mount options: fix reiserfs
[linux-2.6/openmoko-kernel/knife-kernel.git] / arch / mips / mm / tlbex.c
blob218a6cc415e8865cbdb83b59889666fd4f2d8dc7
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.
6 * Synthesize TLB refill handlers at runtime.
8 * Copyright (C) 2004, 2005, 2006, 2008 Thiemo Seufer
9 * Copyright (C) 2005, 2007 Maciej W. Rozycki
10 * Copyright (C) 2006 Ralf Baechle (ralf@linux-mips.org)
12 * ... and the days got worse and worse and now you see
13 * I've gone completly out of my mind.
15 * They're coming to take me a away haha
16 * they're coming to take me a away hoho hihi haha
17 * to the funny farm where code is beautiful all the time ...
19 * (Condolences to Napoleon XIV)
22 #include <linux/kernel.h>
23 #include <linux/types.h>
24 #include <linux/string.h>
25 #include <linux/init.h>
27 #include <asm/mmu_context.h>
28 #include <asm/war.h>
30 #include "uasm.h"
32 static inline int r45k_bvahwbug(void)
34 /* XXX: We should probe for the presence of this bug, but we don't. */
35 return 0;
38 static inline int r4k_250MHZhwbug(void)
40 /* XXX: We should probe for the presence of this bug, but we don't. */
41 return 0;
44 static inline int __maybe_unused bcm1250_m3_war(void)
46 return BCM1250_M3_WAR;
49 static inline int __maybe_unused r10000_llsc_war(void)
51 return R10000_LLSC_WAR;
55 * Found by experiment: At least some revisions of the 4kc throw under
56 * some circumstances a machine check exception, triggered by invalid
57 * values in the index register. Delaying the tlbp instruction until
58 * after the next branch, plus adding an additional nop in front of
59 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
60 * why; it's not an issue caused by the core RTL.
63 static int __init m4kc_tlbp_war(void)
65 return (current_cpu_data.processor_id & 0xffff00) ==
66 (PRID_COMP_MIPS | PRID_IMP_4KC);
69 /* Handle labels (which must be positive integers). */
70 enum label_id {
71 label_second_part = 1,
72 label_leave,
73 #ifdef MODULE_START
74 label_module_alloc,
75 #endif
76 label_vmalloc,
77 label_vmalloc_done,
78 label_tlbw_hazard,
79 label_split,
80 label_nopage_tlbl,
81 label_nopage_tlbs,
82 label_nopage_tlbm,
83 label_smp_pgtable_change,
84 label_r3000_write_probe_fail,
87 UASM_L_LA(_second_part)
88 UASM_L_LA(_leave)
89 #ifdef MODULE_START
90 UASM_L_LA(_module_alloc)
91 #endif
92 UASM_L_LA(_vmalloc)
93 UASM_L_LA(_vmalloc_done)
94 UASM_L_LA(_tlbw_hazard)
95 UASM_L_LA(_split)
96 UASM_L_LA(_nopage_tlbl)
97 UASM_L_LA(_nopage_tlbs)
98 UASM_L_LA(_nopage_tlbm)
99 UASM_L_LA(_smp_pgtable_change)
100 UASM_L_LA(_r3000_write_probe_fail)
103 * For debug purposes.
105 static inline void dump_handler(const u32 *handler, int count)
107 int i;
109 pr_debug("\t.set push\n");
110 pr_debug("\t.set noreorder\n");
112 for (i = 0; i < count; i++)
113 pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
115 pr_debug("\t.set pop\n");
118 /* The only general purpose registers allowed in TLB handlers. */
119 #define K0 26
120 #define K1 27
122 /* Some CP0 registers */
123 #define C0_INDEX 0, 0
124 #define C0_ENTRYLO0 2, 0
125 #define C0_TCBIND 2, 2
126 #define C0_ENTRYLO1 3, 0
127 #define C0_CONTEXT 4, 0
128 #define C0_BADVADDR 8, 0
129 #define C0_ENTRYHI 10, 0
130 #define C0_EPC 14, 0
131 #define C0_XCONTEXT 20, 0
133 #ifdef CONFIG_64BIT
134 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
135 #else
136 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
137 #endif
139 /* The worst case length of the handler is around 18 instructions for
140 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
141 * Maximum space available is 32 instructions for R3000 and 64
142 * instructions for R4000.
144 * We deliberately chose a buffer size of 128, so we won't scribble
145 * over anything important on overflow before we panic.
147 static u32 tlb_handler[128] __initdata;
149 /* simply assume worst case size for labels and relocs */
150 static struct uasm_label labels[128] __initdata;
151 static struct uasm_reloc relocs[128] __initdata;
154 * The R3000 TLB handler is simple.
