xref: /linux/arch/mips/mm/tlbex.c (revision 8c749ce93ee69e789e46b3be98de9e0cbfcf8ed8)
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  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completly out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/cache.h>
30 
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
34 #include <asm/war.h>
35 #include <asm/uasm.h>
36 #include <asm/setup.h>
37 
38 static int mips_xpa_disabled;
39 
40 static int __init xpa_disable(char *s)
41 {
42 	mips_xpa_disabled = 1;
43 
44 	return 1;
45 }
46 
47 __setup("noxpa", xpa_disable);
48 
49 /*
50  * TLB load/store/modify handlers.
51  *
52  * Only the fastpath gets synthesized at runtime, the slowpath for
53  * do_page_fault remains normal asm.
54  */
55 extern void tlb_do_page_fault_0(void);
56 extern void tlb_do_page_fault_1(void);
57 
58 struct work_registers {
59 	int r1;
60 	int r2;
61 	int r3;
62 };
63 
64 struct tlb_reg_save {
65 	unsigned long a;
66 	unsigned long b;
67 } ____cacheline_aligned_in_smp;
68 
69 static struct tlb_reg_save handler_reg_save[NR_CPUS];
70 
71 static inline int r45k_bvahwbug(void)
72 {
73 	/* XXX: We should probe for the presence of this bug, but we don't. */
74 	return 0;
75 }
76 
77 static inline int r4k_250MHZhwbug(void)
78 {
79 	/* XXX: We should probe for the presence of this bug, but we don't. */
80 	return 0;
81 }
82 
83 static inline int __maybe_unused bcm1250_m3_war(void)
84 {
85 	return BCM1250_M3_WAR;
86 }
87 
88 static inline int __maybe_unused r10000_llsc_war(void)
89 {
90 	return R10000_LLSC_WAR;
91 }
92 
93 static int use_bbit_insns(void)
94 {
95 	switch (current_cpu_type()) {
96 	case CPU_CAVIUM_OCTEON:
97 	case CPU_CAVIUM_OCTEON_PLUS:
98 	case CPU_CAVIUM_OCTEON2:
99 	case CPU_CAVIUM_OCTEON3:
100 		return 1;
101 	default:
102 		return 0;
103 	}
104 }
105 
106 static int use_lwx_insns(void)
107 {
108 	switch (current_cpu_type()) {
109 	case CPU_CAVIUM_OCTEON2:
110 	case CPU_CAVIUM_OCTEON3:
111 		return 1;
112 	default:
113 		return 0;
114 	}
115 }
116 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
117     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
118 static bool scratchpad_available(void)
119 {
120 	return true;
121 }
122 static int scratchpad_offset(int i)
123 {
124 	/*
125 	 * CVMSEG starts at address -32768 and extends for
126 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
127 	 */
128 	i += 1; /* Kernel use starts at the top and works down. */
129 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
130 }
131 #else
132 static bool scratchpad_available(void)
133 {
134 	return false;
135 }
136 static int scratchpad_offset(int i)
137 {
138 	BUG();
139 	/* Really unreachable, but evidently some GCC want this. */
140 	return 0;
141 }
142 #endif
143 /*
144  * Found by experiment: At least some revisions of the 4kc throw under
145  * some circumstances a machine check exception, triggered by invalid
146  * values in the index register.  Delaying the tlbp instruction until
147  * after the next branch,  plus adding an additional nop in front of
148  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
149  * why; it's not an issue caused by the core RTL.
150  *
151  */
152 static int m4kc_tlbp_war(void)
153 {
154 	return (current_cpu_data.processor_id & 0xffff00) ==
155 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
156 }
157 
158 /* Handle labels (which must be positive integers). */
159 enum label_id {
160 	label_second_part = 1,
161 	label_leave,
162 	label_vmalloc,
163 	label_vmalloc_done,
164 	label_tlbw_hazard_0,
165 	label_split = label_tlbw_hazard_0 + 8,
166 	label_tlbl_goaround1,
167 	label_tlbl_goaround2,
168 	label_nopage_tlbl,
169 	label_nopage_tlbs,
170 	label_nopage_tlbm,
171 	label_smp_pgtable_change,
172 	label_r3000_write_probe_fail,
173 	label_large_segbits_fault,
174 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
175 	label_tlb_huge_update,
176 #endif
177 };
178 
179 UASM_L_LA(_second_part)
180 UASM_L_LA(_leave)
181 UASM_L_LA(_vmalloc)
182 UASM_L_LA(_vmalloc_done)
183 /* _tlbw_hazard_x is handled differently.  */
184 UASM_L_LA(_split)
185 UASM_L_LA(_tlbl_goaround1)
186 UASM_L_LA(_tlbl_goaround2)
187 UASM_L_LA(_nopage_tlbl)
188 UASM_L_LA(_nopage_tlbs)
189 UASM_L_LA(_nopage_tlbm)
190 UASM_L_LA(_smp_pgtable_change)
191 UASM_L_LA(_r3000_write_probe_fail)
192 UASM_L_LA(_large_segbits_fault)
193 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
194 UASM_L_LA(_tlb_huge_update)
195 #endif
196 
197 static int hazard_instance;
198 
199 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
200 {
201 	switch (instance) {
202 	case 0 ... 7:
203 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
204 		return;
205 	default:
206 		BUG();
207 	}
208 }
209 
210 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
211 {
212 	switch (instance) {
213 	case 0 ... 7:
214 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
215 		break;
216 	default:
217 		BUG();
218 	}
219 }
220 
221 /*
222  * pgtable bits are assigned dynamically depending on processor feature
223  * and statically based on kernel configuration.  This spits out the actual
224  * values the kernel is using.	Required to make sense from disassembled
225  * TLB exception handlers.
226  */
227 static void output_pgtable_bits_defines(void)
228 {
229 #define pr_define(fmt, ...)					\
230 	pr_debug("#define " fmt, ##__VA_ARGS__)
231 
232 	pr_debug("#include <asm/asm.h>\n");
233 	pr_debug("#include <asm/regdef.h>\n");
234 	pr_debug("\n");
235 
236 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
237 	pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
238 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
239 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
240 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
241 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
242 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
243 #endif
244 #if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_MIPSR6)
245 	if (cpu_has_rixi) {
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
248 		pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
249 #endif
250 	}
251 #endif
252 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
253 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
254 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
255 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
256 	pr_debug("\n");
257 }
258 
259 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
260 {
261 	int i;
262 
263 	pr_debug("LEAF(%s)\n", symbol);
264 
265 	pr_debug("\t.set push\n");
266 	pr_debug("\t.set noreorder\n");
267 
268 	for (i = 0; i < count; i++)
269 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
270 
271 	pr_debug("\t.set\tpop\n");
272 
273 	pr_debug("\tEND(%s)\n", symbol);
274 }
275 
276 /* The only general purpose registers allowed in TLB handlers. */
277 #define K0		26
278 #define K1		27
279 
280 /* Some CP0 registers */
281 #define C0_INDEX	0, 0
282 #define C0_ENTRYLO0	2, 0
283 #define C0_TCBIND	2, 2
284 #define C0_ENTRYLO1	3, 0
285 #define C0_CONTEXT	4, 0
286 #define C0_PAGEMASK	5, 0
287 #define C0_BADVADDR	8, 0
288 #define C0_ENTRYHI	10, 0
289 #define C0_EPC		14, 0
290 #define C0_XCONTEXT	20, 0
291 
292 #ifdef CONFIG_64BIT
293 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
294 #else
295 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
296 #endif
297 
298 /* The worst case length of the handler is around 18 instructions for
299  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
300  * Maximum space available is 32 instructions for R3000 and 64
301  * instructions for R4000.
302  *
303  * We deliberately chose a buffer size of 128, so we won't scribble
304  * over anything important on overflow before we panic.
305  */
306 static u32 tlb_handler[128];
307 
308 /* simply assume worst case size for labels and relocs */
309 static struct uasm_label labels[128];
310 static struct uasm_reloc relocs[128];
311 
312 static int check_for_high_segbits;
313 static bool fill_includes_sw_bits;
314 
315 static unsigned int kscratch_used_mask;
316 
317 static inline int __maybe_unused c0_kscratch(void)
318 {
319 	switch (current_cpu_type()) {
320 	case CPU_XLP:
321 	case CPU_XLR:
322 		return 22;
323 	default:
324 		return 31;
325 	}
326 }
327 
328 static int allocate_kscratch(void)
329 {
330 	int r;
331 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
332 
333 	r = ffs(a);
334 
335 	if (r == 0)
336 		return -1;
337 
338 	r--; /* make it zero based */
339 
340 	kscratch_used_mask |= (1 << r);
341 
342 	return r;
343 }
344 
345 static int scratch_reg;
346 static int pgd_reg;
347 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
348 
349 static struct work_registers build_get_work_registers(u32 **p)
350 {
351 	struct work_registers r;
352 
353 	if (scratch_reg >= 0) {
354 		/* Save in CPU local C0_KScratch? */
355 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
356 		r.r1 = K0;
357 		r.r2 = K1;
358 		r.r3 = 1;
359 		return r;
360 	}
361 
362 	if (num_possible_cpus() > 1) {
363 		/* Get smp_processor_id */
364 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
365 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
366 
367 		/* handler_reg_save index in K0 */
368 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
369 
370 		UASM_i_LA(p, K1, (long)&handler_reg_save);
371 		UASM_i_ADDU(p, K0, K0, K1);
372 	} else {
373 		UASM_i_LA(p, K0, (long)&handler_reg_save);
374 	}
375 	/* K0 now points to save area, save $1 and $2  */
376 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
377 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
378 
379 	r.r1 = K1;
380 	r.r2 = 1;
381 	r.r3 = 2;
382 	return r;
383 }
384 
385 static void build_restore_work_registers(u32 **p)
386 {
387 	if (scratch_reg >= 0) {
388 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
389 		return;
390 	}
391 	/* K0 already points to save area, restore $1 and $2  */
392 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
393 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
394 }
395 
396 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
397 
398 /*
399  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
400  * we cannot do r3000 under these circumstances.
