xref: /linux/arch/mips/mm/tlbex.c (revision d195c39052d1da278a00a6744ce59c383b67b191)
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 
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/mmu_context.h>
35 #include <asm/pgtable.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 static inline int __maybe_unused bcm1250_m3_war(void)
87 {
88 	return BCM1250_M3_WAR;
89 }
90 
91 static inline int __maybe_unused r10000_llsc_war(void)
92 {
93 	return R10000_LLSC_WAR;
94 }
95 
96 static int use_bbit_insns(void)
97 {
98 	switch (current_cpu_type()) {
99 	case CPU_CAVIUM_OCTEON:
100 	case CPU_CAVIUM_OCTEON_PLUS:
101 	case CPU_CAVIUM_OCTEON2:
102 	case CPU_CAVIUM_OCTEON3:
103 		return 1;
104 	default:
105 		return 0;
106 	}
107 }
108 
109 static int use_lwx_insns(void)
110 {
111 	switch (current_cpu_type()) {
112 	case CPU_CAVIUM_OCTEON2:
113 	case CPU_CAVIUM_OCTEON3:
114 		return 1;
115 	default:
116 		return 0;
117 	}
118 }
119 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
120     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
121 static bool scratchpad_available(void)
122 {
123 	return true;
124 }
125 static int scratchpad_offset(int i)
126 {
127 	/*
128 	 * CVMSEG starts at address -32768 and extends for
129 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
130 	 */
131 	i += 1; /* Kernel use starts at the top and works down. */
132 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
133 }
134 #else
135 static bool scratchpad_available(void)
136 {
137 	return false;
138 }
139 static int scratchpad_offset(int i)
140 {
141 	BUG();
142 	/* Really unreachable, but evidently some GCC want this. */
143 	return 0;
144 }
145 #endif
146 /*
147  * Found by experiment: At least some revisions of the 4kc throw under
148  * some circumstances a machine check exception, triggered by invalid
149  * values in the index register.  Delaying the tlbp instruction until
150  * after the next branch,  plus adding an additional nop in front of
151  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
152  * why; it's not an issue caused by the core RTL.
153  *
154  */
155 static int m4kc_tlbp_war(void)
156 {
157 	return current_cpu_type() == CPU_4KC;
158 }
159 
160 /* Handle labels (which must be positive integers). */
161 enum label_id {
162 	label_second_part = 1,
163 	label_leave,
164 	label_vmalloc,
165 	label_vmalloc_done,
166 	label_tlbw_hazard_0,
167 	label_split = label_tlbw_hazard_0 + 8,
168 	label_tlbl_goaround1,
169 	label_tlbl_goaround2,
170 	label_nopage_tlbl,
171 	label_nopage_tlbs,
172 	label_nopage_tlbm,
173 	label_smp_pgtable_change,
174 	label_r3000_write_probe_fail,
175 	label_large_segbits_fault,
176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
177 	label_tlb_huge_update,
178 #endif
179 };
180 
181 UASM_L_LA(_second_part)
182 UASM_L_LA(_leave)
183 UASM_L_LA(_vmalloc)
184 UASM_L_LA(_vmalloc_done)
185 /* _tlbw_hazard_x is handled differently.  */
186 UASM_L_LA(_split)
187 UASM_L_LA(_tlbl_goaround1)
188 UASM_L_LA(_tlbl_goaround2)
189 UASM_L_LA(_nopage_tlbl)
190 UASM_L_LA(_nopage_tlbs)
191 UASM_L_LA(_nopage_tlbm)
192 UASM_L_LA(_smp_pgtable_change)
193 UASM_L_LA(_r3000_write_probe_fail)
194 UASM_L_LA(_large_segbits_fault)
195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
196 UASM_L_LA(_tlb_huge_update)
197 #endif
198 
199 static int hazard_instance;
200 
201 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
202 {
203 	switch (instance) {
204 	case 0 ... 7:
205 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
206 		return;
207 	default:
208 		BUG();
209 	}
210 }
211 
212 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
213 {
214 	switch (instance) {
215 	case 0 ... 7:
216 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
217 		break;
218 	default:
219 		BUG();
220 	}
221 }
222 
223 /*
224  * pgtable bits are assigned dynamically depending on processor feature
225  * and statically based on kernel configuration.  This spits out the actual
226  * values the kernel is using.	Required to make sense from disassembled
227  * TLB exception handlers.
228  */
229 static void output_pgtable_bits_defines(void)
230 {
231 #define pr_define(fmt, ...)					\
232 	pr_debug("#define " fmt, ##__VA_ARGS__)
233 
234 	pr_debug("#include <asm/asm.h>\n");
235 	pr_debug("#include <asm/regdef.h>\n");
236 	pr_debug("\n");
237 
238 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
239 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
241 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
242 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
243 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
244 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
245 #endif
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 	if (cpu_has_rixi)
248 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 #endif
250 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
251 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
252 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
253 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
254 	pr_debug("\n");
255 }
256 
257 static inline void dump_handler(const char *symbol, const void *start, const void *end)
258 {
259 	unsigned int count = (end - start) / sizeof(u32);
260 	const u32 *handler = start;
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_PWBASE	5, 5
288 #define C0_PWFIELD	5, 6
289 #define C0_PWSIZE	5, 7
290 #define C0_PWCTL	6, 6
291 #define C0_BADVADDR	8, 0
292 #define C0_PGD		9, 7
293 #define C0_ENTRYHI	10, 0
294 #define C0_EPC		14, 0
295 #define C0_XCONTEXT	20, 0
296 
297 #ifdef CONFIG_64BIT
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
299 #else
300 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
301 #endif
302 
303 /* The worst case length of the handler is around 18 instructions for
304  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
305  * Maximum space available is 32 instructions for R3000 and 64
306  * instructions for R4000.
307  *
308  * We deliberately chose a buffer size of 128, so we won't scribble
309  * over anything important on overflow before we panic.