156 static void __init build_r3000_tlb_refill_handler(void)
158 long pgdc = (long)pgd_current;
159 u32 *p;
161 memset(tlb_handler, 0, sizeof(tlb_handler));
162 p = tlb_handler;
164 uasm_i_mfc0(&p, K0, C0_BADVADDR);
165 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
166 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
167 uasm_i_srl(&p, K0, K0, 22); /* load delay */
168 uasm_i_sll(&p, K0, K0, 2);
169 uasm_i_addu(&p, K1, K1, K0);
170 uasm_i_mfc0(&p, K0, C0_CONTEXT);
171 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
172 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
173 uasm_i_addu(&p, K1, K1, K0);
174 uasm_i_lw(&p, K0, 0, K1);
175 uasm_i_nop(&p); /* load delay */
176 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
177 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
178 uasm_i_tlbwr(&p); /* cp0 delay */
179 uasm_i_jr(&p, K1);
180 uasm_i_rfe(&p); /* branch delay */
182 if (p > tlb_handler + 32)
183 panic("TLB refill handler space exceeded");
185 pr_debug("Wrote TLB refill handler (%u instructions).\n",
186 (unsigned int)(p - tlb_handler));
188 memcpy((void *)ebase, tlb_handler, 0x80);
190 dump_handler((u32 *)ebase, 32);
194 * The R4000 TLB handler is much more complicated. We have two
195 * consecutive handler areas with 32 instructions space each.
196 * Since they aren't used at the same time, we can overflow in the
197 * other one.To keep things simple, we first assume linear space,
198 * then we relocate it to the final handler layout as needed.
200 static u32 final_handler[64] __initdata;
203 * Hazards
205 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
206 * 2. A timing hazard exists for the TLBP instruction.
208 * stalling_instruction
209 * TLBP
211 * The JTLB is being read for the TLBP throughout the stall generated by the
212 * previous instruction. This is not really correct as the stalling instruction
213 * can modify the address used to access the JTLB. The failure symptom is that
214 * the TLBP instruction will use an address created for the stalling instruction
215 * and not the address held in C0_ENHI and thus report the wrong results.
217 * The software work-around is to not allow the instruction preceding the TLBP
218 * to stall - make it an NOP or some other instruction guaranteed not to stall.
220 * Errata 2 will not be fixed. This errata is also on the R5000.
222 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
224 static void __init __maybe_unused build_tlb_probe_entry(u32 **p)
226 switch (current_cpu_type()) {
227 /* Found by experiment: R4600 v2.0 needs this, too. */
228 case CPU_R4600:
229 case CPU_R5000:
230 case CPU_R5000A:
231 case CPU_NEVADA:
232 uasm_i_nop(p);
233 uasm_i_tlbp(p);
234 break;
236 default:
237 uasm_i_tlbp(p);
238 break;
243 * Write random or indexed TLB entry, and care about the hazards from
244 * the preceeding mtc0 and for the following eret.
246 enum tlb_write_entry { tlb_random, tlb_indexed };
248 static void __init build_tlb_write_entry(u32 **p, struct uasm_label **l,
249 struct uasm_reloc **r,
250 enum tlb_write_entry wmode)
252 void(*tlbw)(u32 **) = NULL;
254 switch (wmode) {
255 case tlb_random: tlbw = uasm_i_tlbwr; break;
256 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
259 if (cpu_has_mips_r2) {
260 uasm_i_ehb(p);
261 tlbw(p);
262 return;
265 switch (current_cpu_type()) {
266 case CPU_R4000PC:
267 case CPU_R4000SC:
268 case CPU_R4000MC:
269 case CPU_R4400PC:
270 case CPU_R4400SC:
271 case CPU_R4400MC:
273 * This branch uses up a mtc0 hazard nop slot and saves
274 * two nops after the tlbw instruction.
276 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
277 tlbw(p);
278 uasm_l_tlbw_hazard(l, *p);
279 uasm_i_nop(p);
280 break;
282 case CPU_R4600:
283 case CPU_R4700:
284 case CPU_R5000:
285 case CPU_R5000A:
286 uasm_i_nop(p);
287 tlbw(p);
288 uasm_i_nop(p);
289 break;
291 case CPU_R4300:
292 case CPU_5KC:
293 case CPU_TX49XX:
294 case CPU_AU1000:
295 case CPU_AU1100:
296 case CPU_AU1500:
297 case CPU_AU1550:
298 case CPU_AU1200:
299 case CPU_AU1210:
300 case CPU_AU1250:
301 case CPU_PR4450:
302 uasm_i_nop(p);
303 tlbw(p);
304 break;
306 case CPU_R10000:
307 case CPU_R12000:
308 case CPU_R14000:
309 case CPU_4KC:
310 case CPU_SB1:
311 case CPU_SB1A:
312 case CPU_4KSC:
313 case CPU_20KC:
314 case CPU_25KF:
315 case CPU_BCM3302:
316 case CPU_BCM4710:
317 case CPU_LOONGSON2:
318 if (m4kc_tlbp_war())
319 uasm_i_nop(p);
320 tlbw(p);
321 break;
323 case CPU_NEVADA:
324 uasm_i_nop(p); /* QED specifies 2 nops hazard */
326 * This branch uses up a mtc0 hazard nop slot and saves
327 * a nop after the tlbw instruction.
