xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 96de0e252cedffad61b3cb5e05662c591898e69a)
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 by Thiemo Seufer
9  * Copyright (C) 2005  Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  *
12  * ... and the days got worse and worse and now you see
13  * I've gone completly out of my mind.
14  *
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 ...
18  *
19  * (Condolences to Napoleon XIV)
20  */
21 
22 #include <stdarg.h>
23 
24 #include <linux/mm.h>
25 #include <linux/kernel.h>
26 #include <linux/types.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
29 
30 #include <asm/pgtable.h>
31 #include <asm/cacheflush.h>
32 #include <asm/mmu_context.h>
33 #include <asm/inst.h>
34 #include <asm/elf.h>
35 #include <asm/smp.h>
36 #include <asm/war.h>
37 
38 static inline int r45k_bvahwbug(void)
39 {
40 	/* XXX: We should probe for the presence of this bug, but we don't. */
41 	return 0;
42 }
43 
44 static inline int r4k_250MHZhwbug(void)
45 {
46 	/* XXX: We should probe for the presence of this bug, but we don't. */
47 	return 0;
48 }
49 
50 static inline int __maybe_unused bcm1250_m3_war(void)
51 {
52 	return BCM1250_M3_WAR;
53 }
54 
55 static inline int __maybe_unused r10000_llsc_war(void)
56 {
57 	return R10000_LLSC_WAR;
58 }
59 
60 /*
61  * Found by experiment: At least some revisions of the 4kc throw under
62  * some circumstances a machine check exception, triggered by invalid
63  * values in the index register.  Delaying the tlbp instruction until
64  * after the next branch,  plus adding an additional nop in front of
65  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
66  * why; it's not an issue caused by the core RTL.
67  *
68  */
69 static __init int __attribute__((unused)) m4kc_tlbp_war(void)
70 {
71 	return (current_cpu_data.processor_id & 0xffff00) ==
72 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
73 }
74 
75 /*
76  * A little micro-assembler, intended for TLB refill handler
77  * synthesizing. It is intentionally kept simple, does only support
78  * a subset of instructions, and does not try to hide pipeline effects
79  * like branch delay slots.
80  */
81 
82 enum fields
83 {
84 	RS = 0x001,
85 	RT = 0x002,
86 	RD = 0x004,
87 	RE = 0x008,
88 	SIMM = 0x010,
89 	UIMM = 0x020,
90 	BIMM = 0x040,
91 	JIMM = 0x080,
92 	FUNC = 0x100,
93 	SET = 0x200
94 };
95 
96 #define OP_MASK		0x3f
97 #define OP_SH		26
98 #define RS_MASK		0x1f
99 #define RS_SH		21
100 #define RT_MASK		0x1f
101 #define RT_SH		16
102 #define RD_MASK		0x1f
103 #define RD_SH		11
104 #define RE_MASK		0x1f
105 #define RE_SH		6
106 #define IMM_MASK	0xffff
107 #define IMM_SH		0
108 #define JIMM_MASK	0x3ffffff
109 #define JIMM_SH		0
110 #define FUNC_MASK	0x3f
111 #define FUNC_SH		0
112 #define SET_MASK	0x7
113 #define SET_SH		0
114 
115 enum opcode {
116 	insn_invalid,
117 	insn_addu, insn_addiu, insn_and, insn_andi, insn_beq,
118 	insn_beql, insn_bgez, insn_bgezl, insn_bltz, insn_bltzl,
119 	insn_bne, insn_daddu, insn_daddiu, insn_dmfc0, insn_dmtc0,
120 	insn_dsll, insn_dsll32, insn_dsra, insn_dsrl, insn_dsrl32,
121 	insn_dsubu, insn_eret, insn_j, insn_jal, insn_jr, insn_ld,
122 	insn_ll, insn_lld, insn_lui, insn_lw, insn_mfc0, insn_mtc0,
123 	insn_ori, insn_rfe, insn_sc, insn_scd, insn_sd, insn_sll,
124 	insn_sra, insn_srl, insn_subu, insn_sw, insn_tlbp, insn_tlbwi,
125 	insn_tlbwr, insn_xor, insn_xori
126 };
127 
128 struct insn {
129 	enum opcode opcode;
130 	u32 match;
131 	enum fields fields;
132 };
133 
134 /* This macro sets the non-variable bits of an instruction. */
135 #define M(a, b, c, d, e, f)					\
136 	((a) << OP_SH						\
137 	 | (b) << RS_SH						\
138 	 | (c) << RT_SH						\
139 	 | (d) << RD_SH						\
140 	 | (e) << RE_SH						\
141 	 | (f) << FUNC_SH)
142 
143 static __initdata struct insn insn_table[] = {
144 	{ insn_addiu, M(addiu_op, 0, 0, 0, 0, 0), RS | RT | SIMM },
145 	{ insn_addu, M(spec_op, 0, 0, 0, 0, addu_op), RS | RT | RD },
146 	{ insn_and, M(spec_op, 0, 0, 0, 0, and_op), RS | RT | RD },
147 	{ insn_andi, M(andi_op, 0, 0, 0, 0, 0), RS | RT | UIMM },
148 	{ insn_beq, M(beq_op, 0, 0, 0, 0, 0), RS | RT | BIMM },
149 	{ insn_beql, M(beql_op, 0, 0, 0, 0, 0), RS | RT | BIMM },
150 	{ insn_bgez, M(bcond_op, 0, bgez_op, 0, 0, 0), RS | BIMM },
151 	{ insn_bgezl, M(bcond_op, 0, bgezl_op, 0, 0, 0), RS | BIMM },
152 	{ insn_bltz, M(bcond_op, 0, bltz_op, 0, 0, 0), RS | BIMM },
153 	{ insn_bltzl, M(bcond_op, 0, bltzl_op, 0, 0, 0), RS | BIMM },
154 	{ insn_bne, M(bne_op, 0, 0, 0, 0, 0), RS | RT | BIMM },
155 	{ insn_daddiu, M(daddiu_op, 0, 0, 0, 0, 0), RS | RT | SIMM },
156 	{ insn_daddu, M(spec_op, 0, 0, 0, 0, daddu_op), RS | RT | RD },
157 	{ insn_dmfc0, M(cop0_op, dmfc_op, 0, 0, 0, 0), RT | RD | SET},
158 	{ insn_dmtc0, M(cop0_op, dmtc_op, 0, 0, 0, 0), RT | RD | SET},
159 	{ insn_dsll, M(spec_op, 0, 0, 0, 0, dsll_op), RT | RD | RE },
160 	{ insn_dsll32, M(spec_op, 0, 0, 0, 0, dsll32_op), RT | RD | RE },
161 	{ insn_dsra, M(spec_op, 0, 0, 0, 0, dsra_op), RT | RD | RE },
162 	{ insn_dsrl, M(spec_op, 0, 0, 0, 0, dsrl_op), RT | RD | RE },
163 	{ insn_dsrl32, M(spec_op, 0, 0, 0, 0, dsrl32_op), RT | RD | RE },
164 	{ insn_dsubu, M(spec_op, 0, 0, 0, 0, dsubu_op), RS | RT | RD },
165 	{ insn_eret,  M(cop0_op, cop_op, 0, 0, 0, eret_op),  0 },
166 	{ insn_j,  M(j_op, 0, 0, 0, 0, 0),  JIMM },
167 	{ insn_jal,  M(jal_op, 0, 0, 0, 0, 0),  JIMM },
168 	{ insn_jr,  M(spec_op, 0, 0, 0, 0, jr_op),  RS },
169 	{ insn_ld,  M(ld_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
170 	{ insn_ll,  M(ll_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
171 	{ insn_lld,  M(lld_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
172 	{ insn_lui,  M(lui_op, 0, 0, 0, 0, 0),  RT | SIMM },
173 	{ insn_lw,  M(lw_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
174 	{ insn_mfc0,  M(cop0_op, mfc_op, 0, 0, 0, 0),  RT | RD | SET},
175 	{ insn_mtc0,  M(cop0_op, mtc_op, 0, 0, 0, 0),  RT | RD | SET},
176 	{ insn_ori,  M(ori_op, 0, 0, 0, 0, 0),  RS | RT | UIMM },
177 	{ insn_rfe,  M(cop0_op, cop_op, 0, 0, 0, rfe_op),  0 },
178 	{ insn_sc,  M(sc_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
179 	{ insn_scd,  M(scd_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
180 	{ insn_sd,  M(sd_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
181 	{ insn_sll,  M(spec_op, 0, 0, 0, 0, sll_op),  RT | RD | RE },
182 	{ insn_sra,  M(spec_op, 0, 0, 0, 0, sra_op),  RT | RD | RE },
183 	{ insn_srl,  M(spec_op, 0, 0, 0, 0, srl_op),  RT | RD | RE },
184 	{ insn_subu,  M(spec_op, 0, 0, 0, 0, subu_op),  RS | RT | RD },
185 	{ insn_sw,  M(sw_op, 0, 0, 0, 0, 0),  RS | RT | SIMM },
186 	{ insn_tlbp,  M(cop0_op, cop_op, 0, 0, 0, tlbp_op),  0 },
187 	{ insn_tlbwi,  M(cop0_op, cop_op, 0, 0, 0, tlbwi_op),  0 },
188 	{ insn_tlbwr,  M(cop0_op, cop_op, 0, 0, 0, tlbwr_op),  0 },
189 	{ insn_xor,  M(spec_op, 0, 0, 0, 0, xor_op),  RS | RT | RD },
190 	{ insn_xori,  M(xori_op, 0, 0, 0, 0, 0),  RS | RT | UIMM },
191 	{ insn_invalid, 0, 0 }
192 };
193 
194 #undef M
195 
196 static __init u32 build_rs(u32 arg)
197 {
198 	if (arg & ~RS_MASK)
199 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
200 
201 	return (arg & RS_MASK) << RS_SH;
202 }
203 
204 static __init u32 build_rt(u32 arg)
205 {
206 	if (arg & ~RT_MASK)
207 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
