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
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 * Copyright (C) 2012 MIPS Technologies, Inc.
11 #include <linux/init.h>
12 #include <linux/kernel.h>
13 #include <linux/sched.h>
14 #include <linux/smp.h>
16 #include <linux/module.h>
17 #include <linux/proc_fs.h>
20 #include <asm/cacheops.h>
24 #include <asm/pgtable.h>
25 #include <asm/prefetch.h>
26 #include <asm/bootinfo.h>
27 #include <asm/mipsregs.h>
28 #include <asm/mmu_context.h>
32 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
33 #include <asm/sibyte/sb1250.h>
34 #include <asm/sibyte/sb1250_regs.h>
35 #include <asm/sibyte/sb1250_dma.h>
40 /* Registers used in the assembled routines. */
53 /* Handle labels (which must be positive integers). */
55 label_clear_nopref = 1,
59 label_copy_pref_store,
62 UASM_L_LA(_clear_nopref)
63 UASM_L_LA(_clear_pref)
64 UASM_L_LA(_copy_nopref)
65 UASM_L_LA(_copy_pref_both)
66 UASM_L_LA(_copy_pref_store)
68 /* We need one branch and therefore one relocation per target label. */
69 static struct uasm_label __cpuinitdata labels[5];
70 static struct uasm_reloc __cpuinitdata relocs[5];
72 #define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010)
73 #define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020)
75 static int pref_bias_clear_store __cpuinitdata;
76 static int pref_bias_copy_load __cpuinitdata;
77 static int pref_bias_copy_store __cpuinitdata;
79 static u32 pref_src_mode __cpuinitdata;
80 static u32 pref_dst_mode __cpuinitdata;
82 static int clear_word_size __cpuinitdata;
83 static int copy_word_size __cpuinitdata;
85 static int half_clear_loop_size __cpuinitdata;
86 static int half_copy_loop_size __cpuinitdata;
88 static int cache_line_size __cpuinitdata;
89 #define cache_line_mask() (cache_line_size - 1)
91 static inline void __cpuinit
92 pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
94 if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) {
96 uasm_i_lui(buf, T9, uasm_rel_hi(off));
97 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
99 uasm_i_addiu(buf, T9, ZERO, off);
100 uasm_i_daddu(buf, reg1, reg2, T9);
103 uasm_i_lui(buf, T9, uasm_rel_hi(off));
104 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
105 UASM_i_ADDU(buf, reg1, reg2, T9);
107 UASM_i_ADDIU(buf, reg1, reg2, off);
111 static void __cpuinit set_prefetch_parameters(void)
113 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
118 if (cpu_has_64bit_gp_regs)
124 * The pref's used here are using "streaming" hints, which cause the
125 * copied data to be kicked out of the cache sooner. A page copy often
126 * ends up copying a lot more data than is commonly used, so this seems
127 * to make sense in terms of reducing cache pollution, but I've no real
128 * performance data to back this up.
130 if (cpu_has_prefetch) {
132 * XXX: Most prefetch bias values in here are based on
135 cache_line_size = cpu_dcache_line_size();
136 switch (current_cpu_type()) {
139 /* These processors only support the Pref_Load. */
140 pref_bias_copy_load = 256;
145 * As a workaround for erratum G105 which make the
146 * PrepareForStore hint unusable we fall back to
147 * StoreRetained on the RM9000. Once it is known which
148 * versions of the RM9000 we'll be able to condition-
156 * Those values have been experimentally tuned for an
159 pref_bias_clear_store = 512;
160 pref_bias_copy_load = 256;
161 pref_bias_copy_store = 256;
162 pref_src_mode = Pref_LoadStreamed;
163 pref_dst_mode = Pref_StoreStreamed;
168 pref_bias_clear_store = 128;
169 pref_bias_copy_load = 128;
170 pref_bias_copy_store = 128;
172 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
175 if (current_cpu_type() == CPU_SB1 &&
176 (current_cpu_data.processor_id & 0xff) < 0x02) {
177 pref_src_mode = Pref_Load;
178 pref_dst_mode = Pref_Store;
180 pref_src_mode = Pref_LoadStreamed;
181 pref_dst_mode = Pref_StoreStreamed;
186 pref_bias_clear_store = 128;
187 pref_bias_copy_load = 256;
188 pref_bias_copy_store = 128;
189 pref_src_mode = Pref_LoadStreamed;
190 pref_dst_mode = Pref_PrepareForStore;
194 if (cpu_has_cache_cdex_s)
195 cache_line_size = cpu_scache_line_size();
196 else if (cpu_has_cache_cdex_p)
197 cache_line_size = cpu_dcache_line_size();
200 * Too much unrolling will overflow the available space in
201 * clear_space_array / copy_page_array.
