xref: /linux/arch/mips/mm/tlbex.c (revision 83869019c74cc2d01c96a3be2463a4eebe362224)
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 completely 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/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 #include <linux/pgtable.h>
32 
33 #include <asm/cacheflush.h>
34 #include <asm/cpu-type.h>
35 #include <asm/mmu_context.h>
36 #include <asm/war.h>
37 #include <asm/uasm.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
40 
41 static int mips_xpa_disabled;
42 
43 static int __init xpa_disable(char *s)
44 {
45 	mips_xpa_disabled = 1;
46 
47 	return 1;
48 }
49 
50 __setup("noxpa", xpa_disable);
51 
52 /*
53  * TLB load/store/modify handlers.
54  *
55  * Only the fastpath gets synthesized at runtime, the slowpath for
56  * do_page_fault remains normal asm.
57  */
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
60 
61 struct work_registers {
62 	int r1;
63 	int r2;
64 	int r3;
65 };
66 
67 struct tlb_reg_save {
68 	unsigned long a;
69 	unsigned long b;
70 } ____cacheline_aligned_in_smp;
71 
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
73 
74 static inline int r45k_bvahwbug(void)
75 {
76 	/* XXX: We should probe for the presence of this bug, but we don't. */
77 	return 0;
78 }
79 
80 static inline int r4k_250MHZhwbug(void)
81 {
82 	/* XXX: We should probe for the presence of this bug, but we don't. */
83 	return 0;
84 }
85 
86 extern int sb1250_m3_workaround_needed(void);
87 
88 static inline int __maybe_unused bcm1250_m3_war(void)
89 {
90 	if (IS_ENABLED(CONFIG_SB1_PASS_2_WORKAROUNDS))
91 		return sb1250_m3_workaround_needed();
92 	return 0;
93 }
94 
95 static inline int __maybe_unused r10000_llsc_war(void)
96 {
97 	return IS_ENABLED(CONFIG_WAR_R10000_LLSC);
98 }
99 
100 static int use_bbit_insns(void)
101 {
102 	switch (current_cpu_type()) {
103 	case CPU_CAVIUM_OCTEON:
104 	case CPU_CAVIUM_OCTEON_PLUS:
105 	case CPU_CAVIUM_OCTEON2:
106 	case CPU_CAVIUM_OCTEON3:
107 		return 1;
108 	default:
109 		return 0;
110 	}
111 }
112 
113 static int use_lwx_insns(void)
114 {
115 	switch (current_cpu_type()) {
116 	case CPU_CAVIUM_OCTEON2:
117 	case CPU_CAVIUM_OCTEON3:
118 		return 1;
119 	default:
120 		return 0;
121 	}
122 }
123 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
124     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
125 static bool scratchpad_available(void)
126 {
127 	return true;
128 }
129 static int scratchpad_offset(int i)
130 {
131 	/*
132 	 * CVMSEG starts at address -32768 and extends for
133 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
134 	 */
135 	i += 1; /* Kernel use starts at the top and works down. */
136 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
137 }
138 #else
139 static bool scratchpad_available(void)
140 {
141 	return false;
142 }
143 static int scratchpad_offset(int i)
144 {
145 	BUG();
146 	/* Really unreachable, but evidently some GCC want this. */
147 	return 0;
148 }
149 #endif
150 /*
151  * Found by experiment: At least some revisions of the 4kc throw under
152  * some circumstances a machine check exception, triggered by invalid
153  * values in the index register.  Delaying the tlbp instruction until
154  * after the next branch,  plus adding an additional nop in front of
155  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
156  * why; it's not an issue caused by the core RTL.
157  *
158  */
159 static int m4kc_tlbp_war(void)
160 {
161 	return current_cpu_type() == CPU_4KC;
162 }
163 
164 /* Handle labels (which must be positive integers). */
165 enum label_id {
166 	label_second_part = 1,
167 	label_leave,
168 	label_vmalloc,
169 	label_vmalloc_done,
170 	label_tlbw_hazard_0,
171 	label_split = label_tlbw_hazard_0 + 8,
172 	label_tlbl_goaround1,
173 	label_tlbl_goaround2,
174 	label_nopage_tlbl,
175 	label_nopage_tlbs,
176 	label_nopage_tlbm,
177 	label_smp_pgtable_change,
178 	label_r3000_write_probe_fail,
179 	label_large_segbits_fault,
180 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
181 	label_tlb_huge_update,
182 #endif
183 };
184 
185 UASM_L_LA(_second_part)
186 UASM_L_LA(_leave)
187 UASM_L_LA(_vmalloc)
188 UASM_L_LA(_vmalloc_done)
189 /* _tlbw_hazard_x is handled differently.  */
190 UASM_L_LA(_split)
191 UASM_L_LA(_tlbl_goaround1)
192 UASM_L_LA(_tlbl_goaround2)
193 UASM_L_LA(_nopage_tlbl)
194 UASM_L_LA(_nopage_tlbs)
195 UASM_L_LA(_nopage_tlbm)
196 UASM_L_LA(_smp_pgtable_change)
197 UASM_L_LA(_r3000_write_probe_fail)
198 UASM_L_LA(_large_segbits_fault)
199 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
200 UASM_L_LA(_tlb_huge_update)
201 #endif
202 
203 static int hazard_instance;
204 
205 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
206 {
207 	switch (instance) {
208 	case 0 ... 7:
209 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
210 		return;
211 	default:
212 		BUG();
213 	}
214 }
215 
216 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
217 {
218 	switch (instance) {
219 	case 0 ... 7:
220 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
221 		break;
222 	default:
223 		BUG();
224 	}
225 }
226 
227 /*
228  * pgtable bits are assigned dynamically depending on processor feature
229  * and statically based on kernel configuration.  This spits out the actual
230  * values the kernel is using.	Required to make sense from disassembled
231  * TLB exception handlers.
232  */
233 static void output_pgtable_bits_defines(void)
234 {
235 #define pr_define(fmt, ...)					\
236 	pr_debug("#define " fmt, ##__VA_ARGS__)
237 
238 	pr_debug("#include <asm/asm.h>\n");
239 	pr_debug("#include <asm/regdef.h>\n");
240 	pr_debug("\n");
241 
242 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
243 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
244 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
245 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
246 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
247 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
248 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
249 #endif
250 #ifdef _PAGE_NO_EXEC_SHIFT
251 	if (cpu_has_rixi)
252 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
253 #endif
254 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
255 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
256 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
257 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
258 	pr_debug("\n");
259 }
260 
261 static inline void dump_handler(const char *symbol, const void *start, const void *end)
262 {
263 	unsigned int count = (end - start) / sizeof(u32);
264 	const u32 *handler = start;
265 	int i;
266 
267 	pr_debug("LEAF(%s)\n", symbol);
268 
269 	pr_debug("\t.set push\n");
270 	pr_debug("\t.set noreorder\n");
271 
272 	for (i = 0; i < count; i++)
273 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
274 
275 	pr_debug("\t.set\tpop\n");
276 
277 	pr_debug("\tEND(%s)\n", symbol);
278 }
279 
280 /* The only general purpose registers allowed in TLB handlers. */
281 #define K0		26
282 #define K1		27
283 
284 /* Some CP0 registers */
285 #define C0_INDEX	0, 0
286 #define C0_ENTRYLO0	2, 0
287 #define C0_TCBIND	2, 2
288 #define C0_ENTRYLO1	3, 0
289 #define C0_CONTEXT	4, 0
290 #define C0_PAGEMASK	5, 0
291 #define C0_PWBASE	5, 5
292 #define C0_PWFIELD	5, 6
293 #define C0_PWSIZE	5, 7
294 #define C0_PWCTL	6, 6
295 #define C0_BADVADDR	8, 0
296 #define C0_PGD		9, 7
297 #define C0_ENTRYHI	10, 0
298 #define C0_EPC		14, 0
299 #define C0_XCONTEXT	20, 0
300 
301 #ifdef CONFIG_64BIT
302 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
303 #else
304 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
305 #endif
306 
307 /* The worst case length of the handler is around 18 instructions for
308  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
309  * Maximum space available is 32 instructions for R3000 and 64
310  * instructions for R4000.
311  *
312  * We deliberately chose a buffer size of 128, so we won't scribble
313  * over anything important on overflow before we panic.
314  */
315 static u32 tlb_handler[128];
316 
317 /* simply assume worst case size for labels and relocs */
318 static struct uasm_label labels[128];
319 static struct uasm_reloc relocs[128];
320 
321 static int check_for_high_segbits;
322 static bool fill_includes_sw_bits;
323 
324 static unsigned int kscratch_used_mask;
325 
326 static inline int __maybe_unused c0_kscratch(void)
327 {
328 	return 31;
329 }
330 
331 static int allocate_kscratch(void)
332 {
333 	int r;
334 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
335 
336 	r = ffs(a);
337 
338 	if (r == 0)
339 		return -1;
340 
341 	r--; /* make it zero based */
342 
343 	kscratch_used_mask |= (1 << r);
344 
345 	return r;
346 }
347 
348 static int scratch_reg;
349 int pgd_reg;
350 EXPORT_SYMBOL_GPL(pgd_reg);
351 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
352 
353 static struct work_registers build_get_work_registers(u32 **p)
354 {
355 	struct work_registers r;
356 
357 	if (scratch_reg >= 0) {
358 		/* Save in CPU local C0_KScratch? */
359 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
360 		r.r1 = K0;
361 		r.r2 = K1;
362 		r.r3 = 1;
363 		return r;
364 	}
365 
366 	if (num_possible_cpus() > 1) {
367 		/* Get smp_processor_id */
368 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
369 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
370 
371 		/* handler_reg_save index in K0 */
372 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
373 
374 		UASM_i_LA(p, K1, (long)&handler_reg_save);
375 		UASM_i_ADDU(p, K0, K0, K1);
376 	} else {
377 		UASM_i_LA(p, K0, (long)&handler_reg_save);
378 	}
379 	/* K0 now points to save area, save $1 and $2  */
380 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
381 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
382 
383 	r.r1 = K1;
384 	r.r2 = 1;
385 	r.r3 = 2;
386 	return r;
387 }
388 
389 static void build_restore_work_registers(u32 **p)
390 {
391 	if (scratch_reg >= 0) {
392 		uasm_i_ehb(p);
393 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
394 		return;
395 	}
396 	/* K0 already points to save area, restore $1 and $2  */
397 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
398 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
399 }
400 
401 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
402 
403 /*
404  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
405  * we cannot do r3000 under these circumstances.