208 
209 	return (arg & RT_MASK) << RT_SH;
210 }
211 
212 static __init u32 build_rd(u32 arg)
213 {
214 	if (arg & ~RD_MASK)
215 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
216 
217 	return (arg & RD_MASK) << RD_SH;
218 }
219 
220 static __init u32 build_re(u32 arg)
221 {
222 	if (arg & ~RE_MASK)
223 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
224 
225 	return (arg & RE_MASK) << RE_SH;
226 }
227 
228 static __init u32 build_simm(s32 arg)
229 {
230 	if (arg > 0x7fff || arg < -0x8000)
231 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
232 
233 	return arg & 0xffff;
234 }
235 
236 static __init u32 build_uimm(u32 arg)
237 {
238 	if (arg & ~IMM_MASK)
239 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
240 
241 	return arg & IMM_MASK;
242 }
243 
244 static __init u32 build_bimm(s32 arg)
245 {
246 	if (arg > 0x1ffff || arg < -0x20000)
247 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
248 
249 	if (arg & 0x3)
250 		printk(KERN_WARNING "Invalid TLB synthesizer branch target\n");
251 
252 	return ((arg < 0) ? (1 << 15) : 0) | ((arg >> 2) & 0x7fff);
253 }
254 
255 static __init u32 build_jimm(u32 arg)
256 {
257 	if (arg & ~((JIMM_MASK) << 2))
258 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
259 
260 	return (arg >> 2) & JIMM_MASK;
261 }
262 
263 static __init u32 build_func(u32 arg)
264 {
265 	if (arg & ~FUNC_MASK)
266 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
267 
268 	return arg & FUNC_MASK;
269 }
270 
271 static __init u32 build_set(u32 arg)
272 {
273 	if (arg & ~SET_MASK)
274 		printk(KERN_WARNING "TLB synthesizer field overflow\n");
275 
276 	return arg & SET_MASK;
277 }
278 
279 /*
280  * The order of opcode arguments is implicitly left to right,
281  * starting with RS and ending with FUNC or IMM.
282  */
283 static void __init build_insn(u32 **buf, enum opcode opc, ...)
284 {
285 	struct insn *ip = NULL;
286 	unsigned int i;
287 	va_list ap;
288 	u32 op;
289 
290 	for (i = 0; insn_table[i].opcode != insn_invalid; i++)
291 		if (insn_table[i].opcode == opc) {
292 			ip = &insn_table[i];
293 			break;
294 		}
295 
296 	if (!ip)
297 		panic("Unsupported TLB synthesizer instruction %d", opc);
298 
299 	op = ip->match;
300 	va_start(ap, opc);
301 	if (ip->fields & RS) op |= build_rs(va_arg(ap, u32));
302 	if (ip->fields & RT) op |= build_rt(va_arg(ap, u32));
303 	if (ip->fields & RD) op |= build_rd(va_arg(ap, u32));
304 	if (ip->fields & RE) op |= build_re(va_arg(ap, u32));
305 	if (ip->fields & SIMM) op |= build_simm(va_arg(ap, s32));
306 	if (ip->fields & UIMM) op |= build_uimm(va_arg(ap, u32));
307 	if (ip->fields & BIMM) op |= build_bimm(va_arg(ap, s32));
308 	if (ip->fields & JIMM) op |= build_jimm(va_arg(ap, u32));
309 	if (ip->fields & FUNC) op |= build_func(va_arg(ap, u32));
310 	if (ip->fields & SET) op |= build_set(va_arg(ap, u32));
311 	va_end(ap);
312 
313 	**buf = op;
314 	(*buf)++;
315 }
316 
317 #define I_u1u2u3(op)						\
318 	static inline void __init i##op(u32 **buf, unsigned int a,	\
319 	 	unsigned int b, unsigned int c)			\
320 	{							\
321 		build_insn(buf, insn##op, a, b, c);		\
322 	}
323 
324 #define I_u2u1u3(op)						\
325 	static inline void __init i##op(u32 **buf, unsigned int a,	\
326 	 	unsigned int b, unsigned int c)			\
327 	{							\
328 		build_insn(buf, insn##op, b, a, c);		\
329 	}
330 
331 #define I_u3u1u2(op)						\
332 	static inline void __init i##op(u32 **buf, unsigned int a,	\
333 	 	unsigned int b, unsigned int c)			\
334 	{							\
335 		build_insn(buf, insn##op, b, c, a);		\
336 	}
337 
338 #define I_u1u2s3(op)						\
339 	static inline void __init i##op(u32 **buf, unsigned int a,	\
340 	 	unsigned int b, signed int c)			\
341 	{							\
342 		build_insn(buf, insn##op, a, b, c);		\
343 	}
344 
345 #define I_u2s3u1(op)						\
346 	static inline void __init i##op(u32 **buf, unsigned int a,	\
347 	 	signed int b, unsigned int c)			\
348 	{							\
349 		build_insn(buf, insn##op, c, a, b);		\
350 	}
351 
352 #define I_u2u1s3(op)						\
353 	static inline void __init i##op(u32 **buf, unsigned int a,	\
354 	 	unsigned int b, signed int c)			\
355 	{							\
356 		build_insn(buf, insn##op, b, a, c);		\
357 	}
358 
359 #define I_u1u2(op)						\
360 	static inline void __init i##op(u32 **buf, unsigned int a,	\
361 	 	unsigned int b)					\
362 	{							\
363 		build_insn(buf, insn##op, a, b);		\
364 	}
365 
366 #define I_u1s2(op)						\
367 	static inline void __init i##op(u32 **buf, unsigned int a,	\
368 	 	signed int b)					\
369 	{							\
370 		build_insn(buf, insn##op, a, b);		\
371 	}
372 
373 #define I_u1(op)						\
374 	static inline void __init i##op(u32 **buf, unsigned int a)	\
375 	{							\
376 		build_insn(buf, insn##op, a);			\
377 	}
378 
379 #define I_0(op)							\
380 	static inline void __init i##op(u32 **buf)		\
381 	{							\
382 		build_insn(buf, insn##op);			\
383 	}
384 
385 I_u2u1s3(_addiu);
386 I_u3u1u2(_addu);
387 I_u2u1u3(_andi);
388 I_u3u1u2(_and);
389 I_u1u2s3(_beq);
390 I_u1u2s3(_beql);
391 I_u1s2(_bgez);
392 I_u1s2(_bgezl);
393 I_u1s2(_bltz);
394 I_u1s2(_bltzl);
395 I_u1u2s3(_bne);
396 I_u1u2u3(_dmfc0);
397 I_u1u2u3(_dmtc0);
398 I_u2u1s3(_daddiu);
399 I_u3u1u2(_daddu);
400 I_u2u1u3(_dsll);
401 I_u2u1u3(_dsll32);
402 I_u2u1u3(_dsra);
403 I_u2u1u3(_dsrl);
404 I_u2u1u3(_dsrl32);
405 I_u3u1u2(_dsubu);
406 I_0(_eret);
407 I_u1(_j);
408 I_u1(_jal);
409 I_u1(_jr);
410 I_u2s3u1(_ld);
411 I_u2s3u1(_ll);
412 I_u2s3u1(_lld);
413 I_u1s2(_lui);
414 I_u2s3u1(_lw);
415 I_u1u2u3(_mfc0);
416 I_u1u2u3(_mtc0);
417 I_u2u1u3(_ori);
418 I_0(_rfe);
419 I_u2s3u1(_sc);
420 I_u2s3u1(_scd);
421 I_u2s3u1(_sd);
422 I_u2u1u3(_sll);
423 I_u2u1u3(_sra);
424 I_u2u1u3(_srl);
425 I_u3u1u2(_subu);
426 I_u2s3u1(_sw);
427 I_0(_tlbp);
428 I_0(_tlbwi);
429 I_0(_tlbwr);
430 I_u3u1u2(_xor)
431 I_u2u1u3(_xori);
432 
433 /*
434  * handling labels
435  */
436 
437 enum label_id {
438 	label_invalid,
439 	label_second_part,
440 	label_leave,
441 #ifdef MODULE_START
442 	label_module_alloc,
443 #endif
444 	label_vmalloc,
445 	label_vmalloc_done,
446 	label_tlbw_hazard,
447 	label_split,
448 	label_nopage_tlbl,
449 	label_nopage_tlbs,
450 	label_nopage_tlbm,
451 	label_smp_pgtable_change,
452 	label_r3000_write_probe_fail,
453 };
454 
455 struct label {
456 	u32 *addr;
457 	enum label_id lab;
458 };
459 
460 static __init void build_label(struct label **lab, u32 *addr,
461 			       enum label_id l)
462 {
463 	(*lab)->addr = addr;
464 	(*lab)->lab = l;
465 	(*lab)++;
466 }
467 
468 #define L_LA(lb)						\
469 	static inline void l##lb(struct label **lab, u32 *addr) \
470 	{							\
471 		build_label(lab, addr, label##lb);		\
472 	}
473 
474 L_LA(_second_part)
475 L_LA(_leave)
476 #ifdef MODULE_START
477 L_LA(_module_alloc)
478 #endif
479 L_LA(_vmalloc)
480 L_LA(_vmalloc_done)
481 L_LA(_tlbw_hazard)
482 L_LA(_split)
483 L_LA(_nopage_tlbl)
484 L_LA(_nopage_tlbs)
485 L_LA(_nopage_tlbm)
486 L_LA(_smp_pgtable_change)
487 L_LA(_r3000_write_probe_fail)
488 
489 /* convenience macros for instructions */
490 #ifdef CONFIG_64BIT
491 # define i_LW(buf, rs, rt, off) i_ld(buf, rs, rt, off)
492 # define i_SW(buf, rs, rt, off) i_sd(buf, rs, rt, off)
493 # define i_SLL(buf, rs, rt, sh) i_dsll(buf, rs, rt, sh)
494 # define i_SRA(buf, rs, rt, sh) i_dsra(buf, rs, rt, sh)
495 # define i_SRL(buf, rs, rt, sh) i_dsrl(buf, rs, rt, sh)