310  */
311 static u32 tlb_handler[128];
312 
313 /* simply assume worst case size for labels and relocs */
314 static struct uasm_label labels[128];
315 static struct uasm_reloc relocs[128];
316 
317 static int check_for_high_segbits;
318 static bool fill_includes_sw_bits;
319 
320 static unsigned int kscratch_used_mask;
321 
322 static inline int __maybe_unused c0_kscratch(void)
323 {
324 	switch (current_cpu_type()) {
325 	case CPU_XLP:
326 	case CPU_XLR:
327 		return 22;
328 	default:
329 		return 31;
330 	}
331 }
332 
333 static int allocate_kscratch(void)
334 {
335 	int r;
336 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
337 
338 	r = ffs(a);
339 
340 	if (r == 0)
341 		return -1;
342 
343 	r--; /* make it zero based */
344 
345 	kscratch_used_mask |= (1 << r);
346 
347 	return r;
348 }
349 
350 static int scratch_reg;
351 int pgd_reg;
352 EXPORT_SYMBOL_GPL(pgd_reg);
353 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
354 
355 static struct work_registers build_get_work_registers(u32 **p)
356 {
357 	struct work_registers r;
358 
359 	if (scratch_reg >= 0) {
360 		/* Save in CPU local C0_KScratch? */
361 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
362 		r.r1 = K0;
363 		r.r2 = K1;
364 		r.r3 = 1;
365 		return r;
366 	}
367 
368 	if (num_possible_cpus() > 1) {
369 		/* Get smp_processor_id */
370 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
371 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
372 
373 		/* handler_reg_save index in K0 */
374 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
375 
376 		UASM_i_LA(p, K1, (long)&handler_reg_save);
377 		UASM_i_ADDU(p, K0, K0, K1);
378 	} else {
379 		UASM_i_LA(p, K0, (long)&handler_reg_save);
380 	}
381 	/* K0 now points to save area, save $1 and $2  */
382 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
384 
385 	r.r1 = K1;
386 	r.r2 = 1;
387 	r.r3 = 2;
388 	return r;
389 }
390 
391 static void build_restore_work_registers(u32 **p)
392 {
393 	if (scratch_reg >= 0) {
394 		uasm_i_ehb(p);
395 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
396 		return;
397 	}
398 	/* K0 already points to save area, restore $1 and $2  */
399 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
400 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
401 }
402 
403 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
404 
405 /*
406  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
407  * we cannot do r3000 under these circumstances.
408  *
409  * The R3000 TLB handler is simple.
410  */
411 static void build_r3000_tlb_refill_handler(void)
412 {
413 	long pgdc = (long)pgd_current;
414 	u32 *p;
415 
416 	memset(tlb_handler, 0, sizeof(tlb_handler));
417 	p = tlb_handler;
418 
419 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
420 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
421 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
422 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
423 	uasm_i_sll(&p, K0, K0, 2);
424 	uasm_i_addu(&p, K1, K1, K0);
425 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
426 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
427 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
428 	uasm_i_addu(&p, K1, K1, K0);
429 	uasm_i_lw(&p, K0, 0, K1);
430 	uasm_i_nop(&p); /* load delay */
431 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
432 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
433 	uasm_i_tlbwr(&p); /* cp0 delay */
434 	uasm_i_jr(&p, K1);
435 	uasm_i_rfe(&p); /* branch delay */
436 
437 	if (p > tlb_handler + 32)
438 		panic("TLB refill handler space exceeded");
439 
440 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
441 		 (unsigned int)(p - tlb_handler));
442 
443 	memcpy((void *)ebase, tlb_handler, 0x80);
444 	local_flush_icache_range(ebase, ebase + 0x80);
445 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
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 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
499 			   struct uasm_reloc **r,
500 			   enum tlb_write_entry wmode)
501 {
502 	void(*tlbw)(u32 **) = NULL;
503 
504 	switch (wmode) {
505 	case tlb_random: tlbw = uasm_i_tlbwr; break;
506 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
507 	}
508 
509 	if (cpu_has_mips_r2_r6) {
510 		if (cpu_has_mips_r2_exec_hazard)
511 			uasm_i_ehb(p);
512 		tlbw(p);
513 		return;
514 	}
515 
516 	switch (current_cpu_type()) {
517 	case CPU_R4000PC:
518 	case CPU_R4000SC:
519 	case CPU_R4000MC:
520 	case CPU_R4400PC:
521 	case CPU_R4400SC:
522 	case CPU_R4400MC:
523 		/*
524 		 * This branch uses up a mtc0 hazard nop slot and saves
525 		 * two nops after the tlbw instruction.
526 		 */
527 		uasm_bgezl_hazard(p, r, hazard_instance);
528 		tlbw(p);
529 		uasm_bgezl_label(l, p, hazard_instance);
530 		hazard_instance++;
531 		uasm_i_nop(p);
532 		break;
533 
534 	case CPU_R4600:
535 	case CPU_R4700:
536 		uasm_i_nop(p);
537 		tlbw(p);
538 		uasm_i_nop(p);
539 		break;
540 
541 	case CPU_R5000:
542 	case CPU_NEVADA:
543 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
544 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
545 		tlbw(p);
546 		break;
547 
548 	case CPU_5KC:
549 	case CPU_TX49XX:
550 	case CPU_PR4450:
551 	case CPU_XLR:
552 		uasm_i_nop(p);
553 		tlbw(p);
554 		break;
555 
556 	case CPU_R10000:
557 	case CPU_R12000:
558 	case CPU_R14000:
559 	case CPU_R16000:
560 	case CPU_4KC:
561 	case CPU_4KEC:
562 	case CPU_M14KC:
563 	case CPU_M14KEC:
564 	case CPU_SB1:
565 	case CPU_SB1A:
566 	case CPU_4KSC:
567 	case CPU_20KC:
568 	case CPU_25KF:
569 	case CPU_BMIPS32:
570 	case CPU_BMIPS3300:
571 	case CPU_BMIPS4350:
572 	case CPU_BMIPS4380:
573 	case CPU_BMIPS5000:
574 	case CPU_LOONGSON2EF:
575 	case CPU_LOONGSON64:
576 	case CPU_R5500:
577 		if (m4kc_tlbp_war())
578 			uasm_i_nop(p);
579 		/* fall through */
580 	case CPU_ALCHEMY:
581 		tlbw(p);
582 		break;
583 
584 	case CPU_RM7000:
585 		uasm_i_nop(p);
586 		uasm_i_nop(p);
587 		uasm_i_nop(p);
588 		uasm_i_nop(p);
589 		tlbw(p);
590 		break;
591 
592 	case CPU_VR4111:
593 	case CPU_VR4121:
594 	case CPU_VR4122:
595 	case CPU_VR4181:
596 	case CPU_VR4181A:
597 		uasm_i_nop(p);
598 		uasm_i_nop(p);
599 		tlbw(p);
600 		uasm_i_nop(p);
601 		uasm_i_nop(p);
602 		break;
603 
604 	case CPU_VR4131:
605 	case CPU_VR4133:
606 		uasm_i_nop(p);
607 		uasm_i_nop(p);
608 		tlbw(p);
609 		break;
610 
611 	case CPU_XBURST:
612 		tlbw(p);
613 		uasm_i_nop(p);
614 		break;
615 
616 	default:
617 		panic("No TLB refill handler yet (CPU type: %d)",
618 		      current_cpu_type());
619 		break;
620 	}
621 }
622 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
623 
624 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
625 							unsigned int reg)
626 {
627 	if (_PAGE_GLOBAL_SHIFT == 0) {
628 		/* pte_t is already in EntryLo format */
629 		return;
630 	}
631 
632 	if (cpu_has_rixi && !!_PAGE_NO_EXEC) {
633 		if (fill_includes_sw_bits) {
634 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
635 		} else {
636 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
637 			UASM_i_ROTR(p, reg, reg,
638 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
639 		}
640 	} else {
641 #ifdef CONFIG_PHYS_ADDR_T_64BIT
642 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
643 #else
644 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 #endif
646 	}
647 }
648 
649 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
650 
651 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
652 				   unsigned int tmp, enum label_id lid,
653 				   int restore_scratch)
654 {
655 	if (restore_scratch) {
656 		/*
657 		 * Ensure the MFC0 below observes the value written to the
658 		 * KScratch register by the prior MTC0.