406  *
407  * The R3000 TLB handler is simple.
408  */
409 static void build_r3000_tlb_refill_handler(void)
410 {
411 	long pgdc = (long)pgd_current;
412 	u32 *p;
413 
414 	memset(tlb_handler, 0, sizeof(tlb_handler));
415 	p = tlb_handler;
416 
417 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
418 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
419 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
420 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
421 	uasm_i_sll(&p, K0, K0, 2);
422 	uasm_i_addu(&p, K1, K1, K0);
423 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
424 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
425 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
426 	uasm_i_addu(&p, K1, K1, K0);
427 	uasm_i_lw(&p, K0, 0, K1);
428 	uasm_i_nop(&p); /* load delay */
429 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
430 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
431 	uasm_i_tlbwr(&p); /* cp0 delay */
432 	uasm_i_jr(&p, K1);
433 	uasm_i_rfe(&p); /* branch delay */
434 
435 	if (p > tlb_handler + 32)
436 		panic("TLB refill handler space exceeded");
437 
438 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
439 		 (unsigned int)(p - tlb_handler));
440 
441 	memcpy((void *)ebase, tlb_handler, 0x80);
442 	local_flush_icache_range(ebase, ebase + 0x80);
443 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
444 }
445 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
446 
447 /*
448  * The R4000 TLB handler is much more complicated. We have two
449  * consecutive handler areas with 32 instructions space each.
450  * Since they aren't used at the same time, we can overflow in the
451  * other one.To keep things simple, we first assume linear space,
452  * then we relocate it to the final handler layout as needed.
453  */
454 static u32 final_handler[64];
455 
456 /*
457  * Hazards
458  *
459  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
460  * 2. A timing hazard exists for the TLBP instruction.
461  *
462  *	stalling_instruction
463  *	TLBP
464  *
465  * The JTLB is being read for the TLBP throughout the stall generated by the
466  * previous instruction. This is not really correct as the stalling instruction
467  * can modify the address used to access the JTLB.  The failure symptom is that
468  * the TLBP instruction will use an address created for the stalling instruction
469  * and not the address held in C0_ENHI and thus report the wrong results.
470  *
471  * The software work-around is to not allow the instruction preceding the TLBP
472  * to stall - make it an NOP or some other instruction guaranteed not to stall.
473  *
474  * Errata 2 will not be fixed.	This errata is also on the R5000.
475  *
476  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
477  */
478 static void __maybe_unused build_tlb_probe_entry(u32 **p)
479 {
480 	switch (current_cpu_type()) {
481 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
482 	case CPU_R4600:
483 	case CPU_R4700:
484 	case CPU_R5000:
485 	case CPU_NEVADA:
486 		uasm_i_nop(p);
487 		uasm_i_tlbp(p);
488 		break;
489 
490 	default:
491 		uasm_i_tlbp(p);
492 		break;
493 	}
494 }
495 
496 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
497 			   struct uasm_reloc **r,
498 			   enum tlb_write_entry wmode)
499 {
500 	void(*tlbw)(u32 **) = NULL;
501 
502 	switch (wmode) {
503 	case tlb_random: tlbw = uasm_i_tlbwr; break;
504 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
505 	}
506 
507 	if (cpu_has_mips_r2_r6) {
508 		if (cpu_has_mips_r2_exec_hazard)
509 			uasm_i_ehb(p);
510 		tlbw(p);
511 		return;
512 	}
513 
514 	switch (current_cpu_type()) {
515 	case CPU_R4000PC:
516 	case CPU_R4000SC:
517 	case CPU_R4000MC:
518 	case CPU_R4400PC:
519 	case CPU_R4400SC:
520 	case CPU_R4400MC:
521 		/*
522 		 * This branch uses up a mtc0 hazard nop slot and saves
523 		 * two nops after the tlbw instruction.
524 		 */
525 		uasm_bgezl_hazard(p, r, hazard_instance);
526 		tlbw(p);
527 		uasm_bgezl_label(l, p, hazard_instance);
528 		hazard_instance++;
529 		uasm_i_nop(p);
530 		break;
531 
532 	case CPU_R4600:
533 	case CPU_R4700:
534 		uasm_i_nop(p);
535 		tlbw(p);
536 		uasm_i_nop(p);
537 		break;
538 
539 	case CPU_R5000:
540 	case CPU_NEVADA:
541 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
542 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
543 		tlbw(p);
544 		break;
545 
546 	case CPU_R4300:
547 	case CPU_5KC:
548 	case CPU_TX49XX:
549 	case CPU_PR4450:
550 		uasm_i_nop(p);
551 		tlbw(p);
552 		break;
553 
554 	case CPU_R10000:
555 	case CPU_R12000:
556 	case CPU_R14000:
557 	case CPU_R16000:
558 	case CPU_4KC:
559 	case CPU_4KEC:
560 	case CPU_M14KC:
561 	case CPU_M14KEC:
562 	case CPU_SB1:
563 	case CPU_SB1A:
564 	case CPU_4KSC:
565 	case CPU_20KC:
566 	case CPU_25KF:
567 	case CPU_BMIPS32:
568 	case CPU_BMIPS3300:
569 	case CPU_BMIPS4350:
570 	case CPU_BMIPS4380:
571 	case CPU_BMIPS5000:
572 	case CPU_LOONGSON2EF:
573 	case CPU_LOONGSON64:
574 	case CPU_R5500:
575 		if (m4kc_tlbp_war())
576 			uasm_i_nop(p);
577 		fallthrough;
578 	case CPU_ALCHEMY:
579 		tlbw(p);
580 		break;
581 
582 	case CPU_RM7000:
583 		uasm_i_nop(p);
584 		uasm_i_nop(p);
585 		uasm_i_nop(p);
586 		uasm_i_nop(p);
587 		tlbw(p);
588 		break;
589 
590 	case CPU_VR4111:
591 	case CPU_VR4121:
592 	case CPU_VR4122:
593 	case CPU_VR4181:
594 	case CPU_VR4181A:
595 		uasm_i_nop(p);
596 		uasm_i_nop(p);
597 		tlbw(p);
598 		uasm_i_nop(p);
599 		uasm_i_nop(p);
600 		break;
601 
602 	case CPU_VR4131:
603 	case CPU_VR4133:
604 		uasm_i_nop(p);
605 		uasm_i_nop(p);
606 		tlbw(p);
607 		break;
608 
609 	case CPU_XBURST:
610 		tlbw(p);
611 		uasm_i_nop(p);
612 		break;
613 
614 	default:
615 		panic("No TLB refill handler yet (CPU type: %d)",
616 		      current_cpu_type());
617 		break;
618 	}
619 }
620 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
621 
622 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
623 							unsigned int reg)
624 {
625 	if (_PAGE_GLOBAL_SHIFT == 0) {
626 		/* pte_t is already in EntryLo format */
627 		return;
628 	}
629 
630 	if (cpu_has_rixi && !!_PAGE_NO_EXEC) {
631 		if (fill_includes_sw_bits) {
632 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
633 		} else {
634 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
635 			UASM_i_ROTR(p, reg, reg,
636 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
637 		}
638 	} else {
639 #ifdef CONFIG_PHYS_ADDR_T_64BIT
640 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
641 #else
642 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
643 #endif
644 	}
645 }
646 
647 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
648 
649 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
650 				   unsigned int tmp, enum label_id lid,
651 				   int restore_scratch)
652 {
653 	if (restore_scratch) {
654 		/*
655 		 * Ensure the MFC0 below observes the value written to the
656 		 * KScratch register by the prior MTC0.
657 		 */
658 		if (scratch_reg >= 0)
659 			uasm_i_ehb(p);
660 
661 		/* Reset default page size */
662 		if (PM_DEFAULT_MASK >> 16) {
663 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
664 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
665 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
666 			uasm_il_b(p, r, lid);
667 		} else if (PM_DEFAULT_MASK) {
668 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
669 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
670 			uasm_il_b(p, r, lid);
671 		} else {
672 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
673 			uasm_il_b(p, r, lid);
674 		}
675 		if (scratch_reg >= 0)
676 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
677 		else
678 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
679 	} else {
680 		/* Reset default page size */
681 		if (PM_DEFAULT_MASK >> 16) {
682 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
683 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
684 			uasm_il_b(p, r, lid);
685 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
686 		} else if (PM_DEFAULT_MASK) {
687 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
688 			uasm_il_b(p, r, lid);
689 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
690 		} else {
691 			uasm_il_b(p, r, lid);
692 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
693 		}
694 	}
695 }
696 
697 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
698 				       struct uasm_reloc **r,
699 				       unsigned int tmp,
700 				       enum tlb_write_entry wmode,
701 				       int restore_scratch)
702 {
703 	/* Set huge page tlb entry size */
704 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
705 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
706 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
707 
708 	build_tlb_write_entry(p, l, r, wmode);
709 
710 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
711 }
712 
713 /*
714  * Check if Huge PTE is present, if so then jump to LABEL.
