xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision d6e0cbb1)
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_R4300:
549 	case CPU_5KC:
550 	case CPU_TX49XX:
551 	case CPU_PR4450:
552 	case CPU_XLR:
553 		uasm_i_nop(p);
554 		tlbw(p);
555 		break;
556 
557 	case CPU_R10000:
558 	case CPU_R12000:
559 	case CPU_R14000:
560 	case CPU_R16000:
561 	case CPU_4KC:
562 	case CPU_4KEC:
563 	case CPU_M14KC:
564 	case CPU_M14KEC:
565 	case CPU_SB1:
566 	case CPU_SB1A:
567 	case CPU_4KSC:
568 	case CPU_20KC:
569 	case CPU_25KF:
570 	case CPU_BMIPS32:
571 	case CPU_BMIPS3300:
572 	case CPU_BMIPS4350:
573 	case CPU_BMIPS4380:
574 	case CPU_BMIPS5000:
575 	case CPU_LOONGSON2:
576 	case CPU_LOONGSON3:
577 	case CPU_R5500:
578 		if (m4kc_tlbp_war())
579 			uasm_i_nop(p);
580 		/* fall through */
581 	case CPU_ALCHEMY:
582 		tlbw(p);
583 		break;
584 
585 	case CPU_RM7000:
586 		uasm_i_nop(p);
587 		uasm_i_nop(p);
588 		uasm_i_nop(p);
589 		uasm_i_nop(p);
590 		tlbw(p);
591 		break;
592 
593 	case CPU_VR4111:
594 	case CPU_VR4121:
595 	case CPU_VR4122:
596 	case CPU_VR4181:
597 	case CPU_VR4181A:
598 		uasm_i_nop(p);
599 		uasm_i_nop(p);
600 		tlbw(p);
601 		uasm_i_nop(p);
602 		uasm_i_nop(p);
603 		break;
604 
605 	case CPU_VR4131:
606 	case CPU_VR4133:
607 	case CPU_R5432:
608 		uasm_i_nop(p);
609 		uasm_i_nop(p);
610 		tlbw(p);
611 		break;
612 
613 	case CPU_JZRISC:
614 		tlbw(p);
615 		uasm_i_nop(p);
616 		break;
617 
618 	default:
619 		panic("No TLB refill handler yet (CPU type: %d)",
620 		      current_cpu_type());
621 		break;
622 	}
623 }
624 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
625 
626 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
627 							unsigned int reg)
628 {
629 	if (_PAGE_GLOBAL_SHIFT == 0) {
630 		/* pte_t is already in EntryLo format */
631 		return;
632 	}
633 
634 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
635 		if (fill_includes_sw_bits) {
636 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
637 		} else {
638 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
639 			UASM_i_ROTR(p, reg, reg,
640 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
641 		}
642 	} else {
643 #ifdef CONFIG_PHYS_ADDR_T_64BIT
644 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 #else
646 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
647 #endif
648 	}
649 }
650 
651 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
652 
653 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
654 				   unsigned int tmp, enum label_id lid,
655 				   int restore_scratch)
656 {
657 	if (restore_scratch) {
658 		/* Reset default page size */
659 		if (PM_DEFAULT_MASK >> 16) {
660 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
661 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
662 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
663 			uasm_il_b(p, r, lid);
664 		} else if (PM_DEFAULT_MASK) {
665 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
666 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
667 			uasm_il_b(p, r, lid);
668 		} else {
669 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
670 			uasm_il_b(p, r, lid);
671 		}
672 		if (scratch_reg >= 0) {
673 			uasm_i_ehb(p);
674 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
675 		} else {
676 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
677 		}
678 	} else {
679 		/* Reset default page size */
680 		if (PM_DEFAULT_MASK >> 16) {
681 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
682 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
683 			uasm_il_b(p, r, lid);
684 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
685 		} else if (PM_DEFAULT_MASK) {
686 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
687 			uasm_il_b(p, r, lid);
688 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
689 		} else {
690 			uasm_il_b(p, r, lid);
691 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
692 		}
693 	}
694 }
695 
696 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
697 				       struct uasm_reloc **r,
698 				       unsigned int tmp,
699 				       enum tlb_write_entry wmode,
700 				       int restore_scratch)
701 {
702 	/* Set huge page tlb entry size */
703 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
704 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
705 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
706 
707 	build_tlb_write_entry(p, l, r, wmode);
708 
709 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
710 }
711 
712 /*
713  * Check if Huge PTE is present, if so then jump to LABEL.