659 		 */
660 		if (scratch_reg >= 0)
661 			uasm_i_ehb(p);
662 
663 		/* Reset default page size */
664 		if (PM_DEFAULT_MASK >> 16) {
665 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
666 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
667 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
668 			uasm_il_b(p, r, lid);
669 		} else if (PM_DEFAULT_MASK) {
670 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
671 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
672 			uasm_il_b(p, r, lid);
673 		} else {
674 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
675 			uasm_il_b(p, r, lid);
676 		}
677 		if (scratch_reg >= 0)
678 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
679 		else
680 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
681 	} else {
682 		/* Reset default page size */
683 		if (PM_DEFAULT_MASK >> 16) {
684 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
685 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
686 			uasm_il_b(p, r, lid);
687 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
688 		} else if (PM_DEFAULT_MASK) {
689 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
690 			uasm_il_b(p, r, lid);
691 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
692 		} else {
693 			uasm_il_b(p, r, lid);
694 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
695 		}
696 	}
697 }
698 
699 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
700 				       struct uasm_reloc **r,
701 				       unsigned int tmp,
702 				       enum tlb_write_entry wmode,
703 				       int restore_scratch)
704 {
705 	/* Set huge page tlb entry size */
706 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
707 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
708 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
709 
710 	build_tlb_write_entry(p, l, r, wmode);
711 
712 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
713 }
714 
715 /*
716  * Check if Huge PTE is present, if so then jump to LABEL.
717  */
718 static void
719 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
720 		  unsigned int pmd, int lid)
721 {
722 	UASM_i_LW(p, tmp, 0, pmd);
723 	if (use_bbit_insns()) {
724 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
725 	} else {
726 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
727 		uasm_il_bnez(p, r, tmp, lid);
728 	}
729 }
730 
731 static void build_huge_update_entries(u32 **p, unsigned int pte,
732 				      unsigned int tmp)
733 {
734 	int small_sequence;
735 
736 	/*
737 	 * A huge PTE describes an area the size of the
738 	 * configured huge page size. This is twice the
739 	 * of the large TLB entry size we intend to use.
740 	 * A TLB entry half the size of the configured
741 	 * huge page size is configured into entrylo0
742 	 * and entrylo1 to cover the contiguous huge PTE
743 	 * address space.
744 	 */
745 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
746 
747 	/* We can clobber tmp.	It isn't used after this.*/
748 	if (!small_sequence)
749 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
750 
751 	build_convert_pte_to_entrylo(p, pte);
752 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
753 	/* convert to entrylo1 */
754 	if (small_sequence)
755 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
756 	else
757 		UASM_i_ADDU(p, pte, pte, tmp);
758 
759 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
760 }
761 
762 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
763 				    struct uasm_label **l,
764 				    unsigned int pte,
765 				    unsigned int ptr,
766 				    unsigned int flush)
767 {
768 #ifdef CONFIG_SMP
769 	UASM_i_SC(p, pte, 0, ptr);
770 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
771 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
772 #else
773 	UASM_i_SW(p, pte, 0, ptr);
774 #endif
775 	if (cpu_has_ftlb && flush) {
776 		BUG_ON(!cpu_has_tlbinv);
777 
778 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
779 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
780 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
781 		build_tlb_write_entry(p, l, r, tlb_indexed);
782 
783 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
784 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
785 		build_huge_update_entries(p, pte, ptr);
786 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
787 
788 		return;
789 	}
790 
791 	build_huge_update_entries(p, pte, ptr);
792 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
793 }
794 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
795 
796 #ifdef CONFIG_64BIT
797 /*
798  * TMP and PTR are scratch.
799  * TMP will be clobbered, PTR will hold the pmd entry.
800  */
801 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
802 		      unsigned int tmp, unsigned int ptr)
803 {
804 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
805 	long pgdc = (long)pgd_current;
806 #endif
807 	/*
808 	 * The vmalloc handling is not in the hotpath.
809 	 */
810 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
811 
812 	if (check_for_high_segbits) {
813 		/*
814 		 * The kernel currently implicitely assumes that the
815 		 * MIPS SEGBITS parameter for the processor is
816 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
817 		 * allocate virtual addresses outside the maximum
818 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
819 		 * that doesn't prevent user code from accessing the
820 		 * higher xuseg addresses.  Here, we make sure that
821 		 * everything but the lower xuseg addresses goes down
822 		 * the module_alloc/vmalloc path.