401  *
402  * Declare pgd_current here instead of including mmu_context.h to avoid type
403  * conflicts for tlbmiss_handler_setup_pgd
404  */
405 extern unsigned long pgd_current[];
406 
407 /*
408  * The R3000 TLB handler is simple.
409  */
410 static void build_r3000_tlb_refill_handler(void)
411 {
412 	long pgdc = (long)pgd_current;
413 	u32 *p;
414 
415 	memset(tlb_handler, 0, sizeof(tlb_handler));
416 	p = tlb_handler;
417 
418 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
419 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
420 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
421 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
422 	uasm_i_sll(&p, K0, K0, 2);
423 	uasm_i_addu(&p, K1, K1, K0);
424 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
425 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
426 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
427 	uasm_i_addu(&p, K1, K1, K0);
428 	uasm_i_lw(&p, K0, 0, K1);
429 	uasm_i_nop(&p); /* load delay */
430 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
431 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
432 	uasm_i_tlbwr(&p); /* cp0 delay */
433 	uasm_i_jr(&p, K1);
434 	uasm_i_rfe(&p); /* branch delay */
435 
436 	if (p > tlb_handler + 32)
437 		panic("TLB refill handler space exceeded");
438 
439 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
440 		 (unsigned int)(p - tlb_handler));
441 
442 	memcpy((void *)ebase, tlb_handler, 0x80);
443 	local_flush_icache_range(ebase, ebase + 0x80);
444 
445 	dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
446 }
447 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
448 
449 /*
450  * The R4000 TLB handler is much more complicated. We have two
451  * consecutive handler areas with 32 instructions space each.
452  * Since they aren't used at the same time, we can overflow in the
453  * other one.To keep things simple, we first assume linear space,
454  * then we relocate it to the final handler layout as needed.
455  */
456 static u32 final_handler[64];
457 
458 /*
459  * Hazards
460  *
461  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
462  * 2. A timing hazard exists for the TLBP instruction.
463  *
464  *	stalling_instruction
465  *	TLBP
466  *
467  * The JTLB is being read for the TLBP throughout the stall generated by the
468  * previous instruction. This is not really correct as the stalling instruction
469  * can modify the address used to access the JTLB.  The failure symptom is that
470  * the TLBP instruction will use an address created for the stalling instruction
471  * and not the address held in C0_ENHI and thus report the wrong results.
472  *
473  * The software work-around is to not allow the instruction preceding the TLBP
474  * to stall - make it an NOP or some other instruction guaranteed not to stall.
475  *
476  * Errata 2 will not be fixed.	This errata is also on the R5000.
477  *
478  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
479  */
480 static void __maybe_unused build_tlb_probe_entry(u32 **p)
481 {
482 	switch (current_cpu_type()) {
483 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
484 	case CPU_R4600:
485 	case CPU_R4700:
486 	case CPU_R5000:
487 	case CPU_NEVADA:
488 		uasm_i_nop(p);
489 		uasm_i_tlbp(p);
490 		break;
491 
492 	default:
493 		uasm_i_tlbp(p);
494 		break;
495 	}
496 }
497 
498 /*
499  * Write random or indexed TLB entry, and care about the hazards from
500  * the preceding mtc0 and for the following eret.
501  */
502 enum tlb_write_entry { tlb_random, tlb_indexed };
503 
504 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
505 				  struct uasm_reloc **r,
506 				  enum tlb_write_entry wmode)
507 {
508 	void(*tlbw)(u32 **) = NULL;
509 
510 	switch (wmode) {
511 	case tlb_random: tlbw = uasm_i_tlbwr; break;
512 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
513 	}
514 
515 	if (cpu_has_mips_r2_r6) {
516 		if (cpu_has_mips_r2_exec_hazard)
517 			uasm_i_ehb(p);
518 		tlbw(p);
519 		return;
520 	}
521 
522 	switch (current_cpu_type()) {
523 	case CPU_R4000PC:
524 	case CPU_R4000SC:
525 	case CPU_R4000MC:
526 	case CPU_R4400PC:
527 	case CPU_R4400SC:
528 	case CPU_R4400MC:
529 		/*
530 		 * This branch uses up a mtc0 hazard nop slot and saves
531 		 * two nops after the tlbw instruction.
532 		 */
533 		uasm_bgezl_hazard(p, r, hazard_instance);
534 		tlbw(p);
535 		uasm_bgezl_label(l, p, hazard_instance);
536 		hazard_instance++;
537 		uasm_i_nop(p);
538 		break;
539 
540 	case CPU_R4600:
541 	case CPU_R4700:
542 		uasm_i_nop(p);
543 		tlbw(p);
544 		uasm_i_nop(p);
545 		break;
546 
547 	case CPU_R5000:
548 	case CPU_NEVADA:
549 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
550 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
551 		tlbw(p);
552 		break;
553 
554 	case CPU_R4300:
555 	case CPU_5KC:
556 	case CPU_TX49XX:
557 	case CPU_PR4450:
558 	case CPU_XLR:
559 		uasm_i_nop(p);
560 		tlbw(p);
561 		break;
562 
563 	case CPU_R10000:
564 	case CPU_R12000:
565 	case CPU_R14000:
566 	case CPU_R16000:
567 	case CPU_4KC:
568 	case CPU_4KEC:
569 	case CPU_M14KC:
570 	case CPU_M14KEC:
571 	case CPU_SB1:
572 	case CPU_SB1A:
573 	case CPU_4KSC:
574 	case CPU_20KC:
575 	case CPU_25KF:
576 	case CPU_BMIPS32:
577 	case CPU_BMIPS3300:
578 	case CPU_BMIPS4350:
579 	case CPU_BMIPS4380:
580 	case CPU_BMIPS5000:
581 	case CPU_LOONGSON2:
582 	case CPU_LOONGSON3:
583 	case CPU_R5500:
584 		if (m4kc_tlbp_war())
585 			uasm_i_nop(p);
586 	case CPU_ALCHEMY:
587 		tlbw(p);
588 		break;
589 
590 	case CPU_RM7000:
591 		uasm_i_nop(p);
592 		uasm_i_nop(p);
593 		uasm_i_nop(p);
594 		uasm_i_nop(p);
595 		tlbw(p);
596 		break;
597 
598 	case CPU_VR4111:
599 	case CPU_VR4121:
600 	case CPU_VR4122:
601 	case CPU_VR4181:
602 	case CPU_VR4181A:
603 		uasm_i_nop(p);
604 		uasm_i_nop(p);
605 		tlbw(p);
606 		uasm_i_nop(p);
607 		uasm_i_nop(p);
608 		break;
609 
610 	case CPU_VR4131:
611 	case CPU_VR4133:
612 	case CPU_R5432:
613 		uasm_i_nop(p);
614 		uasm_i_nop(p);
615 		tlbw(p);
616 		break;
617 
618 	case CPU_JZRISC:
619 		tlbw(p);
620 		uasm_i_nop(p);
621 		break;
622 
623 	default:
624 		panic("No TLB refill handler yet (CPU type: %d)",
625 		      current_cpu_type());
626 		break;
627 	}
628 }
629 
630 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
631 							unsigned int reg)
632 {
633 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
634 		if (fill_includes_sw_bits) {
635 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
636 		} else {
637 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
638 			UASM_i_ROTR(p, reg, reg,
639 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
640 		}
641 	} else {
642 #ifdef CONFIG_PHYS_ADDR_T_64BIT
643 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
644 #else
645 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
646 #endif
647 	}
648 }
649 
650 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
651 
652 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
653 				   unsigned int tmp, enum label_id lid,
654 				   int restore_scratch)
655 {
656 	if (restore_scratch) {
657 		/* Reset default page size */
658 		if (PM_DEFAULT_MASK >> 16) {
659 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
660 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
661 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
662 			uasm_il_b(p, r, lid);
663 		} else if (PM_DEFAULT_MASK) {
664 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
665 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
666 			uasm_il_b(p, r, lid);
667 		} else {
668 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
669 			uasm_il_b(p, r, lid);
670 		}
671 		if (scratch_reg >= 0)
672 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
673 		else
674 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
675 	} else {
676 		/* Reset default page size */
677 		if (PM_DEFAULT_MASK >> 16) {
678 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
679 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
680 			uasm_il_b(p, r, lid);
681 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
682 		} else if (PM_DEFAULT_MASK) {
683 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
684 			uasm_il_b(p, r, lid);
685 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
686 		} else {
687 			uasm_il_b(p, r, lid);
688 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
689 		}
690 	}
691 }
692 
693 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
694 				       struct uasm_reloc **r,
695 				       unsigned int tmp,
696 				       enum tlb_write_entry wmode,
697 				       int restore_scratch)
698 {
699 	/* Set huge page tlb entry size */
700 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
701 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
702 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
703 
704 	build_tlb_write_entry(p, l, r, wmode);
705 
706 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
707 }
708 
709 /*
710  * Check if Huge PTE is present, if so then jump to LABEL.