714  */
715 static void
716 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
717 		  unsigned int pmd, int lid)
718 {
719 	UASM_i_LW(p, tmp, 0, pmd);
720 	if (use_bbit_insns()) {
721 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
722 	} else {
723 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
724 		uasm_il_bnez(p, r, tmp, lid);
725 	}
726 }
727 
728 static void build_huge_update_entries(u32 **p, unsigned int pte,
729 				      unsigned int tmp)
730 {
731 	int small_sequence;
732 
733 	/*
734 	 * A huge PTE describes an area the size of the
735 	 * configured huge page size. This is twice the
736 	 * of the large TLB entry size we intend to use.
737 	 * A TLB entry half the size of the configured
738 	 * huge page size is configured into entrylo0
739 	 * and entrylo1 to cover the contiguous huge PTE
740 	 * address space.
741 	 */
742 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
743 
744 	/* We can clobber tmp.	It isn't used after this.*/
745 	if (!small_sequence)
746 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
747 
748 	build_convert_pte_to_entrylo(p, pte);
749 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
750 	/* convert to entrylo1 */
751 	if (small_sequence)
752 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
753 	else
754 		UASM_i_ADDU(p, pte, pte, tmp);
755 
756 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
757 }
758 
759 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
760 				    struct uasm_label **l,
761 				    unsigned int pte,
762 				    unsigned int ptr,
763 				    unsigned int flush)
764 {
765 #ifdef CONFIG_SMP
766 	UASM_i_SC(p, pte, 0, ptr);
767 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
768 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
769 #else
770 	UASM_i_SW(p, pte, 0, ptr);
771 #endif
772 	if (cpu_has_ftlb && flush) {
773 		BUG_ON(!cpu_has_tlbinv);
774 
775 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
776 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
777 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
778 		build_tlb_write_entry(p, l, r, tlb_indexed);
779 
780 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
781 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
782 		build_huge_update_entries(p, pte, ptr);
783 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
784 
785 		return;
786 	}
787 
788 	build_huge_update_entries(p, pte, ptr);
789 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
790 }
791 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
792 
793 #ifdef CONFIG_64BIT
794 /*
795  * TMP and PTR are scratch.
796  * TMP will be clobbered, PTR will hold the pmd entry.
797  */
798 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
799 		      unsigned int tmp, unsigned int ptr)
800 {
801 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
802 	long pgdc = (long)pgd_current;
803 #endif
804 	/*
805 	 * The vmalloc handling is not in the hotpath.
806 	 */
807 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
808 
809 	if (check_for_high_segbits) {
810 		/*
811 		 * The kernel currently implicitely assumes that the
812 		 * MIPS SEGBITS parameter for the processor is
813 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
814 		 * allocate virtual addresses outside the maximum
815 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
816 		 * that doesn't prevent user code from accessing the
817 		 * higher xuseg addresses.  Here, we make sure that
818 		 * everything but the lower xuseg addresses goes down
819 		 * the module_alloc/vmalloc path.
820 		 */
821 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
822 		uasm_il_bnez(p, r, ptr, label_vmalloc);
823 	} else {
824 		uasm_il_bltz(p, r, tmp, label_vmalloc);
825 	}
826 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
827 
828 	if (pgd_reg != -1) {
829 		/* pgd is in pgd_reg */
830 		if (cpu_has_ldpte)
831 			UASM_i_MFC0(p, ptr, C0_PWBASE);
832 		else
833 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
834 	} else {
835 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
836 		/*
837 		 * &pgd << 11 stored in CONTEXT [23..63].
838 		 */
839 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
840 
841 		/* Clear lower 23 bits of context. */
842 		uasm_i_dins(p, ptr, 0, 0, 23);
843 
844 		/* 1 0	1 0 1  << 6  xkphys cached */
845 		uasm_i_ori(p, ptr, ptr, 0x540);
846 		uasm_i_drotr(p, ptr, ptr, 11);
847 #elif defined(CONFIG_SMP)
848 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
849 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
850 		UASM_i_LA_mostly(p, tmp, pgdc);
851 		uasm_i_daddu(p, ptr, ptr, tmp);
852 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
853 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
854 #else
855 		UASM_i_LA_mostly(p, ptr, pgdc);
856 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
857 #endif
858 	}
859 
860 	uasm_l_vmalloc_done(l, *p);
861 
862 	/* get pgd offset in bytes */
863 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
864 
865 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
866 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
867 #ifndef __PAGETABLE_PUD_FOLDED
868 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
869 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
870 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
871 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
872 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
873 #endif
874 #ifndef __PAGETABLE_PMD_FOLDED
875 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
876 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
877 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
878 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
879 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
880 #endif
881 }
882 EXPORT_SYMBOL_GPL(build_get_pmde64);
883 
884 /*
885  * BVADDR is the faulting address, PTR is scratch.