496 # define i_MFC0(buf, rt, rd...) i_dmfc0(buf, rt, rd)
497 # define i_MTC0(buf, rt, rd...) i_dmtc0(buf, rt, rd)
498 # define i_ADDIU(buf, rs, rt, val) i_daddiu(buf, rs, rt, val)
499 # define i_ADDU(buf, rs, rt, rd) i_daddu(buf, rs, rt, rd)
500 # define i_SUBU(buf, rs, rt, rd) i_dsubu(buf, rs, rt, rd)
501 # define i_LL(buf, rs, rt, off) i_lld(buf, rs, rt, off)
502 # define i_SC(buf, rs, rt, off) i_scd(buf, rs, rt, off)
503 #else
504 # define i_LW(buf, rs, rt, off) i_lw(buf, rs, rt, off)
505 # define i_SW(buf, rs, rt, off) i_sw(buf, rs, rt, off)
506 # define i_SLL(buf, rs, rt, sh) i_sll(buf, rs, rt, sh)
507 # define i_SRA(buf, rs, rt, sh) i_sra(buf, rs, rt, sh)
508 # define i_SRL(buf, rs, rt, sh) i_srl(buf, rs, rt, sh)
509 # define i_MFC0(buf, rt, rd...) i_mfc0(buf, rt, rd)
510 # define i_MTC0(buf, rt, rd...) i_mtc0(buf, rt, rd)
511 # define i_ADDIU(buf, rs, rt, val) i_addiu(buf, rs, rt, val)
512 # define i_ADDU(buf, rs, rt, rd) i_addu(buf, rs, rt, rd)
513 # define i_SUBU(buf, rs, rt, rd) i_subu(buf, rs, rt, rd)
514 # define i_LL(buf, rs, rt, off) i_ll(buf, rs, rt, off)
515 # define i_SC(buf, rs, rt, off) i_sc(buf, rs, rt, off)
516 #endif
517 
518 #define i_b(buf, off) i_beq(buf, 0, 0, off)
519 #define i_beqz(buf, rs, off) i_beq(buf, rs, 0, off)
520 #define i_beqzl(buf, rs, off) i_beql(buf, rs, 0, off)
521 #define i_bnez(buf, rs, off) i_bne(buf, rs, 0, off)
522 #define i_bnezl(buf, rs, off) i_bnel(buf, rs, 0, off)
523 #define i_move(buf, a, b) i_ADDU(buf, a, 0, b)
524 #define i_nop(buf) i_sll(buf, 0, 0, 0)
525 #define i_ssnop(buf) i_sll(buf, 0, 0, 1)
526 #define i_ehb(buf) i_sll(buf, 0, 0, 3)
527 
528 #ifdef CONFIG_64BIT
529 static __init int __maybe_unused in_compat_space_p(long addr)
530 {
531 	/* Is this address in 32bit compat space? */
532 	return (((addr) & 0xffffffff00000000L) == 0xffffffff00000000L);
533 }
534 
535 static __init int __maybe_unused rel_highest(long val)
536 {
537 	return ((((val + 0x800080008000L) >> 48) & 0xffff) ^ 0x8000) - 0x8000;
538 }
539 
540 static __init int __maybe_unused rel_higher(long val)
541 {
542 	return ((((val + 0x80008000L) >> 32) & 0xffff) ^ 0x8000) - 0x8000;
543 }
544 #endif
545 
546 static __init int rel_hi(long val)
547 {
548 	return ((((val + 0x8000L) >> 16) & 0xffff) ^ 0x8000) - 0x8000;
549 }
550 
551 static __init int rel_lo(long val)
552 {
553 	return ((val & 0xffff) ^ 0x8000) - 0x8000;
554 }
555 
556 static __init void i_LA_mostly(u32 **buf, unsigned int rs, long addr)
557 {
558 #ifdef CONFIG_64BIT
559 	if (!in_compat_space_p(addr)) {
560 		i_lui(buf, rs, rel_highest(addr));
561 		if (rel_higher(addr))
562 			i_daddiu(buf, rs, rs, rel_higher(addr));
563 		if (rel_hi(addr)) {
564 			i_dsll(buf, rs, rs, 16);
565 			i_daddiu(buf, rs, rs, rel_hi(addr));
566 			i_dsll(buf, rs, rs, 16);
567 		} else
568 			i_dsll32(buf, rs, rs, 0);
569 	} else
570 #endif
571 		i_lui(buf, rs, rel_hi(addr));
572 }
573 
574 static __init void __maybe_unused i_LA(u32 **buf, unsigned int rs,
575 					     long addr)
576 {
577 	i_LA_mostly(buf, rs, addr);
578 	if (rel_lo(addr))
579 		i_ADDIU(buf, rs, rs, rel_lo(addr));
580 }
581 
582 /*
583  * handle relocations
584  */
585 
586 struct reloc {
587 	u32 *addr;
588 	unsigned int type;
589 	enum label_id lab;
590 };
591 
592 static __init void r_mips_pc16(struct reloc **rel, u32 *addr,
593 			       enum label_id l)
594 {
595 	(*rel)->addr = addr;
596 	(*rel)->type = R_MIPS_PC16;
597 	(*rel)->lab = l;
598 	(*rel)++;
599 }
600 
601 static inline void __resolve_relocs(struct reloc *rel, struct label *lab)
602 {
603 	long laddr = (long)lab->addr;
604 	long raddr = (long)rel->addr;
605 
606 	switch (rel->type) {
607 	case R_MIPS_PC16:
608 		*rel->addr |= build_bimm(laddr - (raddr + 4));
609 		break;
610 
611 	default:
612 		panic("Unsupported TLB synthesizer relocation %d",
613 		      rel->type);
614 	}
615 }
616 
617 static __init void resolve_relocs(struct reloc *rel, struct label *lab)
618 {
619 	struct label *l;
620 
621 	for (; rel->lab != label_invalid; rel++)
622 		for (l = lab; l->lab != label_invalid; l++)
623 			if (rel->lab == l->lab)
624 				__resolve_relocs(rel, l);
625 }
626 
627 static __init void move_relocs(struct reloc *rel, u32 *first, u32 *end,
628 			       long off)
629 {
630 	for (; rel->lab != label_invalid; rel++)
631 		if (rel->addr >= first && rel->addr < end)
632 			rel->addr += off;
633 }
634 
635 static __init void move_labels(struct label *lab, u32 *first, u32 *end,
636 			       long off)
637 {
638 	for (; lab->lab != label_invalid; lab++)
639 		if (lab->addr >= first && lab->addr < end)
640 			lab->addr += off;
641 }
642 
643 static __init void copy_handler(struct reloc *rel, struct label *lab,
644 				u32 *first, u32 *end, u32 *target)
645 {
646 	long off = (long)(target - first);
647 
648 	memcpy(target, first, (end - first) * sizeof(u32));
649 
650 	move_relocs(rel, first, end, off);
651 	move_labels(lab, first, end, off);
652 }
653 
654 static __init int __maybe_unused insn_has_bdelay(struct reloc *rel,
655 						       u32 *addr)
656 {
657 	for (; rel->lab != label_invalid; rel++) {
658 		if (rel->addr == addr
659 		    && (rel->type == R_MIPS_PC16
660 			|| rel->type == R_MIPS_26))
661 			return 1;
662 	}
663 
664 	return 0;
665 }
666 
667 /* convenience functions for labeled branches */
668 static void __init __maybe_unused
669 	il_bltz(u32 **p, struct reloc **r, unsigned int reg, enum label_id l)
670 {
671 	r_mips_pc16(r, *p, l);
672 	i_bltz(p, reg, 0);
673 }
674 
675 static void __init __maybe_unused il_b(u32 **p, struct reloc **r,
676 					     enum label_id l)
677 {
678 	r_mips_pc16(r, *p, l);
679 	i_b(p, 0);
680 }
681 
682 static void __init il_beqz(u32 **p, struct reloc **r, unsigned int reg,
683 		    enum label_id l)
684 {
685 	r_mips_pc16(r, *p, l);
686 	i_beqz(p, reg, 0);
687 }
688 
689 static void __init __maybe_unused
690 il_beqzl(u32 **p, struct reloc **r, unsigned int reg, enum label_id l)
691 {
692 	r_mips_pc16(r, *p, l);
693 	i_beqzl(p, reg, 0);
694 }
695 
696 static void __init il_bnez(u32 **p, struct reloc **r, unsigned int reg,
697 		    enum label_id l)
698 {
699 	r_mips_pc16(r, *p, l);
700 	i_bnez(p, reg, 0);
701 }
702 
703 static void __init il_bgezl(u32 **p, struct reloc **r, unsigned int reg,
704 		     enum label_id l)
705 {
706 	r_mips_pc16(r, *p, l);
707 	i_bgezl(p, reg, 0);
708 }
709 
710 static void __init __maybe_unused
711 il_bgez(u32 **p, struct reloc **r, unsigned int reg, enum label_id l)
712 {
713 	r_mips_pc16(r, *p, l);
714 	i_bgez(p, reg, 0);
715 }
716 
717 /* The only general purpose registers allowed in TLB handlers. */
718 #define K0		26
719 #define K1		27
720 
721 /* Some CP0 registers */
722 #define C0_INDEX	0, 0
723 #define C0_ENTRYLO0	2, 0
724 #define C0_TCBIND	2, 2
725 #define C0_ENTRYLO1	3, 0
726 #define C0_CONTEXT	4, 0
727 #define C0_BADVADDR	8, 0
728 #define C0_ENTRYHI	10, 0
729 #define C0_EPC		14, 0
730 #define C0_XCONTEXT	20, 0
731 
732 #ifdef CONFIG_64BIT
733 # define GET_CONTEXT(buf, reg) i_MFC0(buf, reg, C0_XCONTEXT)
734 #else
735 # define GET_CONTEXT(buf, reg) i_MFC0(buf, reg, C0_CONTEXT)
736 #endif
737 
738 /* The worst case length of the handler is around 18 instructions for
739  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
740  * Maximum space available is 32 instructions for R3000 and 64
741  * instructions for R4000.