203 half_clear_loop_size = min(16 * clear_word_size,
204 max(cache_line_size >> 1,
205 4 * clear_word_size));
206 half_copy_loop_size = min(16 * copy_word_size,
207 max(cache_line_size >> 1,
208 4 * copy_word_size));
211 static void __cpuinit build_clear_store(u32 **buf, int off)
213 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
214 uasm_i_sd(buf, ZERO, off, A0);
216 uasm_i_sw(buf, ZERO, off, A0);
220 static inline void __cpuinit build_clear_pref(u32 **buf, int off)
222 if (off & cache_line_mask())
225 if (pref_bias_clear_store) {
226 uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
228 } else if (cache_line_size == (half_clear_loop_size << 1)) {
229 if (cpu_has_cache_cdex_s) {
230 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
231 } else if (cpu_has_cache_cdex_p) {
232 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
239 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
240 uasm_i_lw(buf, ZERO, ZERO, AT);
242 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
247 extern u32 __clear_page_start;
248 extern u32 __clear_page_end;
249 extern u32 __copy_page_start;
250 extern u32 __copy_page_end;
252 void __cpuinit build_clear_page(void)
255 u32 *buf = &__clear_page_start;
256 struct uasm_label *l = labels;
257 struct uasm_reloc *r = relocs;
260 memset(labels, 0, sizeof(labels));
261 memset(relocs, 0, sizeof(relocs));
263 set_prefetch_parameters();
266 * This algorithm makes the following assumptions:
267 * - The prefetch bias is a multiple of 2 words.
268 * - The prefetch bias is less than one page.
270 BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
271 BUG_ON(PAGE_SIZE < pref_bias_clear_store);
273 off = PAGE_SIZE - pref_bias_clear_store;
274 if (off > 0xffff || !pref_bias_clear_store)
275 pg_addiu(&buf, A2, A0, off);
277 uasm_i_ori(&buf, A2, A0, off);
279 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
280 uasm_i_lui(&buf, AT, 0xa000);
282 off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
283 * cache_line_size : 0;
285 build_clear_pref(&buf, -off);
286 off -= cache_line_size;
288 uasm_l_clear_pref(&l, buf);
290 build_clear_pref(&buf, off);
291 build_clear_store(&buf, off);
292 off += clear_word_size;
293 } while (off < half_clear_loop_size);
294 pg_addiu(&buf, A0, A0, 2 * off);
297 build_clear_pref(&buf, off);
298 if (off == -clear_word_size)
299 uasm_il_bne(&buf, &r, A0, A2, label_clear_pref);
300 build_clear_store(&buf, off);
301 off += clear_word_size;
304 if (pref_bias_clear_store) {
305 pg_addiu(&buf, A2, A0, pref_bias_clear_store);
306 uasm_l_clear_nopref(&l, buf);
309 build_clear_store(&buf, off);
310 off += clear_word_size;
311 } while (off < half_clear_loop_size);
312 pg_addiu(&buf, A0, A0, 2 * off);
315 if (off == -clear_word_size)
316 uasm_il_bne(&buf, &r, A0, A2,
318 build_clear_store(&buf, off);
319 off += clear_word_size;
326 BUG_ON(buf > &__clear_page_end);
328 uasm_resolve_relocs(relocs, labels);
330 pr_debug("Synthesized clear page handler (%u instructions).\n",
331 (u32)(buf - &__clear_page_start));
333 pr_debug("\t.set push\n");
334 pr_debug("\t.set noreorder\n");
335 for (i = 0; i < (buf - &__clear_page_start); i++)
336 pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
337 pr_debug("\t.set pop\n");
340 static void __cpuinit build_copy_load(u32 **buf, int reg, int off)
342 if (cpu_has_64bit_gp_regs) {
343 uasm_i_ld(buf, reg, off, A1);
345 uasm_i_lw(buf, reg, off, A1);
349 static void __cpuinit build_copy_store(u32 **buf, int reg, int off)
351 if (cpu_has_64bit_gp_regs) {
352 uasm_i_sd(buf, reg, off, A0);
354 uasm_i_sw(buf, reg, off, A0);
358 static inline void build_copy_load_pref(u32 **buf, int off)
360 if (off & cache_line_mask())
363 if (pref_bias_copy_load)
364 uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1);
367 static inline void build_copy_store_pref(u32 **buf, int off)
369 if (off & cache_line_mask())
372 if (pref_bias_copy_store) {
373 uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
375 } else if (cache_line_size == (half_copy_loop_size << 1)) {
376 if (cpu_has_cache_cdex_s) {
377 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
378 } else if (cpu_has_cache_cdex_p) {
379 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
386 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
387 uasm_i_lw(buf, ZERO, ZERO, AT);
389 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
394 void __cpuinit build_copy_page(void)
397 u32 *buf = &__copy_page_start;
398 struct uasm_label *l = labels;
399 struct uasm_reloc *r = relocs;
402 memset(labels, 0, sizeof(labels));
403 memset(relocs, 0, sizeof(relocs));
405 set_prefetch_parameters();
408 * This algorithm makes the following assumptions:
409 * - All prefetch biases are multiples of 8 words.