886  * PTR will hold the pgd for vmalloc.
887  */
888 static void
889 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
890 			unsigned int bvaddr, unsigned int ptr,
891 			enum vmalloc64_mode mode)
892 {
893 	long swpd = (long)swapper_pg_dir;
894 	int single_insn_swpd;
895 	int did_vmalloc_branch = 0;
896 
897 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
898 
899 	uasm_l_vmalloc(l, *p);
900 
901 	if (mode != not_refill && check_for_high_segbits) {
902 		if (single_insn_swpd) {
903 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
904 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
905 			did_vmalloc_branch = 1;
906 			/* fall through */
907 		} else {
908 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
909 		}
910 	}
911 	if (!did_vmalloc_branch) {
912 		if (single_insn_swpd) {
913 			uasm_il_b(p, r, label_vmalloc_done);
914 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
915 		} else {
916 			UASM_i_LA_mostly(p, ptr, swpd);
917 			uasm_il_b(p, r, label_vmalloc_done);
918 			if (uasm_in_compat_space_p(swpd))
919 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
920 			else
921 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
922 		}
923 	}
924 	if (mode != not_refill && check_for_high_segbits) {
925 		uasm_l_large_segbits_fault(l, *p);
926 		/*
927 		 * We get here if we are an xsseg address, or if we are
928 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
929 		 *
930 		 * Ignoring xsseg (assume disabled so would generate
931 		 * (address errors?), the only remaining possibility
932 		 * is the upper xuseg addresses.  On processors with
933 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
934 		 * addresses would have taken an address error. We try
935 		 * to mimic that here by taking a load/istream page
936 		 * fault.
937 		 */
938 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
939 			uasm_i_sync(p, 0);
940 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
941 		uasm_i_jr(p, ptr);
942 
943 		if (mode == refill_scratch) {
944 			if (scratch_reg >= 0) {
945 				uasm_i_ehb(p);
946 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
947 			} else {
948 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
949 			}
950 		} else {
951 			uasm_i_nop(p);
952 		}
953 	}
954 }
955 
956 #else /* !CONFIG_64BIT */
957 
958 /*
959  * TMP and PTR are scratch.
960  * TMP will be clobbered, PTR will hold the pgd entry.
961  */
962 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
963 {
964 	if (pgd_reg != -1) {
965 		/* pgd is in pgd_reg */
966 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
967 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
968 	} else {
969 		long pgdc = (long)pgd_current;
970 
971 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
972 #ifdef CONFIG_SMP
973 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
974 		UASM_i_LA_mostly(p, tmp, pgdc);
975 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
976 		uasm_i_addu(p, ptr, tmp, ptr);
977 #else
978 		UASM_i_LA_mostly(p, ptr, pgdc);
979 #endif
980 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
981 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
982 	}
983 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
984 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
985 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
986 }
987 EXPORT_SYMBOL_GPL(build_get_pgde32);
988 
989 #endif /* !CONFIG_64BIT */
990 
991 static void build_adjust_context(u32 **p, unsigned int ctx)
992 {
993 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
994 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
995 
996 	switch (current_cpu_type()) {
997 	case CPU_VR41XX:
998 	case CPU_VR4111:
999 	case CPU_VR4121:
1000 	case CPU_VR4122:
1001 	case CPU_VR4131:
1002 	case CPU_VR4181:
1003 	case CPU_VR4181A:
1004 	case CPU_VR4133:
1005 		shift += 2;
1006 		break;
1007 
1008 	default:
1009 		break;
1010 	}
1011 
1012 	if (shift)
1013 		UASM_i_SRL(p, ctx, ctx, shift);
1014 	uasm_i_andi(p, ctx, ctx, mask);
1015 }
1016 
1017 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1018 {
1019 	/*
1020 	 * Bug workaround for the Nevada. It seems as if under certain
1021 	 * circumstances the move from cp0_context might produce a
1022 	 * bogus result when the mfc0 instruction and its consumer are
1023 	 * in a different cacheline or a load instruction, probably any
1024 	 * memory reference, is between them.