742  *
743  * We deliberately chose a buffer size of 128, so we won't scribble
744  * over anything important on overflow before we panic.
745  */
746 static __initdata u32 tlb_handler[128];
747 
748 /* simply assume worst case size for labels and relocs */
749 static __initdata struct label labels[128];
750 static __initdata struct reloc relocs[128];
751 
752 /*
753  * The R3000 TLB handler is simple.
754  */
755 static void __init build_r3000_tlb_refill_handler(void)
756 {
757 	long pgdc = (long)pgd_current;
758 	u32 *p;
759 	int i;
760 
761 	memset(tlb_handler, 0, sizeof(tlb_handler));
762 	p = tlb_handler;
763 
764 	i_mfc0(&p, K0, C0_BADVADDR);
765 	i_lui(&p, K1, rel_hi(pgdc)); /* cp0 delay */
766 	i_lw(&p, K1, rel_lo(pgdc), K1);
767 	i_srl(&p, K0, K0, 22); /* load delay */
768 	i_sll(&p, K0, K0, 2);
769 	i_addu(&p, K1, K1, K0);
770 	i_mfc0(&p, K0, C0_CONTEXT);
771 	i_lw(&p, K1, 0, K1); /* cp0 delay */
772 	i_andi(&p, K0, K0, 0xffc); /* load delay */
773 	i_addu(&p, K1, K1, K0);
774 	i_lw(&p, K0, 0, K1);
775 	i_nop(&p); /* load delay */
776 	i_mtc0(&p, K0, C0_ENTRYLO0);
777 	i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
778 	i_tlbwr(&p); /* cp0 delay */
779 	i_jr(&p, K1);
780 	i_rfe(&p); /* branch delay */
781 
782 	if (p > tlb_handler + 32)
783 		panic("TLB refill handler space exceeded");
784 
785 	pr_info("Synthesized TLB refill handler (%u instructions).\n",
786 		(unsigned int)(p - tlb_handler));
787 
788 	pr_debug("\t.set push\n");
789 	pr_debug("\t.set noreorder\n");
790 	for (i = 0; i < (p - tlb_handler); i++)
791 		pr_debug("\t.word 0x%08x\n", tlb_handler[i]);
792 	pr_debug("\t.set pop\n");
793 
794 	memcpy((void *)ebase, tlb_handler, 0x80);
795 }
796 
797 /*
798  * The R4000 TLB handler is much more complicated. We have two
799  * consecutive handler areas with 32 instructions space each.
800  * Since they aren't used at the same time, we can overflow in the
801  * other one.To keep things simple, we first assume linear space,
802  * then we relocate it to the final handler layout as needed.
803  */
804 static __initdata u32 final_handler[64];
805 
806 /*
807  * Hazards
808  *
809  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
810  * 2. A timing hazard exists for the TLBP instruction.
811  *
812  *      stalling_instruction
813  *      TLBP
814  *
815  * The JTLB is being read for the TLBP throughout the stall generated by the
816  * previous instruction. This is not really correct as the stalling instruction
817  * can modify the address used to access the JTLB.  The failure symptom is that
818  * the TLBP instruction will use an address created for the stalling instruction
819  * and not the address held in C0_ENHI and thus report the wrong results.
820  *
821  * The software work-around is to not allow the instruction preceding the TLBP
822  * to stall - make it an NOP or some other instruction guaranteed not to stall.
823  *
824  * Errata 2 will not be fixed.  This errata is also on the R5000.
825  *
826  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
827  */
828 static __init void __maybe_unused build_tlb_probe_entry(u32 **p)
829 {
830 	switch (current_cpu_type()) {
831 	/* Found by experiment: R4600 v2.0 needs this, too.  */
832 	case CPU_R4600:
833 	case CPU_R5000:
834 	case CPU_R5000A:
835 	case CPU_NEVADA:
836 		i_nop(p);
837 		i_tlbp(p);
838 		break;
839 
840 	default:
841 		i_tlbp(p);
842 		break;
843 	}
844 }
845 
846 /*
847  * Write random or indexed TLB entry, and care about the hazards from
848  * the preceeding mtc0 and for the following eret.
849  */
850 enum tlb_write_entry { tlb_random, tlb_indexed };
851 
852 static __init void build_tlb_write_entry(u32 **p, struct label **l,
853 					 struct reloc **r,
854 					 enum tlb_write_entry wmode)
855 {
856 	void(*tlbw)(u32 **) = NULL;
857 
858 	switch (wmode) {
859 	case tlb_random: tlbw = i_tlbwr; break;
860 	case tlb_indexed: tlbw = i_tlbwi; break;
861 	}
862 
863 	switch (current_cpu_type()) {
864 	case CPU_R4000PC:
865 	case CPU_R4000SC:
866 	case CPU_R4000MC:
867 	case CPU_R4400PC:
868 	case CPU_R4400SC:
869 	case CPU_R4400MC:
870 		/*
871 		 * This branch uses up a mtc0 hazard nop slot and saves
872 		 * two nops after the tlbw instruction.
873 		 */
874 		il_bgezl(p, r, 0, label_tlbw_hazard);
875 		tlbw(p);
876 		l_tlbw_hazard(l, *p);
877 		i_nop(p);
878 		break;
879 
880 	case CPU_R4600:
881 	case CPU_R4700:
882 	case CPU_R5000:
883 	case CPU_R5000A:
884 		i_nop(p);
885 		tlbw(p);
886 		i_nop(p);
887 		break;
888 
889 	case CPU_R4300:
890 	case CPU_5KC:
891 	case CPU_TX49XX:
892 	case CPU_AU1000:
893 	case CPU_AU1100:
894 	case CPU_AU1500:
895 	case CPU_AU1550:
896 	case CPU_AU1200:
897 	case CPU_PR4450:
898 		i_nop(p);
899 		tlbw(p);
900 		break;
901 
902 	case CPU_R10000:
903 	case CPU_R12000:
904 	case CPU_R14000:
905 	case CPU_4KC:
906 	case CPU_SB1:
907 	case CPU_SB1A:
908 	case CPU_4KSC:
909 	case CPU_20KC:
910 	case CPU_25KF:
911 	case CPU_BCM3302:
912 	case CPU_BCM4710:
913 	case CPU_LOONGSON2:
914 		if (m4kc_tlbp_war())
915 			i_nop(p);
916 		tlbw(p);
917 		break;
918 
919 	case CPU_NEVADA:
920 		i_nop(p); /* QED specifies 2 nops hazard */
921 		/*
922 		 * This branch uses up a mtc0 hazard nop slot and saves
923 		 * a nop after the tlbw instruction.