715  */
716 static void
717 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
718 		  unsigned int pmd, int lid)
719 {
720 	UASM_i_LW(p, tmp, 0, pmd);
721 	if (use_bbit_insns()) {
722 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
723 	} else {
724 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
725 		uasm_il_bnez(p, r, tmp, lid);
726 	}
727 }
728 
729 static void build_huge_update_entries(u32 **p, unsigned int pte,
730 				      unsigned int tmp)
731 {
732 	int small_sequence;
733 
734 	/*
735 	 * A huge PTE describes an area the size of the
736 	 * configured huge page size. This is twice the
737 	 * of the large TLB entry size we intend to use.
738 	 * A TLB entry half the size of the configured
739 	 * huge page size is configured into entrylo0
740 	 * and entrylo1 to cover the contiguous huge PTE
741 	 * address space.
742 	 */
743 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
744 
745 	/* We can clobber tmp.	It isn't used after this.*/
746 	if (!small_sequence)
747 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
748 
749 	build_convert_pte_to_entrylo(p, pte);
750 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
751 	/* convert to entrylo1 */
752 	if (small_sequence)
753 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
754 	else
755 		UASM_i_ADDU(p, pte, pte, tmp);
756 
757 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
758 }
759 
760 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
761 				    struct uasm_label **l,
762 				    unsigned int pte,
763 				    unsigned int ptr,
764 				    unsigned int flush)
765 {
766 #ifdef CONFIG_SMP
767 	UASM_i_SC(p, pte, 0, ptr);
768 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
769 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
770 #else
771 	UASM_i_SW(p, pte, 0, ptr);
772 #endif
773 	if (cpu_has_ftlb && flush) {
774 		BUG_ON(!cpu_has_tlbinv);
775 
776 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
777 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
778 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
779 		build_tlb_write_entry(p, l, r, tlb_indexed);
780 
781 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
782 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
783 		build_huge_update_entries(p, pte, ptr);
784 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
785 
786 		return;
787 	}
788 
789 	build_huge_update_entries(p, pte, ptr);
790 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
791 }
792 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
793 
794 #ifdef CONFIG_64BIT
795 /*
796  * TMP and PTR are scratch.
797  * TMP will be clobbered, PTR will hold the pmd entry.
798  */
799 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
800 		      unsigned int tmp, unsigned int ptr)
801 {
802 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
803 	long pgdc = (long)pgd_current;
804 #endif
805 	/*
806 	 * The vmalloc handling is not in the hotpath.
807 	 */
808 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
809 
810 	if (check_for_high_segbits) {
811 		/*
812 		 * The kernel currently implicitely assumes that the
813 		 * MIPS SEGBITS parameter for the processor is
814 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
815 		 * allocate virtual addresses outside the maximum
816 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
817 		 * that doesn't prevent user code from accessing the
818 		 * higher xuseg addresses.  Here, we make sure that
819 		 * everything but the lower xuseg addresses goes down
820 		 * the module_alloc/vmalloc path.
821 		 */
822 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
823 		uasm_il_bnez(p, r, ptr, label_vmalloc);
824 	} else {
825 		uasm_il_bltz(p, r, tmp, label_vmalloc);
826 	}
827 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
828 
829 	if (pgd_reg != -1) {
830 		/* pgd is in pgd_reg */
831 		if (cpu_has_ldpte)
832 			UASM_i_MFC0(p, ptr, C0_PWBASE);
833 		else
834 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
835 	} else {
836 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
837 		/*
838 		 * &pgd << 11 stored in CONTEXT [23..63].
839 		 */
840 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
841 
842 		/* Clear lower 23 bits of context. */
843 		uasm_i_dins(p, ptr, 0, 0, 23);
844 
845 		/* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
846 		uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
847 		uasm_i_drotr(p, ptr, ptr, 11);
848 #elif defined(CONFIG_SMP)
849 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
850 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
851 		UASM_i_LA_mostly(p, tmp, pgdc);
852 		uasm_i_daddu(p, ptr, ptr, tmp);
853 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
854 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
855 #else
856 		UASM_i_LA_mostly(p, ptr, pgdc);
857 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
858 #endif
859 	}
860 
861 	uasm_l_vmalloc_done(l, *p);
862 
863 	/* get pgd offset in bytes */
864 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
865 
866 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
867 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
868 #ifndef __PAGETABLE_PUD_FOLDED
869 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
870 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
871 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
872 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
873 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
874 #endif
875 #ifndef __PAGETABLE_PMD_FOLDED
876 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
877 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
878 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
879 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
880 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
881 #endif
882 }
883 EXPORT_SYMBOL_GPL(build_get_pmde64);
884 
885 /*
886  * BVADDR is the faulting address, PTR is scratch.
887  * PTR will hold the pgd for vmalloc.
888  */
889 static void
890 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
891 			unsigned int bvaddr, unsigned int ptr,
892 			enum vmalloc64_mode mode)
893 {
894 	long swpd = (long)swapper_pg_dir;
895 	int single_insn_swpd;
896 	int did_vmalloc_branch = 0;
897 
898 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
899 
900 	uasm_l_vmalloc(l, *p);
901 
902 	if (mode != not_refill && check_for_high_segbits) {
903 		if (single_insn_swpd) {
904 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
905 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
906 			did_vmalloc_branch = 1;
907 			/* fall through */
908 		} else {
909 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
910 		}
911 	}
912 	if (!did_vmalloc_branch) {
913 		if (single_insn_swpd) {
914 			uasm_il_b(p, r, label_vmalloc_done);
915 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
916 		} else {
917 			UASM_i_LA_mostly(p, ptr, swpd);
918 			uasm_il_b(p, r, label_vmalloc_done);
919 			if (uasm_in_compat_space_p(swpd))
920 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
921 			else
922 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
923 		}
924 	}
925 	if (mode != not_refill && check_for_high_segbits) {
926 		uasm_l_large_segbits_fault(l, *p);
927 
928 		if (mode == refill_scratch && scratch_reg >= 0)
929 			uasm_i_ehb(p);
930 
931 		/*
932 		 * We get here if we are an xsseg address, or if we are
933 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
934 		 *
935 		 * Ignoring xsseg (assume disabled so would generate
936 		 * (address errors?), the only remaining possibility
937 		 * is the upper xuseg addresses.  On processors with
938 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
939 		 * addresses would have taken an address error. We try
940 		 * to mimic that here by taking a load/istream page
941 		 * fault.
942 		 */
943 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
944 			uasm_i_sync(p, 0);
945 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
946 		uasm_i_jr(p, ptr);
947 
948 		if (mode == refill_scratch) {
949 			if (scratch_reg >= 0)
950 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
951 			else
952 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
953 		} else {
954 			uasm_i_nop(p);
955 		}
956 	}
957 }
958 
959 #else /* !CONFIG_64BIT */
960 
961 /*
962  * TMP and PTR are scratch.
963  * TMP will be clobbered, PTR will hold the pgd entry.
964  */
965 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
966 {
967 	if (pgd_reg != -1) {
968 		/* pgd is in pgd_reg */
969 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
970 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
971 	} else {
972 		long pgdc = (long)pgd_current;
973 
974 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
975 #ifdef CONFIG_SMP
976 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
977 		UASM_i_LA_mostly(p, tmp, pgdc);
978 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
979 		uasm_i_addu(p, ptr, tmp, ptr);
980 #else
981 		UASM_i_LA_mostly(p, ptr, pgdc);
982 #endif
983 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
984 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
985 	}
986 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
987 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
988 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
989 }
990 EXPORT_SYMBOL_GPL(build_get_pgde32);
991 
992 #endif /* !CONFIG_64BIT */
993 
994 static void build_adjust_context(u32 **p, unsigned int ctx)
995 {
996 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
997 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
998 
999 	switch (current_cpu_type()) {
1000 	case CPU_VR41XX:
1001 	case CPU_VR4111:
1002 	case CPU_VR4121:
1003 	case CPU_VR4122:
1004 	case CPU_VR4131:
1005 	case CPU_VR4181:
1006 	case CPU_VR4181A:
1007 	case CPU_VR4133:
1008 		shift += 2;
1009 		break;
1010 
1011 	default:
1012 		break;
1013 	}
1014 
1015 	if (shift)
1016 		UASM_i_SRL(p, ctx, ctx, shift);
1017 	uasm_i_andi(p, ctx, ctx, mask);
1018 }
1019 
1020 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1021 {
1022 	/*
1023 	 * Bug workaround for the Nevada. It seems as if under certain
1024 	 * circumstances the move from cp0_context might produce a
1025 	 * bogus result when the mfc0 instruction and its consumer are
1026 	 * in a different cacheline or a load instruction, probably any
1027 	 * memory reference, is between them.
