(&buf, off + copy_word_size);
486 		build_copy_store(&buf, T1, off + copy_word_size);
487 		build_copy_store_pref(&buf, off + 2 * copy_word_size);
488 		build_copy_store(&buf, T2, off + 2 * copy_word_size);
489 		build_copy_store_pref(&buf, off + 3 * copy_word_size);
490 		build_copy_store(&buf, T3, off + 3 * copy_word_size);
491 		off += 4 * copy_word_size;
492 	} while (off < half_copy_loop_size);
493 	pg_addiu(&buf, A1, A1, 2 * off);
494 	pg_addiu(&buf, A0, A0, 2 * off);
495 	off = -off;
496 	do {
497 		build_copy_load_pref(&buf, off);
498 		build_copy_load(&buf, T0, off);
499 		build_copy_load_pref(&buf, off + copy_word_size);
500 		build_copy_load(&buf, T1, off + copy_word_size);
501 		build_copy_load_pref(&buf, off + 2 * copy_word_size);
502 		build_copy_load(&buf, T2, off + 2 * copy_word_size);
503 		build_copy_load_pref(&buf, off + 3 * copy_word_size);
504 		build_copy_load(&buf, T3, off + 3 * copy_word_size);
505 		build_copy_store_pref(&buf, off);
506 		build_copy_store(&buf, T0, off);
507 		build_copy_store_pref(&buf, off + copy_word_size);
508 		build_copy_store(&buf, T1, off + copy_word_size);
509 		build_copy_store_pref(&buf, off + 2 * copy_word_size);
510 		build_copy_store(&buf, T2, off + 2 * copy_word_size);
511 		build_copy_store_pref(&buf, off + 3 * copy_word_size);
512 		if (off == -(4 * copy_word_size))
513 			uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
514 		build_copy_store(&buf, T3, off + 3 * copy_word_size);
515 		off += 4 * copy_word_size;
516 	} while (off < 0);
517 
518 	if (pref_bias_copy_load - pref_bias_copy_store) {
519 		pg_addiu(&buf, A2, A0,
520 			 pref_bias_copy_load - pref_bias_copy_store);
521 		uasm_l_copy_pref_store(&l, buf);
522 		off = 0;
523 		do {
524 			build_copy_load(&buf, T0, off);
525 			build_copy_load(&buf, T1, off + copy_word_size);
526 			build_copy_load(&buf, T2, off + 2 * copy_word_size);
527 			build_copy_load(&buf, T3, off + 3 * copy_word_size);
528 			build_copy_store_pref(&buf, off);
529 			build_copy_store(&buf, T0, off);
530 			build_copy_store_pref(&buf, off + copy_word_size);
531 			build_copy_store(&buf, T1, off + copy_word_size);
532 			build_copy_store_pref(&buf, off + 2 * copy_word_size);
533 			build_copy_store(&buf, T2, off + 2 * copy_word_size);
534 			build_copy_store_pref(&buf, off + 3 * copy_word_size);
535 			build_copy_store(&buf, T3, off + 3 * copy_word_size);
536 			off += 4 * copy_word_size;
537 		} while (off < half_copy_loop_size);
538 		pg_addiu(&buf, A1, A1, 2 * off);
539 		pg_addiu(&buf, A0, A0, 2 * off);
540 		off = -off;
541 		do {
542 			build_copy_load(&buf, T0, off);
543 			build_copy_load(&buf, T1, off + copy_word_size);
544 			build_copy_load(&buf, T2, off + 2 * copy_word_size);
545 			build_copy_load(&buf, T3, off + 3 * copy_word_size);
546 			build_copy_store_pref(&buf, off);
547 			build_copy_store(&buf, T0, off);
548 			build_copy_store_pref(&buf, off + copy_word_size);
549 			build_copy_store(&buf, T1, off + copy_word_size);
550 			build_copy_store_pref(&buf, off + 2 * copy_word_size);
551 			build_copy_store(&buf, T2, off + 2 * copy_word_size);
552 			build_copy_store_pref(&buf, off + 3 * copy_word_size);
553 			if (off == -(4 * copy_word_size))
554 				uasm_il_bne(&buf, &r, A2, A0,
555 					    label_copy_pref_store);
556 			build_copy_store(&buf, T3, off + 3 * copy_word_size);
557 			off += 4 * copy_word_size;
558 		} while (off < 0);
559 	}
560 
561 	if (pref_bias_copy_store) {
562 		pg_addiu(&buf, A2, A0, pref_bias_copy_store);
563 		uasm_l_copy_nopref(&l, buf);
564 		off = 0;
565 		do {
566 			build_copy_load(&buf, T0, off);
567 			build_copy_load(&buf, T1, off + copy_word_size);
568 			build_copy_load(&buf, T2, off + 2 * copy_word_size);
569 			build_copy_load(&buf, T3, off + 3 * copy_word_size);
570 			build_copy_store(&buf, T0, off);
571 			build_copy_store(&buf, T1, off + copy_word_size);
572 			build_copy_store(&buf, T2, off + 2 * copy_word_size);
573 			build_copy_store(&buf, T3, off + 3 * copy_word_size);