924 		 */
925 		il_bgezl(p, r, 0, label_tlbw_hazard);
926 		tlbw(p);
927 		l_tlbw_hazard(l, *p);
928 		break;
929 
930 	case CPU_RM7000:
931 		i_nop(p);
932 		i_nop(p);
933 		i_nop(p);
934 		i_nop(p);
935 		tlbw(p);
936 		break;
937 
938 	case CPU_4KEC:
939 	case CPU_24K:
940 	case CPU_34K:
941 	case CPU_74K:
942 		i_ehb(p);
943 		tlbw(p);
944 		break;
945 
946 	case CPU_RM9000:
947 		/*
948 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
949 		 * use of the JTLB for instructions should not occur for 4
950 		 * cpu cycles and use for data translations should not occur
951 		 * for 3 cpu cycles.
952 		 */
953 		i_ssnop(p);
954 		i_ssnop(p);
955 		i_ssnop(p);
956 		i_ssnop(p);
957 		tlbw(p);
958 		i_ssnop(p);
959 		i_ssnop(p);
960 		i_ssnop(p);
961 		i_ssnop(p);
962 		break;
963 
964 	case CPU_VR4111:
965 	case CPU_VR4121:
966 	case CPU_VR4122:
967 	case CPU_VR4181:
968 	case CPU_VR4181A:
969 		i_nop(p);
970 		i_nop(p);
971 		tlbw(p);
972 		i_nop(p);
973 		i_nop(p);
974 		break;
975 
976 	case CPU_VR4131:
977 	case CPU_VR4133:
978 	case CPU_R5432:
979 		i_nop(p);
980 		i_nop(p);
981 		tlbw(p);
982 		break;
983 
984 	default:
985 		panic("No TLB refill handler yet (CPU type: %d)",
986 		      current_cpu_data.cputype);
987 		break;
988 	}
989 }
990 
991 #ifdef CONFIG_64BIT
992 /*
993  * TMP and PTR are scratch.
994  * TMP will be clobbered, PTR will hold the pmd entry.
995  */
996 static __init void
997 build_get_pmde64(u32 **p, struct label **l, struct reloc **r,
998 		 unsigned int tmp, unsigned int ptr)
999 {
1000 	long pgdc = (long)pgd_current;
1001 
1002 	/*
1003 	 * The vmalloc handling is not in the hotpath.
1004 	 */
1005 	i_dmfc0(p, tmp, C0_BADVADDR);
1006 #ifdef MODULE_START
1007 	il_bltz(p, r, tmp, label_module_alloc);
1008 #else
1009 	il_bltz(p, r, tmp, label_vmalloc);
1010 #endif
1011 	/* No i_nop needed here, since the next insn doesn't touch TMP. */
1012 
1013 #ifdef CONFIG_SMP
1014 # ifdef  CONFIG_MIPS_MT_SMTC
1015 	/*
1016 	 * SMTC uses TCBind value as "CPU" index
1017 	 */
1018 	i_mfc0(p, ptr, C0_TCBIND);
1019 	i_dsrl(p, ptr, ptr, 19);
1020 # else
1021 	/*
1022 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
1023 	 * stored in CONTEXT.
1024 	 */
1025 	i_dmfc0(p, ptr, C0_CONTEXT);
1026 	i_dsrl(p, ptr, ptr, 23);
1027 #endif
1028 	i_LA_mostly(p, tmp, pgdc);
1029 	i_daddu(p, ptr, ptr, tmp);
1030 	i_dmfc0(p, tmp, C0_BADVADDR);
1031 	i_ld(p, ptr, rel_lo(pgdc), ptr);
1032 #else
1033 	i_LA_mostly(p, ptr, pgdc);
1034 	i_ld(p, ptr, rel_lo(pgdc), ptr);
1035 #endif
1036 
1037 	l_vmalloc_done(l, *p);
1038 
1039 	if (PGDIR_SHIFT - 3 < 32)		/* get pgd offset in bytes */
1040 		i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
1041 	else
1042 		i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
1043 
1044 	i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
1045 	i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
1046 	i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
1047 	i_ld(p, ptr, 0, ptr); /* get pmd pointer */
1048 	i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
1049 	i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
1050 	i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
1051 }
1052 
1053 /*
1054  * BVADDR is the faulting address, PTR is scratch.
1055  * PTR will hold the pgd for vmalloc.
1056  */
1057 static __init void
1058 build_get_pgd_vmalloc64(u32 **p, struct label **l, struct reloc **r,
1059 			unsigned int bvaddr, unsigned int ptr)
1060 {
1061 	long swpd = (long)swapper_pg_dir;
1062 
1063 #ifdef MODULE_START
1064 	long modd = (long)module_pg_dir;
1065 
1066 	l_module_alloc(l, *p);
1067 	/*
1068 	 * Assumption:
1069 	 * VMALLOC_START >= 0xc000000000000000UL
1070 	 * MODULE_START >= 0xe000000000000000UL
1071 	 */
1072 	i_SLL(p, ptr, bvaddr, 2);
1073 	il_bgez(p, r, ptr, label_vmalloc);
1074 
1075 	if (in_compat_space_p(MODULE_START) && !rel_lo(MODULE_START)) {
1076 		i_lui(p, ptr, rel_hi(MODULE_START)); /* delay slot */
1077 	} else {
1078 		/* unlikely configuration */
1079 		i_nop(p); /* delay slot */
1080 		i_LA(p, ptr, MODULE_START);
1081 	}
1082 	i_dsubu(p, bvaddr, bvaddr, ptr);
1083 
1084 	if (in_compat_space_p(modd) && !rel_lo(modd)) {
1085 		il_b(p, r, label_vmalloc_done);
1086 		i_lui(p, ptr, rel_hi(modd));
1087 	} else {
1088 		i_LA_mostly(p, ptr, modd);
1089 		il_b(p, r, label_vmalloc_done);
1090 		i_daddiu(p, ptr, ptr, rel_lo(modd));
1091 	}
1092 
1093 	l_vmalloc(l, *p);
1094 	if (in_compat_space_p(MODULE_START) && !rel_lo(MODULE_START) &&
1095 	    MODULE_START << 32 == VMALLOC_START)
1096 		i_dsll32(p, ptr, ptr, 0);	/* typical case */
1097 	else
1098 		i_LA(p, ptr, VMALLOC_START);
1099 #else
1100 	l_vmalloc(l, *p);
1101 	i_LA(p, ptr, VMALLOC_START);
1102 #endif
1103 	i_dsubu(p, bvaddr, bvaddr, ptr);
1104 
1105 	if (in_compat_space_p(swpd) && !rel_lo(swpd)) {
1106 		il_b(p, r, label_vmalloc_done);
1107 		i_lui(p, ptr, rel_hi(swpd));
1108 	} else {
1109 		i_LA_mostly(p, ptr, swpd);
1110 		il_b(p, r, label_vmalloc_done);
1111 		i_daddiu(p, ptr, ptr, rel_lo(swpd));
1112 	}
1113 }
1114 
1115 #else /* !CONFIG_64BIT */
1116 
1117 /*
1118  * TMP and PTR are scratch.
1119  * TMP will be clobbered, PTR will hold the pgd entry.
1120  */
1121 static __init void __maybe_unused
1122 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
1123 {
1124 	long pgdc = (long)pgd_current;
1125 
1126 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
1127 #ifdef CONFIG_SMP
1128 #ifdef  CONFIG_MIPS_MT_SMTC
1129 	/*
1130 	 * SMTC uses TCBind value as "CPU" index
1131 	 */
1132 	i_mfc0(p, ptr, C0_TCBIND);
1133 	i_LA_mostly(p, tmp, pgdc);
1134 	i_srl(p, ptr, ptr, 19);
1135 #else
1136 	/*
1137 	 * smp_processor_id() << 3 is stored in CONTEXT.
1138          */
1139 	i_mfc0(p, ptr, C0_CONTEXT);
1140 	i_LA_mostly(p, tmp, pgdc);
1141 	i_srl(p, ptr, ptr, 23);
1142 #endif
1143 	i_addu(p, ptr, tmp, ptr);
1144 #else
1145 	i_LA_mostly(p, ptr, pgdc);
1146 #endif
1147 	i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
1148 	i_lw(p, ptr, rel_lo(pgdc), ptr);
1149 	i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
1150 	i_sll(p, tmp, tmp, PGD_T_LOG2);
1151 	i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
1152 }
1153 
1154 #endif /* !CONFIG_64BIT */
1155 
1156 static __init void build_adjust_context(u32 **p, unsigned int ctx)
1157 {
1158 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
1159 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1160 
1161 	switch (current_cpu_type()) {
1162 	case CPU_VR41XX:
1163 	case CPU_VR4111:
1164 	case CPU_VR4121:
1165 	case CPU_VR4122:
1166 	case CPU_VR4131:
1167 	case CPU_VR4181:
1168 	case CPU_VR4181A:
1169 	case CPU_VR4133:
1170 		shift += 2;
1171 		break;
1172 
1173 	default:
1174 		break;
1175 	}
1176 
1177 	if (shift)
1178 		i_SRL(p, ctx, ctx, shift);
1179 	i_andi(p, ctx, ctx, mask);
1180 }
1181 
1182 static __init void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1183 {
1184 	/*
1185 	 * Bug workaround for the Nevada. It seems as if under certain
1186 	 * circumstances the move from cp0_context might produce a
1187 	 * bogus result when the mfc0 instruction and its consumer are
1188 	 * in a different cacheline or a load instruction, probably any
1189 	 * memory reference, is between them.