823 		 */
824 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
825 		uasm_il_bnez(p, r, ptr, label_vmalloc);
826 	} else {
827 		uasm_il_bltz(p, r, tmp, label_vmalloc);
828 	}
829 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
830 
831 	if (pgd_reg != -1) {
832 		/* pgd is in pgd_reg */
833 		if (cpu_has_ldpte)
834 			UASM_i_MFC0(p, ptr, C0_PWBASE);
835 		else
836 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
837 	} else {
838 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
839 		/*
840 		 * &pgd << 11 stored in CONTEXT [23..63].
841 		 */
842 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
843 
844 		/* Clear lower 23 bits of context. */
845 		uasm_i_dins(p, ptr, 0, 0, 23);
846 
847 		/* 1 0	1 0 1  << 6  xkphys cached */
848 		uasm_i_ori(p, ptr, ptr, 0x540);
849 		uasm_i_drotr(p, ptr, ptr, 11);
850 #elif defined(CONFIG_SMP)
851 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
852 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
853 		UASM_i_LA_mostly(p, tmp, pgdc);
854 		uasm_i_daddu(p, ptr, ptr, tmp);
855 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
856 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
857 #else
858 		UASM_i_LA_mostly(p, ptr, pgdc);
859 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
860 #endif
861 	}
862 
863 	uasm_l_vmalloc_done(l, *p);
864 
865 	/* get pgd offset in bytes */
866 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
867 
868 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
869 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
870 #ifndef __PAGETABLE_PUD_FOLDED
871 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
872 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
873 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
874 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
875 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
876 #endif
877 #ifndef __PAGETABLE_PMD_FOLDED
878 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
879 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
880 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
881 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
882 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
883 #endif
884 }
885 EXPORT_SYMBOL_GPL(build_get_pmde64);
886 
887 /*
888  * BVADDR is the faulting address, PTR is scratch.
889  * PTR will hold the pgd for vmalloc.
890  */
891 static void
892 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
893 			unsigned int bvaddr, unsigned int ptr,
894 			enum vmalloc64_mode mode)
895 {
896 	long swpd = (long)swapper_pg_dir;
897 	int single_insn_swpd;
898 	int did_vmalloc_branch = 0;
899 
900 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
901 
902 	uasm_l_vmalloc(l, *p);
903 
904 	if (mode != not_refill && check_for_high_segbits) {
905 		if (single_insn_swpd) {
906 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
907 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
908 			did_vmalloc_branch = 1;
909 			/* fall through */
910 		} else {
911 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
912 		}
913 	}
914 	if (!did_vmalloc_branch) {
915 		if (single_insn_swpd) {
916 			uasm_il_b(p, r, label_vmalloc_done);
917 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
918 		} else {
919 			UASM_i_LA_mostly(p, ptr, swpd);
920 			uasm_il_b(p, r, label_vmalloc_done);
921 			if (uasm_in_compat_space_p(swpd))
922 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
923 			else
924 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
925 		}
926 	}
927 	if (mode != not_refill && check_for_high_segbits) {
928 		uasm_l_large_segbits_fault(l, *p);
929 
930 		if (mode == refill_scratch && scratch_reg >= 0)
931 			uasm_i_ehb(p);
932 
933 		/*
934 		 * We get here if we are an xsseg address, or if we are
935 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
936 		 *
937 		 * Ignoring xsseg (assume disabled so would generate
938 		 * (address errors?), the only remaining possibility
939 		 * is the upper xuseg addresses.  On processors with
940 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
941 		 * addresses would have taken an address error. We try
942 		 * to mimic that here by taking a load/istream page
943 		 * fault.
944 		 */
945 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
946 			uasm_i_sync(p, 0);
947 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
948 		uasm_i_jr(p, ptr);
949 
950 		if (mode == refill_scratch) {
951 			if (scratch_reg >= 0)
952 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
953 			else
954 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
955 		} else {
956 			uasm_i_nop(p);
957 		}
958 	}
959 }
960 
961 #else /* !CONFIG_64BIT */
962 
963 /*
964  * TMP and PTR are scratch.
965  * TMP will be clobbered, PTR will hold the pgd entry.
966  */
967 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
968 {
969 	if (pgd_reg != -1) {
970 		/* pgd is in pgd_reg */
971 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
972 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
973 	} else {
974 		long pgdc = (long)pgd_current;
975 
976 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
977 #ifdef CONFIG_SMP
978 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
979 		UASM_i_LA_mostly(p, tmp, pgdc);
980 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
981 		uasm_i_addu(p, ptr, tmp, ptr);
982 #else
983 		UASM_i_LA_mostly(p, ptr, pgdc);
984 #endif
985 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
986 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
987 	}
988 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
989 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
990 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
991 }
992 EXPORT_SYMBOL_GPL(build_get_pgde32);
993 
994 #endif /* !CONFIG_64BIT */
995 
996 static void build_adjust_context(u32 **p, unsigned int ctx)
997 {
998 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
999 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1000 
1001 	switch (current_cpu_type()) {
1002 	case CPU_VR41XX:
1003 	case CPU_VR4111:
1004 	case CPU_VR4121:
1005 	case CPU_VR4122:
1006 	case CPU_VR4131:
1007 	case CPU_VR4181:
1008 	case CPU_VR4181A:
1009 	case CPU_VR4133:
1010 		shift += 2;
1011 		break;
1012 
1013 	default:
1014 		break;
1015 	}
1016 
1017 	if (shift)
1018 		UASM_i_SRL(p, ctx, ctx, shift);
1019 	uasm_i_andi(p, ctx, ctx, mask);
1020 }
1021 
1022 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1023 {
1024 	/*
1025 	 * Bug workaround for the Nevada. It seems as if under certain
1026 	 * circumstances the move from cp0_context might produce a
1027 	 * bogus result when the mfc0 instruction and its consumer are
1028 	 * in a different cacheline or a load instruction, probably any
1029 	 * memory reference, is between them.