1025 	 */
1026 	switch (current_cpu_type()) {
1027 	case CPU_NEVADA:
1028 		UASM_i_LW(p, ptr, 0, ptr);
1029 		GET_CONTEXT(p, tmp); /* get context reg */
1030 		break;
1031 
1032 	default:
1033 		GET_CONTEXT(p, tmp); /* get context reg */
1034 		UASM_i_LW(p, ptr, 0, ptr);
1035 		break;
1036 	}
1037 
1038 	build_adjust_context(p, tmp);
1039 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1040 }
1041 EXPORT_SYMBOL_GPL(build_get_ptep);
1042 
1043 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1044 {
1045 	int pte_off_even = 0;
1046 	int pte_off_odd = sizeof(pte_t);
1047 
1048 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1049 	/* The low 32 bits of EntryLo is stored in pte_high */
1050 	pte_off_even += offsetof(pte_t, pte_high);
1051 	pte_off_odd += offsetof(pte_t, pte_high);
1052 #endif
1053 
1054 	if (IS_ENABLED(CONFIG_XPA)) {
1055 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1056 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1057 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1058 
1059 		if (cpu_has_xpa && !mips_xpa_disabled) {
1060 			uasm_i_lw(p, tmp, 0, ptep);
1061 			uasm_i_ext(p, tmp, tmp, 0, 24);
1062 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1063 		}
1064 
1065 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1066 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1067 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1068 
1069 		if (cpu_has_xpa && !mips_xpa_disabled) {
1070 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1071 			uasm_i_ext(p, tmp, tmp, 0, 24);
1072 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1073 		}
1074 		return;
1075 	}
1076 
1077 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1078 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1079 	if (r45k_bvahwbug())
1080 		build_tlb_probe_entry(p);
1081 	build_convert_pte_to_entrylo(p, tmp);
1082 	if (r4k_250MHZhwbug())
1083 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1084 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1085 	build_convert_pte_to_entrylo(p, ptep);
1086 	if (r45k_bvahwbug())
1087 		uasm_i_mfc0(p, tmp, C0_INDEX);
1088 	if (r4k_250MHZhwbug())
1089 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1090 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1091 }
1092 EXPORT_SYMBOL_GPL(build_update_entries);
1093 
1094 struct mips_huge_tlb_info {
1095 	int huge_pte;
1096 	int restore_scratch;
1097 	bool need_reload_pte;
1098 };
1099 
1100 static struct mips_huge_tlb_info
1101 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1102 			       struct uasm_reloc **r, unsigned int tmp,
1103 			       unsigned int ptr, int c0_scratch_reg)
1104 {
1105 	struct mips_huge_tlb_info rv;
1106 	unsigned int even, odd;
1107 	int vmalloc_branch_delay_filled = 0;
1108 	const int scratch = 1; /* Our extra working register */
1109 
1110 	rv.huge_pte = scratch;
1111 	rv.restore_scratch = 0;
1112 	rv.need_reload_pte = false;
1113 
1114 	if (check_for_high_segbits) {
1115 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1116 
1117 		if (pgd_reg != -1)
1118 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1119 		else
1120 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1121 
1122 		if (c0_scratch_reg >= 0)
1123 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1124 		else
1125 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1126 
1127 		uasm_i_dsrl_safe(p, scratch, tmp,
1128 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1129 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1130 
1131 		if (pgd_reg == -1) {
1132 			vmalloc_branch_delay_filled = 1;
1133 			/* Clear lower 23 bits of context. */
1134 			uasm_i_dins(p, ptr, 0, 0, 23);
1135 		}
1136 	} else {
1137 		if (pgd_reg != -1)
1138 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1139 		else
1140 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1141 
1142 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1143 
1144 		if (c0_scratch_reg >= 0)
1145 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1146 		else
1147 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1148 
1149 		if (pgd_reg == -1)
1150 			/* Clear lower 23 bits of context. */
1151 			uasm_i_dins(p, ptr, 0, 0, 23);
1152 
1153 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1154 	}
1155 
1156 	if (pgd_reg == -1) {
1157 		vmalloc_branch_delay_filled = 1;
1158 		/* 1 0	1 0 1  << 6  xkphys cached */