1028 	 */
1029 	switch (current_cpu_type()) {
1030 	case CPU_NEVADA:
1031 		UASM_i_LW(p, ptr, 0, ptr);
1032 		GET_CONTEXT(p, tmp); /* get context reg */
1033 		break;
1034 
1035 	default:
1036 		GET_CONTEXT(p, tmp); /* get context reg */
1037 		UASM_i_LW(p, ptr, 0, ptr);
1038 		break;
1039 	}
1040 
1041 	build_adjust_context(p, tmp);
1042 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1043 }
1044 EXPORT_SYMBOL_GPL(build_get_ptep);
1045 
1046 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1047 {
1048 	int pte_off_even = 0;
1049 	int pte_off_odd = sizeof(pte_t);
1050 
1051 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1052 	/* The low 32 bits of EntryLo is stored in pte_high */
1053 	pte_off_even += offsetof(pte_t, pte_high);
1054 	pte_off_odd += offsetof(pte_t, pte_high);
1055 #endif
1056 
1057 	if (IS_ENABLED(CONFIG_XPA)) {
1058 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1059 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1060 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1061 
1062 		if (cpu_has_xpa && !mips_xpa_disabled) {
1063 			uasm_i_lw(p, tmp, 0, ptep);
1064 			uasm_i_ext(p, tmp, tmp, 0, 24);
1065 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1066 		}
1067 
1068 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1069 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1070 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1071 
1072 		if (cpu_has_xpa && !mips_xpa_disabled) {
1073 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1074 			uasm_i_ext(p, tmp, tmp, 0, 24);
1075 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1076 		}
1077 		return;
1078 	}
1079 
1080 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1081 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1082 	if (r45k_bvahwbug())
1083 		build_tlb_probe_entry(p);
1084 	build_convert_pte_to_entrylo(p, tmp);
1085 	if (r4k_250MHZhwbug())
1086 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1087 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1088 	build_convert_pte_to_entrylo(p, ptep);
1089 	if (r45k_bvahwbug())
1090 		uasm_i_mfc0(p, tmp, C0_INDEX);
1091 	if (r4k_250MHZhwbug())
1092 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1093 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1094 }
1095 EXPORT_SYMBOL_GPL(build_update_entries);
1096 
1097 struct mips_huge_tlb_info {
1098 	int huge_pte;
1099 	int restore_scratch;
1100 	bool need_reload_pte;
1101 };
1102 
1103 static struct mips_huge_tlb_info
1104 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1105 			       struct uasm_reloc **r, unsigned int tmp,
1106 			       unsigned int ptr, int c0_scratch_reg)
1107 {
1108 	struct mips_huge_tlb_info rv;
1109 	unsigned int even, odd;
1110 	int vmalloc_branch_delay_filled = 0;
1111 	const int scratch = 1; /* Our extra working register */
1112 
1113 	rv.huge_pte = scratch;
1114 	rv.restore_scratch = 0;
1115 	rv.need_reload_pte = false;
1116 
1117 	if (check_for_high_segbits) {
1118 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1119 
1120 		if (pgd_reg != -1)
1121 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1122 		else
1123 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1124 
1125 		if (c0_scratch_reg >= 0)
1126 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1127 		else
1128 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1129 
1130 		uasm_i_dsrl_safe(p, scratch, tmp,
1131 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1132 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1133 
1134 		if (pgd_reg == -1) {
1135 			vmalloc_branch_delay_filled = 1;
1136 			/* Clear lower 23 bits of context. */
1137 			uasm_i_dins(p, ptr, 0, 0, 23);
1138 		}
1139 	} else {
1140 		if (pgd_reg != -1)
1141 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1142 		else
1143 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1144 
1145 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1146 
1147 		if (c0_scratch_reg >= 0)
1148 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1149 		else
1150 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1151 
1152 		if (pgd_reg == -1)
1153 			/* Clear lower 23 bits of context. */
1154 			uasm_i_dins(p, ptr, 0, 0, 23);
1155 
1156 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1157 	}
1158 
1159 	if (pgd_reg == -1) {
1160 		vmalloc_branch_delay_filled = 1;
1161 		/* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
1162 		uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
1163 
1164 		uasm_i_drotr(p, ptr, ptr, 11);
1165 	}
1166 
1167 #ifdef __PAGETABLE_PMD_FOLDED
1168 #define LOC_PTEP scratch
1169 #else
1170 #define LOC_PTEP ptr
1171 #endif
1172 
1173 	if (!vmalloc_branch_delay_filled)
1174 		/* get pgd offset in bytes */
1175 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1176 
1177 	uasm_l_vmalloc_done(l, *p);
1178 
1179 	/*
1180 	 *			   tmp		ptr
1181 	 * fall-through case =	 badvaddr  *pgd_current
1182 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1183 	 */
1184 
1185 	if (vmalloc_branch_delay_filled)
1186 		/* get pgd offset in bytes */
1187 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1188 
1189 #ifdef __PAGETABLE_PMD_FOLDED
1190 	GET_CONTEXT(p, tmp); /* get context reg */
1191 #endif
1192 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1193 
1194 	if (use_lwx_insns()) {
1195 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1196 	} else {
1197 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1198 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1199 	}
1200 
1201 #ifndef __PAGETABLE_PUD_FOLDED
1202 	/* get pud offset in bytes */
1203 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1204 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1205 
1206 	if (use_lwx_insns()) {
1207 		UASM_i_LWX(p, ptr, scratch, ptr);
1208 	} else {
1209 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1210 		UASM_i_LW(p, ptr, 0, ptr);
1211 	}
1212 	/* ptr contains a pointer to PMD entry */
1213 	/* tmp contains the address */
1214 #endif
1215 
1216 #ifndef __PAGETABLE_PMD_FOLDED
1217 	/* get pmd offset in bytes */
1218 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1219 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1220 	GET_CONTEXT(p, tmp); /* get context reg */
1221 
1222 	if (use_lwx_insns()) {
1223 		UASM_i_LWX(p, scratch, scratch, ptr);
1224 	} else {
1225 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1226 		UASM_i_LW(p, scratch, 0, ptr);
1227 	}
1228 #endif
1229 	/* Adjust the context during the load latency. */
1230 	build_adjust_context(p, tmp);
1231 
1232 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1233 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1234 	/*
1235 	 * The in the LWX case we don't want to do the load in the
1236 	 * delay slot.	It cannot issue in the same cycle and may be
1237 	 * speculative and unneeded.
1238 	 */
1239 	if (use_lwx_insns())
1240 		uasm_i_nop(p);
1241 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1242 
1243 
1244 	/* build_update_entries */
1245 	if (use_lwx_insns()) {
1246 		even = ptr;
1247 		odd = tmp;
1248 		UASM_i_LWX(p, even, scratch, tmp);
1249 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1250 		UASM_i_LWX(p, odd, scratch, tmp);
1251 	} else {
1252 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1253 		even = tmp;
1254 		odd = ptr;
1255 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1256 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1257 	}
1258 	if (cpu_has_rixi) {
1259 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1260 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1261 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1262 	} else {
1263 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1264 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1265 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1266 	}
1267 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1268 
1269 	if (c0_scratch_reg >= 0) {
1270 		uasm_i_ehb(p);
1271 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1272 		build_tlb_write_entry(p, l, r, tlb_random);
1273 		uasm_l_leave(l, *p);
1274 		rv.restore_scratch = 1;
1275 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1276 		build_tlb_write_entry(p, l, r, tlb_random);
1277 		uasm_l_leave(l, *p);
1278 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1279 	} else {
1280 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1281 		build_tlb_write_entry(p, l, r, tlb_random);
1282 		uasm_l_leave(l, *p);
1283 		rv.restore_scratch = 1;
1284 	}
1285 
1286 	uasm_i_eret(p); /* return from trap */
1287 
1288 	return rv;
1289 }
1290 
1291 /*
1292  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1293  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1294  * slots before the XTLB refill exception handler which belong to the
1295  * unused TLB refill exception.
1296  */
1297 #define MIPS64_REFILL_INSNS 32
1298 
1299 static void build_r4000_tlb_refill_handler(void)
1300 {
1301 	u32 *p = tlb_handler;
1302 	struct uasm_label *l = labels;
1303 	struct uasm_reloc *r = relocs;
1304 	u32 *f;
1305 	unsigned int final_len;
1306 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1307 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1308 
1309 	memset(tlb_handler, 0, sizeof(tlb_handler));
1310 	memset(labels, 0, sizeof(labels));
1311 	memset(relocs, 0, sizeof(relocs));
1312 	memset(final_handler, 0, sizeof(final_handler));
1313 
1314 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1315 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1316 							  scratch_reg);
1317 		vmalloc_mode = refill_scratch;
1318 	} else {
1319 		htlb_info.huge_pte = K0;
1320 		htlb_info.restore_scratch = 0;
1321 		htlb_info.need_reload_pte = true;
1322 		vmalloc_mode = refill_noscratch;
1323 		/*
1324 		 * create the plain linear handler
1325 		 */
1326 		if (bcm1250_m3_war()) {
1327 			unsigned int segbits = 44;
1328 
1329 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1330 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1331 			uasm_i_xor(&p, K0, K0, K1);
1332 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1333 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1334 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1335 			uasm_i_or(&p, K0, K0, K1);
1336 			uasm_il_bnez(&p, &r, K0, label_leave);
1337 			/* No need for uasm_i_nop */
1338 		}
1339 
1340 #ifdef CONFIG_64BIT
1341 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1342 #else
1343 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1344 #endif
1345 
1346 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1347 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1348 #endif
1349 
1350 		build_get_ptep(&p, K0, K1);
1351 		build_update_entries(&p, K0, K1);
1352 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1353 		uasm_l_leave(&l, p);
1354 		uasm_i_eret(&p); /* return from trap */
1355 	}
1356 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1357 	uasm_l_tlb_huge_update(&l, p);
1358 	if (htlb_info.need_reload_pte)
1359 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1360 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1361 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1362 				   htlb_info.restore_scratch);
1363 #endif
1364 
1365 #ifdef CONFIG_64BIT
1366 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1367 #endif
1368 
1369 	/*
1370 	 * Overflow check: For the 64bit handler, we need at least one
1371 	 * free instruction slot for the wrap-around branch. In worst
1372 	 * case, if the intended insertion point is a delay slot, we
1373 	 * need three, with the second nop'ed and the third being
1374 	 * unused.
1375 	 */
1376 	switch (boot_cpu_type()) {
1377 	default:
1378 		if (sizeof(long) == 4) {
1379 		fallthrough;
1380 	case CPU_LOONGSON2EF:
1381 		/* Loongson2 ebase is different than r4k, we have more space */
1382 			if ((p - tlb_handler) > 64)
1383 				panic("TLB refill handler space exceeded");
1384 			/*
1385 			 * Now fold the handler in the TLB refill handler space.
