xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision a09d2831)
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  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008  Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009  Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  *
13  * ... and the days got worse and worse and now you see
14  * I've gone completly out of my mind.
15  *
16  * They're coming to take me a away haha
17  * they're coming to take me a away hoho hihi haha
18  * to the funny farm where code is beautiful all the time ...
19  *
20  * (Condolences to Napoleon XIV)
21  */
22 
23 #include <linux/bug.h>
24 #include <linux/kernel.h>
25 #include <linux/types.h>
26 #include <linux/smp.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
29 
30 #include <asm/mmu_context.h>
31 #include <asm/war.h>
32 
33 #include "uasm.h"
34 
35 static inline int r45k_bvahwbug(void)
36 {
37 	/* XXX: We should probe for the presence of this bug, but we don't. */
38 	return 0;
39 }
40 
41 static inline int r4k_250MHZhwbug(void)
42 {
43 	/* XXX: We should probe for the presence of this bug, but we don't. */
44 	return 0;
45 }
46 
47 static inline int __maybe_unused bcm1250_m3_war(void)
48 {
49 	return BCM1250_M3_WAR;
50 }
51 
52 static inline int __maybe_unused r10000_llsc_war(void)
53 {
54 	return R10000_LLSC_WAR;
55 }
56 
57 /*
58  * Found by experiment: At least some revisions of the 4kc throw under
59  * some circumstances a machine check exception, triggered by invalid
60  * values in the index register.  Delaying the tlbp instruction until
61  * after the next branch,  plus adding an additional nop in front of
62  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
63  * why; it's not an issue caused by the core RTL.
64  *
65  */
66 static int __cpuinit m4kc_tlbp_war(void)
67 {
68 	return (current_cpu_data.processor_id & 0xffff00) ==
69 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
70 }
71 
72 /* Handle labels (which must be positive integers). */
73 enum label_id {
74 	label_second_part = 1,
75 	label_leave,
76 #ifdef MODULE_START
77 	label_module_alloc,
78 #endif
79 	label_vmalloc,
80 	label_vmalloc_done,
81 	label_tlbw_hazard,
82 	label_split,
83 	label_nopage_tlbl,
84 	label_nopage_tlbs,
85 	label_nopage_tlbm,
86 	label_smp_pgtable_change,
87 	label_r3000_write_probe_fail,
88 #ifdef CONFIG_HUGETLB_PAGE
89 	label_tlb_huge_update,
90 #endif
91 };
92 
93 UASM_L_LA(_second_part)
94 UASM_L_LA(_leave)
95 #ifdef MODULE_START
96 UASM_L_LA(_module_alloc)
97 #endif
98 UASM_L_LA(_vmalloc)
99 UASM_L_LA(_vmalloc_done)
100 UASM_L_LA(_tlbw_hazard)
101 UASM_L_LA(_split)
102 UASM_L_LA(_nopage_tlbl)
103 UASM_L_LA(_nopage_tlbs)
104 UASM_L_LA(_nopage_tlbm)
105 UASM_L_LA(_smp_pgtable_change)
106 UASM_L_LA(_r3000_write_probe_fail)
107 #ifdef CONFIG_HUGETLB_PAGE
108 UASM_L_LA(_tlb_huge_update)
109 #endif
110 
111 /*
112  * For debug purposes.
113  */
114 static inline void dump_handler(const u32 *handler, int count)
115 {
116 	int i;
117 
118 	pr_debug("\t.set push\n");
119 	pr_debug("\t.set noreorder\n");
120 
121 	for (i = 0; i < count; i++)
122 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
123 
124 	pr_debug("\t.set pop\n");
125 }
126 
127 /* The only general purpose registers allowed in TLB handlers. */
128 #define K0		26
129 #define K1		27
130 
131 /* Some CP0 registers */
132 #define C0_INDEX	0, 0
133 #define C0_ENTRYLO0	2, 0
134 #define C0_TCBIND	2, 2
135 #define C0_ENTRYLO1	3, 0
136 #define C0_CONTEXT	4, 0
137 #define C0_PAGEMASK	5, 0
138 #define C0_BADVADDR	8, 0
139 #define C0_ENTRYHI	10, 0
140 #define C0_EPC		14, 0
141 #define C0_XCONTEXT	20, 0
142 
143 #ifdef CONFIG_64BIT
144 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
145 #else
146 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
147 #endif
148 
149 /* The worst case length of the handler is around 18 instructions for
150  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
151  * Maximum space available is 32 instructions for R3000 and 64
152  * instructions for R4000.
153  *
154  * We deliberately chose a buffer size of 128, so we won't scribble
155  * over anything important on overflow before we panic.
156  */
157 static u32 tlb_handler[128] __cpuinitdata;
158 
159 /* simply assume worst case size for labels and relocs */
160 static struct uasm_label labels[128] __cpuinitdata;
161 static struct uasm_reloc relocs[128] __cpuinitdata;
162 
163 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
164 /*
165  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
166  * we cannot do r3000 under these circumstances.
167  */
168 
169 /*
170  * The R3000 TLB handler is simple.