1030 	 */
1031 	switch (current_cpu_type()) {
1032 	case CPU_NEVADA:
1033 		UASM_i_LW(p, ptr, 0, ptr);
1034 		GET_CONTEXT(p, tmp); /* get context reg */
1035 		break;
1036 
1037 	default:
1038 		GET_CONTEXT(p, tmp); /* get context reg */
1039 		UASM_i_LW(p, ptr, 0, ptr);
1040 		break;
1041 	}
1042 
1043 	build_adjust_context(p, tmp);
1044 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1045 }
1046 EXPORT_SYMBOL_GPL(build_get_ptep);
1047 
1048 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1049 {
1050 	int pte_off_even = 0;
1051 	int pte_off_odd = sizeof(pte_t);
1052 
1053 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1054 	/* The low 32 bits of EntryLo is stored in pte_high */
1055 	pte_off_even += offsetof(pte_t, pte_high);
1056 	pte_off_odd += offsetof(pte_t, pte_high);
1057 #endif
1058 
1059 	if (IS_ENABLED(CONFIG_XPA)) {
1060 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1061 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1062 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1063 
1064 		if (cpu_has_xpa && !mips_xpa_disabled) {
1065 			uasm_i_lw(p, tmp, 0, ptep);
1066 			uasm_i_ext(p, tmp, tmp, 0, 24);
1067 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1068 		}
1069 
1070 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1071 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1072 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1073 
1074 		if (cpu_has_xpa && !mips_xpa_disabled) {
1075 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1076 			uasm_i_ext(p, tmp, tmp, 0, 24);
1077 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1078 		}
1079 		return;
1080 	}
1081 
1082 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1083 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1084 	if (r45k_bvahwbug())
1085 		build_tlb_probe_entry(p);
1086 	build_convert_pte_to_entrylo(p, tmp);
1087 	if (r4k_250MHZhwbug())
1088 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1089 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1090 	build_convert_pte_to_entrylo(p, ptep);
1091 	if (r45k_bvahwbug())
1092 		uasm_i_mfc0(p, tmp, C0_INDEX);
1093 	if (r4k_250MHZhwbug())
1094 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1095 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1096 }
1097 EXPORT_SYMBOL_GPL(build_update_entries);
1098 
1099 struct mips_huge_tlb_info {
1100 	int huge_pte;
1101 	int restore_scratch;
1102 	bool need_reload_pte;
1103 };
1104 
1105 static struct mips_huge_tlb_info
1106 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1107 			       struct uasm_reloc **r, unsigned int tmp,
1108 			       unsigned int ptr, int c0_scratch_reg)
1109 {
1110 	struct mips_huge_tlb_info rv;
1111 	unsigned int even, odd;
1112 	int vmalloc_branch_delay_filled = 0;
1113 	const int scratch = 1; /* Our extra working register */
1114 
1115 	rv.huge_pte = scratch;
1116 	rv.restore_scratch = 0;
1117 	rv.need_reload_pte = false;
1118 
1119 	if (check_for_high_segbits) {
1120 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1121 
1122 		if (pgd_reg != -1)
1123 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1124 		else
1125 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1126 
1127 		if (c0_scratch_reg >= 0)
1128 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1129 		else
1130 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1131 
1132 		uasm_i_dsrl_safe(p, scratch, tmp,
1133 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1134 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1135 
1136 		if (pgd_reg == -1) {
1137 			vmalloc_branch_delay_filled = 1;
1138 			/* Clear lower 23 bits of context. */
1139 			uasm_i_dins(p, ptr, 0, 0, 23);
1140 		}
1141 	} else {
1142 		if (pgd_reg != -1)
1143 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1144 		else
1145 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1146 
1147 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1148 
1149 		if (c0_scratch_reg >= 0)
1150 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1151 		else
1152 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1153 
1154 		if (pgd_reg == -1)
1155 			/* Clear lower 23 bits of context. */
1156 			uasm_i_dins(p, ptr, 0, 0, 23);
1157 
1158 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1159 	}
1160 
1161 	if (pgd_reg == -1) {
1162 		vmalloc_branch_delay_filled = 1;
1163 		/* 1 0	1 0 1  << 6  xkphys cached */
1164 		uasm_i_ori(p, ptr, ptr, 0x540);
1165 		uasm_i_drotr(p, ptr, ptr, 11);
1166 	}
1167 
1168 #ifdef __PAGETABLE_PMD_FOLDED
1169 #define LOC_PTEP scratch
1170 #else
1171 #define LOC_PTEP ptr
1172 #endif
1173 
1174 	if (!vmalloc_branch_delay_filled)
1175 		/* get pgd offset in bytes */
1176 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1177 
1178 	uasm_l_vmalloc_done(l, *p);
1179 
1180 	/*
1181 	 *			   tmp		ptr
1182 	 * fall-through case =	 badvaddr  *pgd_current
1183 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1184 	 */
1185 
1186 	if (vmalloc_branch_delay_filled)
1187 		/* get pgd offset in bytes */
1188 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1189 
1190 #ifdef __PAGETABLE_PMD_FOLDED
1191 	GET_CONTEXT(p, tmp); /* get context reg */
1192 #endif
1193 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1194 
1195 	if (use_lwx_insns()) {
1196 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1197 	} else {
1198 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1199 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1200 	}
1201 
1202 #ifndef __PAGETABLE_PUD_FOLDED
1203 	/* get pud offset in bytes */
1204 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1205 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1206 
1207 	if (use_lwx_insns()) {
1208 		UASM_i_LWX(p, ptr, scratch, ptr);
1209 	} else {
1210 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1211 		UASM_i_LW(p, ptr, 0, ptr);
1212 	}
1213 	/* ptr contains a pointer to PMD entry */
1214 	/* tmp contains the address */
1215 #endif
1216 
1217 #ifndef __PAGETABLE_PMD_FOLDED
1218 	/* get pmd offset in bytes */
1219 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1220 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1221 	GET_CONTEXT(p, tmp); /* get context reg */
1222 
1223 	if (use_lwx_insns()) {
1224 		UASM_i_LWX(p, scratch, scratch, ptr);
1225 	} else {
1226 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1227 		UASM_i_LW(p, scratch, 0, ptr);
1228 	}
1229 #endif
1230 	/* Adjust the context during the load latency. */
1231 	build_adjust_context(p, tmp);
1232 
1233 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1234 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1235 	/*
1236 	 * The in the LWX case we don't want to do the load in the
1237 	 * delay slot.	It cannot issue in the same cycle and may be
1238 	 * speculative and unneeded.