1190 	 */
1191 	switch (current_cpu_type()) {
1192 	case CPU_NEVADA:
1193 		i_LW(p, ptr, 0, ptr);
1194 		GET_CONTEXT(p, tmp); /* get context reg */
1195 		break;
1196 
1197 	default:
1198 		GET_CONTEXT(p, tmp); /* get context reg */
1199 		i_LW(p, ptr, 0, ptr);
1200 		break;
1201 	}
1202 
1203 	build_adjust_context(p, tmp);
1204 	i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1205 }
1206 
1207 static __init void build_update_entries(u32 **p, unsigned int tmp,
1208 					unsigned int ptep)
1209 {
1210 	/*
1211 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1212 	 * Kernel is a special case. Only a few CPUs use it.
1213 	 */
1214 #ifdef CONFIG_64BIT_PHYS_ADDR
1215 	if (cpu_has_64bits) {
1216 		i_ld(p, tmp, 0, ptep); /* get even pte */
1217 		i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1218 		i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
1219 		i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
1220 		i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
1221 		i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
1222 	} else {
1223 		int pte_off_even = sizeof(pte_t) / 2;
1224 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1225 
1226 		/* The pte entries are pre-shifted */
1227 		i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1228 		i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
1229 		i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1230 		i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
1231 	}
1232 #else
1233 	i_LW(p, tmp, 0, ptep); /* get even pte */
1234 	i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1235 	if (r45k_bvahwbug())
1236 		build_tlb_probe_entry(p);
1237 	i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
1238 	if (r4k_250MHZhwbug())
1239 		i_mtc0(p, 0, C0_ENTRYLO0);
1240 	i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
1241 	i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
1242 	if (r45k_bvahwbug())
1243 		i_mfc0(p, tmp, C0_INDEX);
1244 	if (r4k_250MHZhwbug())
1245 		i_mtc0(p, 0, C0_ENTRYLO1);
1246 	i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
1247 #endif
1248 }
1249 
1250 static void __init build_r4000_tlb_refill_handler(void)
1251 {
1252 	u32 *p = tlb_handler;
1253 	struct label *l = labels;
1254 	struct reloc *r = relocs;
1255 	u32 *f;
1256 	unsigned int final_len;
1257 	int i;
1258 
1259 	memset(tlb_handler, 0, sizeof(tlb_handler));
1260 	memset(labels, 0, sizeof(labels));
1261 	memset(relocs, 0, sizeof(relocs));
1262 	memset(final_handler, 0, sizeof(final_handler));
1263 
1264 	/*
1265 	 * create the plain linear handler
1266 	 */
1267 	if (bcm1250_m3_war()) {
1268 		i_MFC0(&p, K0, C0_BADVADDR);
1269 		i_MFC0(&p, K1, C0_ENTRYHI);
1270 		i_xor(&p, K0, K0, K1);
1271 		i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1272 		il_bnez(&p, &r, K0, label_leave);
1273 		/* No need for i_nop */
1274 	}
1275 
1276 #ifdef CONFIG_64BIT
1277 	build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1278 #else
1279 	build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1280 #endif
1281 
1282 	build_get_ptep(&p, K0, K1);
1283 	build_update_entries(&p, K0, K1);
1284 	build_tlb_write_entry(&p, &l, &r, tlb_random);
1285 	l_leave(&l, p);
1286 	i_eret(&p); /* return from trap */
1287 
1288 #ifdef CONFIG_64BIT
1289 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
1290 #endif
1291 
1292 	/*
1293 	 * Overflow check: For the 64bit handler, we need at least one
1294 	 * free instruction slot for the wrap-around branch. In worst
1295 	 * case, if the intended insertion point is a delay slot, we
1296 	 * need three, with the second nop'ed and the third being
1297 	 * unused.
1298 	 */
1299 	/* Loongson2 ebase is different than r4k, we have more space */
1300 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1301 	if ((p - tlb_handler) > 64)
1302 		panic("TLB refill handler space exceeded");
1303 #else
1304 	if (((p - tlb_handler) > 63)
1305 	    || (((p - tlb_handler) > 61)
1306 		&& insn_has_bdelay(relocs, tlb_handler + 29)))
1307 		panic("TLB refill handler space exceeded");
1308 #endif
1309 
1310 	/*
1311 	 * Now fold the handler in the TLB refill handler space.
1312 	 */
1313 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1314 	f = final_handler;
1315 	/* Simplest case, just copy the handler. */
1316 	copy_handler(relocs, labels, tlb_handler, p, f);
1317 	final_len = p - tlb_handler;
1318 #else /* CONFIG_64BIT */
1319 	f = final_handler + 32;
1320 	if ((p - tlb_handler) <= 32) {
1321 		/* Just copy the handler. */
1322 		copy_handler(relocs, labels, tlb_handler, p, f);
1323 		final_len = p - tlb_handler;
1324 	} else {
1325 		u32 *split = tlb_handler + 30;
1326 
1327 		/*
1328 		 * Find the split point.
1329 		 */
1330 		if (insn_has_bdelay(relocs, split - 1))
1331 			split--;
1332 
1333 		/* Copy first part of the handler. */
1334 		copy_handler(relocs, labels, tlb_handler, split, f);
1335 		f += split - tlb_handler;
1336 
1337 		/* Insert branch. */
1338 		l_split(&l, final_handler);
1339 		il_b(&f, &r, label_split);
1340 		if (insn_has_bdelay(relocs, split))
1341 			i_nop(&f);
1342 		else {
1343 			copy_handler(relocs, labels, split, split + 1, f);
1344 			move_labels(labels, f, f + 1, -1);
1345 			f++;
1346 			split++;
1347 		}
1348 
1349 		/* Copy the rest of the handler. */
1350 		copy_handler(relocs, labels, split, p, final_handler);
1351 		final_len = (f - (final_handler + 32)) + (p - split);
1352 	}
1353 #endif /* CONFIG_64BIT */
1354 
1355 	resolve_relocs(relocs, labels);
1356 	pr_info("Synthesized TLB refill handler (%u instructions).\n",
1357 		final_len);
1358 
1359 	f = final_handler;
1360 #if defined(CONFIG_64BIT) && !defined(CONFIG_CPU_LOONGSON2)
1361 	if (final_len > 32)
1362 		final_len = 64;
1363 	else
1364 		f = final_handler + 32;
1365 #endif /* CONFIG_64BIT */
1366 	pr_debug("\t.set push\n");
1367 	pr_debug("\t.set noreorder\n");
1368 	for (i = 0; i < final_len; i++)
1369 		pr_debug("\t.word 0x%08x\n", f[i]);
1370 	pr_debug("\t.set pop\n");
1371 
1372 	memcpy((void *)ebase, final_handler, 0x100);
1373 }
1374 
1375 /*
1376  * TLB load/store/modify handlers.
1377  *
1378  * Only the fastpath gets synthesized at runtime, the slowpath for
1379  * do_page_fault remains normal asm.
1380  */
1381 extern void tlb_do_page_fault_0(void);
1382 extern void tlb_do_page_fault_1(void);
1383 
1384 #define __tlb_handler_align \
1385 	__attribute__((__aligned__(1 << CONFIG_MIPS_L1_CACHE_SHIFT)))
1386 
1387 /*
1388  * 128 instructions for the fastpath handler is generous and should
1389  * never be exceeded.