574 			off += 4 * copy_word_size;
575 		} while (off < half_copy_loop_size);
576 		pg_addiu(&buf, A1, A1, 2 * off);
577 		pg_addiu(&buf, A0, A0, 2 * off);
578 		off = -off;
579 		do {
580 			build_copy_load(&buf, T0, off);
581 			build_copy_load(&buf, T1, off + copy_word_size);
582 			build_copy_load(&buf, T2, off + 2 * copy_word_size);
583 			build_copy_load(&buf, T3, off + 3 * copy_word_size);
584 			build_copy_store(&buf, T0, off);
585 			build_copy_store(&buf, T1, off + copy_word_size);
586 			build_copy_store(&buf, T2, off + 2 * copy_word_size);
587 			if (off == -(4 * copy_word_size))
588 				uasm_il_bne(&buf, &r, A2, A0,
589 					    label_copy_nopref);
590 			build_copy_store(&buf, T3, off + 3 * copy_word_size);
591 			off += 4 * copy_word_size;
592 		} while (off < 0);
593 	}
594 
595 	uasm_i_jr(&buf, RA);
596 	uasm_i_nop(&buf);
597 
598 	BUG_ON(buf > copy_page_array + ARRAY_SIZE(copy_page_array));
599 
600 	uasm_resolve_relocs(relocs, labels);
601 
602 	pr_debug("Synthesized copy page handler (%u instructions).\n",
603 		 (u32)(buf - copy_page_array));
604 
605 	pr_debug("\t.set push\n");
606 	pr_debug("\t.set noreorder\n");
607 	for (i = 0; i < (buf - copy_page_array); i++)
608 		pr_debug("\t.word 0x%08x\n", copy_page_array[i]);
609 	pr_debug("\t.set pop\n");
610 }
611 
612 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
613 
614 /*
615  * Pad descriptors to cacheline, since each is exclusively owned by a
616  * particular CPU.
617  */
618 struct dmadscr {
619 	u64 dscr_a;
620 	u64 dscr_b;
621 	u64 pad_a;
622 	u64 pad_b;
623 } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
624 
sb1_dma_init(void)625 void sb1_dma_init(void)
626 {
627 	int i;
628 
629 	for (i = 0; i < DM_NUM_CHANNELS; i++) {
630 		const u64 base_val = CPHYSADDR((unsigned long)&page_descr[i]) |
631 				     V_DM_DSCR_BASE_RINGSZ(1);
632 		void *base_reg = IOADDR(A_DM_REGISTER(i, R_DM_DSCR_BASE));
633 
634 		__raw_writeq(base_val, base_reg);
635 		__raw_writeq(base_val | M_DM_DSCR_BASE_RESET, base_reg);
636 		__raw_writeq(base_val | M_DM_DSCR_BASE_ENABL, base_reg);
637 	}
638 }
639 
clear_page(void * page)640 void clear_page(void *page)
641 {
642 	u64 to_phys = CPHYSADDR((unsigned long)page);
643 	unsigned int cpu = smp_processor_id();
644 
645 	/* if the page is not in KSEG0, use old way */
646 	if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
647 		return clear_page_cpu(page);
648 
649 	page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
650 				 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
651 	page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
652 	__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
653 
654 	/*
655 	 * Don't really want to do it this way, but there's no
656 	 * reliable way to delay completion detection.
657 	 */
658 	while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
659 		 & M_DM_DSCR_BASE_INTERRUPT))
660 		;
661 	__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
662 }
663 
copy_page(void * to,void * from)664 void copy_page(void *to, void *from)
665 {
666 	u64 from_phys = CPHYSADDR((unsigned long)from);
667 	u64 to_phys = CPHYSADDR((unsigned long)to);
668 	unsigned int cpu = smp_processor_id();
669 
670 	/* if any page is not in KSEG0, use old way */
671 	if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
672 	    || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
673 		return copy_page_cpu(to, from);
674 
675 	page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
676 				 M_DM_DSCRA_INTERRUPT;
677 	page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
678 	__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
679 
680 	/*
681 	 * Don't really want to do it this way, but there's no
682 	 * reliable way to delay completion detection.
683 	 */
684 	while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
685 		 & M_DM_DSCR_BASE_INTERRUPT))
686 		;
687 	__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
688 }
689 
690 #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */
691