711  */
712 static void
713 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
714 		  unsigned int pmd, int lid)
715 {
716 	UASM_i_LW(p, tmp, 0, pmd);
717 	if (use_bbit_insns()) {
718 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
719 	} else {
720 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
721 		uasm_il_bnez(p, r, tmp, lid);
722 	}
723 }
724 
725 static void build_huge_update_entries(u32 **p, unsigned int pte,
726 				      unsigned int tmp)
727 {
728 	int small_sequence;
729 
730 	/*
731 	 * A huge PTE describes an area the size of the
732 	 * configured huge page size. This is twice the
733 	 * of the large TLB entry size we intend to use.
734 	 * A TLB entry half the size of the configured
735 	 * huge page size is configured into entrylo0
736 	 * and entrylo1 to cover the contiguous huge PTE
737 	 * address space.
738 	 */
739 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
740 
741 	/* We can clobber tmp.	It isn't used after this.*/
742 	if (!small_sequence)
743 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
744 
745 	build_convert_pte_to_entrylo(p, pte);
746 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
747 	/* convert to entrylo1 */
748 	if (small_sequence)
749 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
750 	else
751 		UASM_i_ADDU(p, pte, pte, tmp);
752 
753 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
754 }
755 
756 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
757 				    struct uasm_label **l,
758 				    unsigned int pte,
759 				    unsigned int ptr)
760 {
761 #ifdef CONFIG_SMP
762 	UASM_i_SC(p, pte, 0, ptr);
763 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
764 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
765 #else
766 	UASM_i_SW(p, pte, 0, ptr);
767 #endif
768 	build_huge_update_entries(p, pte, ptr);
769 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
770 }
771 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
772 
773 #ifdef CONFIG_64BIT
774 /*
775  * TMP and PTR are scratch.
776  * TMP will be clobbered, PTR will hold the pmd entry.
777  */
778 static void
779 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
780 		 unsigned int tmp, unsigned int ptr)
781 {
782 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
783 	long pgdc = (long)pgd_current;
784 #endif
785 	/*
786 	 * The vmalloc handling is not in the hotpath.
787 	 */
788 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
789 
790 	if (check_for_high_segbits) {
791 		/*
792 		 * The kernel currently implicitely assumes that the
793 		 * MIPS SEGBITS parameter for the processor is
794 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
795 		 * allocate virtual addresses outside the maximum
796 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
797 		 * that doesn't prevent user code from accessing the
798 		 * higher xuseg addresses.  Here, we make sure that
799 		 * everything but the lower xuseg addresses goes down
800 		 * the module_alloc/vmalloc path.
801 		 */
802 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
803 		uasm_il_bnez(p, r, ptr, label_vmalloc);
804 	} else {
805 		uasm_il_bltz(p, r, tmp, label_vmalloc);
806 	}
807 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
808 
809 	if (pgd_reg != -1) {
810 		/* pgd is in pgd_reg */
811 		UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
812 	} else {
813 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
814 		/*
815 		 * &pgd << 11 stored in CONTEXT [23..63].
816 		 */
817 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
818 
819 		/* Clear lower 23 bits of context. */
820 		uasm_i_dins(p, ptr, 0, 0, 23);
821 
822 		/* 1 0	1 0 1  << 6  xkphys cached */
823 		uasm_i_ori(p, ptr, ptr, 0x540);
824 		uasm_i_drotr(p, ptr, ptr, 11);
825 #elif defined(CONFIG_SMP)
826 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
827 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
828 		UASM_i_LA_mostly(p, tmp, pgdc);
829 		uasm_i_daddu(p, ptr, ptr, tmp);
830 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
831 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
832 #else
833 		UASM_i_LA_mostly(p, ptr, pgdc);
834 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
835 #endif
836 	}
837 
838 	uasm_l_vmalloc_done(l, *p);
839 
840 	/* get pgd offset in bytes */
841 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
842 
843 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
844 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
845 #ifndef __PAGETABLE_PMD_FOLDED
846 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
847 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
848 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
849 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
850 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
851 #endif
852 }
853 
854 /*
855  * BVADDR is the faulting address, PTR is scratch.
856  * PTR will hold the pgd for vmalloc.
857  */
858 static void
859 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
860 			unsigned int bvaddr, unsigned int ptr,
861 			enum vmalloc64_mode mode)
862 {
863 	long swpd = (long)swapper_pg_dir;
864 	int single_insn_swpd;
865 	int did_vmalloc_branch = 0;
866 
867 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
868 
869 	uasm_l_vmalloc(l, *p);
870 
871 	if (mode != not_refill && check_for_high_segbits) {
872 		if (single_insn_swpd) {
873 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
874 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
875 			did_vmalloc_branch = 1;
876 			/* fall through */
877 		} else {
878 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
879 		}
880 	}
881 	if (!did_vmalloc_branch) {
882 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
883 			uasm_il_b(p, r, label_vmalloc_done);
884 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
885 		} else {
886 			UASM_i_LA_mostly(p, ptr, swpd);
887 			uasm_il_b(p, r, label_vmalloc_done);
888 			if (uasm_in_compat_space_p(swpd))
889 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
890 			else
891 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
892 		}
893 	}
894 	if (mode != not_refill && check_for_high_segbits) {
895 		uasm_l_large_segbits_fault(l, *p);
896 		/*
897 		 * We get here if we are an xsseg address, or if we are
898 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
899 		 *
900 		 * Ignoring xsseg (assume disabled so would generate
901 		 * (address errors?), the only remaining possibility
902 		 * is the upper xuseg addresses.  On processors with
903 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
904 		 * addresses would have taken an address error. We try
905 		 * to mimic that here by taking a load/istream page
906 		 * fault.
907 		 */
908 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
909 		uasm_i_jr(p, ptr);
910 
911 		if (mode == refill_scratch) {
912 			if (scratch_reg >= 0)
913 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
914 			else
915 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
916 		} else {
917 			uasm_i_nop(p);
918 		}
919 	}
920 }
921 
922 #else /* !CONFIG_64BIT */
923 
924 /*
925  * TMP and PTR are scratch.
926  * TMP will be clobbered, PTR will hold the pgd entry.
927  */
928 static void __maybe_unused
929 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
930 {
931 	if (pgd_reg != -1) {
932 		/* pgd is in pgd_reg */
933 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
934 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
935 	} else {
936 		long pgdc = (long)pgd_current;
937 
938 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
939 #ifdef CONFIG_SMP
940 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
941 		UASM_i_LA_mostly(p, tmp, pgdc);
942 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
943 		uasm_i_addu(p, ptr, tmp, ptr);
944 #else
945 		UASM_i_LA_mostly(p, ptr, pgdc);
946 #endif
947 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
948 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
949 	}
950 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
951 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
952 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
953 }
954 
955 #endif /* !CONFIG_64BIT */
956 
957 static void build_adjust_context(u32 **p, unsigned int ctx)
958 {
959 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
960 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
961 
962 	switch (current_cpu_type()) {
963 	case CPU_VR41XX:
964 	case CPU_VR4111:
965 	case CPU_VR4121:
966 	case CPU_VR4122:
967 	case CPU_VR4131:
968 	case CPU_VR4181:
969 	case CPU_VR4181A:
970 	case CPU_VR4133:
971 		shift += 2;
972 		break;
973 
974 	default:
975 		break;
976 	}
977 
978 	if (shift)
979 		UASM_i_SRL(p, ctx, ctx, shift);
980 	uasm_i_andi(p, ctx, ctx, mask);
981 }
982 
983 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
984 {
985 	/*
986 	 * Bug workaround for the Nevada. It seems as if under certain
987 	 * circumstances the move from cp0_context might produce a
988 	 * bogus result when the mfc0 instruction and its consumer are
989 	 * in a different cacheline or a load instruction, probably any
990 	 * memory reference, is between them.
991 	 */
992 	switch (current_cpu_type()) {
993 	case CPU_NEVADA:
994 		UASM_i_LW(p, ptr, 0, ptr);
995 		GET_CONTEXT(p, tmp); /* get context reg */
996 		break;
997 
998 	default:
999 		GET_CONTEXT(p, tmp); /* get context reg */
1000 		UASM_i_LW(p, ptr, 0, ptr);
1001 		break;
1002 	}
1003 
1004 	build_adjust_context(p, tmp);
1005 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1006 }
1007 
1008 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1009 {
1010 	/*
1011 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1012 	 * Kernel is a special case. Only a few CPUs use it.