1390  */
1391 #define FASTPATH_SIZE 128
1392 
1393 u32 __tlb_handler_align handle_tlbl[FASTPATH_SIZE];
1394 u32 __tlb_handler_align handle_tlbs[FASTPATH_SIZE];
1395 u32 __tlb_handler_align handle_tlbm[FASTPATH_SIZE];
1396 
1397 static void __init
1398 iPTE_LW(u32 **p, struct label **l, unsigned int pte, unsigned int ptr)
1399 {
1400 #ifdef CONFIG_SMP
1401 # ifdef CONFIG_64BIT_PHYS_ADDR
1402 	if (cpu_has_64bits)
1403 		i_lld(p, pte, 0, ptr);
1404 	else
1405 # endif
1406 		i_LL(p, pte, 0, ptr);
1407 #else
1408 # ifdef CONFIG_64BIT_PHYS_ADDR
1409 	if (cpu_has_64bits)
1410 		i_ld(p, pte, 0, ptr);
1411 	else
1412 # endif
1413 		i_LW(p, pte, 0, ptr);
1414 #endif
1415 }
1416 
1417 static void __init
1418 iPTE_SW(u32 **p, struct reloc **r, unsigned int pte, unsigned int ptr,
1419 	unsigned int mode)
1420 {
1421 #ifdef CONFIG_64BIT_PHYS_ADDR
1422 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1423 #endif
1424 
1425 	i_ori(p, pte, pte, mode);
1426 #ifdef CONFIG_SMP
1427 # ifdef CONFIG_64BIT_PHYS_ADDR
1428 	if (cpu_has_64bits)
1429 		i_scd(p, pte, 0, ptr);
1430 	else
1431 # endif
1432 		i_SC(p, pte, 0, ptr);
1433 
1434 	if (r10000_llsc_war())
1435 		il_beqzl(p, r, pte, label_smp_pgtable_change);
1436 	else
1437 		il_beqz(p, r, pte, label_smp_pgtable_change);
1438 
1439 # ifdef CONFIG_64BIT_PHYS_ADDR
1440 	if (!cpu_has_64bits) {
1441 		/* no i_nop needed */
1442 		i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1443 		i_ori(p, pte, pte, hwmode);
1444 		i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1445 		il_beqz(p, r, pte, label_smp_pgtable_change);
1446 		/* no i_nop needed */
1447 		i_lw(p, pte, 0, ptr);
1448 	} else
1449 		i_nop(p);
1450 # else
1451 	i_nop(p);
1452 # endif
1453 #else
1454 # ifdef CONFIG_64BIT_PHYS_ADDR
1455 	if (cpu_has_64bits)
1456 		i_sd(p, pte, 0, ptr);
1457 	else
1458 # endif
1459 		i_SW(p, pte, 0, ptr);
1460 
1461 # ifdef CONFIG_64BIT_PHYS_ADDR
1462 	if (!cpu_has_64bits) {
1463 		i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1464 		i_ori(p, pte, pte, hwmode);
1465 		i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1466 		i_lw(p, pte, 0, ptr);
1467 	}
1468 # endif
1469 #endif
1470 }
1471 
1472 /*
1473  * Check if PTE is present, if not then jump to LABEL. PTR points to
1474  * the page table where this PTE is located, PTE will be re-loaded
1475  * with it's original value.
1476  */
1477 static void __init
1478 build_pte_present(u32 **p, struct label **l, struct reloc **r,
1479 		  unsigned int pte, unsigned int ptr, enum label_id lid)
1480 {
1481 	i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1482 	i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1483 	il_bnez(p, r, pte, lid);
1484 	iPTE_LW(p, l, pte, ptr);
1485 }
1486 
1487 /* Make PTE valid, store result in PTR. */
1488 static void __init
1489 build_make_valid(u32 **p, struct reloc **r, unsigned int pte,
1490 		 unsigned int ptr)
1491 {
1492 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1493 
1494 	iPTE_SW(p, r, pte, ptr, mode);
1495 }
1496 
1497 /*
1498  * Check if PTE can be written to, if not branch to LABEL. Regardless
1499  * restore PTE with value from PTR when done.
1500  */
1501 static void __init
1502 build_pte_writable(u32 **p, struct label **l, struct reloc **r,
1503 		   unsigned int pte, unsigned int ptr, enum label_id lid)
1504 {
1505 	i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1506 	i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1507 	il_bnez(p, r, pte, lid);
1508 	iPTE_LW(p, l, pte, ptr);
1509 }
1510 
1511 /* Make PTE writable, update software status bits as well, then store
1512  * at PTR.
1513  */
1514 static void __init
1515 build_make_write(u32 **p, struct reloc **r, unsigned int pte,
1516 		 unsigned int ptr)
1517 {
1518 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1519 			     | _PAGE_DIRTY);
1520 
1521 	iPTE_SW(p, r, pte, ptr, mode);
1522 }
1523 
1524 /*
1525  * Check if PTE can be modified, if not branch to LABEL. Regardless
1526  * restore PTE with value from PTR when done.
1527  */
1528 static void __init
1529 build_pte_modifiable(u32 **p, struct label **l, struct reloc **r,
1530 		     unsigned int pte, unsigned int ptr, enum label_id lid)
1531 {
1532 	i_andi(p, pte, pte, _PAGE_WRITE);
1533 	il_beqz(p, r, pte, lid);
1534 	iPTE_LW(p, l, pte, ptr);
1535 }
1536 
1537 /*
1538  * R3000 style TLB load/store/modify handlers.
1539  */
1540 
1541 /*
1542  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1543  * Then it returns.
1544  */
1545 static void __init
1546 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1547 {
1548 	i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1549 	i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1550 	i_tlbwi(p);
1551 	i_jr(p, tmp);
1552 	i_rfe(p); /* branch delay */
1553 }
1554 
1555 /*
1556  * This places the pte into ENTRYLO0 and writes it with tlbwi
1557  * or tlbwr as appropriate.  This is because the index register
1558  * may have the probe fail bit set as a result of a trap on a
1559  * kseg2 access, i.e. without refill.  Then it returns.
1560  */
1561 static void __init
1562 build_r3000_tlb_reload_write(u32 **p, struct label **l, struct reloc **r,
1563 			     unsigned int pte, unsigned int tmp)
1564 {
1565 	i_mfc0(p, tmp, C0_INDEX);
1566 	i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1567 	il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1568 	i_mfc0(p, tmp, C0_EPC); /* branch delay */
1569 	i_tlbwi(p); /* cp0 delay */
1570 	i_jr(p, tmp);
1571 	i_rfe(p); /* branch delay */
1572 	l_r3000_write_probe_fail(l, *p);
1573 	i_tlbwr(p); /* cp0 delay */
1574 	i_jr(p, tmp);
1575 	i_rfe(p); /* branch delay */
1576 }
1577 
1578 static void __init
1579 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1580 				   unsigned int ptr)
1581 {
1582 	long pgdc = (long)pgd_current;
1583 
1584 	i_mfc0(p, pte, C0_BADVADDR);
1585 	i_lui(p, ptr, rel_hi(pgdc)); /* cp0 delay */
1586 	i_lw(p, ptr, rel_lo(pgdc), ptr);
1587 	i_srl(p, pte, pte, 22); /* load delay */
1588 	i_sll(p, pte, pte, 2);
1589 	i_addu(p, ptr, ptr, pte);
1590 	i_mfc0(p, pte, C0_CONTEXT);
1591 	i_lw(p, ptr, 0, ptr); /* cp0 delay */
1592 	i_andi(p, pte, pte, 0xffc); /* load delay */
1593 	i_addu(p, ptr, ptr, pte);
1594 	i_lw(p, pte, 0, ptr);
1595 	i_tlbp(p); /* load delay */
1596 }
1597 
1598 static void __init build_r3000_tlb_load_handler(void)
1599 {
1600 	u32 *p = handle_tlbl;
1601 	struct label *l = labels;
1602 	struct reloc *r = relocs;
1603 	int i;
1604 
1605 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1606 	memset(labels, 0, sizeof(labels));
1607 	memset(relocs, 0, sizeof(relocs));
1608 
1609 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1610 	build_pte_present(&p, &l, &r, K0, K1, label_nopage_tlbl);
1611 	i_nop(&p); /* load delay */
1612 	build_make_valid(&p, &r, K0, K1);
1613 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1614 
1615 	l_nopage_tlbl(&l, p);
1616 	i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1617 	i_nop(&p);
1618 
1619 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1620 		panic("TLB load handler fastpath space exceeded");
1621 
1622 	resolve_relocs(relocs, labels);
1623 	pr_info("Synthesized TLB load handler fastpath (%u instructions).\n",
1624 		(unsigned int)(p - handle_tlbl));
1625 
1626 	pr_debug("\t.set push\n");
1627 	pr_debug("\t.set noreorder\n");
1628 	for (i = 0; i < (p - handle_tlbl); i++)
1629 		pr_debug("\t.word 0x%08x\n", handle_tlbl[i]);
1630 	pr_debug("\t.set pop\n");
1631 }
1632 
1633 static void __init build_r3000_tlb_store_handler(void)
1634 {
1635 	u32 *p = handle_tlbs;
1636 	struct label *l = labels;
1637 	struct reloc *r = relocs;
1638 	int i;
1639 
1640 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1641 	memset(labels, 0, sizeof(labels));
1642 	memset(relocs, 0, sizeof(relocs));
1643 
1644 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1645 	build_pte_writable(&p, &l, &r, K0, K1, label_nopage_tlbs);
1646 	i_nop(&p); /* load delay */
1647 	build_make_write(&p, &r, K0, K1);
1648 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1649 
1650 	l_nopage_tlbs(&l, p);
1651 	i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1652 	i_nop(&p);
1653 
1654 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1655 		panic("TLB store handler fastpath space exceeded");
1656 
1657 	resolve_relocs(relocs, labels);
1658 	pr_info("Synthesized TLB store handler fastpath (%u instructions).\n",
1659 		(unsigned int)(p - handle_tlbs));
1660 
1661 	pr_debug("\t.set push\n");
1662 	pr_debug("\t.set noreorder\n");
1663 	for (i = 0; i < (p - handle_tlbs); i++)
1664 		pr_debug("\t.word 0x%08x\n", handle_tlbs[i]);
1665 	pr_debug("\t.set pop\n");
1666 }
1667 
1668 static void __init build_r3000_tlb_modify_handler(void)
1669 {
1670 	u32 *p = handle_tlbm;
1671 	struct label *l = labels;
1672 	struct reloc *r = relocs;
1673 	int i;
1674 
1675 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1676 	memset(labels, 0, sizeof(labels));
1677 	memset(relocs, 0, sizeof(relocs));
1678 
1679 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1680 	build_pte_modifiable(&p, &l, &r, K0, K1, label_nopage_tlbm);
1681 	i_nop(&p); /* load delay */
1682 	build_make_write(&p, &r, K0, K1);
1683 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1684 
1685 	l_nopage_tlbm(&l, p);
1686 	i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1687 	i_nop(&p);
1688 
1689 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1690 		panic("TLB modify handler fastpath space exceeded");
1691 
1692 	resolve_relocs(relocs, labels);
1693 	pr_info("Synthesized TLB modify handler fastpath (%u instructions).\n",
1694 		(unsigned int)(p - handle_tlbm));
1695 
1696 	pr_debug("\t.set push\n");
1697 	pr_debug("\t.set noreorder\n");
1698 	for (i = 0; i < (p - handle_tlbm); i++)
1699 		pr_debug("\t.word 0x%08x\n", handle_tlbm[i]);
1700 	pr_debug("\t.set pop\n");
1701 }
1702 
1703 /*
1704  * R4000 style TLB load/store/modify handlers.