171  */
172 static void __cpuinit build_r3000_tlb_refill_handler(void)
173 {
174 	long pgdc = (long)pgd_current;
175 	u32 *p;
176 
177 	memset(tlb_handler, 0, sizeof(tlb_handler));
178 	p = tlb_handler;
179 
180 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
181 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
182 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
183 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
184 	uasm_i_sll(&p, K0, K0, 2);
185 	uasm_i_addu(&p, K1, K1, K0);
186 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
187 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
188 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
189 	uasm_i_addu(&p, K1, K1, K0);
190 	uasm_i_lw(&p, K0, 0, K1);
191 	uasm_i_nop(&p); /* load delay */
192 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
193 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
194 	uasm_i_tlbwr(&p); /* cp0 delay */
195 	uasm_i_jr(&p, K1);
196 	uasm_i_rfe(&p); /* branch delay */
197 
198 	if (p > tlb_handler + 32)
199 		panic("TLB refill handler space exceeded");
200 
201 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
202 		 (unsigned int)(p - tlb_handler));
203 
204 	memcpy((void *)ebase, tlb_handler, 0x80);
205 
206 	dump_handler((u32 *)ebase, 32);
207 }
208 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
209 
210 /*
211  * The R4000 TLB handler is much more complicated. We have two
212  * consecutive handler areas with 32 instructions space each.
213  * Since they aren't used at the same time, we can overflow in the
214  * other one.To keep things simple, we first assume linear space,
215  * then we relocate it to the final handler layout as needed.
216  */
217 static u32 final_handler[64] __cpuinitdata;
218 
219 /*
220  * Hazards
221  *
222  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
223  * 2. A timing hazard exists for the TLBP instruction.
224  *
225  *      stalling_instruction
226  *      TLBP
227  *
228  * The JTLB is being read for the TLBP throughout the stall generated by the
229  * previous instruction. This is not really correct as the stalling instruction
230  * can modify the address used to access the JTLB.  The failure symptom is that
231  * the TLBP instruction will use an address created for the stalling instruction
232  * and not the address held in C0_ENHI and thus report the wrong results.
233  *
234  * The software work-around is to not allow the instruction preceding the TLBP
235  * to stall - make it an NOP or some other instruction guaranteed not to stall.
236  *
237  * Errata 2 will not be fixed.  This errata is also on the R5000.
238  *
239  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
240  */
241 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
242 {
243 	switch (current_cpu_type()) {
244 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
245 	case CPU_R4600:
246 	case CPU_R4700:
247 	case CPU_R5000:
248 	case CPU_R5000A:
249 	case CPU_NEVADA:
250 		uasm_i_nop(p);
251 		uasm_i_tlbp(p);
252 		break;
253 
254 	default:
255 		uasm_i_tlbp(p);
256 		break;
257 	}
258 }
259 
260 /*
261  * Write random or indexed TLB entry, and care about the hazards from
262  * the preceeding mtc0 and for the following eret.
263  */
264 enum tlb_write_entry { tlb_random, tlb_indexed };
265 
266 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
267 					 struct uasm_reloc **r,
268 					 enum tlb_write_entry wmode)
269 {
270 	void(*tlbw)(u32 **) = NULL;
271 
272 	switch (wmode) {
273 	case tlb_random: tlbw = uasm_i_tlbwr; break;
274 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
275 	}
276 
277 	if (cpu_has_mips_r2) {
278 		if (cpu_has_mips_r2_exec_hazard)
279 			uasm_i_ehb(p);
280 		tlbw(p);
281 		return;
282 	}
283 
284 	switch (current_cpu_type()) {
285 	case CPU_R4000PC:
286 	case CPU_R4000SC:
287 	case CPU_R4000MC:
288 	case CPU_R4400PC:
289 	case CPU_R4400SC:
290 	case CPU_R4400MC:
291 		/*
292 		 * This branch uses up a mtc0 hazard nop slot and saves
293 		 * two nops after the tlbw instruction.
294 		 */
295 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
296 		tlbw(p);
297 		uasm_l_tlbw_hazard(l, *p);
298 		uasm_i_nop(p);
299 		break;
300 
301 	case CPU_R4600:
302 	case CPU_R4700:
303 	case CPU_R5000:
304 	case CPU_R5000A:
305 		uasm_i_nop(p);
306 		tlbw(p);
307 		uasm_i_nop(p);
308 		break;
309 
310 	case CPU_R4300:
311 	case CPU_5KC:
312 	case CPU_TX49XX:
313 	case CPU_PR4450:
314 		uasm_i_nop(p);
315 		tlbw(p);
316 		break;
317 
318 	case CPU_R10000:
319 	case CPU_R12000:
320 	case CPU_R14000:
321 	case CPU_4KC:
322 	case CPU_4KEC:
323 	case CPU_SB1:
324 	case CPU_SB1A:
325 	case CPU_4KSC:
326 	case CPU_20KC:
327 	case CPU_25KF:
328 	case CPU_BCM3302:
329 	case CPU_BCM4710:
330 	case CPU_LOONGSON2:
331 	case CPU_BCM6338:
332 	case CPU_BCM6345:
333 	case CPU_BCM6348:
334 	case CPU_BCM6358:
335 	case CPU_R5500:
336 		if (m4kc_tlbp_war())
337 			uasm_i_nop(p);
338 	case CPU_ALCHEMY:
339 		tlbw(p);
340 		break;
341 
342 	case CPU_NEVADA:
343 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
344 		/*
345 		 * This branch uses up a mtc0 hazard nop slot and saves
346 		 * a nop after the tlbw instruction.
347 		 */
348 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
349 		tlbw(p);
350 		uasm_l_tlbw_hazard(l, *p);
351 		break;
352 
353 	case CPU_RM7000:
354 		uasm_i_nop(p);
355 		uasm_i_nop(p);
356 		uasm_i_nop(p);
357 		uasm_i_nop(p);
358 		tlbw(p);
359 		break;
360 
361 	case CPU_RM9000:
362 		/*
363 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
364 		 * use of the JTLB for instructions should not occur for 4
365 		 * cpu cycles and use for data translations should not occur
366 		 * for 3 cpu cycles.