1013 	 */
1014 	if (config_enabled(CONFIG_PHYS_ADDR_T_64BIT) && !cpu_has_64bits) {
1015 		int pte_off_even = sizeof(pte_t) / 2;
1016 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1017 #ifdef CONFIG_XPA
1018 		const int scratch = 1; /* Our extra working register */
1019 
1020 		uasm_i_addu(p, scratch, 0, ptep);
1021 #endif
1022 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1023 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* odd pte */
1024 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1025 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1026 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1027 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1);
1028 #ifdef CONFIG_XPA
1029 		uasm_i_lw(p, tmp, 0, scratch);
1030 		uasm_i_lw(p, ptep, sizeof(pte_t), scratch);
1031 		uasm_i_lui(p, scratch, 0xff);
1032 		uasm_i_ori(p, scratch, scratch, 0xffff);
1033 		uasm_i_and(p, tmp, scratch, tmp);
1034 		uasm_i_and(p, ptep, scratch, ptep);
1035 		uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1036 		uasm_i_mthc0(p, ptep, C0_ENTRYLO1);
1037 #endif
1038 		return;
1039 	}
1040 
1041 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1042 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1043 	if (r45k_bvahwbug())
1044 		build_tlb_probe_entry(p);
1045 	build_convert_pte_to_entrylo(p, tmp);
1046 	if (r4k_250MHZhwbug())
1047 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1048 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1049 	build_convert_pte_to_entrylo(p, ptep);
1050 	if (r45k_bvahwbug())
1051 		uasm_i_mfc0(p, tmp, C0_INDEX);
1052 	if (r4k_250MHZhwbug())
1053 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1054 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1055 }
1056 
1057 struct mips_huge_tlb_info {
1058 	int huge_pte;
1059 	int restore_scratch;
1060 	bool need_reload_pte;
1061 };
1062 
1063 static struct mips_huge_tlb_info
1064 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1065 			       struct uasm_reloc **r, unsigned int tmp,
1066 			       unsigned int ptr, int c0_scratch_reg)
1067 {
1068 	struct mips_huge_tlb_info rv;
1069 	unsigned int even, odd;
1070 	int vmalloc_branch_delay_filled = 0;
1071 	const int scratch = 1; /* Our extra working register */
1072 
1073 	rv.huge_pte = scratch;
1074 	rv.restore_scratch = 0;
1075 	rv.need_reload_pte = false;
1076 
1077 	if (check_for_high_segbits) {
1078 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1079 
1080 		if (pgd_reg != -1)
1081 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1082 		else
1083 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1084 
1085 		if (c0_scratch_reg >= 0)
1086 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1087 		else
1088 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1089 
1090 		uasm_i_dsrl_safe(p, scratch, tmp,
1091 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1092 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1093 
1094 		if (pgd_reg == -1) {
1095 			vmalloc_branch_delay_filled = 1;
1096 			/* Clear lower 23 bits of context. */
1097 			uasm_i_dins(p, ptr, 0, 0, 23);
1098 		}
1099 	} else {
1100 		if (pgd_reg != -1)
1101 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1102 		else
1103 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1104 
1105 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1106 
1107 		if (c0_scratch_reg >= 0)
1108 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1109 		else
1110 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1111 
1112 		if (pgd_reg == -1)
1113 			/* Clear lower 23 bits of context. */
1114 			uasm_i_dins(p, ptr, 0, 0, 23);
1115 
1116 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1117 	}
1118 
1119 	if (pgd_reg == -1) {
1120 		vmalloc_branch_delay_filled = 1;
1121 		/* 1 0	1 0 1  << 6  xkphys cached */
1122 		uasm_i_ori(p, ptr, ptr, 0x540);
1123 		uasm_i_drotr(p, ptr, ptr, 11);
1124 	}
1125 
1126 #ifdef __PAGETABLE_PMD_FOLDED
1127 #define LOC_PTEP scratch
1128 #else
1129 #define LOC_PTEP ptr
1130 #endif
1131 
1132 	if (!vmalloc_branch_delay_filled)
1133 		/* get pgd offset in bytes */
1134 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1135 
1136 	uasm_l_vmalloc_done(l, *p);
1137 
1138 	/*
1139 	 *			   tmp		ptr
1140 	 * fall-through case =	 badvaddr  *pgd_current
1141 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1142 	 */
1143 
1144 	if (vmalloc_branch_delay_filled)
1145 		/* get pgd offset in bytes */
1146 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1147 
1148 #ifdef __PAGETABLE_PMD_FOLDED
1149 	GET_CONTEXT(p, tmp); /* get context reg */
1150 #endif
1151 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1152 
1153 	if (use_lwx_insns()) {
1154 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1155 	} else {
1156 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1157 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1158 	}
1159 
1160 #ifndef __PAGETABLE_PMD_FOLDED
1161 	/* get pmd offset in bytes */
1162 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1163 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1164 	GET_CONTEXT(p, tmp); /* get context reg */
1165 
1166 	if (use_lwx_insns()) {
1167 		UASM_i_LWX(p, scratch, scratch, ptr);
1168 	} else {
1169 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1170 		UASM_i_LW(p, scratch, 0, ptr);
1171 	}
1172 #endif
1173 	/* Adjust the context during the load latency. */
1174 	build_adjust_context(p, tmp);
1175 
1176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1177 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1178 	/*
1179 	 * The in the LWX case we don't want to do the load in the
1180 	 * delay slot.	It cannot issue in the same cycle and may be
1181 	 * speculative and unneeded.
1182 	 */
1183 	if (use_lwx_insns())
1184 		uasm_i_nop(p);
1185 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1186 
1187 
1188 	/* build_update_entries */
1189 	if (use_lwx_insns()) {
1190 		even = ptr;
1191 		odd = tmp;
1192 		UASM_i_LWX(p, even, scratch, tmp);
1193 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1194 		UASM_i_LWX(p, odd, scratch, tmp);
1195 	} else {
1196 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1197 		even = tmp;
1198 		odd = ptr;
1199 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1200 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1201 	}
1202 	if (cpu_has_rixi) {
1203 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1204 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1205 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1206 	} else {
1207 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1208 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1209 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1210 	}
1211 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1212 
1213 	if (c0_scratch_reg >= 0) {
1214 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1215 		build_tlb_write_entry(p, l, r, tlb_random);
1216 		uasm_l_leave(l, *p);
1217 		rv.restore_scratch = 1;
1218 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1219 		build_tlb_write_entry(p, l, r, tlb_random);
1220 		uasm_l_leave(l, *p);
1221 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1222 	} else {
1223 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1224 		build_tlb_write_entry(p, l, r, tlb_random);
1225 		uasm_l_leave(l, *p);
1226 		rv.restore_scratch = 1;
1227 	}
1228 
1229 	uasm_i_eret(p); /* return from trap */
1230 
1231 	return rv;
1232 }
1233 
1234 /*
1235  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1236  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1237  * slots before the XTLB refill exception handler which belong to the
1238  * unused TLB refill exception.
1239  */
1240 #define MIPS64_REFILL_INSNS 32
1241 
1242 static void build_r4000_tlb_refill_handler(void)
1243 {
1244 	u32 *p = tlb_handler;
1245 	struct uasm_label *l = labels;
1246 	struct uasm_reloc *r = relocs;
1247 	u32 *f;
1248 	unsigned int final_len;
1249 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1250 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1251 
1252 	memset(tlb_handler, 0, sizeof(tlb_handler));
1253 	memset(labels, 0, sizeof(labels));
1254 	memset(relocs, 0, sizeof(relocs));
1255 	memset(final_handler, 0, sizeof(final_handler));
1256 
1257 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1258 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1259 							  scratch_reg);
1260 		vmalloc_mode = refill_scratch;
1261 	} else {
1262 		htlb_info.huge_pte = K0;
1263 		htlb_info.restore_scratch = 0;
1264 		htlb_info.need_reload_pte = true;
1265 		vmalloc_mode = refill_noscratch;
1266 		/*
1267 		 * create the plain linear handler
1268 		 */
1269 		if (bcm1250_m3_war()) {
1270 			unsigned int segbits = 44;
1271 
1272 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1273 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1274 			uasm_i_xor(&p, K0, K0, K1);
1275 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1276 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1277 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1278 			uasm_i_or(&p, K0, K0, K1);
1279 			uasm_il_bnez(&p, &r, K0, label_leave);
1280 			/* No need for uasm_i_nop */
1281 		}
1282 
1283 #ifdef CONFIG_64BIT
1284 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1285 #else
1286 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1287 #endif
1288 
1289 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1290 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1291 #endif
1292 
1293 		build_get_ptep(&p, K0, K1);
1294 		build_update_entries(&p, K0, K1);
1295 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1296 		uasm_l_leave(&l, p);
1297 		uasm_i_eret(&p); /* return from trap */
1298 	}
1299 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1300 	uasm_l_tlb_huge_update(&l, p);
1301 	if (htlb_info.need_reload_pte)
1302 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1303 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1304 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1305 				   htlb_info.restore_scratch);
1306 #endif
1307 
1308 #ifdef CONFIG_64BIT
1309 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1310 #endif
1311 
1312 	/*
1313 	 * Overflow check: For the 64bit handler, we need at least one
1314 	 * free instruction slot for the wrap-around branch. In worst
1315 	 * case, if the intended insertion point is a delay slot, we
1316 	 * need three, with the second nop'ed and the third being
1317 	 * unused.