1705  */
1706 static void __init
1707 build_r4000_tlbchange_handler_head(u32 **p, struct label **l,
1708 				   struct reloc **r, unsigned int pte,
1709 				   unsigned int ptr)
1710 {
1711 #ifdef CONFIG_64BIT
1712 	build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1713 #else
1714 	build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1715 #endif
1716 
1717 	i_MFC0(p, pte, C0_BADVADDR);
1718 	i_LW(p, ptr, 0, ptr);
1719 	i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1720 	i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1721 	i_ADDU(p, ptr, ptr, pte);
1722 
1723 #ifdef CONFIG_SMP
1724 	l_smp_pgtable_change(l, *p);
1725 # endif
1726 	iPTE_LW(p, l, pte, ptr); /* get even pte */
1727 	if (!m4kc_tlbp_war())
1728 		build_tlb_probe_entry(p);
1729 }
1730 
1731 static void __init
1732 build_r4000_tlbchange_handler_tail(u32 **p, struct label **l,
1733 				   struct reloc **r, unsigned int tmp,
1734 				   unsigned int ptr)
1735 {
1736 	i_ori(p, ptr, ptr, sizeof(pte_t));
1737 	i_xori(p, ptr, ptr, sizeof(pte_t));
1738 	build_update_entries(p, tmp, ptr);
1739 	build_tlb_write_entry(p, l, r, tlb_indexed);
1740 	l_leave(l, *p);
1741 	i_eret(p); /* return from trap */
1742 
1743 #ifdef CONFIG_64BIT
1744 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1745 #endif
1746 }
1747 
1748 static void __init build_r4000_tlb_load_handler(void)
1749 {
1750 	u32 *p = handle_tlbl;
1751 	struct label *l = labels;
1752 	struct reloc *r = relocs;
1753 	int i;
1754 
1755 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1756 	memset(labels, 0, sizeof(labels));
1757 	memset(relocs, 0, sizeof(relocs));
1758 
1759 	if (bcm1250_m3_war()) {
1760 		i_MFC0(&p, K0, C0_BADVADDR);
1761 		i_MFC0(&p, K1, C0_ENTRYHI);
1762 		i_xor(&p, K0, K0, K1);
1763 		i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1764 		il_bnez(&p, &r, K0, label_leave);
1765 		/* No need for i_nop */
1766 	}
1767 
1768 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1769 	build_pte_present(&p, &l, &r, K0, K1, label_nopage_tlbl);
1770 	if (m4kc_tlbp_war())
1771 		build_tlb_probe_entry(&p);
1772 	build_make_valid(&p, &r, K0, K1);
1773 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1774 
1775 	l_nopage_tlbl(&l, p);
1776 	i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1777 	i_nop(&p);
1778 
1779 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1780 		panic("TLB load handler fastpath space exceeded");
1781 
1782 	resolve_relocs(relocs, labels);
1783 	pr_info("Synthesized TLB load handler fastpath (%u instructions).\n",
1784 		(unsigned int)(p - handle_tlbl));
1785 
1786 	pr_debug("\t.set push\n");
1787 	pr_debug("\t.set noreorder\n");
1788 	for (i = 0; i < (p - handle_tlbl); i++)
1789 		pr_debug("\t.word 0x%08x\n", handle_tlbl[i]);
1790 	pr_debug("\t.set pop\n");
1791 }
1792 
1793 static void __init build_r4000_tlb_store_handler(void)
1794 {
1795 	u32 *p = handle_tlbs;
1796 	struct label *l = labels;
1797 	struct reloc *r = relocs;
1798 	int i;
1799 
1800 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1801 	memset(labels, 0, sizeof(labels));
1802 	memset(relocs, 0, sizeof(relocs));
1803 
1804 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1805 	build_pte_writable(&p, &l, &r, K0, K1, label_nopage_tlbs);
1806 	if (m4kc_tlbp_war())
1807 		build_tlb_probe_entry(&p);
1808 	build_make_write(&p, &r, K0, K1);
1809 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1810 
1811 	l_nopage_tlbs(&l, p);
1812 	i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1813 	i_nop(&p);
1814 
1815 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1816 		panic("TLB store handler fastpath space exceeded");
1817 
1818 	resolve_relocs(relocs, labels);
1819 	pr_info("Synthesized TLB store handler fastpath (%u instructions).\n",
1820 		(unsigned int)(p - handle_tlbs));
1821 
1822 	pr_debug("\t.set push\n");
1823 	pr_debug("\t.set noreorder\n");
1824 	for (i = 0; i < (p - handle_tlbs); i++)
1825 		pr_debug("\t.word 0x%08x\n", handle_tlbs[i]);
1826 	pr_debug("\t.set pop\n");
1827 }
1828 
1829 static void __init build_r4000_tlb_modify_handler(void)
1830 {
1831 	u32 *p = handle_tlbm;
1832 	struct label *l = labels;
1833 	struct reloc *r = relocs;
1834 	int i;
1835 
1836 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1837 	memset(labels, 0, sizeof(labels));
1838 	memset(relocs, 0, sizeof(relocs));
1839 
1840 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1841 	build_pte_modifiable(&p, &l, &r, K0, K1, label_nopage_tlbm);
1842 	if (m4kc_tlbp_war())
1843 		build_tlb_probe_entry(&p);
1844 	/* Present and writable bits set, set accessed and dirty bits. */
1845 	build_make_write(&p, &r, K0, K1);
1846 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1847 
1848 	l_nopage_tlbm(&l, p);
1849 	i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1850 	i_nop(&p);
1851 
1852 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1853 		panic("TLB modify handler fastpath space exceeded");
1854 
1855 	resolve_relocs(relocs, labels);
1856 	pr_info("Synthesized TLB modify handler fastpath (%u instructions).\n",
1857 		(unsigned int)(p - handle_tlbm));
1858 
1859 	pr_debug("\t.set push\n");
1860 	pr_debug("\t.set noreorder\n");
1861 	for (i = 0; i < (p - handle_tlbm); i++)
1862 		pr_debug("\t.word 0x%08x\n", handle_tlbm[i]);
1863 	pr_debug("\t.set pop\n");
1864 }
1865 
1866 void __init build_tlb_refill_handler(void)
1867 {
1868 	/*
1869 	 * The refill handler is generated per-CPU, multi-node systems
1870 	 * may have local storage for it. The other handlers are only
1871 	 * needed once.
1872 	 */
1873 	static int run_once = 0;
1874 
1875 	switch (current_cpu_type()) {
1876 	case CPU_R2000:
1877 	case CPU_R3000:
1878 	case CPU_R3000A:
1879 	case CPU_R3081E:
1880 	case CPU_TX3912:
1881 	case CPU_TX3922:
1882 	case CPU_TX3927:
1883 		build_r3000_tlb_refill_handler();
1884 		if (!run_once) {
1885 			build_r3000_tlb_load_handler();
1886 			build_r3000_tlb_store_handler();
1887 			build_r3000_tlb_modify_handler();
1888 			run_once++;
1889 		}
1890 		break;
1891 
1892 	case CPU_R6000:
1893 	case CPU_R6000A:
1894 		panic("No R6000 TLB refill handler yet");
1895 		break;
1896 
1897 	case CPU_R8000:
1898 		panic("No R8000 TLB refill handler yet");
1899 		break;
1900 
1901 	default:
1902 		build_r4000_tlb_refill_handler();
1903 		if (!run_once) {
1904 			build_r4000_tlb_load_handler();
1905 			build_r4000_tlb_store_handler();
1906 			build_r4000_tlb_modify_handler();
1907 			run_once++;
1908 		}
1909 	}
1910 }
1911 
1912 void __init flush_tlb_handlers(void)
1913 {
1914 	flush_icache_range((unsigned long)handle_tlbl,
1915 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1916 	flush_icache_range((unsigned long)handle_tlbs,
1917 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1918 	flush_icache_range((unsigned long)handle_tlbm,
1919 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
1920 }
1921