1386 			 */
1387 			f = final_handler;
1388 			/* Simplest case, just copy the handler. */
1389 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1390 			final_len = p - tlb_handler;
1391 			break;
1392 		} else {
1393 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1394 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1395 				&& uasm_insn_has_bdelay(relocs,
1396 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1397 				panic("TLB refill handler space exceeded");
1398 			/*
1399 			 * Now fold the handler in the TLB refill handler space.
1400 			 */
1401 			f = final_handler + MIPS64_REFILL_INSNS;
1402 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1403 				/* Just copy the handler. */
1404 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1405 				final_len = p - tlb_handler;
1406 			} else {
1407 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1408 				const enum label_id ls = label_tlb_huge_update;
1409 #else
1410 				const enum label_id ls = label_vmalloc;
1411 #endif
1412 				u32 *split;
1413 				int ov = 0;
1414 				int i;
1415 
1416 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1417 					;
1418 				BUG_ON(i == ARRAY_SIZE(labels));
1419 				split = labels[i].addr;
1420 
1421 				/*
1422 				 * See if we have overflown one way or the other.
1423 				 */
1424 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1425 				    split < p - MIPS64_REFILL_INSNS)
1426 					ov = 1;
1427 
1428 				if (ov) {
1429 					/*
1430 					 * Split two instructions before the end.  One
1431 					 * for the branch and one for the instruction
1432 					 * in the delay slot.
1433 					 */
1434 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1435 
1436 					/*
1437 					 * If the branch would fall in a delay slot,
1438 					 * we must back up an additional instruction
1439 					 * so that it is no longer in a delay slot.
1440 					 */
1441 					if (uasm_insn_has_bdelay(relocs, split - 1))
1442 						split--;
1443 				}
1444 				/* Copy first part of the handler. */
1445 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1446 				f += split - tlb_handler;
1447 
1448 				if (ov) {
1449 					/* Insert branch. */
1450 					uasm_l_split(&l, final_handler);
1451 					uasm_il_b(&f, &r, label_split);
1452 					if (uasm_insn_has_bdelay(relocs, split))
1453 						uasm_i_nop(&f);
1454 					else {
1455 						uasm_copy_handler(relocs, labels,
1456 								  split, split + 1, f);
1457 						uasm_move_labels(labels, f, f + 1, -1);
1458 						f++;
1459 						split++;
1460 					}
1461 				}
1462 
1463 				/* Copy the rest of the handler. */
1464 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1465 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1466 					    (p - split);
1467 			}
1468 		}
1469 		break;
1470 	}
1471 
1472 	uasm_resolve_relocs(relocs, labels);
1473 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1474 		 final_len);
1475 
1476 	memcpy((void *)ebase, final_handler, 0x100);
1477 	local_flush_icache_range(ebase, ebase + 0x100);
1478 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1479 }
1480 
1481 static void setup_pw(void)
1482 {
1483 	unsigned int pwctl;
1484 	unsigned long pgd_i, pgd_w;
1485 #ifndef __PAGETABLE_PMD_FOLDED
1486 	unsigned long pmd_i, pmd_w;
1487 #endif
1488 	unsigned long pt_i, pt_w;
1489 	unsigned long pte_i, pte_w;
1490 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1491 	unsigned long psn;
1492 
1493 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1494 #endif
1495 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1496 #ifndef __PAGETABLE_PMD_FOLDED
1497 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1498 
1499 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1500 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1501 #else
1502 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1503 #endif
1504 
1505 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1506 	pt_w  = PAGE_SHIFT - 3;
1507 
1508 	pte_i = ilog2(_PAGE_GLOBAL);
1509 	pte_w = 0;
1510 	pwctl = 1 << 30; /* Set PWDirExt */
1511 
1512 #ifndef __PAGETABLE_PMD_FOLDED
1513 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1514 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1515 #else
1516 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1517 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1518 #endif
1519 
1520 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1521 	pwctl |= (1 << 6 | psn);
1522 #endif
1523 	write_c0_pwctl(pwctl);
1524 	write_c0_kpgd((long)swapper_pg_dir);
1525 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1526 }
1527 
1528 static void build_loongson3_tlb_refill_handler(void)
1529 {
1530 	u32 *p = tlb_handler;
1531 	struct uasm_label *l = labels;
1532 	struct uasm_reloc *r = relocs;
1533 
1534 	memset(labels, 0, sizeof(labels));
1535 	memset(relocs, 0, sizeof(relocs));
1536 	memset(tlb_handler, 0, sizeof(tlb_handler));
1537 
1538 	if (check_for_high_segbits) {
1539 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1540 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1541 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1542 		uasm_i_nop(&p);
1543 
1544 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1545 		uasm_i_nop(&p);
1546 		uasm_l_vmalloc(&l, p);
1547 	}
1548 
1549 	uasm_i_dmfc0(&p, K1, C0_PGD);
1550 
1551 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1552 #ifndef __PAGETABLE_PMD_FOLDED
1553 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1554 #endif
1555 	uasm_i_ldpte(&p, K1, 0);      /* even */
1556 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1557 	uasm_i_tlbwr(&p);
1558 
1559 	/* restore page mask */
1560 	if (PM_DEFAULT_MASK >> 16) {
1561 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1562 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1563 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1564 	} else if (PM_DEFAULT_MASK) {
1565 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1566 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1567 	} else {
1568 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1569 	}
1570 
1571 	uasm_i_eret(&p);
1572 
1573 	if (check_for_high_segbits) {
1574 		uasm_l_large_segbits_fault(&l, p);
1575 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1576 		uasm_i_jr(&p, K1);
1577 		uasm_i_nop(&p);
1578 	}
1579 
1580 	uasm_resolve_relocs(relocs, labels);
1581 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1582 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1583 	dump_handler("loongson3_tlb_refill",
1584 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1585 }
1586 
1587 static void build_setup_pgd(void)
1588 {
1589 	const int a0 = 4;
1590 	const int __maybe_unused a1 = 5;
1591 	const int __maybe_unused a2 = 6;
1592 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1593 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1594 	long pgdc = (long)pgd_current;
1595 #endif
1596 
1597 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1598 	memset(labels, 0, sizeof(labels));
1599 	memset(relocs, 0, sizeof(relocs));
1600 	pgd_reg = allocate_kscratch();
1601 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1602 	if (pgd_reg == -1) {
1603 		struct uasm_label *l = labels;
1604 		struct uasm_reloc *r = relocs;
1605 
1606 		/* PGD << 11 in c0_Context */
1607 		/*
1608 		 * If it is a ckseg0 address, convert to a physical
1609 		 * address.  Shifting right by 29 and adding 4 will
1610 		 * result in zero for these addresses.
1611 		 *
1612 		 */
1613 		UASM_i_SRA(&p, a1, a0, 29);
1614 		UASM_i_ADDIU(&p, a1, a1, 4);
1615 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1616 		uasm_i_nop(&p);
1617 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1618 		uasm_l_tlbl_goaround1(&l, p);
1619 		UASM_i_SLL(&p, a0, a0, 11);
1620 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1621 		uasm_i_jr(&p, 31);
1622 		uasm_i_ehb(&p);
1623 	} else {
1624 		/* PGD in c0_KScratch */
1625 		if (cpu_has_ldpte)
1626 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1627 		else
1628 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1629 		uasm_i_jr(&p, 31);
1630 		uasm_i_ehb(&p);
1631 	}
1632 #else
1633 #ifdef CONFIG_SMP
1634 	/* Save PGD to pgd_current[smp_processor_id()] */
1635 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1636 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1637 	UASM_i_LA_mostly(&p, a2, pgdc);
1638 	UASM_i_ADDU(&p, a2, a2, a1);
1639 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1640 #else
1641 	UASM_i_LA_mostly(&p, a2, pgdc);
1642 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1643 #endif /* SMP */
1644 
1645 	/* if pgd_reg is allocated, save PGD also to scratch register */
1646 	if (pgd_reg != -1) {
1647 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1648 		uasm_i_jr(&p, 31);
1649 		uasm_i_ehb(&p);
1650 	} else {
1651 		uasm_i_jr(&p, 31);
1652 		uasm_i_nop(&p);
1653 	}
1654 #endif
1655 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1656 		panic("tlbmiss_handler_setup_pgd space exceeded");
1657 
1658 	uasm_resolve_relocs(relocs, labels);
1659 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1660 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1661 
1662 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1663 					tlbmiss_handler_setup_pgd_end);
1664 }
1665 
1666 static void
1667 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1668 {
1669 #ifdef CONFIG_SMP
1670 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1671 		uasm_i_sync(p, 0);
1672 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1673 	if (cpu_has_64bits)
1674 		uasm_i_lld(p, pte, 0, ptr);
1675 	else
1676 # endif
1677 		UASM_i_LL(p, pte, 0, ptr);
1678 #else
1679 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1680 	if (cpu_has_64bits)
1681 		uasm_i_ld(p, pte, 0, ptr);
1682 	else
1683 # endif
1684 		UASM_i_LW(p, pte, 0, ptr);
1685 #endif
1686 }
1687 
1688 static void
1689 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1690 	unsigned int mode, unsigned int scratch)
1691 {
1692 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1693 	unsigned int swmode = mode & ~hwmode;
1694 
1695 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1696 		uasm_i_lui(p, scratch, swmode >> 16);
1697 		uasm_i_or(p, pte, pte, scratch);
1698 		BUG_ON(swmode & 0xffff);
1699 	} else {
1700 		uasm_i_ori(p, pte, pte, mode);
1701 	}
1702 
1703 #ifdef CONFIG_SMP
1704 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1705 	if (cpu_has_64bits)
1706 		uasm_i_scd(p, pte, 0, ptr);
1707 	else
1708 # endif
1709 		UASM_i_SC(p, pte, 0, ptr);
1710 
1711 	if (r10000_llsc_war())
1712 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1713 	else
1714 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1715 
1716 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1717 	if (!cpu_has_64bits) {
1718 		/* no uasm_i_nop needed */
1719 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1720 		uasm_i_ori(p, pte, pte, hwmode);
1721 		BUG_ON(hwmode & ~0xffff);
1722 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1723 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1724 		/* no uasm_i_nop needed */
1725 		uasm_i_lw(p, pte, 0, ptr);
1726 	} else
1727 		uasm_i_nop(p);
1728 # else
1729 	uasm_i_nop(p);
1730 # endif
1731 #else
1732 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1733 	if (cpu_has_64bits)
1734 		uasm_i_sd(p, pte, 0, ptr);
1735 	else
1736 # endif
1737 		UASM_i_SW(p, pte, 0, ptr);
1738 
1739 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1740 	if (!cpu_has_64bits) {
1741 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1742 		uasm_i_ori(p, pte, pte, hwmode);
1743 		BUG_ON(hwmode & ~0xffff);
1744 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1745 		uasm_i_lw(p, pte, 0, ptr);
1746 	}
1747 # endif
1748 #endif
1749 }
1750 
1751 /*
1752  * Check if PTE is present, if not then jump to LABEL. PTR points to
1753  * the page table where this PTE is located, PTE will be re-loaded
1754  * with it's original value.