329 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
330 tlbw(p);
331 uasm_l_tlbw_hazard(l, *p);
332 break;
334 case CPU_RM7000:
335 uasm_i_nop(p);
336 uasm_i_nop(p);
337 uasm_i_nop(p);
338 uasm_i_nop(p);
339 tlbw(p);
340 break;
342 case CPU_RM9000:
344 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
345 * use of the JTLB for instructions should not occur for 4
346 * cpu cycles and use for data translations should not occur
347 * for 3 cpu cycles.
349 uasm_i_ssnop(p);
350 uasm_i_ssnop(p);
351 uasm_i_ssnop(p);
352 uasm_i_ssnop(p);
353 tlbw(p);
354 uasm_i_ssnop(p);
355 uasm_i_ssnop(p);
356 uasm_i_ssnop(p);
357 uasm_i_ssnop(p);
358 break;
360 case CPU_VR4111:
361 case CPU_VR4121:
362 case CPU_VR4122:
363 case CPU_VR4181:
364 case CPU_VR4181A:
365 uasm_i_nop(p);
366 uasm_i_nop(p);
367 tlbw(p);
368 uasm_i_nop(p);
369 uasm_i_nop(p);
370 break;
372 case CPU_VR4131:
373 case CPU_VR4133:
374 case CPU_R5432:
375 uasm_i_nop(p);
376 uasm_i_nop(p);
377 tlbw(p);
378 break;
380 default:
381 panic("No TLB refill handler yet (CPU type: %d)",
382 current_cpu_data.cputype);
383 break;
387 #ifdef CONFIG_64BIT
389 * TMP and PTR are scratch.
390 * TMP will be clobbered, PTR will hold the pmd entry.
392 static void __init
393 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
394 unsigned int tmp, unsigned int ptr)
396 long pgdc = (long)pgd_current;
399 * The vmalloc handling is not in the hotpath.
401 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
402 #ifdef MODULE_START
403 uasm_il_bltz(p, r, tmp, label_module_alloc);
404 #else
405 uasm_il_bltz(p, r, tmp, label_vmalloc);
406 #endif
407 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
409 #ifdef CONFIG_SMP
410 # ifdef CONFIG_MIPS_MT_SMTC
412 * SMTC uses TCBind value as "CPU" index
414 uasm_i_mfc0(p, ptr, C0_TCBIND);
415 uasm_i_dsrl(p, ptr, ptr, 19);
416 # else
418 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
419 * stored in CONTEXT.
421 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
422 uasm_i_dsrl(p, ptr, ptr, 23);
423 #endif
424 UASM_i_LA_mostly(p, tmp, pgdc);
425 uasm_i_daddu(p, ptr, ptr, tmp);
426 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
427 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
428 #else
429 UASM_i_LA_mostly(p, ptr, pgdc);
430 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
431 #endif
433 uasm_l_vmalloc_done(l, *p);
435 if (PGDIR_SHIFT - 3 < 32) /* get pgd offset in bytes */
436 uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
437 else
438 uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
440 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
441 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
442 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
443 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
444 uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
445 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
446 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
450 * BVADDR is the faulting address, PTR is scratch.
451 * PTR will hold the pgd for vmalloc.