1318 	 */
1319 	switch (boot_cpu_type()) {
1320 	default:
1321 		if (sizeof(long) == 4) {
1322 	case CPU_LOONGSON2:
1323 		/* Loongson2 ebase is different than r4k, we have more space */
1324 			if ((p - tlb_handler) > 64)
1325 				panic("TLB refill handler space exceeded");
1326 			/*
1327 			 * Now fold the handler in the TLB refill handler space.
1328 			 */
1329 			f = final_handler;
1330 			/* Simplest case, just copy the handler. */
1331 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1332 			final_len = p - tlb_handler;
1333 			break;
1334 		} else {
1335 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1336 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1337 				&& uasm_insn_has_bdelay(relocs,
1338 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1339 				panic("TLB refill handler space exceeded");
1340 			/*
1341 			 * Now fold the handler in the TLB refill handler space.
1342 			 */
1343 			f = final_handler + MIPS64_REFILL_INSNS;
1344 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1345 				/* Just copy the handler. */
1346 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1347 				final_len = p - tlb_handler;
1348 			} else {
1349 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1350 				const enum label_id ls = label_tlb_huge_update;
1351 #else
1352 				const enum label_id ls = label_vmalloc;
1353 #endif
1354 				u32 *split;
1355 				int ov = 0;
1356 				int i;
1357 
1358 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1359 					;
1360 				BUG_ON(i == ARRAY_SIZE(labels));
1361 				split = labels[i].addr;
1362 
1363 				/*
1364 				 * See if we have overflown one way or the other.
1365 				 */
1366 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1367 				    split < p - MIPS64_REFILL_INSNS)
1368 					ov = 1;
1369 
1370 				if (ov) {
1371 					/*
1372 					 * Split two instructions before the end.  One
1373 					 * for the branch and one for the instruction
1374 					 * in the delay slot.
1375 					 */
1376 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1377 
1378 					/*
1379 					 * If the branch would fall in a delay slot,
1380 					 * we must back up an additional instruction
1381 					 * so that it is no longer in a delay slot.
1382 					 */
1383 					if (uasm_insn_has_bdelay(relocs, split - 1))
1384 						split--;
1385 				}
1386 				/* Copy first part of the handler. */
1387 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1388 				f += split - tlb_handler;
1389 
1390 				if (ov) {
1391 					/* Insert branch. */
1392 					uasm_l_split(&l, final_handler);
1393 					uasm_il_b(&f, &r, label_split);
1394 					if (uasm_insn_has_bdelay(relocs, split))
1395 						uasm_i_nop(&f);
1396 					else {
1397 						uasm_copy_handler(relocs, labels,
1398 								  split, split + 1, f);
1399 						uasm_move_labels(labels, f, f + 1, -1);
1400 						f++;
1401 						split++;
1402 					}
1403 				}
1404 
1405 				/* Copy the rest of the handler. */
1406 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1407 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1408 					    (p - split);
1409 			}
1410 		}
1411 		break;
1412 	}
1413 
1414 	uasm_resolve_relocs(relocs, labels);
1415 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1416 		 final_len);
1417 
1418 	memcpy((void *)ebase, final_handler, 0x100);
1419 	local_flush_icache_range(ebase, ebase + 0x100);
1420 
1421 	dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1422 }
1423 
1424 extern u32 handle_tlbl[], handle_tlbl_end[];
1425 extern u32 handle_tlbs[], handle_tlbs_end[];
1426 extern u32 handle_tlbm[], handle_tlbm_end[];
1427 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1428 extern u32 tlbmiss_handler_setup_pgd_end[];
1429 
1430 static void build_setup_pgd(void)
1431 {
1432 	const int a0 = 4;
1433 	const int __maybe_unused a1 = 5;
1434 	const int __maybe_unused a2 = 6;
1435 	u32 *p = tlbmiss_handler_setup_pgd_start;
1436 	const int tlbmiss_handler_setup_pgd_size =
1437 		tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1438 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1439 	long pgdc = (long)pgd_current;
1440 #endif
1441 
1442 	memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1443 					sizeof(tlbmiss_handler_setup_pgd[0]));
1444 	memset(labels, 0, sizeof(labels));
1445 	memset(relocs, 0, sizeof(relocs));
1446 	pgd_reg = allocate_kscratch();
1447 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1448 	if (pgd_reg == -1) {
1449 		struct uasm_label *l = labels;
1450 		struct uasm_reloc *r = relocs;
1451 
1452 		/* PGD << 11 in c0_Context */
1453 		/*
1454 		 * If it is a ckseg0 address, convert to a physical
1455 		 * address.  Shifting right by 29 and adding 4 will
1456 		 * result in zero for these addresses.
1457 		 *
1458 		 */
1459 		UASM_i_SRA(&p, a1, a0, 29);
1460 		UASM_i_ADDIU(&p, a1, a1, 4);
1461 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1462 		uasm_i_nop(&p);
1463 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1464 		uasm_l_tlbl_goaround1(&l, p);
1465 		UASM_i_SLL(&p, a0, a0, 11);
1466 		uasm_i_jr(&p, 31);
1467 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1468 	} else {
1469 		/* PGD in c0_KScratch */
1470 		uasm_i_jr(&p, 31);
1471 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1472 	}
1473 #else
1474 #ifdef CONFIG_SMP
1475 	/* Save PGD to pgd_current[smp_processor_id()] */
1476 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1477 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1478 	UASM_i_LA_mostly(&p, a2, pgdc);
1479 	UASM_i_ADDU(&p, a2, a2, a1);
1480 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1481 #else
1482 	UASM_i_LA_mostly(&p, a2, pgdc);
1483 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1484 #endif /* SMP */
1485 	uasm_i_jr(&p, 31);
1486 
1487 	/* if pgd_reg is allocated, save PGD also to scratch register */
1488 	if (pgd_reg != -1)
1489 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1490 	else
1491 		uasm_i_nop(&p);
1492 #endif
1493 	if (p >= tlbmiss_handler_setup_pgd_end)
1494 		panic("tlbmiss_handler_setup_pgd space exceeded");
1495 
1496 	uasm_resolve_relocs(relocs, labels);
1497 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1498 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1499 
1500 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1501 					tlbmiss_handler_setup_pgd_size);
1502 }
1503 
1504 static void
1505 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1506 {
1507 #ifdef CONFIG_SMP
1508 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1509 	if (cpu_has_64bits)
1510 		uasm_i_lld(p, pte, 0, ptr);
1511 	else
1512 # endif
1513 		UASM_i_LL(p, pte, 0, ptr);
1514 #else
1515 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1516 	if (cpu_has_64bits)
1517 		uasm_i_ld(p, pte, 0, ptr);
1518 	else
1519 # endif
1520 		UASM_i_LW(p, pte, 0, ptr);
1521 #endif
1522 }
1523 
1524 static void
1525 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1526 	unsigned int mode)
1527 {
1528 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1529 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1530 
1531 	if (!cpu_has_64bits) {
1532 		const int scratch = 1; /* Our extra working register */
1533 
1534 		uasm_i_lui(p, scratch, (mode >> 16));
1535 		uasm_i_or(p, pte, pte, scratch);
1536 	} else
1537 #endif
1538 	uasm_i_ori(p, pte, pte, mode);
1539 #ifdef CONFIG_SMP
1540 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1541 	if (cpu_has_64bits)
1542 		uasm_i_scd(p, pte, 0, ptr);
1543 	else
1544 # endif
1545 		UASM_i_SC(p, pte, 0, ptr);
1546 
1547 	if (r10000_llsc_war())
1548 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1549 	else
1550 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1551 
1552 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1553 	if (!cpu_has_64bits) {
1554 		/* no uasm_i_nop needed */
1555 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1556 		uasm_i_ori(p, pte, pte, hwmode);
1557 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1558 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1559 		/* no uasm_i_nop needed */
1560 		uasm_i_lw(p, pte, 0, ptr);
1561 	} else
1562 		uasm_i_nop(p);
1563 # else
1564 	uasm_i_nop(p);
1565 # endif
1566 #else
1567 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1568 	if (cpu_has_64bits)
1569 		uasm_i_sd(p, pte, 0, ptr);
1570 	else
1571 # endif
1572 		UASM_i_SW(p, pte, 0, ptr);
1573 
1574 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1575 	if (!cpu_has_64bits) {
1576 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1577 		uasm_i_ori(p, pte, pte, hwmode);
1578 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1579 		uasm_i_lw(p, pte, 0, ptr);
1580 	}
1581 # endif
1582 #endif
1583 }
1584 
1585 /*
1586  * Check if PTE is present, if not then jump to LABEL. PTR points to
1587  * the page table where this PTE is located, PTE will be re-loaded
1588  * with it's original value.