367 		 */
368 		uasm_i_ssnop(p);
369 		uasm_i_ssnop(p);
370 		uasm_i_ssnop(p);
371 		uasm_i_ssnop(p);
372 		tlbw(p);
373 		uasm_i_ssnop(p);
374 		uasm_i_ssnop(p);
375 		uasm_i_ssnop(p);
376 		uasm_i_ssnop(p);
377 		break;
378 
379 	case CPU_VR4111:
380 	case CPU_VR4121:
381 	case CPU_VR4122:
382 	case CPU_VR4181:
383 	case CPU_VR4181A:
384 		uasm_i_nop(p);
385 		uasm_i_nop(p);
386 		tlbw(p);
387 		uasm_i_nop(p);
388 		uasm_i_nop(p);
389 		break;
390 
391 	case CPU_VR4131:
392 	case CPU_VR4133:
393 	case CPU_R5432:
394 		uasm_i_nop(p);
395 		uasm_i_nop(p);
396 		tlbw(p);
397 		break;
398 
399 	default:
400 		panic("No TLB refill handler yet (CPU type: %d)",
401 		      current_cpu_data.cputype);
402 		break;
403 	}
404 }
405 
406 #ifdef CONFIG_HUGETLB_PAGE
407 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
408 						 struct uasm_label **l,
409 						 struct uasm_reloc **r,
410 						 unsigned int tmp,
411 						 enum tlb_write_entry wmode)
412 {
413 	/* Set huge page tlb entry size */
414 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
415 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
416 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
417 
418 	build_tlb_write_entry(p, l, r, wmode);
419 
420 	/* Reset default page size */
421 	if (PM_DEFAULT_MASK >> 16) {
422 		uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
423 		uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
424 		uasm_il_b(p, r, label_leave);
425 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
426 	} else if (PM_DEFAULT_MASK) {
427 		uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
428 		uasm_il_b(p, r, label_leave);
429 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
430 	} else {
431 		uasm_il_b(p, r, label_leave);
432 		uasm_i_mtc0(p, 0, C0_PAGEMASK);
433 	}
434 }
435 
436 /*
437  * Check if Huge PTE is present, if so then jump to LABEL.
438  */
439 static void __cpuinit
440 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
441 		unsigned int pmd, int lid)
442 {
443 	UASM_i_LW(p, tmp, 0, pmd);
444 	uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
445 	uasm_il_bnez(p, r, tmp, lid);
446 }
447 
448 static __cpuinit void build_huge_update_entries(u32 **p,
449 						unsigned int pte,
450 						unsigned int tmp)
451 {
452 	int small_sequence;
453 
454 	/*
455 	 * A huge PTE describes an area the size of the
456 	 * configured huge page size. This is twice the
457 	 * of the large TLB entry size we intend to use.
458 	 * A TLB entry half the size of the configured
459 	 * huge page size is configured into entrylo0
460 	 * and entrylo1 to cover the contiguous huge PTE
461 	 * address space.
462 	 */
463 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
464 
465 	/* We can clobber tmp.  It isn't used after this.*/
466 	if (!small_sequence)
467 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
468 
469 	UASM_i_SRL(p, pte, pte, 6); /* convert to entrylo */
470 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* load it */
471 	/* convert to entrylo1 */
472 	if (small_sequence)
473 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
474 	else
475 		UASM_i_ADDU(p, pte, pte, tmp);
476 
477 	uasm_i_mtc0(p, pte, C0_ENTRYLO1); /* load it */
478 }
479 
480 static __cpuinit void build_huge_handler_tail(u32 **p,
481 					      struct uasm_reloc **r,
482 					      struct uasm_label **l,
483 					      unsigned int pte,
484 					      unsigned int ptr)
485 {
486 #ifdef CONFIG_SMP
487 	UASM_i_SC(p, pte, 0, ptr);
488 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
489 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
490 #else
491 	UASM_i_SW(p, pte, 0, ptr);
492 #endif
493 	build_huge_update_entries(p, pte, ptr);
494 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed);
495 }
496 #endif /* CONFIG_HUGETLB_PAGE */
497 
498 #ifdef CONFIG_64BIT
499 /*
500  * TMP and PTR are scratch.
501  * TMP will be clobbered, PTR will hold the pmd entry.
502  */
503 static void __cpuinit
504 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
505 		 unsigned int tmp, unsigned int ptr)
506 {
507 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
508 	long pgdc = (long)pgd_current;
509 #endif
510 	/*
511 	 * The vmalloc handling is not in the hotpath.
512 	 */
513 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
514 	uasm_il_bltz(p, r, tmp, label_vmalloc);
515 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
516 
517 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
518 	/*
519 	 * &pgd << 11 stored in CONTEXT [23..63].
520 	 */
521 	UASM_i_MFC0(p, ptr, C0_CONTEXT);
522 	uasm_i_dins(p, ptr, 0, 0, 23); /* Clear lower 23 bits of context. */
523 	uasm_i_ori(p, ptr, ptr, 0x540); /* 1 0  1 0 1  << 6  xkphys cached */
524 	uasm_i_drotr(p, ptr, ptr, 11);
525 #elif defined(CONFIG_SMP)
526 # ifdef  CONFIG_MIPS_MT_SMTC
527 	/*
528 	 * SMTC uses TCBind value as "CPU" index
529 	 */
530 	uasm_i_mfc0(p, ptr, C0_TCBIND);
531 	uasm_i_dsrl(p, ptr, ptr, 19);
532 # else
533 	/*
534 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
535 	 * stored in CONTEXT.
536 	 */
537 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
538 	uasm_i_dsrl(p, ptr, ptr, 23);
539 # endif
540 	UASM_i_LA_mostly(p, tmp, pgdc);
541 	uasm_i_daddu(p, ptr, ptr, tmp);
542 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
543 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
544 #else
545 	UASM_i_LA_mostly(p, ptr, pgdc);
546 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
547 #endif
548 
549 	uasm_l_vmalloc_done(l, *p);
550 
551 	if (PGDIR_SHIFT - 3 < 32)		/* get pgd offset in bytes */
552 		uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
553 	else
554 		uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
555 
556 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
557 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
558 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
559 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
560 	uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
561 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
562 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
563 }
564 
565 /*
566  * BVADDR is the faulting address, PTR is scratch.
567  * PTR will hold the pgd for vmalloc.
568  */
569 static void __cpuinit
570 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
571 			unsigned int bvaddr, unsigned int ptr)
572 {
573 	long swpd = (long)swapper_pg_dir;
574 
575 	uasm_l_vmalloc(l, *p);
576 
577 	if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
578 		uasm_il_b(p, r, label_vmalloc_done);
579 		uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
580 	} else {
581 		UASM_i_LA_mostly(p, ptr, swpd);
582 		uasm_il_b(p, r, label_vmalloc_done);
583 		if (uasm_in_compat_space_p(swpd))
584 			uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
585 		else
586 			uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
587 	}
588 }
589 
590 #else /* !CONFIG_64BIT */
591 
592 /*
593  * TMP and PTR are scratch.
594  * TMP will be clobbered, PTR will hold the pgd entry.
595  */
596 static void __cpuinit __maybe_unused
597 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
598 {
599 	long pgdc = (long)pgd_current;
600 
601 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
602 #ifdef CONFIG_SMP
603 #ifdef  CONFIG_MIPS_MT_SMTC
604 	/*
605 	 * SMTC uses TCBind value as "CPU" index
606 	 */
607 	uasm_i_mfc0(p, ptr, C0_TCBIND);
608 	UASM_i_LA_mostly(p, tmp, pgdc);
609 	uasm_i_srl(p, ptr, ptr, 19);
610 #else
611 	/*
612 	 * smp_processor_id() << 3 is stored in CONTEXT.
613          */
614 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
615 	UASM_i_LA_mostly(p, tmp, pgdc);
616 	uasm_i_srl(p, ptr, ptr, 23);
617 #endif
618 	uasm_i_addu(p, ptr, tmp, ptr);
619 #else
620 	UASM_i_LA_mostly(p, ptr, pgdc);
621 #endif
622 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
623 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
624 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
625 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
626 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
627 }
628 
629 #endif /* !CONFIG_64BIT */
630 
631 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
632 {
633 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
634 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
635 
636 	switch (current_cpu_type()) {
637 	case CPU_VR41XX:
638 	case CPU_VR4111:
639 	case CPU_VR4121:
640 	case CPU_VR4122:
641 	case CPU_VR4131:
642 	case CPU_VR4181:
643 	case CPU_VR4181A:
644 	case CPU_VR4133:
645 		shift += 2;
646 		break;
647 
648 	default:
649 		break;
650 	}
651 
652 	if (shift)
653 		UASM_i_SRL(p, ctx, ctx, shift);
654 	uasm_i_andi(p, ctx, ctx, mask);
655 }
656 
657 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
658 {
659 	/*
660 	 * Bug workaround for the Nevada. It seems as if under certain
661 	 * circumstances the move from cp0_context might produce a
662 	 * bogus result when the mfc0 instruction and its consumer are
663 	 * in a different cacheline or a load instruction, probably any
664 	 * memory reference, is between them.
665 	 */
666 	switch (current_cpu_type()) {
667 	case CPU_NEVADA:
668 		UASM_i_LW(p, ptr, 0, ptr);
669 		GET_CONTEXT(p, tmp); /* get context reg */
670 		break;
671 
672 	default:
673 		GET_CONTEXT(p, tmp); /* get context reg */
674 		UASM_i_LW(p, ptr, 0, ptr);
675 		break;
676 	}
677 
678 	build_adjust_context(p, tmp);
679 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
680 }
681 
682 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
683 					unsigned int ptep)
684 {
685 	/*
686 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
687 	 * Kernel is a special case. Only a few CPUs use it.
688 	 */
689 #ifdef CONFIG_64BIT_PHYS_ADDR
690 	if (cpu_has_64bits) {
691 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
692 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
693 		uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
694 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
695 		uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
696 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
697 	} else {
698 		int pte_off_even = sizeof(pte_t) / 2;
699 		int pte_off_odd = pte_off_even + sizeof(pte_t);
700 
701 		/* The pte entries are pre-shifted */
702 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
703 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
704 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
705 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
706 	}
707 #else
708 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
709 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
710 	if (r45k_bvahwbug())
711 		build_tlb_probe_entry(p);
712 	UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
713 	if (r4k_250MHZhwbug())
714 		uasm_i_mtc0(p, 0, C0_ENTRYLO0);
715 	uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
716 	UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
717 	if (r45k_bvahwbug())
718 		uasm_i_mfc0(p, tmp, C0_INDEX);
719 	if (r4k_250MHZhwbug())
720 		uasm_i_mtc0(p, 0, C0_ENTRYLO1);
721 	uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
722 #endif
723 }
724 
725 /*
726  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
727  * because EXL == 0.  If we wrap, we can also use the 32 instruction
728  * slots before the XTLB refill exception handler which belong to the
729  * unused TLB refill exception.
730  */
731 #define MIPS64_REFILL_INSNS 32
732 
733 static void __cpuinit build_r4000_tlb_refill_handler(void)
734 {
735 	u32 *p = tlb_handler;
736 	struct uasm_label *l = labels;
737 	struct uasm_reloc *r = relocs;
738 	u32 *f;
739 	unsigned int final_len;
740 
741 	memset(tlb_handler, 0, sizeof(tlb_handler));
742 	memset(labels, 0, sizeof(labels));
743 	memset(relocs, 0, sizeof(relocs));
744 	memset(final_handler, 0, sizeof(final_handler));
745 
746 	/*
747 	 * create the plain linear handler
748 	 */
749 	if (bcm1250_m3_war()) {
750 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
751 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
752 		uasm_i_xor(&p, K0, K0, K1);
753 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
754 		uasm_il_bnez(&p, &r, K0, label_leave);
755 		/* No need for uasm_i_nop */
756 	}
757 
758 #ifdef CONFIG_64BIT
759 	build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
760 #else
761 	build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
762 #endif
763 
764 #ifdef CONFIG_HUGETLB_PAGE
765 	build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
766 #endif
767 
768 	build_get_ptep(&p, K0, K1);
769 	build_update_entries(&p, K0, K1);
770 	build_tlb_write_entry(&p, &l, &r, tlb_random);
771 	uasm_l_leave(&l, p);
772 	uasm_i_eret(&p); /* return from trap */
773 
774 #ifdef CONFIG_HUGETLB_PAGE
775 	uasm_l_tlb_huge_update(&l, p);
776 	UASM_i_LW(&p, K0, 0, K1);
777 	build_huge_update_entries(&p, K0, K1);
778 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random);
779 #endif
780 
781 #ifdef CONFIG_64BIT
782 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
783 #endif
784 
785 	/*
786 	 * Overflow check: For the 64bit handler, we need at least one
787 	 * free instruction slot for the wrap-around branch. In worst
788 	 * case, if the intended insertion point is a delay slot, we
789 	 * need three, with the second nop'ed and the third being
790 	 * unused.
791 	 */
792 	/* Loongson2 ebase is different than r4k, we have more space */
793 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
794 	if ((p - tlb_handler) > 64)
795 		panic("TLB refill handler space exceeded");
796 #else
797 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
798 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
799 		&& uasm_insn_has_bdelay(relocs,
800 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
801 		panic("TLB refill handler space exceeded");
802 #endif
803 
804 	/*
805 	 * Now fold the handler in the TLB refill handler space.
806 	 */
807 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
808 	f = final_handler;
809 	/* Simplest case, just copy the handler. */
810 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
811 	final_len = p - tlb_handler;
812 #else /* CONFIG_64BIT */
813 	f = final_handler + MIPS64_REFILL_INSNS;
814 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
815 		/* Just copy the handler. */
816 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
817 		final_len = p - tlb_handler;
818 	} else {
819 #if defined(CONFIG_HUGETLB_PAGE)
820 		const enum label_id ls = label_tlb_huge_update;
821 #elif defined(MODULE_START)
822 		const enum label_id ls = label_module_alloc;
823 #else
824 		const enum label_id ls = label_vmalloc;
825 #endif
826 		u32 *split;
827 		int ov = 0;
828 		int i;
829 
830 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
831 			;
832 		BUG_ON(i == ARRAY_SIZE(labels));
833 		split = labels[i].addr;
834 
835 		/*
836 		 * See if we have overflown one way or the other.
837 		 */
838 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
839 		    split < p - MIPS64_REFILL_INSNS)
840 			ov = 1;
841 
842 		if (ov) {
843 			/*
844 			 * Split two instructions before the end.  One
845 			 * for the branch and one for the instruction
846 			 * in the delay slot.
847 			 */
848 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
849 
850 			/*
851 			 * If the branch would fall in a delay slot,
852 			 * we must back up an additional instruction
853 			 * so that it is no longer in a delay slot.
854 			 */
855 			if (uasm_insn_has_bdelay(relocs, split - 1))
856 				split--;
857 		}
858 		/* Copy first part of the handler. */
859 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
860 		f += split - tlb_handler;
861 
862 		if (ov) {
863 			/* Insert branch. */
864 			uasm_l_split(&l, final_handler);
865 			uasm_il_b(&f, &r, label_split);
866 			if (uasm_insn_has_bdelay(relocs, split))
867 				uasm_i_nop(&f);
868 			else {
869 				uasm_copy_handler(relocs, labels,
870 						  split, split + 1, f);
871 				uasm_move_labels(labels, f, f + 1, -1);
872 				f++;
873 				split++;
874 			}
875 		}
876 
877 		/* Copy the rest of the handler. */
878 		uasm_copy_handler(relocs, labels, split, p, final_handler);
879 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
880 			    (p - split);
881 	}
882 #endif /* CONFIG_64BIT */
883 
884 	uasm_resolve_relocs(relocs, labels);
885 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
886 		 final_len);
887 
888 	memcpy((void *)ebase, final_handler, 0x100);
889 
890 	dump_handler((u32 *)ebase, 64);
891 }
892 
893 /*
894  * TLB load/store/modify handlers.
895  *
896  * Only the fastpath gets synthesized at runtime, the slowpath for
897  * do_page_fault remains normal asm.
898  */
899 extern void tlb_do_page_fault_0(void);
900 extern void tlb_do_page_fault_1(void);
901 
902 /*
903  * 128 instructions for the fastpath handler is generous and should
904  * never be exceeded.
905  */
906 #define FASTPATH_SIZE 128
907 
908 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
909 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
910 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
911 
912 static void __cpuinit
913 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
914 {
915 #ifdef CONFIG_SMP
916 # ifdef CONFIG_64BIT_PHYS_ADDR
917 	if (cpu_has_64bits)
918 		uasm_i_lld(p, pte, 0, ptr);
919 	else
920 # endif
921 		UASM_i_LL(p, pte, 0, ptr);
922 #else
923 # ifdef CONFIG_64BIT_PHYS_ADDR
924 	if (cpu_has_64bits)
925 		uasm_i_ld(p, pte, 0, ptr);
926 	else
927 # endif
928 		UASM_i_LW(p, pte, 0, ptr);
929 #endif
930 }
931 
932 static void __cpuinit
933 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
934 	unsigned int mode)
935 {
936 #ifdef CONFIG_64BIT_PHYS_ADDR
937 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
938 #endif
939 
940 	uasm_i_ori(p, pte, pte, mode);
941 #ifdef CONFIG_SMP
942 # ifdef CONFIG_64BIT_PHYS_ADDR
943 	if (cpu_has_64bits)
944 		uasm_i_scd(p, pte, 0, ptr);
945 	else
946 # endif
947 		UASM_i_SC(p, pte, 0, ptr);
948 
949 	if (r10000_llsc_war())
950 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
951 	else
952 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
953 
954 # ifdef CONFIG_64BIT_PHYS_ADDR
955 	if (!cpu_has_64bits) {
956 		/* no uasm_i_nop needed */
957 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
958 		uasm_i_ori(p, pte, pte, hwmode);
959 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
960 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
961 		/* no uasm_i_nop needed */
962 		uasm_i_lw(p, pte, 0, ptr);
963 	} else
964 		uasm_i_nop(p);
965 # else
966 	uasm_i_nop(p);
967 # endif
968 #else
969 # ifdef CONFIG_64BIT_PHYS_ADDR
970 	if (cpu_has_64bits)
971 		uasm_i_sd(p, pte, 0, ptr);
972 	else
973 # endif
974 		UASM_i_SW(p, pte, 0, ptr);
975 
976 # ifdef CONFIG_64BIT_PHYS_ADDR
977 	if (!cpu_has_64bits) {
978 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
979 		uasm_i_ori(p, pte, pte, hwmode);
980 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
981 		uasm_i_lw(p, pte, 0, ptr);
982 	}
983 # endif
984 #endif
985 }
986 
987 /*
988  * Check if PTE is present, if not then jump to LABEL. PTR points to
989  * the page table where this PTE is located, PTE will be re-loaded
990  * with it's original value.
991  */
992 static void __cpuinit
993 build_pte_present(u32 **p, struct uasm_reloc **r,
994 		  unsigned int pte, unsigned int ptr, enum label_id lid)
995 {
996 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
997 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
998 	uasm_il_bnez(p, r, pte, lid);
999 	iPTE_LW(p, pte, ptr);
1000 }
1001 
1002 /* Make PTE valid, store result in PTR. */
1003 static void __cpuinit
1004 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1005 		 unsigned int ptr)
1006 {
1007 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1008 
1009 	iPTE_SW(p, r, pte, ptr, mode);
1010 }
1011 
1012 /*
1013  * Check if PTE can be written to, if not branch to LABEL. Regardless
1014  * restore PTE with value from PTR when done.
1015  */
1016 static void __cpuinit
1017 build_pte_writable(u32 **p, struct uasm_reloc **r,
1018 		   unsigned int pte, unsigned int ptr, enum label_id lid)
1019 {
1020 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1021 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1022 	uasm_il_bnez(p, r, pte, lid);
1023 	iPTE_LW(p, pte, ptr);
1024 }
1025 
1026 /* Make PTE writable, update software status bits as well, then store
1027  * at PTR.
1028  */
1029 static void __cpuinit
1030 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1031 		 unsigned int ptr)
1032 {
1033 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1034 			     | _PAGE_DIRTY);
1035 
1036 	iPTE_SW(p, r, pte, ptr, mode);
1037 }
1038 
1039 /*
1040  * Check if PTE can be modified, if not branch to LABEL. Regardless
1041  * restore PTE with value from PTR when done.
1042  */
1043 static void __cpuinit
1044 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1045 		     unsigned int pte, unsigned int ptr, enum label_id lid)
1046 {
1047 	uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1048 	uasm_il_beqz(p, r, pte, lid);
1049 	iPTE_LW(p, pte, ptr);
1050 }
1051 
1052 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1053 /*
1054  * R3000 style TLB load/store/modify handlers.
1055  */
1056 
1057 /*
1058  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1059  * Then it returns.
1060  */
1061 static void __cpuinit
1062 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1063 {
1064 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1065 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1066 	uasm_i_tlbwi(p);
1067 	uasm_i_jr(p, tmp);
1068 	uasm_i_rfe(p); /* branch delay */
1069 }
1070 
1071 /*
1072  * This places the pte into ENTRYLO0 and writes it with tlbwi
1073  * or tlbwr as appropriate.  This is because the index register
1074  * may have the probe fail bit set as a result of a trap on a
1075  * kseg2 access, i.e. without refill.  Then it returns.
1076  */
1077 static void __cpuinit
1078 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1079 			     struct uasm_reloc **r, unsigned int pte,
1080 			     unsigned int tmp)
1081 {
1082 	uasm_i_mfc0(p, tmp, C0_INDEX);
1083 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1084 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1085 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1086 	uasm_i_tlbwi(p); /* cp0 delay */
1087 	uasm_i_jr(p, tmp);
1088 	uasm_i_rfe(p); /* branch delay */
1089 	uasm_l_r3000_write_probe_fail(l, *p);
1090 	uasm_i_tlbwr(p); /* cp0 delay */
1091 	uasm_i_jr(p, tmp);
1092 	uasm_i_rfe(p); /* branch delay */
1093 }
1094 
1095 static void __cpuinit
1096 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1097 				   unsigned int ptr)
1098 {
1099 	long pgdc = (long)pgd_current;
1100 
1101 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1102 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1103 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1104 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1105 	uasm_i_sll(p, pte, pte, 2);
1106 	uasm_i_addu(p, ptr, ptr, pte);
1107 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1108 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1109 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1110 	uasm_i_addu(p, ptr, ptr, pte);
1111 	uasm_i_lw(p, pte, 0, ptr);
1112 	uasm_i_tlbp(p); /* load delay */
1113 }
1114 
1115 static void __cpuinit build_r3000_tlb_load_handler(void)
1116 {
1117 	u32 *p = handle_tlbl;
1118 	struct uasm_label *l = labels;
1119 	struct uasm_reloc *r = relocs;
1120 
1121 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1122 	memset(labels, 0, sizeof(labels));
1123 	memset(relocs, 0, sizeof(relocs));
1124 
1125 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1126 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1127 	uasm_i_nop(&p); /* load delay */
1128 	build_make_valid(&p, &r, K0, K1);
1129 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1130 
1131 	uasm_l_nopage_tlbl(&l, p);
1132 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1133 	uasm_i_nop(&p);
1134 
1135 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1136 		panic("TLB load handler fastpath space exceeded");
1137 
1138 	uasm_resolve_relocs(relocs, labels);
1139 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1140 		 (unsigned int)(p - handle_tlbl));
1141 
1142 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1143 }
1144 
1145 static void __cpuinit build_r3000_tlb_store_handler(void)
1146 {
1147 	u32 *p = handle_tlbs;
1148 	struct uasm_label *l = labels;
1149 	struct uasm_reloc *r = relocs;
1150 
1151 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1152 	memset(labels, 0, sizeof(labels));
1153 	memset(relocs, 0, sizeof(relocs));
1154 
1155 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1156 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1157 	uasm_i_nop(&p); /* load delay */
1158 	build_make_write(&p, &r, K0, K1);
1159 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1160 
1161 	uasm_l_nopage_tlbs(&l, p);
1162 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1163 	uasm_i_nop(&p);
1164 
1165 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1166 		panic("TLB store handler fastpath space exceeded");
1167 
1168 	uasm_resolve_relocs(relocs, labels);
1169 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1170 		 (unsigned int)(p - handle_tlbs));
1171 
1172 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1173 }
1174 
1175 static void __cpuinit build_r3000_tlb_modify_handler(void)
1176 {
1177 	u32 *p = handle_tlbm;
1178 	struct uasm_label *l = labels;
1179 	struct uasm_reloc *r = relocs;
1180 
1181 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1182 	memset(labels, 0, sizeof(labels));
1183 	memset(relocs, 0, sizeof(relocs));
1184 
1185 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1186 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1187 	uasm_i_nop(&p); /* load delay */
1188 	build_make_write(&p, &r, K0, K1);
1189 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1190 
1191 	uasm_l_nopage_tlbm(&l, p);
1192 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1193 	uasm_i_nop(&p);
1194 
1195 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1196 		panic("TLB modify handler fastpath space exceeded");
1197 
1198 	uasm_resolve_relocs(relocs, labels);
1199 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1200 		 (unsigned int)(p - handle_tlbm));
1201 
1202 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1203 }
1204 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1205 
1206 /*
1207  * R4000 style TLB load/store/modify handlers.
1208  */
1209 static void __cpuinit
1210 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1211 				   struct uasm_reloc **r, unsigned int pte,
1212 				   unsigned int ptr)
1213 {
1214 #ifdef CONFIG_64BIT
1215 	build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1216 #else
1217 	build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1218 #endif
1219 
1220 #ifdef CONFIG_HUGETLB_PAGE
1221 	/*
1222 	 * For huge tlb entries, pmd doesn't contain an address but
1223 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1224 	 * see if we need to jump to huge tlb processing.
1225 	 */
1226 	build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1227 #endif
1228 
1229 	UASM_i_MFC0(p, pte, C0_BADVADDR);
1230 	UASM_i_LW(p, ptr, 0, ptr);
1231 	UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1232 	uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1233 	UASM_i_ADDU(p, ptr, ptr, pte);
1234 
1235 #ifdef CONFIG_SMP
1236 	uasm_l_smp_pgtable_change(l, *p);
1237 #endif
1238 	iPTE_LW(p, pte, ptr); /* get even pte */
1239 	if (!m4kc_tlbp_war())
1240 		build_tlb_probe_entry(p);
1241 }
1242 
1243 static void __cpuinit
1244 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1245 				   struct uasm_reloc **r, unsigned int tmp,
1246 				   unsigned int ptr)
1247 {
1248 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1249 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1250 	build_update_entries(p, tmp, ptr);
1251 	build_tlb_write_entry(p, l, r, tlb_indexed);
1252 	uasm_l_leave(l, *p);
1253 	uasm_i_eret(p); /* return from trap */
1254 
1255 #ifdef CONFIG_64BIT
1256 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1257 #endif
1258 }
1259 
1260 static void __cpuinit build_r4000_tlb_load_handler(void)
1261 {
1262 	u32 *p = handle_tlbl;
1263 	struct uasm_label *l = labels;
1264 	struct uasm_reloc *r = relocs;
1265 
1266 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1267 	memset(labels, 0, sizeof(labels));
1268 	memset(relocs, 0, sizeof(relocs));
1269 
1270 	if (bcm1250_m3_war()) {
1271 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
1272 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1273 		uasm_i_xor(&p, K0, K0, K1);
1274 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1275 		uasm_il_bnez(&p, &r, K0, label_leave);
1276 		/* No need for uasm_i_nop */
1277 	}
1278 
1279 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1280 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1281 	if (m4kc_tlbp_war())
1282 		build_tlb_probe_entry(&p);
1283 	build_make_valid(&p, &r, K0, K1);
1284 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1285 
1286 #ifdef CONFIG_HUGETLB_PAGE
1287 	/*
1288 	 * This is the entry point when build_r4000_tlbchange_handler_head
1289 	 * spots a huge page.