410 * - The prefetch biases are less than one page.
411 * - The store prefetch bias isn't greater than the load
414 BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
415 BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
416 BUG_ON(PAGE_SIZE < pref_bias_copy_load);
417 BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
419 off = PAGE_SIZE - pref_bias_copy_load;
420 if (off > 0xffff || !pref_bias_copy_load)
421 pg_addiu(&buf, A2, A0, off);
423 uasm_i_ori(&buf, A2, A0, off);
425 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
426 uasm_i_lui(&buf, AT, 0xa000);
428 off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
431 build_copy_load_pref(&buf, -off);
432 off -= cache_line_size;
434 off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
437 build_copy_store_pref(&buf, -off);
438 off -= cache_line_size;
440 uasm_l_copy_pref_both(&l, buf);
442 build_copy_load_pref(&buf, off);
443 build_copy_load(&buf, T0, off);
444 build_copy_load_pref(&buf, off + copy_word_size);
445 build_copy_load(&buf, T1, off + copy_word_size);
446 build_copy_load_pref(&buf, off + 2 * copy_word_size);
447 build_copy_load(&buf, T2, off + 2 * copy_word_size);
448 build_copy_load_pref(&buf, off + 3 * copy_word_size);
449 build_copy_load(&buf, T3, off + 3 * copy_word_size);
450 build_copy_store_pref(&buf, off);
451 build_copy_store(&buf, T0, off);
452 build_copy_store_pref(&buf, off + copy_word_size);
453 build_copy_store(&buf, T1, off + copy_word_size);
454 build_copy_store_pref(&buf, off + 2 * copy_word_size);
455 build_copy_store(&buf, T2, off + 2 * copy_word_size);
456 build_copy_store_pref(&buf, off + 3 * copy_word_size);
457 build_copy_store(&buf, T3, off + 3 * copy_word_size);
458 off += 4 * copy_word_size;
459 } while (off < half_copy_loop_size);
460 pg_addiu(&buf, A1, A1, 2 * off);
461 pg_addiu(&buf, A0, A0, 2 * off);
464 build_copy_load_pref(&buf, off);
465 build_copy_load(&buf, T0, off);
466 build_copy_load_pref(&buf, off + copy_word_size);
467 build_copy_load(&buf, T1, off + copy_word_size);
468 build_copy_load_pref(&buf, off + 2 * copy_word_size);
469 build_copy_load(&buf, T2, off + 2 * copy_word_size);
470 build_copy_load_pref(&buf, off + 3 * copy_word_size);
471 build_copy_load(&buf, T3, off + 3 * copy_word_size);
472 build_copy_store_pref(&buf, off);
473 build_copy_store(&buf, T0, off);
474 build_copy_store_pref(&buf, off + copy_word_size);
475 build_copy_store(&buf, T1, off + copy_word_size);
476 build_copy_store_pref(&buf, off + 2 * copy_word_size);
477 build_copy_store(&buf, T2, off + 2 * copy_word_size);
478 build_copy_store_pref(&buf, off + 3 * copy_word_size);
479 if (off == -(4 * copy_word_size))
480 uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
481 build_copy_store(&buf, T3, off + 3 * copy_word_size);
482 off += 4 * copy_word_size;
485 if (pref_bias_copy_load - pref_bias_copy_store) {
486 pg_addiu(&buf, A2, A0,
487 pref_bias_copy_load - pref_bias_copy_store);
488 uasm_l_copy_pref_store(&l, buf);
491 build_copy_load(&buf, T0, off);
492 build_copy_load(&buf, T1, off + copy_word_size);
493 build_copy_load(&buf, T2, off + 2 * copy_word_size);
494 build_copy_load(&buf, T3, off + 3 * copy_word_size);
495 build_copy_store_pref(&buf, off);
496 build_copy_store(&buf, T0, off);
497 build_copy_store_pref(&buf, off + copy_word_size);
498 build_copy_store(&buf, T1, off + copy_word_size);
499 build_copy_store_pref(&buf, off + 2 * copy_word_size);
500 build_copy_store(&buf, T2, off + 2 * copy_word_size);
501 build_copy_store_pref(&buf, off + 3 * copy_word_size);
502 build_copy_store(&buf, T3, off + 3 * copy_word_size);
503 off += 4 * copy_word_size;
504 } while (off < half_copy_loop_size);
505 pg_addiu(&buf, A1, A1, 2 * off);
506 pg_addiu(&buf, A0, A0, 2 * off);
509 build_copy_load(&buf, T0, off);
510 build_copy_load(&buf, T1, off + copy_word_size);
511 build_copy_load(&buf, T2, off + 2 * copy_word_size);
512 build_copy_load(&buf, T3, off + 3 * copy_word_size);
513 build_copy_store_pref(&buf, off);
514 build_copy_store(&buf, T0, off);
515 build_copy_store_pref(&buf, off + copy_word_size);
516 build_copy_store(&buf, T1, off + copy_word_size);
517 build_copy_store_pref(&buf, off + 2 * copy_word_size);