1589  */
1590 static void
1591 build_pte_present(u32 **p, struct uasm_reloc **r,
1592 		  int pte, int ptr, int scratch, enum label_id lid)
1593 {
1594 	int t = scratch >= 0 ? scratch : pte;
1595 	int cur = pte;
1596 
1597 	if (cpu_has_rixi) {
1598 		if (use_bbit_insns()) {
1599 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1600 			uasm_i_nop(p);
1601 		} else {
1602 			if (_PAGE_PRESENT_SHIFT) {
1603 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1604 				cur = t;
1605 			}
1606 			uasm_i_andi(p, t, cur, 1);
1607 			uasm_il_beqz(p, r, t, lid);
1608 			if (pte == t)
1609 				/* You lose the SMP race :-(*/
1610 				iPTE_LW(p, pte, ptr);
1611 		}
1612 	} else {
1613 		if (_PAGE_PRESENT_SHIFT) {
1614 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1615 			cur = t;
1616 		}
1617 		uasm_i_andi(p, t, cur,
1618 			(_PAGE_PRESENT | _PAGE_READ) >> _PAGE_PRESENT_SHIFT);
1619 		uasm_i_xori(p, t, t,
1620 			(_PAGE_PRESENT | _PAGE_READ) >> _PAGE_PRESENT_SHIFT);
1621 		uasm_il_bnez(p, r, t, lid);
1622 		if (pte == t)
1623 			/* You lose the SMP race :-(*/
1624 			iPTE_LW(p, pte, ptr);
1625 	}
1626 }
1627 
1628 /* Make PTE valid, store result in PTR. */
1629 static void
1630 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1631 		 unsigned int ptr)
1632 {
1633 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1634 
1635 	iPTE_SW(p, r, pte, ptr, mode);
1636 }
1637 
1638 /*
1639  * Check if PTE can be written to, if not branch to LABEL. Regardless
1640  * restore PTE with value from PTR when done.
1641  */
1642 static void
1643 build_pte_writable(u32 **p, struct uasm_reloc **r,
1644 		   unsigned int pte, unsigned int ptr, int scratch,
1645 		   enum label_id lid)
1646 {
1647 	int t = scratch >= 0 ? scratch : pte;
1648 	int cur = pte;
1649 
1650 	if (_PAGE_PRESENT_SHIFT) {
1651 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1652 		cur = t;
1653 	}
1654 	uasm_i_andi(p, t, cur,
1655 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1656 	uasm_i_xori(p, t, t,
1657 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1658 	uasm_il_bnez(p, r, t, lid);
1659 	if (pte == t)
1660 		/* You lose the SMP race :-(*/
1661 		iPTE_LW(p, pte, ptr);
1662 	else
1663 		uasm_i_nop(p);
1664 }
1665 
1666 /* Make PTE writable, update software status bits as well, then store
1667  * at PTR.
1668  */
1669 static void
1670 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1671 		 unsigned int ptr)
1672 {
1673 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1674 			     | _PAGE_DIRTY);
1675 
1676 	iPTE_SW(p, r, pte, ptr, mode);
1677 }
1678 
1679 /*
1680  * Check if PTE can be modified, if not branch to LABEL. Regardless
1681  * restore PTE with value from PTR when done.
1682  */
1683 static void
1684 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1685 		     unsigned int pte, unsigned int ptr, int scratch,
1686 		     enum label_id lid)
1687 {
1688 	if (use_bbit_insns()) {
1689 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1690 		uasm_i_nop(p);
1691 	} else {
1692 		int t = scratch >= 0 ? scratch : pte;
1693 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1694 		uasm_i_andi(p, t, t, 1);
1695 		uasm_il_beqz(p, r, t, lid);
1696 		if (pte == t)
1697 			/* You lose the SMP race :-(*/
1698 			iPTE_LW(p, pte, ptr);
1699 	}
1700 }
1701 
1702 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1703 
1704 
1705 /*
1706  * R3000 style TLB load/store/modify handlers.
1707  */
1708 
1709 /*
1710  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1711  * Then it returns.
1712  */
1713 static void
1714 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1715 {
1716 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1717 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1718 	uasm_i_tlbwi(p);
1719 	uasm_i_jr(p, tmp);
1720 	uasm_i_rfe(p); /* branch delay */
1721 }
1722 
1723 /*
1724  * This places the pte into ENTRYLO0 and writes it with tlbwi
1725  * or tlbwr as appropriate.  This is because the index register
1726  * may have the probe fail bit set as a result of a trap on a
1727  * kseg2 access, i.e. without refill.  Then it returns.
1728  */
1729 static void
1730 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1731 			     struct uasm_reloc **r, unsigned int pte,
1732 			     unsigned int tmp)
1733 {
1734 	uasm_i_mfc0(p, tmp, C0_INDEX);
1735 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1736 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1737 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1738 	uasm_i_tlbwi(p); /* cp0 delay */
1739 	uasm_i_jr(p, tmp);
1740 	uasm_i_rfe(p); /* branch delay */
1741 	uasm_l_r3000_write_probe_fail(l, *p);
1742 	uasm_i_tlbwr(p); /* cp0 delay */
1743 	uasm_i_jr(p, tmp);
1744 	uasm_i_rfe(p); /* branch delay */
1745 }
1746 
1747 static void
1748 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1749 				   unsigned int ptr)
1750 {
1751 	long pgdc = (long)pgd_current;
1752 
1753 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1754 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1755 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1756 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1757 	uasm_i_sll(p, pte, pte, 2);
1758 	uasm_i_addu(p, ptr, ptr, pte);
1759 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1760 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1761 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1762 	uasm_i_addu(p, ptr, ptr, pte);
1763 	uasm_i_lw(p, pte, 0, ptr);
1764 	uasm_i_tlbp(p); /* load delay */
1765 }
1766 
1767 static void build_r3000_tlb_load_handler(void)
1768 {
1769 	u32 *p = handle_tlbl;
1770 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1771 	struct uasm_label *l = labels;
1772 	struct uasm_reloc *r = relocs;
1773 
1774 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1775 	memset(labels, 0, sizeof(labels));
1776 	memset(relocs, 0, sizeof(relocs));
1777 
1778 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1779 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1780 	uasm_i_nop(&p); /* load delay */
1781 	build_make_valid(&p, &r, K0, K1);
1782 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1783 
1784 	uasm_l_nopage_tlbl(&l, p);
1785 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1786 	uasm_i_nop(&p);
1787 
1788 	if (p >= handle_tlbl_end)
1789 		panic("TLB load handler fastpath space exceeded");
1790 
1791 	uasm_resolve_relocs(relocs, labels);
1792 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1793 		 (unsigned int)(p - handle_tlbl));
1794 
1795 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1796 }
1797 
1798 static void build_r3000_tlb_store_handler(void)
1799 {
1800 	u32 *p = handle_tlbs;
1801 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1802 	struct uasm_label *l = labels;
1803 	struct uasm_reloc *r = relocs;
1804 
1805 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1806 	memset(labels, 0, sizeof(labels));
1807 	memset(relocs, 0, sizeof(relocs));
1808 
1809 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1810 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1811 	uasm_i_nop(&p); /* load delay */
1812 	build_make_write(&p, &r, K0, K1);
1813 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1814 
1815 	uasm_l_nopage_tlbs(&l, p);
1816 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1817 	uasm_i_nop(&p);
1818 
1819 	if (p >= handle_tlbs_end)
1820 		panic("TLB store handler fastpath space exceeded");
1821 
1822 	uasm_resolve_relocs(relocs, labels);
1823 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1824 		 (unsigned int)(p - handle_tlbs));
1825 
1826 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1827 }
1828 
1829 static void build_r3000_tlb_modify_handler(void)
1830 {
1831 	u32 *p = handle_tlbm;
1832 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1833 	struct uasm_label *l = labels;
1834 	struct uasm_reloc *r = relocs;
1835 
1836 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1837 	memset(labels, 0, sizeof(labels));
1838 	memset(relocs, 0, sizeof(relocs));
1839 
1840 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1841 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1842 	uasm_i_nop(&p); /* load delay */
1843 	build_make_write(&p, &r, K0, K1);
1844 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1845 
1846 	uasm_l_nopage_tlbm(&l, p);
1847 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1848 	uasm_i_nop(&p);
1849 
1850 	if (p >= handle_tlbm_end)
1851 		panic("TLB modify handler fastpath space exceeded");
1852 
1853 	uasm_resolve_relocs(relocs, labels);
1854 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1855 		 (unsigned int)(p - handle_tlbm));
1856 
1857 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1858 }
1859 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1860 
1861 /*
1862  * R4000 style TLB load/store/modify handlers.