453 static void __init
454 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
455 unsigned int bvaddr, unsigned int ptr)
457 long swpd = (long)swapper_pg_dir;
459 #ifdef MODULE_START
460 long modd = (long)module_pg_dir;
462 uasm_l_module_alloc(l, *p);
464 * Assumption:
465 * VMALLOC_START >= 0xc000000000000000UL
466 * MODULE_START >= 0xe000000000000000UL
468 UASM_i_SLL(p, ptr, bvaddr, 2);
469 uasm_il_bgez(p, r, ptr, label_vmalloc);
471 if (uasm_in_compat_space_p(MODULE_START) &&
472 !uasm_rel_lo(MODULE_START)) {
473 uasm_i_lui(p, ptr, uasm_rel_hi(MODULE_START)); /* delay slot */
474 } else {
475 /* unlikely configuration */
476 uasm_i_nop(p); /* delay slot */
477 UASM_i_LA(p, ptr, MODULE_START);
479 uasm_i_dsubu(p, bvaddr, bvaddr, ptr);
481 if (uasm_in_compat_space_p(modd) && !uasm_rel_lo(modd)) {
482 uasm_il_b(p, r, label_vmalloc_done);
483 uasm_i_lui(p, ptr, uasm_rel_hi(modd));
484 } else {
485 UASM_i_LA_mostly(p, ptr, modd);
486 uasm_il_b(p, r, label_vmalloc_done);
487 if (uasm_in_compat_space_p(modd))
488 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(modd));
489 else
490 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(modd));
493 uasm_l_vmalloc(l, *p);
494 if (uasm_in_compat_space_p(MODULE_START) &&
495 !uasm_rel_lo(MODULE_START) &&
496 MODULE_START << 32 == VMALLOC_START)
497 uasm_i_dsll32(p, ptr, ptr, 0); /* typical case */
498 else
499 UASM_i_LA(p, ptr, VMALLOC_START);
500 #else
501 uasm_l_vmalloc(l, *p);
502 UASM_i_LA(p, ptr, VMALLOC_START);
503 #endif
504 uasm_i_dsubu(p, bvaddr, bvaddr, ptr);
506 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
507 uasm_il_b(p, r, label_vmalloc_done);
508 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
509 } else {
510 UASM_i_LA_mostly(p, ptr, swpd);
511 uasm_il_b(p, r, label_vmalloc_done);
512 if (uasm_in_compat_space_p(swpd))
513 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
514 else
515 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
519 #else /* !CONFIG_64BIT */
522 * TMP and PTR are scratch.
523 * TMP will be clobbered, PTR will hold the pgd entry.
525 static void __init __maybe_unused
526 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
528 long pgdc = (long)pgd_current;
530 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
531 #ifdef CONFIG_SMP
532 #ifdef CONFIG_MIPS_MT_SMTC
534 * SMTC uses TCBind value as "CPU" index
536 uasm_i_mfc0(p, ptr, C0_TCBIND);
537 UASM_i_LA_mostly(p, tmp, pgdc);
538 uasm_i_srl(p, ptr, ptr, 19);
539 #else
541 * smp_processor_id() << 3 is stored in CONTEXT.
543 uasm_i_mfc0(p, ptr, C0_CONTEXT);
544 UASM_i_LA_mostly(p, tmp, pgdc);
545 uasm_i_srl(p, ptr, ptr, 23);
546 #endif
547 uasm_i_addu(p, ptr, tmp, ptr);
548 #else
549 UASM_i_LA_mostly(p, ptr, pgdc);
550 #endif
551 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
552 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
553 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
554 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
555 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
558 #endif /* !CONFIG_64BIT */
560 static void __init build_adjust_context(u32 **p, unsigned int ctx)
562 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
563 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
565 switch (current_cpu_type()) {
566 case CPU_VR41XX:
567 case CPU_VR4111:
568 case CPU_VR4121:
569 case CPU_VR4122:
570 case CPU_VR4131:
571 case CPU_VR4181:
572 case CPU_VR4181A:
573 case CPU_VR4133:
574 shift += 2;
575 break;
577 default:
578 break;
581 if (shift)
582 UASM_i_SRL(p, ctx, ctx, shift);
583 uasm_i_andi(p, ctx, ctx, mask);
586 static void __init build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
589 * Bug workaround for the Nevada. It seems as if under certain
590 * circumstances the move from cp0_context might produce a
591 * bogus result when the mfc0 instruction and its consumer are
592 * in a different cacheline or a load instruction, probably any
593 * memory reference, is between them.
595 switch (current_cpu_type()) {
596 case CPU_NEVADA:
597 UASM_i_LW(p, ptr, 0, ptr);
598 GET_CONTEXT(p, tmp); /* get context reg */
599 break;
601 default:
602 GET_CONTEXT(p, tmp); /* get context reg */
603 UASM_i_LW(p, ptr, 0, ptr);
604 break;
607 build_adjust_context(p, tmp);
608 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
611 static void __init build_update_entries(u32 **p, unsigned int tmp,
612 unsigned int ptep)
615 * 64bit address support (36bit on a 32bit CPU) in a 32bit
616 * Kernel is a special case. Only a few CPUs use it.