1239 	 */
1240 	if (use_lwx_insns())
1241 		uasm_i_nop(p);
1242 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1243 
1244 
1245 	/* build_update_entries */
1246 	if (use_lwx_insns()) {
1247 		even = ptr;
1248 		odd = tmp;
1249 		UASM_i_LWX(p, even, scratch, tmp);
1250 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1251 		UASM_i_LWX(p, odd, scratch, tmp);
1252 	} else {
1253 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1254 		even = tmp;
1255 		odd = ptr;
1256 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1257 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1258 	}
1259 	if (cpu_has_rixi) {
1260 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1261 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1262 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1263 	} else {
1264 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1265 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1266 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1267 	}
1268 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1269 
1270 	if (c0_scratch_reg >= 0) {
1271 		uasm_i_ehb(p);
1272 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1273 		build_tlb_write_entry(p, l, r, tlb_random);
1274 		uasm_l_leave(l, *p);
1275 		rv.restore_scratch = 1;
1276 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1277 		build_tlb_write_entry(p, l, r, tlb_random);
1278 		uasm_l_leave(l, *p);
1279 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1280 	} else {
1281 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1282 		build_tlb_write_entry(p, l, r, tlb_random);
1283 		uasm_l_leave(l, *p);
1284 		rv.restore_scratch = 1;
1285 	}
1286 
1287 	uasm_i_eret(p); /* return from trap */
1288 
1289 	return rv;
1290 }
1291 
1292 /*
1293  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1294  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1295  * slots before the XTLB refill exception handler which belong to the
1296  * unused TLB refill exception.
1297  */
1298 #define MIPS64_REFILL_INSNS 32
1299 
1300 static void build_r4000_tlb_refill_handler(void)
1301 {
1302 	u32 *p = tlb_handler;
1303 	struct uasm_label *l = labels;
1304 	struct uasm_reloc *r = relocs;
1305 	u32 *f;
1306 	unsigned int final_len;
1307 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1308 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1309 
1310 	memset(tlb_handler, 0, sizeof(tlb_handler));
1311 	memset(labels, 0, sizeof(labels));
1312 	memset(relocs, 0, sizeof(relocs));
1313 	memset(final_handler, 0, sizeof(final_handler));
1314 
1315 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1316 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1317 							  scratch_reg);
1318 		vmalloc_mode = refill_scratch;
1319 	} else {
1320 		htlb_info.huge_pte = K0;
1321 		htlb_info.restore_scratch = 0;
1322 		htlb_info.need_reload_pte = true;
1323 		vmalloc_mode = refill_noscratch;
1324 		/*
1325 		 * create the plain linear handler
1326 		 */
1327 		if (bcm1250_m3_war()) {
1328 			unsigned int segbits = 44;
1329 
1330 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1331 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1332 			uasm_i_xor(&p, K0, K0, K1);
1333 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1334 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1335 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1336 			uasm_i_or(&p, K0, K0, K1);
1337 			uasm_il_bnez(&p, &r, K0, label_leave);
1338 			/* No need for uasm_i_nop */
1339 		}
1340 
1341 #ifdef CONFIG_64BIT
1342 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1343 #else
1344 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1345 #endif
1346 
1347 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1348 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1349 #endif
1350 
1351 		build_get_ptep(&p, K0, K1);
1352 		build_update_entries(&p, K0, K1);
1353 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1354 		uasm_l_leave(&l, p);
1355 		uasm_i_eret(&p); /* return from trap */
1356 	}
1357 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1358 	uasm_l_tlb_huge_update(&l, p);
1359 	if (htlb_info.need_reload_pte)
1360 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1361 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1362 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1363 				   htlb_info.restore_scratch);
1364 #endif
1365 
1366 #ifdef CONFIG_64BIT
1367 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1368 #endif
1369 
1370 	/*
1371 	 * Overflow check: For the 64bit handler, we need at least one
1372 	 * free instruction slot for the wrap-around branch. In worst
1373 	 * case, if the intended insertion point is a delay slot, we
1374 	 * need three, with the second nop'ed and the third being
1375 	 * unused.
1376 	 */
1377 	switch (boot_cpu_type()) {
1378 	default:
1379 		if (sizeof(long) == 4) {
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 
2167 		case CPU_CAVIUM_OCTEON:
2168 		case CPU_CAVIUM_OCTEON_PLUS:
2169 		case CPU_CAVIUM_OCTEON2:
2170 				break;
2171 			}
2172 		}
2173 
2174 		/* Examine  entrylo 0 or 1 based on ptr. */
2175 		if (use_bbit_insns()) {
2176 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2177 		} else {
2178 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2179 			uasm_i_beqz(&p, wr.r3, 8);
2180 		}
2181 		/* load it in the delay slot*/
2182 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2183 		/* load it if ptr is odd */
2184 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2185 		/*
2186 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2187 		 * XI must have triggered it.
2188 		 */
2189 		if (use_bbit_insns()) {
2190 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2191 			uasm_i_nop(&p);
2192 			uasm_l_tlbl_goaround1(&l, p);
2193 		} else {
2194 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2195 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2196 			uasm_i_nop(&p);
2197 		}
2198 		uasm_l_tlbl_goaround1(&l, p);
2199 	}
2200 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2201 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2202 
2203 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2204 	/*
2205 	 * This is the entry point when build_r4000_tlbchange_handler_head
2206 	 * spots a huge page.
2207 	 */
2208 	uasm_l_tlb_huge_update(&l, p);
2209 	iPTE_LW(&p, wr.r1, wr.r2);
2210 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2211 	build_tlb_probe_entry(&p);
2212 
2213 	if (cpu_has_rixi && !cpu_has_rixiex) {
2214 		/*
2215 		 * If the page is not _PAGE_VALID, RI or XI could not
2216 		 * have triggered it.  Skip the expensive test..
2217 		 */
2218 		if (use_bbit_insns()) {
2219 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2220 				      label_tlbl_goaround2);
2221 		} else {
2222 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2223 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2224 		}
2225 		uasm_i_nop(&p);
2226 
2227 		/*
2228 		 * Warn if something may race with us & replace the TLB entry
2229 		 * before we read it here. Everything with such races should
2230 		 * also have dedicated RiXi exception handlers, so this
2231 		 * shouldn't be hit.