1863  */
1864 static struct work_registers
1865 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1866 				   struct uasm_reloc **r)
1867 {
1868 	struct work_registers wr = build_get_work_registers(p);
1869 
1870 #ifdef CONFIG_64BIT
1871 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1872 #else
1873 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1874 #endif
1875 
1876 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1877 	/*
1878 	 * For huge tlb entries, pmd doesn't contain an address but
1879 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1880 	 * see if we need to jump to huge tlb processing.
1881 	 */
1882 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1883 #endif
1884 
1885 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1886 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1887 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1888 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1889 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1890 
1891 #ifdef CONFIG_SMP
1892 	uasm_l_smp_pgtable_change(l, *p);
1893 #endif
1894 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1895 	if (!m4kc_tlbp_war()) {
1896 		build_tlb_probe_entry(p);
1897 		if (cpu_has_htw) {
1898 			/* race condition happens, leaving */
1899 			uasm_i_ehb(p);
1900 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
1901 			uasm_il_bltz(p, r, wr.r3, label_leave);
1902 			uasm_i_nop(p);
1903 		}
1904 	}
1905 	return wr;
1906 }
1907 
1908 static void
1909 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1910 				   struct uasm_reloc **r, unsigned int tmp,
1911 				   unsigned int ptr)
1912 {
1913 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1914 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1915 	build_update_entries(p, tmp, ptr);
1916 	build_tlb_write_entry(p, l, r, tlb_indexed);
1917 	uasm_l_leave(l, *p);
1918 	build_restore_work_registers(p);
1919 	uasm_i_eret(p); /* return from trap */
1920 
1921 #ifdef CONFIG_64BIT
1922 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1923 #endif
1924 }
1925 
1926 static void build_r4000_tlb_load_handler(void)
1927 {
1928 	u32 *p = handle_tlbl;
1929 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1930 	struct uasm_label *l = labels;
1931 	struct uasm_reloc *r = relocs;
1932 	struct work_registers wr;
1933 
1934 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1935 	memset(labels, 0, sizeof(labels));
1936 	memset(relocs, 0, sizeof(relocs));
1937 
1938 	if (bcm1250_m3_war()) {
1939 		unsigned int segbits = 44;
1940 
1941 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1942 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1943 		uasm_i_xor(&p, K0, K0, K1);
1944 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1945 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1946 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1947 		uasm_i_or(&p, K0, K0, K1);
1948 		uasm_il_bnez(&p, &r, K0, label_leave);
1949 		/* No need for uasm_i_nop */
1950 	}
1951 
1952 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1953 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1954 	if (m4kc_tlbp_war())
1955 		build_tlb_probe_entry(&p);
1956 
1957 	if (cpu_has_rixi && !cpu_has_rixiex) {
1958 		/*
1959 		 * If the page is not _PAGE_VALID, RI or XI could not
1960 		 * have triggered it.  Skip the expensive test..
1961 		 */
1962 		if (use_bbit_insns()) {
1963 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1964 				      label_tlbl_goaround1);
1965 		} else {
1966 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1967 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1968 		}
1969 		uasm_i_nop(&p);
1970 
1971 		uasm_i_tlbr(&p);
1972 
1973 		switch (current_cpu_type()) {
1974 		default:
1975 			if (cpu_has_mips_r2_exec_hazard) {
1976 				uasm_i_ehb(&p);
1977 
1978 		case CPU_CAVIUM_OCTEON:
1979 		case CPU_CAVIUM_OCTEON_PLUS:
1980 		case CPU_CAVIUM_OCTEON2:
1981 				break;
1982 			}
1983 		}
1984 
1985 		/* Examine  entrylo 0 or 1 based on ptr. */
1986 		if (use_bbit_insns()) {
1987 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1988 		} else {
1989 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1990 			uasm_i_beqz(&p, wr.r3, 8);
1991 		}
1992 		/* load it in the delay slot*/
1993 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1994 		/* load it if ptr is odd */
1995 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1996 		/*
1997 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1998 		 * XI must have triggered it.
1999 		 */
2000 		if (use_bbit_insns()) {
2001 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2002 			uasm_i_nop(&p);
2003 			uasm_l_tlbl_goaround1(&l, p);
2004 		} else {
2005 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2006 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2007 			uasm_i_nop(&p);
2008 		}
2009 		uasm_l_tlbl_goaround1(&l, p);
2010 	}
2011 	build_make_valid(&p, &r, wr.r1, wr.r2);
2012 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2013 
2014 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2015 	/*
2016 	 * This is the entry point when build_r4000_tlbchange_handler_head
2017 	 * spots a huge page.
2018 	 */
2019 	uasm_l_tlb_huge_update(&l, p);
2020 	iPTE_LW(&p, wr.r1, wr.r2);
2021 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2022 	build_tlb_probe_entry(&p);
2023 
2024 	if (cpu_has_rixi && !cpu_has_rixiex) {
2025 		/*
2026 		 * If the page is not _PAGE_VALID, RI or XI could not
2027 		 * have triggered it.  Skip the expensive test..
2028 		 */
2029 		if (use_bbit_insns()) {
2030 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2031 				      label_tlbl_goaround2);
2032 		} else {
2033 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2034 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2035 		}
2036 		uasm_i_nop(&p);
2037 
2038 		uasm_i_tlbr(&p);
2039 
2040 		switch (current_cpu_type()) {
2041 		default:
2042 			if (cpu_has_mips_r2_exec_hazard) {
2043 				uasm_i_ehb(&p);
2044 
2045 		case CPU_CAVIUM_OCTEON:
2046 		case CPU_CAVIUM_OCTEON_PLUS:
2047 		case CPU_CAVIUM_OCTEON2:
2048 				break;
2049 			}
2050 		}
2051 
2052 		/* Examine  entrylo 0 or 1 based on ptr. */
2053 		if (use_bbit_insns()) {
2054 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2055 		} else {
2056 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2057 			uasm_i_beqz(&p, wr.r3, 8);
2058 		}
2059 		/* load it in the delay slot*/
2060 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2061 		/* load it if ptr is odd */
2062 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2063 		/*
2064 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2065 		 * XI must have triggered it.
2066 		 */
2067 		if (use_bbit_insns()) {
2068 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2069 		} else {
2070 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2071 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2072 		}
2073 		if (PM_DEFAULT_MASK == 0)
2074 			uasm_i_nop(&p);
2075 		/*
2076 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2077 		 * it is restored in build_huge_tlb_write_entry.
2078 		 */
2079 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2080 
2081 		uasm_l_tlbl_goaround2(&l, p);
2082 	}
2083 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2084 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2085 #endif
2086 
2087 	uasm_l_nopage_tlbl(&l, p);
2088 	build_restore_work_registers(&p);
2089 #ifdef CONFIG_CPU_MICROMIPS
2090 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2091 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2092 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2093 		uasm_i_jr(&p, K0);
2094 	} else
2095 #endif
2096 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2097 	uasm_i_nop(&p);
2098 
2099 	if (p >= handle_tlbl_end)
2100 		panic("TLB load handler fastpath space exceeded");
2101 
2102 	uasm_resolve_relocs(relocs, labels);
2103 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2104 		 (unsigned int)(p - handle_tlbl));
2105 
2106 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2107 }
2108 
2109 static void build_r4000_tlb_store_handler(void)
2110 {
2111 	u32 *p = handle_tlbs;
2112 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2113 	struct uasm_label *l = labels;
2114 	struct uasm_reloc *r = relocs;
2115 	struct work_registers wr;
2116 
2117 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2118 	memset(labels, 0, sizeof(labels));
2119 	memset(relocs, 0, sizeof(relocs));
2120 
2121 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2122 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2123 	if (m4kc_tlbp_war())
2124 		build_tlb_probe_entry(&p);
2125 	build_make_write(&p, &r, wr.r1, wr.r2);
2126 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2127 
2128 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2129 	/*
2130 	 * This is the entry point when
2131 	 * build_r4000_tlbchange_handler_head spots a huge page.
2132 	 */
2133 	uasm_l_tlb_huge_update(&l, p);
2134 	iPTE_LW(&p, wr.r1, wr.r2);
2135 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2136 	build_tlb_probe_entry(&p);
2137 	uasm_i_ori(&p, wr.r1, wr.r1,
2138 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2139 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2140 #endif
2141 
2142 	uasm_l_nopage_tlbs(&l, p);
2143 	build_restore_work_registers(&p);
2144 #ifdef CONFIG_CPU_MICROMIPS
2145 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2146 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2147 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2148 		uasm_i_jr(&p, K0);
2149 	} else
2150 #endif
2151 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2152 	uasm_i_nop(&p);
2153 
2154 	if (p >= handle_tlbs_end)
2155 		panic("TLB store handler fastpath space exceeded");
2156 
2157 	uasm_resolve_relocs(relocs, labels);
2158 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2159 		 (unsigned int)(p - handle_tlbs));
2160 
2161 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2162 }
2163 
2164 static void build_r4000_tlb_modify_handler(void)
2165 {
2166 	u32 *p = handle_tlbm;
2167 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2168 	struct uasm_label *l = labels;
2169 	struct uasm_reloc *r = relocs;
2170 	struct work_registers wr;
2171 
2172 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2173 	memset(labels, 0, sizeof(labels));
2174 	memset(relocs, 0, sizeof(relocs));
2175 
2176 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2177 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2178 	if (m4kc_tlbp_war())
2179 		build_tlb_probe_entry(&p);
2180 	/* Present and writable bits set, set accessed and dirty bits. */
2181 	build_make_write(&p, &r, wr.r1, wr.r2);
2182 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2183 
2184 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2185 	/*
2186 	 * This is the entry point when
2187 	 * build_r4000_tlbchange_handler_head spots a huge page.
2188 	 */
2189 	uasm_l_tlb_huge_update(&l, p);
2190 	iPTE_LW(&p, wr.r1, wr.r2);
2191 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2192 	build_tlb_probe_entry(&p);
2193 	uasm_i_ori(&p, wr.r1, wr.r1,
2194 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2195 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2196 #endif
2197 
2198 	uasm_l_nopage_tlbm(&l, p);
2199 	build_restore_work_registers(&p);
2200 #ifdef CONFIG_CPU_MICROMIPS
2201 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2202 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2203 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2204 		uasm_i_jr(&p, K0);
2205 	} else
2206 #endif
2207 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2208 	uasm_i_nop(&p);
2209 
2210 	if (p >= handle_tlbm_end)
2211 		panic("TLB modify handler fastpath space exceeded");
2212 
2213 	uasm_resolve_relocs(relocs, labels);
2214 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2215 		 (unsigned int)(p - handle_tlbm));
2216 
2217 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2218 }
2219 
2220 static void flush_tlb_handlers(void)
2221 {
2222 	local_flush_icache_range((unsigned long)handle_tlbl,
2223 			   (unsigned long)handle_tlbl_end);
2224 	local_flush_icache_range((unsigned long)handle_tlbs,
2225 			   (unsigned long)handle_tlbs_end);
2226 	local_flush_icache_range((unsigned long)handle_tlbm,
2227 			   (unsigned long)handle_tlbm_end);
2228 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2229 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2230 }
2231 
2232 static void print_htw_config(void)
2233 {
2234 	unsigned long config;
2235 	unsigned int pwctl;
2236 	const int field = 2 * sizeof(unsigned long);
2237 
2238 	config = read_c0_pwfield();
2239 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2240 		field, config,
2241 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2242 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2243 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2244 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2245 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2246 
2247 	config = read_c0_pwsize();
2248 	pr_debug("PWSize  (0x%0*lx): GDW: 0x%02lx  UDW: 0x%02lx  MDW: 0x%02lx  PTW: 0x%02lx  PTEW: 0x%02lx\n",
2249 		field, config,
2250 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2251 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2252 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2253 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2254 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2255 
2256 	pwctl = read_c0_pwctl();
2257 	pr_debug("PWCtl   (0x%x): PWEn: 0x%x  DPH: 0x%x  HugePg: 0x%x  Psn: 0x%x\n",
2258 		pwctl,
2259 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2260 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2261 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2262 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2263 }
2264 
2265 static void config_htw_params(void)
2266 {
2267 	unsigned long pwfield, pwsize, ptei;
2268 	unsigned int config;
2269 
2270 	/*
2271 	 * We are using 2-level page tables, so we only need to
2272 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2273 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2274 	 * write values less than 0xc in these fields because the entire
2275 	 * write will be dropped. As a result of which, we must preserve
2276 	 * the original reset values and overwrite only what we really want.
2277 	 */
2278 
2279 	pwfield = read_c0_pwfield();
2280 	/* re-initialize the GDI field */
2281 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2282 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2283 	/* re-initialize the PTI field including the even/odd bit */
2284 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2285 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2286 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2287 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2288 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2289 	}
2290 	/* Set the PTEI right shift */
2291 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2292 	pwfield |= ptei;
2293 	write_c0_pwfield(pwfield);
2294 	/* Check whether the PTEI value is supported */
2295 	back_to_back_c0_hazard();
2296 	pwfield = read_c0_pwfield();
2297 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2298 		!= ptei) {
2299 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2300 			ptei);
2301 		/*
2302 		 * Drop option to avoid HTW being enabled via another path
2303 		 * (eg htw_reset())
2304 		 */
2305 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2306 		return;
2307 	}
2308 
2309 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2310 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2311 	if (CONFIG_PGTABLE_LEVELS >= 3)
2312 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2313 
2314 	/* If XPA has been enabled, PTEs are 64-bit in size. */
2315 	if (config_enabled(CONFIG_64BITS) || (read_c0_pagegrain() & PG_ELPA))
2316 		pwsize |= 1;
2317 
2318 	write_c0_pwsize(pwsize);
2319 
2320 	/* Make sure everything is set before we enable the HTW */
2321 	back_to_back_c0_hazard();
2322 
2323 	/* Enable HTW and disable the rest of the pwctl fields */
2324 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2325 	write_c0_pwctl(config);
2326 	pr_info("Hardware Page Table Walker enabled\n");
2327 
2328 	print_htw_config();
2329 }
2330 
2331 static void config_xpa_params(void)
2332 {
2333 #ifdef CONFIG_XPA
2334 	unsigned int pagegrain;
2335 
2336 	if (mips_xpa_disabled) {
2337 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2338 		return;
2339 	}
2340 
2341 	pagegrain = read_c0_pagegrain();
2342 	write_c0_pagegrain(pagegrain | PG_ELPA);
2343 	back_to_back_c0_hazard();
2344 	pagegrain = read_c0_pagegrain();
2345 
2346 	if (pagegrain & PG_ELPA)
2347 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2348 	else
2349 		panic("Extended Physical Addressing (XPA) disabled");
2350 #endif
2351 }
2352 
2353 static void check_pabits(void)
2354 {
2355 	unsigned long entry;
2356 	unsigned pabits, fillbits;
2357 
2358 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2359 		/*
2360 		 * We'll only be making use of the fact that we can rotate bits
2361 		 * into the fill if the CPU supports RIXI, so don't bother
2362 		 * probing this for CPUs which don't.
2363 		 */
2364 		return;
2365 	}
2366 
2367 	write_c0_entrylo0(~0ul);
2368 	back_to_back_c0_hazard();
2369 	entry = read_c0_entrylo0();
2370 
2371 	/* clear all non-PFN bits */
2372 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2373 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2374 
2375 	/* find a lower bound on PABITS, and upper bound on fill bits */
2376 	pabits = fls_long(entry) + 6;
2377 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2378 
2379 	/* minus the RI & XI bits */
2380 	fillbits -= min_t(unsigned, fillbits, 2);
2381 
2382 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2383 		fill_includes_sw_bits = true;
2384 
2385 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2386 }
2387 
2388 void build_tlb_refill_handler(void)
2389 {
2390 	/*
2391 	 * The refill handler is generated per-CPU, multi-node systems
2392 	 * may have local storage for it. The other handlers are only
2393 	 * needed once.
2394 	 */
2395 	static int run_once = 0;
2396 
2397 	output_pgtable_bits_defines();
2398 	check_pabits();
2399 
2400 #ifdef CONFIG_64BIT
2401 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2402 #endif
2403 
2404 	switch (current_cpu_type()) {
2405 	case CPU_R2000:
2406 	case CPU_R3000:
2407 	case CPU_R3000A:
2408 	case CPU_R3081E:
2409 	case CPU_TX3912:
2410 	case CPU_TX3922:
2411 	case CPU_TX3927:
2412 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2413 		if (cpu_has_local_ebase)
2414 			build_r3000_tlb_refill_handler();
2415 		if (!run_once) {
2416 			if (!cpu_has_local_ebase)
2417 				build_r3000_tlb_refill_handler();
2418 			build_setup_pgd();
2419 			build_r3000_tlb_load_handler();
2420 			build_r3000_tlb_store_handler();
2421 			build_r3000_tlb_modify_handler();
2422 			flush_tlb_handlers();
2423 			run_once++;
2424 		}
2425 #else
2426 		panic("No R3000 TLB refill handler");
2427 #endif
2428 		break;
2429 
2430 	case CPU_R6000:
2431 	case CPU_R6000A:
2432 		panic("No R6000 TLB refill handler yet");
2433 		break;
2434 
2435 	case CPU_R8000:
2436 		panic("No R8000 TLB refill handler yet");
2437 		break;
2438 
2439 	default:
2440 		if (!run_once) {
2441 			scratch_reg = allocate_kscratch();
2442 			build_setup_pgd();
2443 			build_r4000_tlb_load_handler();
2444 			build_r4000_tlb_store_handler();
2445 			build_r4000_tlb_modify_handler();
2446 			if (!cpu_has_local_ebase)
2447 				build_r4000_tlb_refill_handler();
2448 			flush_tlb_handlers();
2449 			run_once++;
2450 		}
2451 		if (cpu_has_local_ebase)
2452 			build_r4000_tlb_refill_handler();
2453 		if (cpu_has_xpa)
2454 			config_xpa_params();
2455 		if (cpu_has_htw)
2456 			config_htw_params();
2457 	}
2458 }
2459