1159 		uasm_i_ori(p, ptr, ptr, 0x540);
1160 		uasm_i_drotr(p, ptr, ptr, 11);
1161 	}
1162 
1163 #ifdef __PAGETABLE_PMD_FOLDED
1164 #define LOC_PTEP scratch
1165 #else
1166 #define LOC_PTEP ptr
1167 #endif
1168 
1169 	if (!vmalloc_branch_delay_filled)
1170 		/* get pgd offset in bytes */
1171 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1172 
1173 	uasm_l_vmalloc_done(l, *p);
1174 
1175 	/*
1176 	 *			   tmp		ptr
1177 	 * fall-through case =	 badvaddr  *pgd_current
1178 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1179 	 */
1180 
1181 	if (vmalloc_branch_delay_filled)
1182 		/* get pgd offset in bytes */
1183 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1184 
1185 #ifdef __PAGETABLE_PMD_FOLDED
1186 	GET_CONTEXT(p, tmp); /* get context reg */
1187 #endif
1188 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1189 
1190 	if (use_lwx_insns()) {
1191 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1192 	} else {
1193 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1194 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1195 	}
1196 
1197 #ifndef __PAGETABLE_PUD_FOLDED
1198 	/* get pud offset in bytes */
1199 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1200 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1201 
1202 	if (use_lwx_insns()) {
1203 		UASM_i_LWX(p, ptr, scratch, ptr);
1204 	} else {
1205 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1206 		UASM_i_LW(p, ptr, 0, ptr);
1207 	}
1208 	/* ptr contains a pointer to PMD entry */
1209 	/* tmp contains the address */
1210 #endif
1211 
1212 #ifndef __PAGETABLE_PMD_FOLDED
1213 	/* get pmd offset in bytes */
1214 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1215 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1216 	GET_CONTEXT(p, tmp); /* get context reg */
1217 
1218 	if (use_lwx_insns()) {
1219 		UASM_i_LWX(p, scratch, scratch, ptr);
1220 	} else {
1221 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1222 		UASM_i_LW(p, scratch, 0, ptr);
1223 	}
1224 #endif
1225 	/* Adjust the context during the load latency. */
1226 	build_adjust_context(p, tmp);
1227 
1228 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1229 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1230 	/*
1231 	 * The in the LWX case we don't want to do the load in the
1232 	 * delay slot.	It cannot issue in the same cycle and may be
1233 	 * speculative and unneeded.
1234 	 */
1235 	if (use_lwx_insns())
1236 		uasm_i_nop(p);
1237 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1238 
1239 
1240 	/* build_update_entries */
1241 	if (use_lwx_insns()) {
1242 		even = ptr;
1243 		odd = tmp;
1244 		UASM_i_LWX(p, even, scratch, tmp);
1245 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1246 		UASM_i_LWX(p, odd, scratch, tmp);
1247 	} else {
1248 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1249 		even = tmp;
1250 		odd = ptr;
1251 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1252 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1253 	}
1254 	if (cpu_has_rixi) {
1255 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1256 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1257 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1258 	} else {
1259 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1260 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1261 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1262 	}
1263 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1264 
1265 	if (c0_scratch_reg >= 0) {
1266 		uasm_i_ehb(p);
1267 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1268 		build_tlb_write_entry(p, l, r, tlb_random);
1269 		uasm_l_leave(l, *p);
1270 		rv.restore_scratch = 1;
1271 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1272 		build_tlb_write_entry(p, l, r, tlb_random);
1273 		uasm_l_leave(l, *p);
1274 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1275 	} else {
1276 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1277 		build_tlb_write_entry(p, l, r, tlb_random);
1278 		uasm_l_leave(l, *p);
1279 		rv.restore_scratch = 1;
1280 	}
1281 
1282 	uasm_i_eret(p); /* return from trap */
1283 
1284 	return rv;
1285 }
1286 
1287 /*
1288  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1289  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1290  * slots before the XTLB refill exception handler which belong to the
1291  * unused TLB refill exception.