618 #ifdef CONFIG_64BIT_PHYS_ADDR
619 if (cpu_has_64bits) {
620 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
621 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
622 uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
623 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
624 uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
625 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
626 } else {
627 int pte_off_even = sizeof(pte_t) / 2;
628 int pte_off_odd = pte_off_even + sizeof(pte_t);
630 /* The pte entries are pre-shifted */
631 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
632 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
633 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
634 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
636 #else
637 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
638 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
639 if (r45k_bvahwbug())
640 build_tlb_probe_entry(p);
641 UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
642 if (r4k_250MHZhwbug())
643 uasm_i_mtc0(p, 0, C0_ENTRYLO0);
644 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
645 UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
646 if (r45k_bvahwbug())
647 uasm_i_mfc0(p, tmp, C0_INDEX);
648 if (r4k_250MHZhwbug())
649 uasm_i_mtc0(p, 0, C0_ENTRYLO1);
650 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
651 #endif
654 static void __init build_r4000_tlb_refill_handler(void)
656 u32 *p = tlb_handler;
657 struct uasm_label *l = labels;
658 struct uasm_reloc *r = relocs;
659 u32 *f;
660 unsigned int final_len;
662 memset(tlb_handler, 0, sizeof(tlb_handler));
663 memset(labels, 0, sizeof(labels));
664 memset(relocs, 0, sizeof(relocs));
665 memset(final_handler, 0, sizeof(final_handler));
668 * create the plain linear handler
670 if (bcm1250_m3_war()) {
671 UASM_i_MFC0(&p, K0, C0_BADVADDR);
672 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
673 uasm_i_xor(&p, K0, K0, K1);
674 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
675 uasm_il_bnez(&p, &r, K0, label_leave);
676 /* No need for uasm_i_nop */
679 #ifdef CONFIG_64BIT
680 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
681 #else
682 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
683 #endif
685 build_get_ptep(&p, K0, K1);
686 build_update_entries(&p, K0, K1);
687 build_tlb_write_entry(&p, &l, &r, tlb_random);
688 uasm_l_leave(&l, p);
689 uasm_i_eret(&p); /* return from trap */
691 #ifdef CONFIG_64BIT
692 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
693 #endif
696 * Overflow check: For the 64bit handler, we need at least one
697 * free instruction slot for the wrap-around branch. In worst
698 * case, if the intended insertion point is a delay slot, we
699 * need three, with the second nop'ed and the third being
700 * unused.
702 /* Loongson2 ebase is different than r4k, we have more space */
703 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
704 if ((p - tlb_handler) > 64)
705 panic("TLB refill handler space exceeded");
706 #else
707 if (((p - tlb_handler) > 63)
708 || (((p - tlb_handler) > 61)
709 && uasm_insn_has_bdelay(relocs, tlb_handler + 29)))
710 panic("TLB refill handler space exceeded");
711 #endif
714 * Now fold the handler in the TLB refill handler space.
716 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
717 f = final_handler;
718 /* Simplest case, just copy the handler. */
719 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
720 final_len = p - tlb_handler;
721 #else /* CONFIG_64BIT */
722 f = final_handler + 32;
723 if ((p - tlb_handler) <= 32) {
724 /* Just copy the handler. */
725 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
726 final_len = p - tlb_handler;
727 } else {
728 u32 *split = tlb_handler + 30;
731 * Find the split point.
733 if (uasm_insn_has_bdelay(relocs, split - 1))
734 split--;
736 /* Copy first part of the handler. */
737 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
738 f += split - tlb_handler;
740 /* Insert branch. */
741 uasm_l_split(&l, final_handler);
742 uasm_il_b(&f, &r, label_split);
743 if (uasm_insn_has_bdelay(relocs, split))
744 uasm_i_nop(&f);
745 else {
746 uasm_copy_handler(relocs, labels, split, split + 1, f);
747 uasm_move_labels(labels, f, f + 1, -1);
748 f++;
749 split++;
752 /* Copy the rest of the handler. */
753 uasm_copy_handler(relocs, labels, split, p, final_handler);
754 final_len = (f - (final_handler + 32)) + (p - split);
756 #endif /* CONFIG_64BIT */
758 uasm_resolve_relocs(relocs, labels);
759 pr_debug("Wrote TLB refill handler (%u instructions).\n",
760 final_len);
762 memcpy((void *)ebase, final_handler, 0x100);
764 dump_handler((u32 *)ebase, 64);
768 * TLB load/store/modify handlers.
770 * Only the fastpath gets synthesized at runtime, the slowpath for
771 * do_page_fault remains normal asm.
773 extern void tlb_do_page_fault_0(void);
774 extern void tlb_do_page_fault_1(void);
777 * 128 instructions for the fastpath handler is generous and should
778 * never be exceeded.