1755  */
1756 static void
1757 build_pte_present(u32 **p, struct uasm_reloc **r,
1758 		  int pte, int ptr, int scratch, enum label_id lid)
1759 {
1760 	int t = scratch >= 0 ? scratch : pte;
1761 	int cur = pte;
1762 
1763 	if (cpu_has_rixi) {
1764 		if (use_bbit_insns()) {
1765 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1766 			uasm_i_nop(p);
1767 		} else {
1768 			if (_PAGE_PRESENT_SHIFT) {
1769 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1770 				cur = t;
1771 			}
1772 			uasm_i_andi(p, t, cur, 1);
1773 			uasm_il_beqz(p, r, t, lid);
1774 			if (pte == t)
1775 				/* You lose the SMP race :-(*/
1776 				iPTE_LW(p, pte, ptr);
1777 		}
1778 	} else {
1779 		if (_PAGE_PRESENT_SHIFT) {
1780 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1781 			cur = t;
1782 		}
1783 		uasm_i_andi(p, t, cur,
1784 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1785 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1786 		uasm_il_bnez(p, r, t, lid);
1787 		if (pte == t)
1788 			/* You lose the SMP race :-(*/
1789 			iPTE_LW(p, pte, ptr);
1790 	}
1791 }
1792 
1793 /* Make PTE valid, store result in PTR. */
1794 static void
1795 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1796 		 unsigned int ptr, unsigned int scratch)
1797 {
1798 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1799 
1800 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1801 }
1802 
1803 /*
1804  * Check if PTE can be written to, if not branch to LABEL. Regardless
1805  * restore PTE with value from PTR when done.
1806  */
1807 static void
1808 build_pte_writable(u32 **p, struct uasm_reloc **r,
1809 		   unsigned int pte, unsigned int ptr, int scratch,
1810 		   enum label_id lid)
1811 {
1812 	int t = scratch >= 0 ? scratch : pte;
1813 	int cur = pte;
1814 
1815 	if (_PAGE_PRESENT_SHIFT) {
1816 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1817 		cur = t;
1818 	}
1819 	uasm_i_andi(p, t, cur,
1820 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1821 	uasm_i_xori(p, t, t,
1822 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1823 	uasm_il_bnez(p, r, t, lid);
1824 	if (pte == t)
1825 		/* You lose the SMP race :-(*/
1826 		iPTE_LW(p, pte, ptr);
1827 	else
1828 		uasm_i_nop(p);
1829 }
1830 
1831 /* Make PTE writable, update software status bits as well, then store
1832  * at PTR.
1833  */
1834 static void
1835 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1836 		 unsigned int ptr, unsigned int scratch)
1837 {
1838 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1839 			     | _PAGE_DIRTY);
1840 
1841 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1842 }
1843 
1844 /*
1845  * Check if PTE can be modified, if not branch to LABEL. Regardless
1846  * restore PTE with value from PTR when done.
1847  */
1848 static void
1849 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1850 		     unsigned int pte, unsigned int ptr, int scratch,
1851 		     enum label_id lid)
1852 {
1853 	if (use_bbit_insns()) {
1854 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1855 		uasm_i_nop(p);
1856 	} else {
1857 		int t = scratch >= 0 ? scratch : pte;
1858 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1859 		uasm_i_andi(p, t, t, 1);
1860 		uasm_il_beqz(p, r, t, lid);
1861 		if (pte == t)
1862 			/* You lose the SMP race :-(*/
1863 			iPTE_LW(p, pte, ptr);
1864 	}
1865 }
1866 
1867 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1868 
1869 
1870 /*
1871  * R3000 style TLB load/store/modify handlers.
1872  */
1873 
1874 /*
1875  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1876  * Then it returns.
1877  */
1878 static void
1879 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1880 {
1881 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1882 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1883 	uasm_i_tlbwi(p);
1884 	uasm_i_jr(p, tmp);
1885 	uasm_i_rfe(p); /* branch delay */
1886 }
1887 
1888 /*
1889  * This places the pte into ENTRYLO0 and writes it with tlbwi
1890  * or tlbwr as appropriate.  This is because the index register
1891  * may have the probe fail bit set as a result of a trap on a
1892  * kseg2 access, i.e. without refill.  Then it returns.
1893  */
1894 static void
1895 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1896 			     struct uasm_reloc **r, unsigned int pte,
1897 			     unsigned int tmp)
1898 {
1899 	uasm_i_mfc0(p, tmp, C0_INDEX);
1900 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1901 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1902 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1903 	uasm_i_tlbwi(p); /* cp0 delay */
1904 	uasm_i_jr(p, tmp);
1905 	uasm_i_rfe(p); /* branch delay */
1906 	uasm_l_r3000_write_probe_fail(l, *p);
1907 	uasm_i_tlbwr(p); /* cp0 delay */
1908 	uasm_i_jr(p, tmp);
1909 	uasm_i_rfe(p); /* branch delay */
1910 }
1911 
1912 static void
1913 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1914 				   unsigned int ptr)
1915 {
1916 	long pgdc = (long)pgd_current;
1917 
1918 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1919 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1920 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1921 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1922 	uasm_i_sll(p, pte, pte, 2);
1923 	uasm_i_addu(p, ptr, ptr, pte);
1924 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1925 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1926 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1927 	uasm_i_addu(p, ptr, ptr, pte);
1928 	uasm_i_lw(p, pte, 0, ptr);
1929 	uasm_i_tlbp(p); /* load delay */
1930 }
1931 
1932 static void build_r3000_tlb_load_handler(void)
1933 {
1934 	u32 *p = (u32 *)handle_tlbl;
1935 	struct uasm_label *l = labels;
1936 	struct uasm_reloc *r = relocs;
1937 
1938 	memset(p, 0, handle_tlbl_end - (char *)p);
1939 	memset(labels, 0, sizeof(labels));
1940 	memset(relocs, 0, sizeof(relocs));
1941 
1942 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1943 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1944 	uasm_i_nop(&p); /* load delay */
1945 	build_make_valid(&p, &r, K0, K1, -1);
1946 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1947 
1948 	uasm_l_nopage_tlbl(&l, p);
1949 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1950 	uasm_i_nop(&p);
1951 
1952 	if (p >= (u32 *)handle_tlbl_end)
1953 		panic("TLB load handler fastpath space exceeded");
1954 
1955 	uasm_resolve_relocs(relocs, labels);
1956 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1957 		 (unsigned int)(p - (u32 *)handle_tlbl));
1958 
1959 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1960 }
1961 
1962 static void build_r3000_tlb_store_handler(void)
1963 {
1964 	u32 *p = (u32 *)handle_tlbs;
1965 	struct uasm_label *l = labels;
1966 	struct uasm_reloc *r = relocs;
1967 
1968 	memset(p, 0, handle_tlbs_end - (char *)p);
1969 	memset(labels, 0, sizeof(labels));
1970 	memset(relocs, 0, sizeof(relocs));
1971 
1972 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1973 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1974 	uasm_i_nop(&p); /* load delay */
1975 	build_make_write(&p, &r, K0, K1, -1);
1976 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1977 
1978 	uasm_l_nopage_tlbs(&l, p);
1979 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1980 	uasm_i_nop(&p);
1981 
1982 	if (p >= (u32 *)handle_tlbs_end)
1983 		panic("TLB store handler fastpath space exceeded");
1984 
1985 	uasm_resolve_relocs(relocs, labels);
1986 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1987 		 (unsigned int)(p - (u32 *)handle_tlbs));
1988 
1989 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1990 }
1991 
1992 static void build_r3000_tlb_modify_handler(void)
1993 {
1994 	u32 *p = (u32 *)handle_tlbm;
1995 	struct uasm_label *l = labels;
1996 	struct uasm_reloc *r = relocs;
1997 
1998 	memset(p, 0, handle_tlbm_end - (char *)p);
1999 	memset(labels, 0, sizeof(labels));
2000 	memset(relocs, 0, sizeof(relocs));
2001 
2002 	build_r3000_tlbchange_handler_head(&p, K0, K1);
2003 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
2004 	uasm_i_nop(&p); /* load delay */
2005 	build_make_write(&p, &r, K0, K1, -1);
2006 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
2007 
2008 	uasm_l_nopage_tlbm(&l, p);
2009 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2010 	uasm_i_nop(&p);
2011 
2012 	if (p >= (u32 *)handle_tlbm_end)
2013 		panic("TLB modify handler fastpath space exceeded");
2014 
2015 	uasm_resolve_relocs(relocs, labels);
2016 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2017 		 (unsigned int)(p - (u32 *)handle_tlbm));
2018 
2019 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2020 }
2021 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2022 
2023 static bool cpu_has_tlbex_tlbp_race(void)
2024 {
2025 	/*
2026 	 * When a Hardware Table Walker is running it can replace TLB entries
2027 	 * at any time, leading to a race between it & the CPU.