518 build_copy_store(&buf, T2, off + 2 * copy_word_size);
519 build_copy_store_pref(&buf, off + 3 * copy_word_size);
520 if (off == -(4 * copy_word_size))
521 uasm_il_bne(&buf, &r, A2, A0,
522 label_copy_pref_store);
523 build_copy_store(&buf, T3, off + 3 * copy_word_size);
524 off += 4 * copy_word_size;
528 if (pref_bias_copy_store) {
529 pg_addiu(&buf, A2, A0, pref_bias_copy_store);
530 uasm_l_copy_nopref(&l, buf);
533 build_copy_load(&buf, T0, off);
534 build_copy_load(&buf, T1, off + copy_word_size);
535 build_copy_load(&buf, T2, off + 2 * copy_word_size);
536 build_copy_load(&buf, T3, off + 3 * copy_word_size);
537 build_copy_store(&buf, T0, off);
538 build_copy_store(&buf, T1, off + copy_word_size);
539 build_copy_store(&buf, T2, off + 2 * copy_word_size);
540 build_copy_store(&buf, T3, off + 3 * copy_word_size);
541 off += 4 * copy_word_size;
542 } while (off < half_copy_loop_size);
543 pg_addiu(&buf, A1, A1, 2 * off);
544 pg_addiu(&buf, A0, A0, 2 * off);
547 build_copy_load(&buf, T0, off);
548 build_copy_load(&buf, T1, off + copy_word_size);
549 build_copy_load(&buf, T2, off + 2 * copy_word_size);
550 build_copy_load(&buf, T3, off + 3 * copy_word_size);
551 build_copy_store(&buf, T0, off);
552 build_copy_store(&buf, T1, off + copy_word_size);
553 build_copy_store(&buf, T2, off + 2 * copy_word_size);
554 if (off == -(4 * copy_word_size))
555 uasm_il_bne(&buf, &r, A2, A0,
557 build_copy_store(&buf, T3, off + 3 * copy_word_size);
558 off += 4 * copy_word_size;
565 BUG_ON(buf > &__copy_page_end);
567 uasm_resolve_relocs(relocs, labels);
569 pr_debug("Synthesized copy page handler (%u instructions).\n",
570 (u32)(buf - &__copy_page_start));
572 pr_debug("\t.set push\n");
573 pr_debug("\t.set noreorder\n");
574 for (i = 0; i < (buf - &__copy_page_start); i++)
575 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
576 pr_debug("\t.set pop\n");
579 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
580 extern void clear_page_cpu(void *page);
581 extern void copy_page_cpu(void *to, void *from);
584 * Pad descriptors to cacheline, since each is exclusively owned by a
592 } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
594 void sb1_dma_init(void)
598 for (i = 0; i < DM_NUM_CHANNELS; i++) {
599 const u64 base_val = CPHYSADDR((unsigned long)&page_descr[i]) |
600 V_DM_DSCR_BASE_RINGSZ(1);
601 void *base_reg = IOADDR(A_DM_REGISTER(i, R_DM_DSCR_BASE));
603 __raw_writeq(base_val, base_reg);
604 __raw_writeq(base_val | M_DM_DSCR_BASE_RESET, base_reg);
605 __raw_writeq(base_val | M_DM_DSCR_BASE_ENABL, base_reg);
609 void clear_page(void *page)
611 u64 to_phys = CPHYSADDR((unsigned long)page);
612 unsigned int cpu = smp_processor_id();
614 /* if the page is not in KSEG0, use old way */
615 if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
616 return clear_page_cpu(page);
618 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
619 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
620 page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
621 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
624 * Don't really want to do it this way, but there's no
625 * reliable way to delay completion detection.
627 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
628 & M_DM_DSCR_BASE_INTERRUPT))
630 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
633 void copy_page(void *to, void *from)
635 u64 from_phys = CPHYSADDR((unsigned long)from);
636 u64 to_phys = CPHYSADDR((unsigned long)to);
637 unsigned int cpu = smp_processor_id();
639 /* if any page is not in KSEG0, use old way */
640 if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
641 || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
642 return copy_page_cpu(to, from);
644 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
645 M_DM_DSCRA_INTERRUPT;
646 page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
647 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
650 * Don't really want to do it this way, but there's no
651 * reliable way to delay completion detection.
653 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
654 & M_DM_DSCR_BASE_INTERRUPT))
656 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
659 #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */