xref: /linux/arch/mips/mm/tlbex.c (revision 2277ab4a1df50e05bc732fe9488d4e902bb8399a)
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 /*
164  * The R3000 TLB handler is simple.
165  */
166 static void __cpuinit build_r3000_tlb_refill_handler(void)
167 {
168 	long pgdc = (long)pgd_current;
169 	u32 *p;
170 
171 	memset(tlb_handler, 0, sizeof(tlb_handler));
172 	p = tlb_handler;
173 
174 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
175 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
176 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
177 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
178 	uasm_i_sll(&p, K0, K0, 2);
179 	uasm_i_addu(&p, K1, K1, K0);
180 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
181 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
182 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
183 	uasm_i_addu(&p, K1, K1, K0);
184 	uasm_i_lw(&p, K0, 0, K1);
185 	uasm_i_nop(&p); /* load delay */
186 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
187 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
188 	uasm_i_tlbwr(&p); /* cp0 delay */
189 	uasm_i_jr(&p, K1);
190 	uasm_i_rfe(&p); /* branch delay */
191 
192 	if (p > tlb_handler + 32)
193 		panic("TLB refill handler space exceeded");
194 
195 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
196 		 (unsigned int)(p - tlb_handler));
197 
198 	memcpy((void *)ebase, tlb_handler, 0x80);
199 
200 	dump_handler((u32 *)ebase, 32);
201 }
202 
203 /*
204  * The R4000 TLB handler is much more complicated. We have two
205  * consecutive handler areas with 32 instructions space each.
206  * Since they aren't used at the same time, we can overflow in the
207  * other one.To keep things simple, we first assume linear space,
208  * then we relocate it to the final handler layout as needed.
209  */
210 static u32 final_handler[64] __cpuinitdata;
211 
212 /*
213  * Hazards
214  *
215  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
216  * 2. A timing hazard exists for the TLBP instruction.
217  *
218  *      stalling_instruction
219  *      TLBP
220  *
221  * The JTLB is being read for the TLBP throughout the stall generated by the
222  * previous instruction. This is not really correct as the stalling instruction
223  * can modify the address used to access the JTLB.  The failure symptom is that
224  * the TLBP instruction will use an address created for the stalling instruction
225  * and not the address held in C0_ENHI and thus report the wrong results.
226  *
227  * The software work-around is to not allow the instruction preceding the TLBP
228  * to stall - make it an NOP or some other instruction guaranteed not to stall.
229  *
230  * Errata 2 will not be fixed.  This errata is also on the R5000.
231  *
232  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
233  */
234 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
235 {
236 	switch (current_cpu_type()) {
237 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
238 	case CPU_R4600:
239 	case CPU_R4700:
240 	case CPU_R5000:
241 	case CPU_R5000A:
242 	case CPU_NEVADA:
243 		uasm_i_nop(p);
244 		uasm_i_tlbp(p);
245 		break;
246 
247 	default:
248 		uasm_i_tlbp(p);
249 		break;
250 	}
251 }
252 
253 /*
254  * Write random or indexed TLB entry, and care about the hazards from
255  * the preceeding mtc0 and for the following eret.
256  */
257 enum tlb_write_entry { tlb_random, tlb_indexed };
258 
259 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
260 					 struct uasm_reloc **r,
261 					 enum tlb_write_entry wmode)
262 {
263 	void(*tlbw)(u32 **) = NULL;
264 
265 	switch (wmode) {
266 	case tlb_random: tlbw = uasm_i_tlbwr; break;
267 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
268 	}
269 
270 	if (cpu_has_mips_r2) {
271 		if (cpu_has_mips_r2_exec_hazard)
272 			uasm_i_ehb(p);
273 		tlbw(p);
274 		return;
275 	}
276 
277 	switch (current_cpu_type()) {
278 	case CPU_R4000PC:
279 	case CPU_R4000SC:
280 	case CPU_R4000MC:
281 	case CPU_R4400PC:
282 	case CPU_R4400SC:
283 	case CPU_R4400MC:
284 		/*
285 		 * This branch uses up a mtc0 hazard nop slot and saves
286 		 * two nops after the tlbw instruction.
287 		 */
288 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
289 		tlbw(p);
290 		uasm_l_tlbw_hazard(l, *p);
291 		uasm_i_nop(p);
292 		break;
293 
294 	case CPU_R4600:
295 	case CPU_R4700:
296 	case CPU_R5000:
297 	case CPU_R5000A:
298 		uasm_i_nop(p);
299 		tlbw(p);
300 		uasm_i_nop(p);
301 		break;
302 
303 	case CPU_R4300:
304 	case CPU_5KC:
305 	case CPU_TX49XX:
306 	case CPU_PR4450:
307 		uasm_i_nop(p);
308 		tlbw(p);
309 		break;
310 
311 	case CPU_R10000:
312 	case CPU_R12000:
313 	case CPU_R14000:
314 	case CPU_4KC:
315 	case CPU_4KEC:
316 	case CPU_SB1:
317 	case CPU_SB1A:
318 	case CPU_4KSC:
319 	case CPU_20KC:
320 	case CPU_25KF:
321 	case CPU_BCM3302:
322 	case CPU_BCM4710:
323 	case CPU_LOONGSON2:
324 	case CPU_R5500:
325 		if (m4kc_tlbp_war())
326 			uasm_i_nop(p);
327 	case CPU_ALCHEMY:
328 		tlbw(p);
329 		break;
330 
331 	case CPU_NEVADA:
332 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
333 		/*
334 		 * This branch uses up a mtc0 hazard nop slot and saves
335 		 * a nop after the tlbw instruction.
336 		 */
337 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
338 		tlbw(p);
339 		uasm_l_tlbw_hazard(l, *p);
340 		break;
341 
342 	case CPU_RM7000:
343 		uasm_i_nop(p);
344 		uasm_i_nop(p);
345 		uasm_i_nop(p);
346 		uasm_i_nop(p);
347 		tlbw(p);
348 		break;
349 
350 	case CPU_RM9000:
351 		/*
352 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
353 		 * use of the JTLB for instructions should not occur for 4
354 		 * cpu cycles and use for data translations should not occur
355 		 * for 3 cpu cycles.
356 		 */
357 		uasm_i_ssnop(p);
358 		uasm_i_ssnop(p);
359 		uasm_i_ssnop(p);
360 		uasm_i_ssnop(p);
361 		tlbw(p);
362 		uasm_i_ssnop(p);
363 		uasm_i_ssnop(p);
364 		uasm_i_ssnop(p);
365 		uasm_i_ssnop(p);
366 		break;
367 
368 	case CPU_VR4111:
369 	case CPU_VR4121:
370 	case CPU_VR4122:
371 	case CPU_VR4181:
372 	case CPU_VR4181A:
373 		uasm_i_nop(p);
374 		uasm_i_nop(p);
375 		tlbw(p);
376 		uasm_i_nop(p);
377 		uasm_i_nop(p);
378 		break;
379 
380 	case CPU_VR4131:
381 	case CPU_VR4133:
382 	case CPU_R5432:
383 		uasm_i_nop(p);
384 		uasm_i_nop(p);
385 		tlbw(p);
386 		break;
387 
388 	default:
389 		panic("No TLB refill handler yet (CPU type: %d)",
390 		      current_cpu_data.cputype);
391 		break;
392 	}
393 }
394 
395 #ifdef CONFIG_HUGETLB_PAGE
396 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
397 						 struct uasm_label **l,
398 						 struct uasm_reloc **r,
399 						 unsigned int tmp,
400 						 enum tlb_write_entry wmode)
401 {
402 	/* Set huge page tlb entry size */
403 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
404 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
405 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
406 
407 	build_tlb_write_entry(p, l, r, wmode);
408 
409 	/* Reset default page size */
410 	if (PM_DEFAULT_MASK >> 16) {
411 		uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
412 		uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
413 		uasm_il_b(p, r, label_leave);
414 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
415 	} else if (PM_DEFAULT_MASK) {
416 		uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
417 		uasm_il_b(p, r, label_leave);
418 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
419 	} else {
420 		uasm_il_b(p, r, label_leave);
421 		uasm_i_mtc0(p, 0, C0_PAGEMASK);
422 	}
423 }
424 
425 /*
426  * Check if Huge PTE is present, if so then jump to LABEL.
427  */
428 static void __cpuinit
429 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
430 		unsigned int pmd, int lid)
431 {
432 	UASM_i_LW(p, tmp, 0, pmd);
433 	uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
434 	uasm_il_bnez(p, r, tmp, lid);
435 }
436 
437 static __cpuinit void build_huge_update_entries(u32 **p,
438 						unsigned int pte,
439 						unsigned int tmp)
440 {
441 	int small_sequence;
442 
443 	/*
444 	 * A huge PTE describes an area the size of the
445 	 * configured huge page size. This is twice the
446 	 * of the large TLB entry size we intend to use.
447 	 * A TLB entry half the size of the configured
448 	 * huge page size is configured into entrylo0
449 	 * and entrylo1 to cover the contiguous huge PTE
450 	 * address space.
451 	 */
452 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
453 
454 	/* We can clobber tmp.  It isn't used after this.*/
455 	if (!small_sequence)
456 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
457 
458 	UASM_i_SRL(p, pte, pte, 6); /* convert to entrylo */
459 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* load it */
460 	/* convert to entrylo1 */
461 	if (small_sequence)
462 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
463 	else
464 		UASM_i_ADDU(p, pte, pte, tmp);
465 
466 	uasm_i_mtc0(p, pte, C0_ENTRYLO1); /* load it */
467 }
468 
469 static __cpuinit void build_huge_handler_tail(u32 **p,
470 					      struct uasm_reloc **r,
471 					      struct uasm_label **l,
472 					      unsigned int pte,
473 					      unsigned int ptr)
474 {
475 #ifdef CONFIG_SMP
476 	UASM_i_SC(p, pte, 0, ptr);
477 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
478 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
479 #else
480 	UASM_i_SW(p, pte, 0, ptr);
481 #endif
482 	build_huge_update_entries(p, pte, ptr);
483 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed);
484 }
485 #endif /* CONFIG_HUGETLB_PAGE */
486 
487 #ifdef CONFIG_64BIT
488 /*
489  * TMP and PTR are scratch.
490  * TMP will be clobbered, PTR will hold the pmd entry.
491  */
492 static void __cpuinit
493 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
494 		 unsigned int tmp, unsigned int ptr)
495 {
496 	long pgdc = (long)pgd_current;
497 
498 	/*
499 	 * The vmalloc handling is not in the hotpath.
500 	 */
501 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
502 #ifdef MODULE_START
503 	uasm_il_bltz(p, r, tmp, label_module_alloc);
504 #else
505 	uasm_il_bltz(p, r, tmp, label_vmalloc);
506 #endif
507 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
508 
509 #ifdef CONFIG_SMP
510 # ifdef  CONFIG_MIPS_MT_SMTC
511 	/*
512 	 * SMTC uses TCBind value as "CPU" index
513 	 */
514 	uasm_i_mfc0(p, ptr, C0_TCBIND);
515 	uasm_i_dsrl(p, ptr, ptr, 19);
516 # else
517 	/*
518 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
519 	 * stored in CONTEXT.
520 	 */
521 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
522 	uasm_i_dsrl(p, ptr, ptr, 23);
523 #endif
524 	UASM_i_LA_mostly(p, tmp, pgdc);
525 	uasm_i_daddu(p, ptr, ptr, tmp);
526 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
527 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
528 #else
529 	UASM_i_LA_mostly(p, ptr, pgdc);
530 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
531 #endif
532 
533 	uasm_l_vmalloc_done(l, *p);
534 
535 	if (PGDIR_SHIFT - 3 < 32)		/* get pgd offset in bytes */
536 		uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
537 	else
538 		uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
539 
540 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
541 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
542 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
543 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
544 	uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
545 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
546 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
547 }
548 
549 /*
550  * BVADDR is the faulting address, PTR is scratch.
551  * PTR will hold the pgd for vmalloc.
552  */
553 static void __cpuinit
554 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
555 			unsigned int bvaddr, unsigned int ptr)
556 {
557 	long swpd = (long)swapper_pg_dir;
558 
559 #ifdef MODULE_START
560 	long modd = (long)module_pg_dir;
561 
562 	uasm_l_module_alloc(l, *p);
563 	/*
564 	 * Assumption:
565 	 * VMALLOC_START >= 0xc000000000000000UL
566 	 * MODULE_START >= 0xe000000000000000UL
567 	 */
568 	UASM_i_SLL(p, ptr, bvaddr, 2);
569 	uasm_il_bgez(p, r, ptr, label_vmalloc);
570 
571 	if (uasm_in_compat_space_p(MODULE_START) &&
572 	    !uasm_rel_lo(MODULE_START)) {
573 		uasm_i_lui(p, ptr, uasm_rel_hi(MODULE_START)); /* delay slot */
574 	} else {
575 		/* unlikely configuration */
576 		uasm_i_nop(p); /* delay slot */
577 		UASM_i_LA(p, ptr, MODULE_START);
578 	}
579 	uasm_i_dsubu(p, bvaddr, bvaddr, ptr);
580 
581 	if (uasm_in_compat_space_p(modd) && !uasm_rel_lo(modd)) {
582 		uasm_il_b(p, r, label_vmalloc_done);
583 		uasm_i_lui(p, ptr, uasm_rel_hi(modd));
584 	} else {
585 		UASM_i_LA_mostly(p, ptr, modd);
586 		uasm_il_b(p, r, label_vmalloc_done);
587 		if (uasm_in_compat_space_p(modd))
588 			uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(modd));
589 		else
590 			uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(modd));
591 	}
592 
593 	uasm_l_vmalloc(l, *p);
594 	if (uasm_in_compat_space_p(MODULE_START) &&
595 	    !uasm_rel_lo(MODULE_START) &&
596 	    MODULE_START << 32 == VMALLOC_START)
597 		uasm_i_dsll32(p, ptr, ptr, 0);	/* typical case */
598 	else
599 		UASM_i_LA(p, ptr, VMALLOC_START);
600 #else
601 	uasm_l_vmalloc(l, *p);
602 	UASM_i_LA(p, ptr, VMALLOC_START);
603 #endif
604 	uasm_i_dsubu(p, bvaddr, bvaddr, ptr);
605 
606 	if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
607 		uasm_il_b(p, r, label_vmalloc_done);
608 		uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
609 	} else {
610 		UASM_i_LA_mostly(p, ptr, swpd);
611 		uasm_il_b(p, r, label_vmalloc_done);
612 		if (uasm_in_compat_space_p(swpd))
613 			uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
614 		else
615 			uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
616 	}
617 }
618 
619 #else /* !CONFIG_64BIT */
620 
621 /*
622  * TMP and PTR are scratch.
623  * TMP will be clobbered, PTR will hold the pgd entry.
624  */
625 static void __cpuinit __maybe_unused
626 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
627 {
628 	long pgdc = (long)pgd_current;
629 
630 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
631 #ifdef CONFIG_SMP
632 #ifdef  CONFIG_MIPS_MT_SMTC
633 	/*
634 	 * SMTC uses TCBind value as "CPU" index
635 	 */
636 	uasm_i_mfc0(p, ptr, C0_TCBIND);
637 	UASM_i_LA_mostly(p, tmp, pgdc);
638 	uasm_i_srl(p, ptr, ptr, 19);
639 #else
640 	/*
641 	 * smp_processor_id() << 3 is stored in CONTEXT.
642          */
643 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
644 	UASM_i_LA_mostly(p, tmp, pgdc);
645 	uasm_i_srl(p, ptr, ptr, 23);
646 #endif
647 	uasm_i_addu(p, ptr, tmp, ptr);
648 #else
649 	UASM_i_LA_mostly(p, ptr, pgdc);
650 #endif
651 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
652 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
653 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
654 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
655 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
656 }
657 
658 #endif /* !CONFIG_64BIT */
659 
660 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
661 {
662 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
663 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
664 
665 	switch (current_cpu_type()) {
666 	case CPU_VR41XX:
667 	case CPU_VR4111:
668 	case CPU_VR4121:
669 	case CPU_VR4122:
670 	case CPU_VR4131:
671 	case CPU_VR4181:
672 	case CPU_VR4181A:
673 	case CPU_VR4133:
674 		shift += 2;
675 		break;
676 
677 	default:
678 		break;
679 	}
680 
681 	if (shift)
682 		UASM_i_SRL(p, ctx, ctx, shift);
683 	uasm_i_andi(p, ctx, ctx, mask);
684 }
685 
686 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
687 {
688 	/*
689 	 * Bug workaround for the Nevada. It seems as if under certain
690 	 * circumstances the move from cp0_context might produce a
691 	 * bogus result when the mfc0 instruction and its consumer are
692 	 * in a different cacheline or a load instruction, probably any
693 	 * memory reference, is between them.
694 	 */
695 	switch (current_cpu_type()) {
696 	case CPU_NEVADA:
697 		UASM_i_LW(p, ptr, 0, ptr);
698 		GET_CONTEXT(p, tmp); /* get context reg */
699 		break;
700 
701 	default:
702 		GET_CONTEXT(p, tmp); /* get context reg */
703 		UASM_i_LW(p, ptr, 0, ptr);
704 		break;
705 	}
706 
707 	build_adjust_context(p, tmp);
708 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
709 }
710 
711 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
712 					unsigned int ptep)
713 {
714 	/*
715 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
716 	 * Kernel is a special case. Only a few CPUs use it.
717 	 */
718 #ifdef CONFIG_64BIT_PHYS_ADDR
719 	if (cpu_has_64bits) {
720 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
721 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
722 		uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
723 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
724 		uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
725 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
726 	} else {
727 		int pte_off_even = sizeof(pte_t) / 2;
728 		int pte_off_odd = pte_off_even + sizeof(pte_t);
729 
730 		/* The pte entries are pre-shifted */
731 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
732 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
733 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
734 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
735 	}
736 #else
737 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
738 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
739 	if (r45k_bvahwbug())
740 		build_tlb_probe_entry(p);
741 	UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
742 	if (r4k_250MHZhwbug())
743 		uasm_i_mtc0(p, 0, C0_ENTRYLO0);
744 	uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
745 	UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
746 	if (r45k_bvahwbug())
747 		uasm_i_mfc0(p, tmp, C0_INDEX);
748 	if (r4k_250MHZhwbug())
749 		uasm_i_mtc0(p, 0, C0_ENTRYLO1);
750 	uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
751 #endif
752 }
753 
754 /*
755  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
756  * because EXL == 0.  If we wrap, we can also use the 32 instruction
757  * slots before the XTLB refill exception handler which belong to the
758  * unused TLB refill exception.
759  */
760 #define MIPS64_REFILL_INSNS 32
761 
762 static void __cpuinit build_r4000_tlb_refill_handler(void)
763 {
764 	u32 *p = tlb_handler;
765 	struct uasm_label *l = labels;
766 	struct uasm_reloc *r = relocs;
767 	u32 *f;
768 	unsigned int final_len;
769 
770 	memset(tlb_handler, 0, sizeof(tlb_handler));
771 	memset(labels, 0, sizeof(labels));
772 	memset(relocs, 0, sizeof(relocs));
773 	memset(final_handler, 0, sizeof(final_handler));
774 
775 	/*
776 	 * create the plain linear handler
777 	 */
778 	if (bcm1250_m3_war()) {
779 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
780 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
781 		uasm_i_xor(&p, K0, K0, K1);
782 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
783 		uasm_il_bnez(&p, &r, K0, label_leave);
784 		/* No need for uasm_i_nop */
785 	}
786 
787 #ifdef CONFIG_64BIT
788 	build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
789 #else
790 	build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
791 #endif
792 
793 #ifdef CONFIG_HUGETLB_PAGE
794 	build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
795 #endif
796 
797 	build_get_ptep(&p, K0, K1);
798 	build_update_entries(&p, K0, K1);
799 	build_tlb_write_entry(&p, &l, &r, tlb_random);
800 	uasm_l_leave(&l, p);
801 	uasm_i_eret(&p); /* return from trap */
802 
803 #ifdef CONFIG_HUGETLB_PAGE
804 	uasm_l_tlb_huge_update(&l, p);
805 	UASM_i_LW(&p, K0, 0, K1);
806 	build_huge_update_entries(&p, K0, K1);
807 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random);
808 #endif
809 
810 #ifdef CONFIG_64BIT
811 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
812 #endif
813 
814 	/*
815 	 * Overflow check: For the 64bit handler, we need at least one
816 	 * free instruction slot for the wrap-around branch. In worst
817 	 * case, if the intended insertion point is a delay slot, we
818 	 * need three, with the second nop'ed and the third being
819 	 * unused.
820 	 */
821 	/* Loongson2 ebase is different than r4k, we have more space */
822 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
823 	if ((p - tlb_handler) > 64)
824 		panic("TLB refill handler space exceeded");
825 #else
826 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
827 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
828 		&& uasm_insn_has_bdelay(relocs,
829 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
830 		panic("TLB refill handler space exceeded");
831 #endif
832 
833 	/*
834 	 * Now fold the handler in the TLB refill handler space.
835 	 */
836 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
837 	f = final_handler;
838 	/* Simplest case, just copy the handler. */
839 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
840 	final_len = p - tlb_handler;
841 #else /* CONFIG_64BIT */
842 	f = final_handler + MIPS64_REFILL_INSNS;
843 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
844 		/* Just copy the handler. */
845 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
846 		final_len = p - tlb_handler;
847 	} else {
848 #if defined(CONFIG_HUGETLB_PAGE)
849 		const enum label_id ls = label_tlb_huge_update;
850 #elif defined(MODULE_START)
851 		const enum label_id ls = label_module_alloc;
852 #else
853 		const enum label_id ls = label_vmalloc;
854 #endif
855 		u32 *split;
856 		int ov = 0;
857 		int i;
858 
859 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
860 			;
861 		BUG_ON(i == ARRAY_SIZE(labels));
862 		split = labels[i].addr;
863 
864 		/*
865 		 * See if we have overflown one way or the other.
866 		 */
867 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
868 		    split < p - MIPS64_REFILL_INSNS)
869 			ov = 1;
870 
871 		if (ov) {
872 			/*
873 			 * Split two instructions before the end.  One
874 			 * for the branch and one for the instruction
875 			 * in the delay slot.
876 			 */
877 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
878 
879 			/*
880 			 * If the branch would fall in a delay slot,
881 			 * we must back up an additional instruction
882 			 * so that it is no longer in a delay slot.
883 			 */
884 			if (uasm_insn_has_bdelay(relocs, split - 1))
885 				split--;
886 		}
887 		/* Copy first part of the handler. */
888 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
889 		f += split - tlb_handler;
890 
891 		if (ov) {
892 			/* Insert branch. */
893 			uasm_l_split(&l, final_handler);
894 			uasm_il_b(&f, &r, label_split);
895 			if (uasm_insn_has_bdelay(relocs, split))
896 				uasm_i_nop(&f);
897 			else {
898 				uasm_copy_handler(relocs, labels,
899 						  split, split + 1, f);
900 				uasm_move_labels(labels, f, f + 1, -1);
901 				f++;
902 				split++;
903 			}
904 		}
905 
906 		/* Copy the rest of the handler. */
907 		uasm_copy_handler(relocs, labels, split, p, final_handler);
908 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
909 			    (p - split);
910 	}
911 #endif /* CONFIG_64BIT */
912 
913 	uasm_resolve_relocs(relocs, labels);
914 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
915 		 final_len);
916 
917 	memcpy((void *)ebase, final_handler, 0x100);
918 
919 	dump_handler((u32 *)ebase, 64);
920 }
921 
922 /*
923  * TLB load/store/modify handlers.
924  *
925  * Only the fastpath gets synthesized at runtime, the slowpath for
926  * do_page_fault remains normal asm.
927  */
928 extern void tlb_do_page_fault_0(void);
929 extern void tlb_do_page_fault_1(void);
930 
931 /*
932  * 128 instructions for the fastpath handler is generous and should
933  * never be exceeded.
934  */
935 #define FASTPATH_SIZE 128
936 
937 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
938 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
939 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
940 
941 static void __cpuinit
942 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
943 {
944 #ifdef CONFIG_SMP
945 # ifdef CONFIG_64BIT_PHYS_ADDR
946 	if (cpu_has_64bits)
947 		uasm_i_lld(p, pte, 0, ptr);
948 	else
949 # endif
950 		UASM_i_LL(p, pte, 0, ptr);
951 #else
952 # ifdef CONFIG_64BIT_PHYS_ADDR
953 	if (cpu_has_64bits)
954 		uasm_i_ld(p, pte, 0, ptr);
955 	else
956 # endif
957 		UASM_i_LW(p, pte, 0, ptr);
958 #endif
959 }
960 
961 static void __cpuinit
962 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
963 	unsigned int mode)
964 {
965 #ifdef CONFIG_64BIT_PHYS_ADDR
966 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
967 #endif
968 
969 	uasm_i_ori(p, pte, pte, mode);
970 #ifdef CONFIG_SMP
971 # ifdef CONFIG_64BIT_PHYS_ADDR
972 	if (cpu_has_64bits)
973 		uasm_i_scd(p, pte, 0, ptr);
974 	else
975 # endif
976 		UASM_i_SC(p, pte, 0, ptr);
977 
978 	if (r10000_llsc_war())
979 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
980 	else
981 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
982 
983 # ifdef CONFIG_64BIT_PHYS_ADDR
984 	if (!cpu_has_64bits) {
985 		/* no uasm_i_nop needed */
986 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
987 		uasm_i_ori(p, pte, pte, hwmode);
988 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
989 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
990 		/* no uasm_i_nop needed */
991 		uasm_i_lw(p, pte, 0, ptr);
992 	} else
993 		uasm_i_nop(p);
994 # else
995 	uasm_i_nop(p);
996 # endif
997 #else
998 # ifdef CONFIG_64BIT_PHYS_ADDR
999 	if (cpu_has_64bits)
1000 		uasm_i_sd(p, pte, 0, ptr);
1001 	else
1002 # endif
1003 		UASM_i_SW(p, pte, 0, ptr);
1004 
1005 # ifdef CONFIG_64BIT_PHYS_ADDR
1006 	if (!cpu_has_64bits) {
1007 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1008 		uasm_i_ori(p, pte, pte, hwmode);
1009 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1010 		uasm_i_lw(p, pte, 0, ptr);
1011 	}
1012 # endif
1013 #endif
1014 }
1015 
1016 /*
1017  * Check if PTE is present, if not then jump to LABEL. PTR points to
1018  * the page table where this PTE is located, PTE will be re-loaded
1019  * with it's original value.
1020  */
1021 static void __cpuinit
1022 build_pte_present(u32 **p, struct uasm_reloc **r,
1023 		  unsigned int pte, unsigned int ptr, enum label_id lid)
1024 {
1025 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1026 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1027 	uasm_il_bnez(p, r, pte, lid);
1028 	iPTE_LW(p, pte, ptr);
1029 }
1030 
1031 /* Make PTE valid, store result in PTR. */
1032 static void __cpuinit
1033 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1034 		 unsigned int ptr)
1035 {
1036 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1037 
1038 	iPTE_SW(p, r, pte, ptr, mode);
1039 }
1040 
1041 /*
1042  * Check if PTE can be written to, if not branch to LABEL. Regardless
1043  * restore PTE with value from PTR when done.
1044  */
1045 static void __cpuinit
1046 build_pte_writable(u32 **p, struct uasm_reloc **r,
1047 		   unsigned int pte, unsigned int ptr, enum label_id lid)
1048 {
1049 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1050 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1051 	uasm_il_bnez(p, r, pte, lid);
1052 	iPTE_LW(p, pte, ptr);
1053 }
1054 
1055 /* Make PTE writable, update software status bits as well, then store
1056  * at PTR.
1057  */
1058 static void __cpuinit
1059 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1060 		 unsigned int ptr)
1061 {
1062 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1063 			     | _PAGE_DIRTY);
1064 
1065 	iPTE_SW(p, r, pte, ptr, mode);
1066 }
1067 
1068 /*
1069  * Check if PTE can be modified, if not branch to LABEL. Regardless
1070  * restore PTE with value from PTR when done.
1071  */
1072 static void __cpuinit
1073 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1074 		     unsigned int pte, unsigned int ptr, enum label_id lid)
1075 {
1076 	uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1077 	uasm_il_beqz(p, r, pte, lid);
1078 	iPTE_LW(p, pte, ptr);
1079 }
1080 
1081 /*
1082  * R3000 style TLB load/store/modify handlers.
1083  */
1084 
1085 /*
1086  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1087  * Then it returns.
1088  */
1089 static void __cpuinit
1090 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1091 {
1092 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1093 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1094 	uasm_i_tlbwi(p);
1095 	uasm_i_jr(p, tmp);
1096 	uasm_i_rfe(p); /* branch delay */
1097 }
1098 
1099 /*
1100  * This places the pte into ENTRYLO0 and writes it with tlbwi
1101  * or tlbwr as appropriate.  This is because the index register
1102  * may have the probe fail bit set as a result of a trap on a
1103  * kseg2 access, i.e. without refill.  Then it returns.
1104  */
1105 static void __cpuinit
1106 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1107 			     struct uasm_reloc **r, unsigned int pte,
1108 			     unsigned int tmp)
1109 {
1110 	uasm_i_mfc0(p, tmp, C0_INDEX);
1111 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1112 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1113 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1114 	uasm_i_tlbwi(p); /* cp0 delay */
1115 	uasm_i_jr(p, tmp);
1116 	uasm_i_rfe(p); /* branch delay */
1117 	uasm_l_r3000_write_probe_fail(l, *p);
1118 	uasm_i_tlbwr(p); /* cp0 delay */
1119 	uasm_i_jr(p, tmp);
1120 	uasm_i_rfe(p); /* branch delay */
1121 }
1122 
1123 static void __cpuinit
1124 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1125 				   unsigned int ptr)
1126 {
1127 	long pgdc = (long)pgd_current;
1128 
1129 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1130 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1131 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1132 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1133 	uasm_i_sll(p, pte, pte, 2);
1134 	uasm_i_addu(p, ptr, ptr, pte);
1135 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1136 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1137 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1138 	uasm_i_addu(p, ptr, ptr, pte);
1139 	uasm_i_lw(p, pte, 0, ptr);
1140 	uasm_i_tlbp(p); /* load delay */
1141 }
1142 
1143 static void __cpuinit build_r3000_tlb_load_handler(void)
1144 {
1145 	u32 *p = handle_tlbl;
1146 	struct uasm_label *l = labels;
1147 	struct uasm_reloc *r = relocs;
1148 
1149 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1150 	memset(labels, 0, sizeof(labels));
1151 	memset(relocs, 0, sizeof(relocs));
1152 
1153 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1154 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1155 	uasm_i_nop(&p); /* load delay */
1156 	build_make_valid(&p, &r, K0, K1);
1157 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1158 
1159 	uasm_l_nopage_tlbl(&l, p);
1160 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1161 	uasm_i_nop(&p);
1162 
1163 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1164 		panic("TLB load handler fastpath space exceeded");
1165 
1166 	uasm_resolve_relocs(relocs, labels);
1167 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1168 		 (unsigned int)(p - handle_tlbl));
1169 
1170 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1171 }
1172 
1173 static void __cpuinit build_r3000_tlb_store_handler(void)
1174 {
1175 	u32 *p = handle_tlbs;
1176 	struct uasm_label *l = labels;
1177 	struct uasm_reloc *r = relocs;
1178 
1179 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1180 	memset(labels, 0, sizeof(labels));
1181 	memset(relocs, 0, sizeof(relocs));
1182 
1183 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1184 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1185 	uasm_i_nop(&p); /* load delay */
1186 	build_make_write(&p, &r, K0, K1);
1187 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1188 
1189 	uasm_l_nopage_tlbs(&l, p);
1190 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1191 	uasm_i_nop(&p);
1192 
1193 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1194 		panic("TLB store handler fastpath space exceeded");
1195 
1196 	uasm_resolve_relocs(relocs, labels);
1197 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1198 		 (unsigned int)(p - handle_tlbs));
1199 
1200 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1201 }
1202 
1203 static void __cpuinit build_r3000_tlb_modify_handler(void)
1204 {
1205 	u32 *p = handle_tlbm;
1206 	struct uasm_label *l = labels;
1207 	struct uasm_reloc *r = relocs;
1208 
1209 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1210 	memset(labels, 0, sizeof(labels));
1211 	memset(relocs, 0, sizeof(relocs));
1212 
1213 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1214 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1215 	uasm_i_nop(&p); /* load delay */
1216 	build_make_write(&p, &r, K0, K1);
1217 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1218 
1219 	uasm_l_nopage_tlbm(&l, p);
1220 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1221 	uasm_i_nop(&p);
1222 
1223 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1224 		panic("TLB modify handler fastpath space exceeded");
1225 
1226 	uasm_resolve_relocs(relocs, labels);
1227 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1228 		 (unsigned int)(p - handle_tlbm));
1229 
1230 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1231 }
1232 
1233 /*
1234  * R4000 style TLB load/store/modify handlers.
1235  */
1236 static void __cpuinit
1237 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1238 				   struct uasm_reloc **r, unsigned int pte,
1239 				   unsigned int ptr)
1240 {
1241 #ifdef CONFIG_64BIT
1242 	build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1243 #else
1244 	build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1245 #endif
1246 
1247 #ifdef CONFIG_HUGETLB_PAGE
1248 	/*
1249 	 * For huge tlb entries, pmd doesn't contain an address but
1250 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1251 	 * see if we need to jump to huge tlb processing.
1252 	 */
1253 	build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1254 #endif
1255 
1256 	UASM_i_MFC0(p, pte, C0_BADVADDR);
1257 	UASM_i_LW(p, ptr, 0, ptr);
1258 	UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1259 	uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1260 	UASM_i_ADDU(p, ptr, ptr, pte);
1261 
1262 #ifdef CONFIG_SMP
1263 	uasm_l_smp_pgtable_change(l, *p);
1264 #endif
1265 	iPTE_LW(p, pte, ptr); /* get even pte */
1266 	if (!m4kc_tlbp_war())
1267 		build_tlb_probe_entry(p);
1268 }
1269 
1270 static void __cpuinit
1271 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1272 				   struct uasm_reloc **r, unsigned int tmp,
1273 				   unsigned int ptr)
1274 {
1275 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1276 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1277 	build_update_entries(p, tmp, ptr);
1278 	build_tlb_write_entry(p, l, r, tlb_indexed);
1279 	uasm_l_leave(l, *p);
1280 	uasm_i_eret(p); /* return from trap */
1281 
1282 #ifdef CONFIG_64BIT
1283 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1284 #endif
1285 }
1286 
1287 static void __cpuinit build_r4000_tlb_load_handler(void)
1288 {
1289 	u32 *p = handle_tlbl;
1290 	struct uasm_label *l = labels;
1291 	struct uasm_reloc *r = relocs;
1292 
1293 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1294 	memset(labels, 0, sizeof(labels));
1295 	memset(relocs, 0, sizeof(relocs));
1296 
1297 	if (bcm1250_m3_war()) {
1298 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
1299 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1300 		uasm_i_xor(&p, K0, K0, K1);
1301 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1302 		uasm_il_bnez(&p, &r, K0, label_leave);
1303 		/* No need for uasm_i_nop */
1304 	}
1305 
1306 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1307 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1308 	if (m4kc_tlbp_war())
1309 		build_tlb_probe_entry(&p);
1310 	build_make_valid(&p, &r, K0, K1);
1311 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1312 
1313 #ifdef CONFIG_HUGETLB_PAGE
1314 	/*
1315 	 * This is the entry point when build_r4000_tlbchange_handler_head
1316 	 * spots a huge page.
1317 	 */
1318 	uasm_l_tlb_huge_update(&l, p);
1319 	iPTE_LW(&p, K0, K1);
1320 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1321 	build_tlb_probe_entry(&p);
1322 	uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1323 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1324 #endif
1325 
1326 	uasm_l_nopage_tlbl(&l, p);
1327 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1328 	uasm_i_nop(&p);
1329 
1330 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1331 		panic("TLB load handler fastpath space exceeded");
1332 
1333 	uasm_resolve_relocs(relocs, labels);
1334 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1335 		 (unsigned int)(p - handle_tlbl));
1336 
1337 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1338 }
1339 
1340 static void __cpuinit build_r4000_tlb_store_handler(void)
1341 {
1342 	u32 *p = handle_tlbs;
1343 	struct uasm_label *l = labels;
1344 	struct uasm_reloc *r = relocs;
1345 
1346 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1347 	memset(labels, 0, sizeof(labels));
1348 	memset(relocs, 0, sizeof(relocs));
1349 
1350 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1351 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1352 	if (m4kc_tlbp_war())
1353 		build_tlb_probe_entry(&p);
1354 	build_make_write(&p, &r, K0, K1);
1355 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1356 
1357 #ifdef CONFIG_HUGETLB_PAGE
1358 	/*
1359 	 * This is the entry point when
1360 	 * build_r4000_tlbchange_handler_head spots a huge page.
1361 	 */
1362 	uasm_l_tlb_huge_update(&l, p);
1363 	iPTE_LW(&p, K0, K1);
1364 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1365 	build_tlb_probe_entry(&p);
1366 	uasm_i_ori(&p, K0, K0,
1367 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1368 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1369 #endif
1370 
1371 	uasm_l_nopage_tlbs(&l, p);
1372 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1373 	uasm_i_nop(&p);
1374 
1375 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1376 		panic("TLB store handler fastpath space exceeded");
1377 
1378 	uasm_resolve_relocs(relocs, labels);
1379 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1380 		 (unsigned int)(p - handle_tlbs));
1381 
1382 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1383 }
1384 
1385 static void __cpuinit build_r4000_tlb_modify_handler(void)
1386 {
1387 	u32 *p = handle_tlbm;
1388 	struct uasm_label *l = labels;
1389 	struct uasm_reloc *r = relocs;
1390 
1391 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1392 	memset(labels, 0, sizeof(labels));
1393 	memset(relocs, 0, sizeof(relocs));
1394 
1395 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1396 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1397 	if (m4kc_tlbp_war())
1398 		build_tlb_probe_entry(&p);
1399 	/* Present and writable bits set, set accessed and dirty bits. */
1400 	build_make_write(&p, &r, K0, K1);
1401 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1402 
1403 #ifdef CONFIG_HUGETLB_PAGE
1404 	/*
1405 	 * This is the entry point when
1406 	 * build_r4000_tlbchange_handler_head spots a huge page.
1407 	 */
1408 	uasm_l_tlb_huge_update(&l, p);
1409 	iPTE_LW(&p, K0, K1);
1410 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1411 	build_tlb_probe_entry(&p);
1412 	uasm_i_ori(&p, K0, K0,
1413 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1414 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1415 #endif
1416 
1417 	uasm_l_nopage_tlbm(&l, p);
1418 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1419 	uasm_i_nop(&p);
1420 
1421 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1422 		panic("TLB modify handler fastpath space exceeded");
1423 
1424 	uasm_resolve_relocs(relocs, labels);
1425 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1426 		 (unsigned int)(p - handle_tlbm));
1427 
1428 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1429 }
1430 
1431 void __cpuinit build_tlb_refill_handler(void)
1432 {
1433 	/*
1434 	 * The refill handler is generated per-CPU, multi-node systems
1435 	 * may have local storage for it. The other handlers are only
1436 	 * needed once.
1437 	 */
1438 	static int run_once = 0;
1439 
1440 	switch (current_cpu_type()) {
1441 	case CPU_R2000:
1442 	case CPU_R3000:
1443 	case CPU_R3000A:
1444 	case CPU_R3081E:
1445 	case CPU_TX3912:
1446 	case CPU_TX3922:
1447 	case CPU_TX3927:
1448 		build_r3000_tlb_refill_handler();
1449 		if (!run_once) {
1450 			build_r3000_tlb_load_handler();
1451 			build_r3000_tlb_store_handler();
1452 			build_r3000_tlb_modify_handler();
1453 			run_once++;
1454 		}
1455 		break;
1456 
1457 	case CPU_R6000:
1458 	case CPU_R6000A:
1459 		panic("No R6000 TLB refill handler yet");
1460 		break;
1461 
1462 	case CPU_R8000:
1463 		panic("No R8000 TLB refill handler yet");
1464 		break;
1465 
1466 	default:
1467 		build_r4000_tlb_refill_handler();
1468 		if (!run_once) {
1469 			build_r4000_tlb_load_handler();
1470 			build_r4000_tlb_store_handler();
1471 			build_r4000_tlb_modify_handler();
1472 			run_once++;
1473 		}
1474 	}
1475 }
1476 
1477 void __cpuinit flush_tlb_handlers(void)
1478 {
1479 	local_flush_icache_range((unsigned long)handle_tlbl,
1480 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1481 	local_flush_icache_range((unsigned long)handle_tlbs,
1482 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1483 	local_flush_icache_range((unsigned long)handle_tlbm,
1484 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
1485 }
1486