2028 	 */
2029 	if (cpu_has_htw)
2030 		return true;
2031 
2032 	/*
2033 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2034 	 * replaced at any time by a sibling performing a write to the FTLB.
2035 	 */
2036 	if (cpu_has_shared_ftlb_ram)
2037 		return true;
2038 
2039 	/* In all other cases there ought to be no race condition to handle */
2040 	return false;
2041 }
2042 
2043 /*
2044  * R4000 style TLB load/store/modify handlers.
2045  */
2046 static struct work_registers
2047 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2048 				   struct uasm_reloc **r)
2049 {
2050 	struct work_registers wr = build_get_work_registers(p);
2051 
2052 #ifdef CONFIG_64BIT
2053 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2054 #else
2055 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2056 #endif
2057 
2058 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2059 	/*
2060 	 * For huge tlb entries, pmd doesn't contain an address but
2061 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2062 	 * see if we need to jump to huge tlb processing.
2063 	 */
2064 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2065 #endif
2066 
2067 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2068 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2069 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2070 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2071 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2072 
2073 #ifdef CONFIG_SMP
2074 	uasm_l_smp_pgtable_change(l, *p);
2075 #endif
2076 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2077 	if (!m4kc_tlbp_war()) {
2078 		build_tlb_probe_entry(p);
2079 		if (cpu_has_tlbex_tlbp_race()) {
2080 			/* race condition happens, leaving */
2081 			uasm_i_ehb(p);
2082 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2083 			uasm_il_bltz(p, r, wr.r3, label_leave);
2084 			uasm_i_nop(p);
2085 		}
2086 	}
2087 	return wr;
2088 }
2089 
2090 static void
2091 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2092 				   struct uasm_reloc **r, unsigned int tmp,
2093 				   unsigned int ptr)
2094 {
2095 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2096 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2097 	build_update_entries(p, tmp, ptr);
2098 	build_tlb_write_entry(p, l, r, tlb_indexed);
2099 	uasm_l_leave(l, *p);
2100 	build_restore_work_registers(p);
2101 	uasm_i_eret(p); /* return from trap */
2102 
2103 #ifdef CONFIG_64BIT
2104 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2105 #endif
2106 }
2107 
2108 static void build_r4000_tlb_load_handler(void)
2109 {
2110 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2111 	struct uasm_label *l = labels;
2112 	struct uasm_reloc *r = relocs;
2113 	struct work_registers wr;
2114 
2115 	memset(p, 0, handle_tlbl_end - (char *)p);
2116 	memset(labels, 0, sizeof(labels));
2117 	memset(relocs, 0, sizeof(relocs));
2118 
2119 	if (bcm1250_m3_war()) {
2120 		unsigned int segbits = 44;
2121 
2122 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2123 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2124 		uasm_i_xor(&p, K0, K0, K1);
2125 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2126 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2127 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2128 		uasm_i_or(&p, K0, K0, K1);
2129 		uasm_il_bnez(&p, &r, K0, label_leave);
2130 		/* No need for uasm_i_nop */
2131 	}
2132 
2133 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2134 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2135 	if (m4kc_tlbp_war())
2136 		build_tlb_probe_entry(&p);
2137 
2138 	if (cpu_has_rixi && !cpu_has_rixiex) {
2139 		/*
2140 		 * If the page is not _PAGE_VALID, RI or XI could not
2141 		 * have triggered it.  Skip the expensive test..
2142 		 */
2143 		if (use_bbit_insns()) {
2144 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2145 				      label_tlbl_goaround1);
2146 		} else {
2147 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2148 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2149 		}
2150 		uasm_i_nop(&p);
2151 
2152 		/*
2153 		 * Warn if something may race with us & replace the TLB entry
2154 		 * before we read it here. Everything with such races should
2155 		 * also have dedicated RiXi exception handlers, so this
2156 		 * shouldn't be hit.
2157 		 */
2158 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2159 
2160 		uasm_i_tlbr(&p);
2161 
2162 		switch (current_cpu_type()) {
2163 		default:
2164 			if (cpu_has_mips_r2_exec_hazard) {
2165 				uasm_i_ehb(&p);
2166 			fallthrough;
2167 
2168 		case CPU_CAVIUM_OCTEON:
2169 		case CPU_CAVIUM_OCTEON_PLUS:
2170 		case CPU_CAVIUM_OCTEON2:
2171 				break;
2172 			}
2173 		}
2174 
2175 		/* Examine  entrylo 0 or 1 based on ptr. */
2176 		if (use_bbit_insns()) {
2177 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2178 		} else {
2179 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2180 			uasm_i_beqz(&p, wr.r3, 8);
2181 		}
2182 		/* load it in the delay slot*/
2183 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2184 		/* load it if ptr is odd */
2185 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2186 		/*
2187 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2188 		 * XI must have triggered it.
2189 		 */
2190 		if (use_bbit_insns()) {
2191 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2192 			uasm_i_nop(&p);
2193 			uasm_l_tlbl_goaround1(&l, p);
2194 		} else {
2195 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2196 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2197 			uasm_i_nop(&p);
2198 		}
2199 		uasm_l_tlbl_goaround1(&l, p);
2200 	}
2201 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2202 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2203 
2204 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2205 	/*
2206 	 * This is the entry point when build_r4000_tlbchange_handler_head
2207 	 * spots a huge page.
2208 	 */
2209 	uasm_l_tlb_huge_update(&l, p);
2210 	iPTE_LW(&p, wr.r1, wr.r2);
2211 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2212 	build_tlb_probe_entry(&p);
2213 
2214 	if (cpu_has_rixi && !cpu_has_rixiex) {
2215 		/*
2216 		 * If the page is not _PAGE_VALID, RI or XI could not
2217 		 * have triggered it.  Skip the expensive test..
2218 		 */
2219 		if (use_bbit_insns()) {
2220 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2221 				      label_tlbl_goaround2);
2222 		} else {
2223 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2224 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2225 		}
2226 		uasm_i_nop(&p);
2227 
2228 		/*
2229 		 * Warn if something may race with us & replace the TLB entry
2230 		 * before we read it here. Everything with such races should
2231 		 * also have dedicated RiXi exception handlers, so this
2232 		 * shouldn't be hit.
2233 		 */
2234 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2235 
2236 		uasm_i_tlbr(&p);
2237 
2238 		switch (current_cpu_type()) {
2239 		default:
2240 			if (cpu_has_mips_r2_exec_hazard) {
2241 				uasm_i_ehb(&p);
2242 
2243 		case CPU_CAVIUM_OCTEON:
2244 		case CPU_CAVIUM_OCTEON_PLUS:
2245 		case CPU_CAVIUM_OCTEON2:
2246 				break;
2247 			}
2248 		}
2249 
2250 		/* Examine  entrylo 0 or 1 based on ptr. */
2251 		if (use_bbit_insns()) {
2252 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2253 		} else {
2254 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2255 			uasm_i_beqz(&p, wr.r3, 8);
2256 		}
2257 		/* load it in the delay slot*/
2258 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2259 		/* load it if ptr is odd */
2260 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2261 		/*
2262 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2263 		 * XI must have triggered it.
2264 		 */
2265 		if (use_bbit_insns()) {
2266 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2267 		} else {
2268 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2269 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2270 		}
2271 		if (PM_DEFAULT_MASK == 0)
2272 			uasm_i_nop(&p);
2273 		/*
2274 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2275 		 * it is restored in build_huge_tlb_write_entry.
2276 		 */
2277 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2278 
2279 		uasm_l_tlbl_goaround2(&l, p);
2280 	}
2281 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2282 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2283 #endif
2284 
2285 	uasm_l_nopage_tlbl(&l, p);
2286 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2287 		uasm_i_sync(&p, 0);
2288 	build_restore_work_registers(&p);
2289 #ifdef CONFIG_CPU_MICROMIPS
2290 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2291 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2292 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2293 		uasm_i_jr(&p, K0);
2294 	} else
2295 #endif
2296 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2297 	uasm_i_nop(&p);
2298 
2299 	if (p >= (u32 *)handle_tlbl_end)
2300 		panic("TLB load handler fastpath space exceeded");
2301 
2302 	uasm_resolve_relocs(relocs, labels);
2303 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2304 		 (unsigned int)(p - (u32 *)handle_tlbl));
2305 
2306 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2307 }
2308 
2309 static void build_r4000_tlb_store_handler(void)
2310 {
2311 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2312 	struct uasm_label *l = labels;
2313 	struct uasm_reloc *r = relocs;
2314 	struct work_registers wr;
2315 
2316 	memset(p, 0, handle_tlbs_end - (char *)p);
2317 	memset(labels, 0, sizeof(labels));
2318 	memset(relocs, 0, sizeof(relocs));
2319 
2320 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2321 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2322 	if (m4kc_tlbp_war())
2323 		build_tlb_probe_entry(&p);
2324 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2325 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2326 
2327 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2328 	/*
2329 	 * This is the entry point when
2330 	 * build_r4000_tlbchange_handler_head spots a huge page.