780 #define FASTPATH_SIZE 128
782 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
783 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
784 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
786 static void __init
787 iPTE_LW(u32 **p, struct uasm_label **l, unsigned int pte, unsigned int ptr)
789 #ifdef CONFIG_SMP
790 # ifdef CONFIG_64BIT_PHYS_ADDR
791 if (cpu_has_64bits)
792 uasm_i_lld(p, pte, 0, ptr);
793 else
794 # endif
795 UASM_i_LL(p, pte, 0, ptr);
796 #else
797 # ifdef CONFIG_64BIT_PHYS_ADDR
798 if (cpu_has_64bits)
799 uasm_i_ld(p, pte, 0, ptr);
800 else
801 # endif
802 UASM_i_LW(p, pte, 0, ptr);
803 #endif
806 static void __init
807 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
808 unsigned int mode)
810 #ifdef CONFIG_64BIT_PHYS_ADDR
811 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
812 #endif
814 uasm_i_ori(p, pte, pte, mode);
815 #ifdef CONFIG_SMP
816 # ifdef CONFIG_64BIT_PHYS_ADDR
817 if (cpu_has_64bits)
818 uasm_i_scd(p, pte, 0, ptr);
819 else
820 # endif
821 UASM_i_SC(p, pte, 0, ptr);
823 if (r10000_llsc_war())
824 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
825 else
826 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
828 # ifdef CONFIG_64BIT_PHYS_ADDR
829 if (!cpu_has_64bits) {
830 /* no uasm_i_nop needed */
831 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
832 uasm_i_ori(p, pte, pte, hwmode);
833 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
834 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
835 /* no uasm_i_nop needed */
836 uasm_i_lw(p, pte, 0, ptr);
837 } else
838 uasm_i_nop(p);
839 # else
840 uasm_i_nop(p);
841 # endif
842 #else
843 # ifdef CONFIG_64BIT_PHYS_ADDR
844 if (cpu_has_64bits)
845 uasm_i_sd(p, pte, 0, ptr);
846 else
847 # endif
848 UASM_i_SW(p, pte, 0, ptr);
850 # ifdef CONFIG_64BIT_PHYS_ADDR
851 if (!cpu_has_64bits) {
852 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
853 uasm_i_ori(p, pte, pte, hwmode);
854 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
855 uasm_i_lw(p, pte, 0, ptr);
857 # endif
858 #endif
862 * Check if PTE is present, if not then jump to LABEL. PTR points to
863 * the page table where this PTE is located, PTE will be re-loaded
864 * with it's original value.
866 static void __init
867 build_pte_present(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
868 unsigned int pte, unsigned int ptr, enum label_id lid)
870 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
871 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
872 uasm_il_bnez(p, r, pte, lid);
873 iPTE_LW(p, l, pte, ptr);
876 /* Make PTE valid, store result in PTR. */
877 static void __init
878 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
879 unsigned int ptr)
881 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
883 iPTE_SW(p, r, pte, ptr, mode);
887 * Check if PTE can be written to, if not branch to LABEL. Regardless
888 * restore PTE with value from PTR when done.
890 static void __init
891 build_pte_writable(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
892 unsigned int pte, unsigned int ptr, enum label_id lid)
894 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
895 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
896 uasm_il_bnez(p, r, pte, lid);
897 iPTE_LW(p, l, pte, ptr);
900 /* Make PTE writable, update software status bits as well, then store
901 * at PTR.
903 static void __init
904 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
905 unsigned int ptr)
907 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
908 | _PAGE_DIRTY);
910 iPTE_SW(p, r, pte, ptr, mode);
914 * Check if PTE can be modified, if not branch to LABEL. Regardless
915 * restore PTE with value from PTR when done.
917 static void __init
918 build_pte_modifiable(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
919 unsigned int pte, unsigned int ptr, enum label_id lid)
921 uasm_i_andi(p, pte, pte, _PAGE_WRITE);
922 uasm_il_beqz(p, r, pte, lid);
923 iPTE_LW(p, l, pte, ptr);
927 * R3000 style TLB load/store/modify handlers.
931 * This places the pte into ENTRYLO0 and writes it with tlbwi.
932 * Then it returns.
934 static void __init
935 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
937 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
938 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
939 uasm_i_tlbwi(p);
940 uasm_i_jr(p, tmp);
941 uasm_i_rfe(p); /* branch delay */
945 * This places the pte into ENTRYLO0 and writes it with tlbwi
946 * or tlbwr as appropriate. This is because the index register
947 * may have the probe fail bit set as a result of a trap on a
948 * kseg2 access, i.e. without refill. Then it returns.
950 static void __init
951 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
952 struct uasm_reloc **r, unsigned int pte,
953 unsigned int tmp)
955 uasm_i_mfc0(p, tmp, C0_INDEX);
956 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
957 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
958 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
959 uasm_i_tlbwi(p); /* cp0 delay */
960 uasm_i_jr(p, tmp);
961 uasm_i_rfe(p); /* branch delay */
962 uasm_l_r3000_write_probe_fail(l, *p);
963 uasm_i_tlbwr(p); /* cp0 delay */
964 uasm_i_jr(p, tmp);
965 uasm_i_rfe(p); /* branch delay */
968 static void __init
969 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
970 unsigned int ptr)
972 long pgdc = (long)pgd_current;
974 uasm_i_mfc0(p, pte, C0_BADVADDR);
975 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
976 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
977 uasm_i_srl(p, pte, pte, 22); /* load delay */
978 uasm_i_sll(p, pte, pte, 2);
979 uasm_i_addu(p, ptr, ptr, pte);
980 uasm_i_mfc0(p, pte, C0_CONTEXT);
981 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
982 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
983 uasm_i_addu(p, ptr, ptr, pte);
984 uasm_i_lw(p, pte, 0, ptr);
985 uasm_i_tlbp(p); /* load delay */
988 static void __init build_r3000_tlb_load_handler(void)
990 u32 *p = handle_tlbl;
991 struct uasm_label *l = labels;
992 struct uasm_reloc *r = relocs;
994 memset(handle_tlbl, 0, sizeof(handle_tlbl));
995 memset(labels, 0, sizeof(labels));
996 memset(relocs, 0, sizeof(relocs));
998 build_r3000_tlbchange_handler_head(&p, K0, K1);
999 build_pte_present(&p, &l, &r, K0, K1, label_nopage_tlbl);
1000 uasm_i_nop(&p); /* load delay */
1001 build_make_valid(&p, &r, K0, K1);
1002 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1004 uasm_l_nopage_tlbl(&l, p);
1005 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1006 uasm_i_nop(&p);
1008 if ((p - handle_tlbl) > FASTPATH_SIZE)
1009 panic("TLB load handler fastpath space exceeded");
1011 uasm_resolve_relocs(relocs, labels);
1012 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1013 (unsigned int)(p - handle_tlbl));
1015 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1018 static void __init build_r3000_tlb_store_handler(void)
1020 u32 *p = handle_tlbs;
1021 struct uasm_label *l = labels;
1022 struct uasm_reloc *r = relocs;
1024 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1025 memset(labels, 0, sizeof(labels));
1026 memset(relocs, 0, sizeof(relocs));
1028 build_r3000_tlbchange_handler_head(&p, K0, K1);
1029 build_pte_writable(&p, &l, &r, K0, K1, label_nopage_tlbs);
1030 uasm_i_nop(&p); /* load delay */
1031 build_make_write(&p, &r, K0, K1);
1032 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1034 uasm_l_nopage_tlbs(&l, p);
1035 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1036 uasm_i_nop(&p);
1038 if ((p - handle_tlbs) > FASTPATH_SIZE)
1039 panic("TLB store handler fastpath space exceeded");
1041 uasm_resolve_relocs(relocs, labels);
1042 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1043 (unsigned int)(p - handle_tlbs));
1045 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1048 static void __init build_r3000_tlb_modify_handler(void)
1050 u32 *p = handle_tlbm;
1051 struct uasm_label *l = labels;
1052 struct uasm_reloc *r = relocs;
1054 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1055 memset(labels, 0, sizeof(labels));
1056 memset(relocs, 0, sizeof(relocs));
1058 build_r3000_tlbchange_handler_head(&p, K0, K1);
1059 build_pte_modifiable(&p, &l, &r, K0, K1, label_nopage_tlbm);
1060 uasm_i_nop(&p); /* load delay */
1061 build_make_write(&p, &r, K0, K1);
1062 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1064 uasm_l_nopage_tlbm(&l, p);
1065 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1066 uasm_i_nop(&p);
1068 if ((p - handle_tlbm) > FASTPATH_SIZE)
1069 panic("TLB modify handler fastpath space exceeded");
1071 uasm_resolve_relocs(relocs, labels);
1072 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1073 (unsigned int)(p - handle_tlbm));
1075 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1079 * R4000 style TLB load/store/modify handlers.