2232 		 */
2233 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2234 
2235 		uasm_i_tlbr(&p);
2236 
2237 		switch (current_cpu_type()) {
2238 		default:
2239 			if (cpu_has_mips_r2_exec_hazard) {
2240 				uasm_i_ehb(&p);
2241 
2242 		case CPU_CAVIUM_OCTEON:
2243 		case CPU_CAVIUM_OCTEON_PLUS:
2244 		case CPU_CAVIUM_OCTEON2:
2245 				break;
2246 			}
2247 		}
2248 
2249 		/* Examine  entrylo 0 or 1 based on ptr. */
2250 		if (use_bbit_insns()) {
2251 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2252 		} else {
2253 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2254 			uasm_i_beqz(&p, wr.r3, 8);
2255 		}
2256 		/* load it in the delay slot*/
2257 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2258 		/* load it if ptr is odd */
2259 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2260 		/*
2261 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2262 		 * XI must have triggered it.
2263 		 */
2264 		if (use_bbit_insns()) {
2265 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2266 		} else {
2267 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2268 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2269 		}
2270 		if (PM_DEFAULT_MASK == 0)
2271 			uasm_i_nop(&p);
2272 		/*
2273 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2274 		 * it is restored in build_huge_tlb_write_entry.
2275 		 */
2276 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2277 
2278 		uasm_l_tlbl_goaround2(&l, p);
2279 	}
2280 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2281 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2282 #endif
2283 
2284 	uasm_l_nopage_tlbl(&l, p);
2285 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2286 		uasm_i_sync(&p, 0);
2287 	build_restore_work_registers(&p);
2288 #ifdef CONFIG_CPU_MICROMIPS
2289 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2290 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2291 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2292 		uasm_i_jr(&p, K0);
2293 	} else
2294 #endif
2295 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2296 	uasm_i_nop(&p);
2297 
2298 	if (p >= (u32 *)handle_tlbl_end)
2299 		panic("TLB load handler fastpath space exceeded");
2300 
2301 	uasm_resolve_relocs(relocs, labels);
2302 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2303 		 (unsigned int)(p - (u32 *)handle_tlbl));
2304 
2305 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2306 }
2307 
2308 static void build_r4000_tlb_store_handler(void)
2309 {
2310 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2311 	struct uasm_label *l = labels;
2312 	struct uasm_reloc *r = relocs;
2313 	struct work_registers wr;
2314 
2315 	memset(p, 0, handle_tlbs_end - (char *)p);
2316 	memset(labels, 0, sizeof(labels));
2317 	memset(relocs, 0, sizeof(relocs));
2318 
2319 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2320 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2321 	if (m4kc_tlbp_war())
2322 		build_tlb_probe_entry(&p);
2323 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2324 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2325 
2326 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2327 	/*
2328 	 * This is the entry point when
2329 	 * build_r4000_tlbchange_handler_head spots a huge page.
2330 	 */
2331 	uasm_l_tlb_huge_update(&l, p);
2332 	iPTE_LW(&p, wr.r1, wr.r2);
2333 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2334 	build_tlb_probe_entry(&p);
2335 	uasm_i_ori(&p, wr.r1, wr.r1,
2336 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2337 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2338 #endif
2339 
2340 	uasm_l_nopage_tlbs(&l, p);
2341 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2342 		uasm_i_sync(&p, 0);
2343 	build_restore_work_registers(&p);
2344 #ifdef CONFIG_CPU_MICROMIPS
2345 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2346 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2347 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2348 		uasm_i_jr(&p, K0);
2349 	} else
2350 #endif
2351 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2352 	uasm_i_nop(&p);
2353 
2354 	if (p >= (u32 *)handle_tlbs_end)
2355 		panic("TLB store handler fastpath space exceeded");
2356 
2357 	uasm_resolve_relocs(relocs, labels);
2358 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2359 		 (unsigned int)(p - (u32 *)handle_tlbs));
2360 
2361 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2362 }
2363 
2364 static void build_r4000_tlb_modify_handler(void)
2365 {
2366 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2367 	struct uasm_label *l = labels;
2368 	struct uasm_reloc *r = relocs;
2369 	struct work_registers wr;
2370 
2371 	memset(p, 0, handle_tlbm_end - (char *)p);
2372 	memset(labels, 0, sizeof(labels));
2373 	memset(relocs, 0, sizeof(relocs));
2374 
2375 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2376 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2377 	if (m4kc_tlbp_war())
2378 		build_tlb_probe_entry(&p);
2379 	/* Present and writable bits set, set accessed and dirty bits. */
2380 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2381 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2382 
2383 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2384 	/*
2385 	 * This is the entry point when
2386 	 * build_r4000_tlbchange_handler_head spots a huge page.
2387 	 */
2388 	uasm_l_tlb_huge_update(&l, p);
2389 	iPTE_LW(&p, wr.r1, wr.r2);
2390 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2391 	build_tlb_probe_entry(&p);
2392 	uasm_i_ori(&p, wr.r1, wr.r1,
2393 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2394 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2395 #endif
2396 
2397 	uasm_l_nopage_tlbm(&l, p);
2398 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2399 		uasm_i_sync(&p, 0);
2400 	build_restore_work_registers(&p);
2401 #ifdef CONFIG_CPU_MICROMIPS
2402 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2403 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2404 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2405 		uasm_i_jr(&p, K0);
2406 	} else
2407 #endif
2408 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2409 	uasm_i_nop(&p);
2410 
2411 	if (p >= (u32 *)handle_tlbm_end)
2412 		panic("TLB modify handler fastpath space exceeded");
2413 
2414 	uasm_resolve_relocs(relocs, labels);
2415 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2416 		 (unsigned int)(p - (u32 *)handle_tlbm));
2417 
2418 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2419 }
2420 
2421 static void flush_tlb_handlers(void)
2422 {
2423 	local_flush_icache_range((unsigned long)handle_tlbl,
2424 			   (unsigned long)handle_tlbl_end);
2425 	local_flush_icache_range((unsigned long)handle_tlbs,
2426 			   (unsigned long)handle_tlbs_end);
2427 	local_flush_icache_range((unsigned long)handle_tlbm,
2428 			   (unsigned long)handle_tlbm_end);
2429 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2430 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2431 }
2432 
2433 static void print_htw_config(void)
2434 {
2435 	unsigned long config;
2436 	unsigned int pwctl;
2437 	const int field = 2 * sizeof(unsigned long);
2438 
2439 	config = read_c0_pwfield();
2440 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2441 		field, config,
2442 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2443 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2444 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2445 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2446 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2447 
2448 	config = read_c0_pwsize();
2449 	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",
2450 		field, config,
2451 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2452 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2453 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2454 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2455 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2456 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2457 
2458 	pwctl = read_c0_pwctl();
2459 	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",
2460 		pwctl,
2461 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2462 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2463 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2464 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2465 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2466 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2467 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2468 }
2469 
2470 static void config_htw_params(void)
2471 {
2472 	unsigned long pwfield, pwsize, ptei;
2473 	unsigned int config;
2474 
2475 	/*
2476 	 * We are using 2-level page tables, so we only need to
2477 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2478 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2479 	 * write values less than 0xc in these fields because the entire
2480 	 * write will be dropped. As a result of which, we must preserve
2481 	 * the original reset values and overwrite only what we really want.