2331 	 */
2332 	uasm_l_tlb_huge_update(&l, p);
2333 	iPTE_LW(&p, wr.r1, wr.r2);
2334 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2335 	build_tlb_probe_entry(&p);
2336 	uasm_i_ori(&p, wr.r1, wr.r1,
2337 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2338 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2339 #endif
2340 
2341 	uasm_l_nopage_tlbs(&l, p);
2342 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2343 		uasm_i_sync(&p, 0);
2344 	build_restore_work_registers(&p);
2345 #ifdef CONFIG_CPU_MICROMIPS
2346 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2347 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2348 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2349 		uasm_i_jr(&p, K0);
2350 	} else
2351 #endif
2352 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2353 	uasm_i_nop(&p);
2354 
2355 	if (p >= (u32 *)handle_tlbs_end)
2356 		panic("TLB store handler fastpath space exceeded");
2357 
2358 	uasm_resolve_relocs(relocs, labels);
2359 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2360 		 (unsigned int)(p - (u32 *)handle_tlbs));
2361 
2362 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2363 }
2364 
2365 static void build_r4000_tlb_modify_handler(void)
2366 {
2367 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2368 	struct uasm_label *l = labels;
2369 	struct uasm_reloc *r = relocs;
2370 	struct work_registers wr;
2371 
2372 	memset(p, 0, handle_tlbm_end - (char *)p);
2373 	memset(labels, 0, sizeof(labels));
2374 	memset(relocs, 0, sizeof(relocs));
2375 
2376 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2377 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2378 	if (m4kc_tlbp_war())
2379 		build_tlb_probe_entry(&p);
2380 	/* Present and writable bits set, set accessed and dirty bits. */
2381 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2382 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2383 
2384 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2385 	/*
2386 	 * This is the entry point when
2387 	 * build_r4000_tlbchange_handler_head spots a huge page.
2388 	 */
2389 	uasm_l_tlb_huge_update(&l, p);
2390 	iPTE_LW(&p, wr.r1, wr.r2);
2391 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2392 	build_tlb_probe_entry(&p);
2393 	uasm_i_ori(&p, wr.r1, wr.r1,
2394 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2395 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2396 #endif
2397 
2398 	uasm_l_nopage_tlbm(&l, p);
2399 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2400 		uasm_i_sync(&p, 0);
2401 	build_restore_work_registers(&p);
2402 #ifdef CONFIG_CPU_MICROMIPS
2403 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2404 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2405 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2406 		uasm_i_jr(&p, K0);
2407 	} else
2408 #endif
2409 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2410 	uasm_i_nop(&p);
2411 
2412 	if (p >= (u32 *)handle_tlbm_end)
2413 		panic("TLB modify handler fastpath space exceeded");
2414 
2415 	uasm_resolve_relocs(relocs, labels);
2416 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2417 		 (unsigned int)(p - (u32 *)handle_tlbm));
2418 
2419 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2420 }
2421 
2422 static void flush_tlb_handlers(void)
2423 {
2424 	local_flush_icache_range((unsigned long)handle_tlbl,
2425 			   (unsigned long)handle_tlbl_end);
2426 	local_flush_icache_range((unsigned long)handle_tlbs,
2427 			   (unsigned long)handle_tlbs_end);
2428 	local_flush_icache_range((unsigned long)handle_tlbm,
2429 			   (unsigned long)handle_tlbm_end);
2430 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2431 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2432 }
2433 
2434 static void print_htw_config(void)
2435 {
2436 	unsigned long config;
2437 	unsigned int pwctl;
2438 	const int field = 2 * sizeof(unsigned long);
2439 
2440 	config = read_c0_pwfield();
2441 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2442 		field, config,
2443 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2444 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2445 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2446 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2447 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2448 
2449 	config = read_c0_pwsize();
2450 	pr_debug("PWSize  (0x%0*lx): PS: 0x%lx  GDW: 0x%02lx  UDW: 0x%02lx  MDW: 0x%02lx  PTW: 0x%02lx  PTEW: 0x%02lx\n",
2451 		field, config,
2452 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2453 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2454 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2455 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2456 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2457 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2458 
2459 	pwctl = read_c0_pwctl();
2460 	pr_debug("PWCtl   (0x%x): PWEn: 0x%x  XK: 0x%x  XS: 0x%x  XU: 0x%x  DPH: 0x%x  HugePg: 0x%x  Psn: 0x%x\n",
2461 		pwctl,
2462 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2463 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2464 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2465 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2466 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2467 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2468 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2469 }
2470 
2471 static void config_htw_params(void)
2472 {
2473 	unsigned long pwfield, pwsize, ptei;
2474 	unsigned int config;
2475 
2476 	/*
2477 	 * We are using 2-level page tables, so we only need to
2478 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2479 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2480 	 * write values less than 0xc in these fields because the entire
2481 	 * write will be dropped. As a result of which, we must preserve
2482 	 * the original reset values and overwrite only what we really want.
2483 	 */
2484 
2485 	pwfield = read_c0_pwfield();
2486 	/* re-initialize the GDI field */
2487 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2488 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2489 	/* re-initialize the PTI field including the even/odd bit */
2490 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2491 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2492 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2493 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2494 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2495 	}
2496 	/* Set the PTEI right shift */
2497 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2498 	pwfield |= ptei;
2499 	write_c0_pwfield(pwfield);
2500 	/* Check whether the PTEI value is supported */
2501 	back_to_back_c0_hazard();
2502 	pwfield = read_c0_pwfield();
2503 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2504 		!= ptei) {
2505 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2506 			ptei);
2507 		/*
2508 		 * Drop option to avoid HTW being enabled via another path
2509 		 * (eg htw_reset())
2510 		 */
2511 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2512 		return;
2513 	}
2514 
2515 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2516 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2517 	if (CONFIG_PGTABLE_LEVELS >= 3)
2518 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2519 
2520 	/* Set pointer size to size of directory pointers */
2521 	if (IS_ENABLED(CONFIG_64BIT))
2522 		pwsize |= MIPS_PWSIZE_PS_MASK;
2523 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2524 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2525 			& MIPS_PWSIZE_PTEW_MASK;
2526 
2527 	write_c0_pwsize(pwsize);
2528 
2529 	/* Make sure everything is set before we enable the HTW */
2530 	back_to_back_c0_hazard();
2531 
2532 	/*
2533 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2534 	 * the pwctl fields.
2535 	 */
2536 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2537 	if (IS_ENABLED(CONFIG_64BIT))
2538 		config |= MIPS_PWCTL_XU_MASK;
2539 	write_c0_pwctl(config);
2540 	pr_info("Hardware Page Table Walker enabled\n");
2541 
2542 	print_htw_config();
2543 }
2544 
2545 static void config_xpa_params(void)
2546 {
2547 #ifdef CONFIG_XPA
2548 	unsigned int pagegrain;
2549 
2550 	if (mips_xpa_disabled) {
2551 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2552 		return;
2553 	}
2554 
2555 	pagegrain = read_c0_pagegrain();
2556 	write_c0_pagegrain(pagegrain | PG_ELPA);
2557 	back_to_back_c0_hazard();
2558 	pagegrain = read_c0_pagegrain();
2559 
2560 	if (pagegrain & PG_ELPA)
2561 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2562 	else
2563 		panic("Extended Physical Addressing (XPA) disabled");
2564 #endif
2565 }
2566 
2567 static void check_pabits(void)
2568 {
2569 	unsigned long entry;
2570 	unsigned pabits, fillbits;
2571 
2572 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2573 		/*
2574 		 * We'll only be making use of the fact that we can rotate bits
2575 		 * into the fill if the CPU supports RIXI, so don't bother
2576 		 * probing this for CPUs which don't.
2577 		 */
2578 		return;
2579 	}
2580 
2581 	write_c0_entrylo0(~0ul);
2582 	back_to_back_c0_hazard();
2583 	entry = read_c0_entrylo0();
2584 
2585 	/* clear all non-PFN bits */
2586 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2587 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2588 
2589 	/* find a lower bound on PABITS, and upper bound on fill bits */
2590 	pabits = fls_long(entry) + 6;
2591 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2592 
2593 	/* minus the RI & XI bits */
2594 	fillbits -= min_t(unsigned, fillbits, 2);
2595 
2596 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2597 		fill_includes_sw_bits = true;
2598 
2599 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2600 }
2601 
2602 void build_tlb_refill_handler(void)
2603 {
2604 	/*
2605 	 * The refill handler is generated per-CPU, multi-node systems
2606 	 * may have local storage for it. The other handlers are only
2607 	 * needed once.
2608 	 */
2609 	static int run_once = 0;
2610 
2611 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2612 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2613 
2614 	output_pgtable_bits_defines();
2615 	check_pabits();
2616 
2617 #ifdef CONFIG_64BIT
2618 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2619 #endif
2620 
2621 	if (cpu_has_3kex) {
2622 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2623 		if (!run_once) {
2624 			build_setup_pgd();
2625 			build_r3000_tlb_refill_handler();
2626 			build_r3000_tlb_load_handler();
2627 			build_r3000_tlb_store_handler();
2628 			build_r3000_tlb_modify_handler();
2629 			flush_tlb_handlers();
2630 			run_once++;
2631 		}
2632 #else
2633 		panic("No R3000 TLB refill handler");
2634 #endif
2635 		return;
2636 	}
2637 
2638 	if (cpu_has_ldpte)
2639 		setup_pw();
2640 
2641 	if (!run_once) {
2642 		scratch_reg = allocate_kscratch();
2643 		build_setup_pgd();
2644 		build_r4000_tlb_load_handler();
2645 		build_r4000_tlb_store_handler();
2646 		build_r4000_tlb_modify_handler();
2647 		if (cpu_has_ldpte)
2648 			build_loongson3_tlb_refill_handler();
2649 		else
2650 			build_r4000_tlb_refill_handler();
2651 		flush_tlb_handlers();
2652 		run_once++;
2653 	}
2654 	if (cpu_has_xpa)
2655 		config_xpa_params();
2656 	if (cpu_has_htw)
2657 		config_htw_params();
2658 }
2659