1081 static void __init
1082 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1083 struct uasm_reloc **r, unsigned int pte,
1084 unsigned int ptr)
1086 #ifdef CONFIG_64BIT
1087 build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1088 #else
1089 build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1090 #endif
1092 UASM_i_MFC0(p, pte, C0_BADVADDR);
1093 UASM_i_LW(p, ptr, 0, ptr);
1094 UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1095 uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1096 UASM_i_ADDU(p, ptr, ptr, pte);
1098 #ifdef CONFIG_SMP
1099 uasm_l_smp_pgtable_change(l, *p);
1100 #endif
1101 iPTE_LW(p, l, pte, ptr); /* get even pte */
1102 if (!m4kc_tlbp_war())
1103 build_tlb_probe_entry(p);
1106 static void __init
1107 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1108 struct uasm_reloc **r, unsigned int tmp,
1109 unsigned int ptr)
1111 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1112 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1113 build_update_entries(p, tmp, ptr);
1114 build_tlb_write_entry(p, l, r, tlb_indexed);
1115 uasm_l_leave(l, *p);
1116 uasm_i_eret(p); /* return from trap */
1118 #ifdef CONFIG_64BIT
1119 build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1120 #endif
1123 static void __init build_r4000_tlb_load_handler(void)
1125 u32 *p = handle_tlbl;
1126 struct uasm_label *l = labels;
1127 struct uasm_reloc *r = relocs;
1129 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1130 memset(labels, 0, sizeof(labels));
1131 memset(relocs, 0, sizeof(relocs));
1133 if (bcm1250_m3_war()) {
1134 UASM_i_MFC0(&p, K0, C0_BADVADDR);
1135 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1136 uasm_i_xor(&p, K0, K0, K1);
1137 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1138 uasm_il_bnez(&p, &r, K0, label_leave);
1139 /* No need for uasm_i_nop */
1142 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1143 build_pte_present(&p, &l, &r, K0, K1, label_nopage_tlbl);
1144 if (m4kc_tlbp_war())
1145 build_tlb_probe_entry(&p);
1146 build_make_valid(&p, &r, K0, K1);
1147 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1149 uasm_l_nopage_tlbl(&l, p);
1150 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1151 uasm_i_nop(&p);
1153 if ((p - handle_tlbl) > FASTPATH_SIZE)
1154 panic("TLB load handler fastpath space exceeded");
1156 uasm_resolve_relocs(relocs, labels);
1157 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1158 (unsigned int)(p - handle_tlbl));
1160 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1163 static void __init build_r4000_tlb_store_handler(void)
1165 u32 *p = handle_tlbs;
1166 struct uasm_label *l = labels;
1167 struct uasm_reloc *r = relocs;
1169 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1170 memset(labels, 0, sizeof(labels));
1171 memset(relocs, 0, sizeof(relocs));
1173 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1174 build_pte_writable(&p, &l, &r, K0, K1, label_nopage_tlbs);
1175 if (m4kc_tlbp_war())
1176 build_tlb_probe_entry(&p);
1177 build_make_write(&p, &r, K0, K1);
1178 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1180 uasm_l_nopage_tlbs(&l, p);
1181 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1182 uasm_i_nop(&p);
1184 if ((p - handle_tlbs) > FASTPATH_SIZE)
1185 panic("TLB store handler fastpath space exceeded");
1187 uasm_resolve_relocs(relocs, labels);
1188 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1189 (unsigned int)(p - handle_tlbs));
1191 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1194 static void __init build_r4000_tlb_modify_handler(void)
1196 u32 *p = handle_tlbm;
1197 struct uasm_label *l = labels;
1198 struct uasm_reloc *r = relocs;
1200 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1201 memset(labels, 0, sizeof(labels));
1202 memset(relocs, 0, sizeof(relocs));
1204 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1205 build_pte_modifiable(&p, &l, &r, K0, K1, label_nopage_tlbm);
1206 if (m4kc_tlbp_war())
1207 build_tlb_probe_entry(&p);
1208 /* Present and writable bits set, set accessed and dirty bits. */
1209 build_make_write(&p, &r, K0, K1);
1210 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1212 uasm_l_nopage_tlbm(&l, p);
1213 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1214 uasm_i_nop(&p);
1216 if ((p - handle_tlbm) > FASTPATH_SIZE)
1217 panic("TLB modify handler fastpath space exceeded");
1219 uasm_resolve_relocs(relocs, labels);
1220 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1221 (unsigned int)(p - handle_tlbm));
1223 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1226 void __init build_tlb_refill_handler(void)
1229 * The refill handler is generated per-CPU, multi-node systems
1230 * may have local storage for it. The other handlers are only
1231 * needed once.
1233 static int run_once = 0;
1235 switch (current_cpu_type()) {
1236 case CPU_R2000:
1237 case CPU_R3000:
1238 case CPU_R3000A:
1239 case CPU_R3081E:
1240 case CPU_TX3912:
1241 case CPU_TX3922:
1242 case CPU_TX3927:
1243 build_r3000_tlb_refill_handler();
1244 if (!run_once) {
1245 build_r3000_tlb_load_handler();
1246 build_r3000_tlb_store_handler();
1247 build_r3000_tlb_modify_handler();
1248 run_once++;
1250 break;
1252 case CPU_R6000:
1253 case CPU_R6000A:
1254 panic("No R6000 TLB refill handler yet");
1255 break;
1257 case CPU_R8000:
1258 panic("No R8000 TLB refill handler yet");
1259 break;
1261 default:
1262 build_r4000_tlb_refill_handler();
1263 if (!run_once) {
1264 build_r4000_tlb_load_handler();
1265 build_r4000_tlb_store_handler();
1266 build_r4000_tlb_modify_handler();
1267 run_once++;
1272 void __init flush_tlb_handlers(void)
1274 flush_icache_range((unsigned long)handle_tlbl,
1275 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1276 flush_icache_range((unsigned long)handle_tlbs,
1277 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1278 flush_icache_range((unsigned long)handle_tlbm,
1279 (unsigned long)handle_tlbm + sizeof(handle_tlbm));