2482 	 */
2483 
2484 	pwfield = read_c0_pwfield();
2485 	/* re-initialize the GDI field */
2486 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2487 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2488 	/* re-initialize the PTI field including the even/odd bit */
2489 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2490 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2491 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2492 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2493 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2494 	}
2495 	/* Set the PTEI right shift */
2496 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2497 	pwfield |= ptei;
2498 	write_c0_pwfield(pwfield);
2499 	/* Check whether the PTEI value is supported */
2500 	back_to_back_c0_hazard();
2501 	pwfield = read_c0_pwfield();
2502 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2503 		!= ptei) {
2504 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2505 			ptei);
2506 		/*
2507 		 * Drop option to avoid HTW being enabled via another path
2508 		 * (eg htw_reset())
2509 		 */
2510 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2511 		return;
2512 	}
2513 
2514 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2515 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2516 	if (CONFIG_PGTABLE_LEVELS >= 3)
2517 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2518 
2519 	/* Set pointer size to size of directory pointers */
2520 	if (IS_ENABLED(CONFIG_64BIT))
2521 		pwsize |= MIPS_PWSIZE_PS_MASK;
2522 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2523 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2524 			& MIPS_PWSIZE_PTEW_MASK;
2525 
2526 	write_c0_pwsize(pwsize);
2527 
2528 	/* Make sure everything is set before we enable the HTW */
2529 	back_to_back_c0_hazard();
2530 
2531 	/*
2532 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2533 	 * the pwctl fields.
2534 	 */
2535 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2536 	if (IS_ENABLED(CONFIG_64BIT))
2537 		config |= MIPS_PWCTL_XU_MASK;
2538 	write_c0_pwctl(config);
2539 	pr_info("Hardware Page Table Walker enabled\n");
2540 
2541 	print_htw_config();
2542 }
2543 
2544 static void config_xpa_params(void)
2545 {
2546 #ifdef CONFIG_XPA
2547 	unsigned int pagegrain;
2548 
2549 	if (mips_xpa_disabled) {
2550 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2551 		return;
2552 	}
2553 
2554 	pagegrain = read_c0_pagegrain();
2555 	write_c0_pagegrain(pagegrain | PG_ELPA);
2556 	back_to_back_c0_hazard();
2557 	pagegrain = read_c0_pagegrain();
2558 
2559 	if (pagegrain & PG_ELPA)
2560 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2561 	else
2562 		panic("Extended Physical Addressing (XPA) disabled");
2563 #endif
2564 }
2565 
2566 static void check_pabits(void)
2567 {
2568 	unsigned long entry;
2569 	unsigned pabits, fillbits;
2570 
2571 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2572 		/*
2573 		 * We'll only be making use of the fact that we can rotate bits
2574 		 * into the fill if the CPU supports RIXI, so don't bother
2575 		 * probing this for CPUs which don't.
2576 		 */
2577 		return;
2578 	}
2579 
2580 	write_c0_entrylo0(~0ul);
2581 	back_to_back_c0_hazard();
2582 	entry = read_c0_entrylo0();
2583 
2584 	/* clear all non-PFN bits */
2585 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2586 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2587 
2588 	/* find a lower bound on PABITS, and upper bound on fill bits */
2589 	pabits = fls_long(entry) + 6;
2590 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2591 
2592 	/* minus the RI & XI bits */
2593 	fillbits -= min_t(unsigned, fillbits, 2);
2594 
2595 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2596 		fill_includes_sw_bits = true;
2597 
2598 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2599 }
2600 
2601 void build_tlb_refill_handler(void)
2602 {
2603 	/*
2604 	 * The refill handler is generated per-CPU, multi-node systems
2605 	 * may have local storage for it. The other handlers are only
2606 	 * needed once.
2607 	 */
2608 	static int run_once = 0;
2609 
2610 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2611 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2612 
2613 	output_pgtable_bits_defines();
2614 	check_pabits();
2615 
2616 #ifdef CONFIG_64BIT
2617 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2618 #endif
2619 
2620 	if (cpu_has_3kex) {
2621 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2622 		if (!run_once) {
2623 			build_setup_pgd();
2624 			build_r3000_tlb_refill_handler();
2625 			build_r3000_tlb_load_handler();
2626 			build_r3000_tlb_store_handler();
2627 			build_r3000_tlb_modify_handler();
2628 			flush_tlb_handlers();
2629 			run_once++;
2630 		}
2631 #else
2632 		panic("No R3000 TLB refill handler");
2633 #endif
2634 		return;
2635 	}
2636 
2637 	if (cpu_has_ldpte)
2638 		setup_pw();
2639 
2640 	if (!run_once) {
2641 		scratch_reg = allocate_kscratch();
2642 		build_setup_pgd();
2643 		build_r4000_tlb_load_handler();
2644 		build_r4000_tlb_store_handler();
2645 		build_r4000_tlb_modify_handler();
2646 		if (cpu_has_ldpte)
2647 			build_loongson3_tlb_refill_handler();
2648 		else
2649 			build_r4000_tlb_refill_handler();
2650 		flush_tlb_handlers();
2651 		run_once++;
2652 	}
2653 	if (cpu_has_xpa)
2654 		config_xpa_params();
2655 	if (cpu_has_htw)
2656 		config_htw_params();
2657 }
2658