1290 	 */
1291 	uasm_l_tlb_huge_update(&l, p);
1292 	iPTE_LW(&p, K0, K1);
1293 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1294 	build_tlb_probe_entry(&p);
1295 	uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1296 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1297 #endif
1298 
1299 	uasm_l_nopage_tlbl(&l, p);
1300 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1301 	uasm_i_nop(&p);
1302 
1303 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1304 		panic("TLB load handler fastpath space exceeded");
1305 
1306 	uasm_resolve_relocs(relocs, labels);
1307 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1308 		 (unsigned int)(p - handle_tlbl));
1309 
1310 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1311 }
1312 
1313 static void __cpuinit build_r4000_tlb_store_handler(void)
1314 {
1315 	u32 *p = handle_tlbs;
1316 	struct uasm_label *l = labels;
1317 	struct uasm_reloc *r = relocs;
1318 
1319 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1320 	memset(labels, 0, sizeof(labels));
1321 	memset(relocs, 0, sizeof(relocs));
1322 
1323 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1324 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1325 	if (m4kc_tlbp_war())
1326 		build_tlb_probe_entry(&p);
1327 	build_make_write(&p, &r, K0, K1);
1328 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1329 
1330 #ifdef CONFIG_HUGETLB_PAGE
1331 	/*
1332 	 * This is the entry point when
1333 	 * build_r4000_tlbchange_handler_head spots a huge page.
1334 	 */
1335 	uasm_l_tlb_huge_update(&l, p);
1336 	iPTE_LW(&p, K0, K1);
1337 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1338 	build_tlb_probe_entry(&p);
1339 	uasm_i_ori(&p, K0, K0,
1340 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1341 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1342 #endif
1343 
1344 	uasm_l_nopage_tlbs(&l, p);
1345 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1346 	uasm_i_nop(&p);
1347 
1348 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1349 		panic("TLB store handler fastpath space exceeded");
1350 
1351 	uasm_resolve_relocs(relocs, labels);
1352 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1353 		 (unsigned int)(p - handle_tlbs));
1354 
1355 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1356 }
1357 
1358 static void __cpuinit build_r4000_tlb_modify_handler(void)
1359 {
1360 	u32 *p = handle_tlbm;
1361 	struct uasm_label *l = labels;
1362 	struct uasm_reloc *r = relocs;
1363 
1364 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1365 	memset(labels, 0, sizeof(labels));
1366 	memset(relocs, 0, sizeof(relocs));
1367 
1368 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1369 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1370 	if (m4kc_tlbp_war())
1371 		build_tlb_probe_entry(&p);
1372 	/* Present and writable bits set, set accessed and dirty bits. */
1373 	build_make_write(&p, &r, K0, K1);
1374 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1375 
1376 #ifdef CONFIG_HUGETLB_PAGE
1377 	/*
1378 	 * This is the entry point when
1379 	 * build_r4000_tlbchange_handler_head spots a huge page.
1380 	 */
1381 	uasm_l_tlb_huge_update(&l, p);
1382 	iPTE_LW(&p, K0, K1);
1383 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1384 	build_tlb_probe_entry(&p);
1385 	uasm_i_ori(&p, K0, K0,
1386 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1387 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1388 #endif
1389 
1390 	uasm_l_nopage_tlbm(&l, p);
1391 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1392 	uasm_i_nop(&p);
1393 
1394 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1395 		panic("TLB modify handler fastpath space exceeded");
1396 
1397 	uasm_resolve_relocs(relocs, labels);
1398 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1399 		 (unsigned int)(p - handle_tlbm));
1400 
1401 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1402 }
1403 
1404 void __cpuinit build_tlb_refill_handler(void)
1405 {
1406 	/*
1407 	 * The refill handler is generated per-CPU, multi-node systems
1408 	 * may have local storage for it. The other handlers are only
1409 	 * needed once.
1410 	 */
1411 	static int run_once = 0;
1412 
1413 	switch (current_cpu_type()) {
1414 	case CPU_R2000:
1415 	case CPU_R3000:
1416 	case CPU_R3000A:
1417 	case CPU_R3081E:
1418 	case CPU_TX3912:
1419 	case CPU_TX3922:
1420 	case CPU_TX3927:
1421 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1422 		build_r3000_tlb_refill_handler();
1423 		if (!run_once) {
1424 			build_r3000_tlb_load_handler();
1425 			build_r3000_tlb_store_handler();
1426 			build_r3000_tlb_modify_handler();
1427 			run_once++;
1428 		}
1429 #else
1430 		panic("No R3000 TLB refill handler");
1431 #endif
1432 		break;
1433 
1434 	case CPU_R6000:
1435 	case CPU_R6000A:
1436 		panic("No R6000 TLB refill handler yet");
1437 		break;
1438 
1439 	case CPU_R8000:
1440 		panic("No R8000 TLB refill handler yet");
1441 		break;
1442 
1443 	default:
1444 		build_r4000_tlb_refill_handler();
1445 		if (!run_once) {
1446 			build_r4000_tlb_load_handler();
1447 			build_r4000_tlb_store_handler();
1448 			build_r4000_tlb_modify_handler();
1449 			run_once++;
1450 		}
1451 	}
1452 }
1453 
1454 void __cpuinit flush_tlb_handlers(void)
1455 {
1456 	local_flush_icache_range((unsigned long)handle_tlbl,
1457 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1458 	local_flush_icache_range((unsigned long)handle_tlbs,
1459 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1460 	local_flush_icache_range((unsigned long)handle_tlbm,
1461 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
1462 }
1463