1292  */
1293 #define MIPS64_REFILL_INSNS 32
1294 
1295 static void build_r4000_tlb_refill_handler(void)
1296 {
1297 	u32 *p = tlb_handler;
1298 	struct uasm_label *l = labels;
1299 	struct uasm_reloc *r = relocs;
1300 	u32 *f;
1301 	unsigned int final_len;
1302 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1303 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1304 
1305 	memset(tlb_handler, 0, sizeof(tlb_handler));
1306 	memset(labels, 0, sizeof(labels));
1307 	memset(relocs, 0, sizeof(relocs));
1308 	memset(final_handler, 0, sizeof(final_handler));
1309 
1310 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1311 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1312 							  scratch_reg);
1313 		vmalloc_mode = refill_scratch;
1314 	} else {
1315 		htlb_info.huge_pte = K0;
1316 		htlb_info.restore_scratch = 0;
1317 		htlb_info.need_reload_pte = true;
1318 		vmalloc_mode = refill_noscratch;
1319 		/*
1320 		 * create the plain linear handler
1321 		 */
1322 		if (bcm1250_m3_war()) {
1323 			unsigned int segbits = 44;
1324 
1325 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1326 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1327 			uasm_i_xor(&p, K0, K0, K1);
1328 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1329 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1330 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1331 			uasm_i_or(&p, K0, K0, K1);
1332 			uasm_il_bnez(&p, &r, K0, label_leave);
1333 			/* No need for uasm_i_nop */
1334 		}
1335 
1336 #ifdef CONFIG_64BIT
1337 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1338 #else
1339 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1340 #endif
1341 
1342 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1343 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1344 #endif
1345 
1346 		build_get_ptep(&p, K0, K1);
1347 		build_update_entries(&p, K0, K1);
1348 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1349 		uasm_l_leave(&l, p);
1350 		uasm_i_eret(&p); /* return from trap */
1351 	}
1352 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1353 	uasm_l_tlb_huge_update(&l, p);
1354 	if (htlb_info.need_reload_pte)
1355 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1356 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1357 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1358 				   htlb_info.restore_scratch);
1359 #endif
1360 
1361 #ifdef CONFIG_64BIT
1362 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1363 #endif
1364 
1365 	/*
1366 	 * Overflow check: For the 64bit handler, we need at least one
1367 	 * free instruction slot for the wrap-around branch. In worst
1368 	 * case, if the intended insertion point is a delay slot, we
1369 	 * need three, with the second nop'ed and the third being
1370 	 * unused.
1371 	 */
1372 	switch (boot_cpu_type()) {
1373 	default:
1374 		if (sizeof(long) == 4) {
1375 	case CPU_LOONGSON2:
1376 		/* Loongson2 ebase is different than r4k, we have more space */
1377 			if ((p - tlb_handler) > 64)
1378 				panic("TLB refill handler space exceeded");
1379 			/*
1380 			 * Now fold the handler in the TLB refill handler space.
1381 			 */
1382 			f = final_handler;
1383 			/* Simplest case, just copy the handler. */
1384 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1385 			final_len = p - tlb_handler;
1386 			break;
1387 		} else {
1388 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1389 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1390 				&& uasm_insn_has_bdelay(relocs,
1391 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1392 				panic("TLB refill handler space exceeded");
1393 			/*
1394 			 * Now fold the handler in the TLB refill handler space.
1395 			 */
1396 			f = final_handler + MIPS64_REFILL_INSNS;
1397 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1398 				/* Just copy the handler. */
1399 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1400 				final_len = p - tlb_handler;
1401 			} else {
1402 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1403 				const enum label_id ls = label_tlb_huge_update;
1404 #else
1405 				const enum label_id ls = label_vmalloc;
1406 #endif
1407 				u32 *split;
1408 				int ov = 0;
1409 				int i;
1410 
1411 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1412 					;
1413 				BUG_ON(i == ARRAY_SIZE(labels));
1414 				split = labels[i].addr;
1415 
1416 				/*
1417 				 * See if we have overflown one way or the other.
1418 				 */
1419 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1420 				    split < p - MIPS64_REFILL_INSNS)
1421 					ov = 1;
1422 
1423 				if (ov) {
1424 					/*
1425 					 * Split two instructions before the end.  One
1426 					 * for the branch and one for the instruction
1427 					 * in the delay slot.
1428 					 */
1429 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1430 
1431 					/*
1432 					 * If the branch would fall in a delay slot,
1433 					 * we must back up an additional instruction
1434 					 * so that it is no longer in a delay slot.
1435 					 */
1436 					if (uasm_insn_has_bdelay(relocs, split - 1))
1437 						split--;
1438 				}
1439 				/* Copy first part of the handler. */
1440 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1441 				f += split - tlb_handler;
1442 
1443 				if (ov) {
1444 					/* Insert branch. */
1445 					uasm_l_split(&l, final_handler);
1446 					uasm_il_b(&f, &r, label_split);
1447 					if (uasm_insn_has_bdelay(relocs, split))
1448 						uasm_i_nop(&f);
1449 					else {
1450 						uasm_copy_handler(relocs, labels,
1451 								  split, split + 1, f);
1452 						uasm_move_labels(labels, f, f + 1, -1);
1453 						f++;
1454 						split++;
1455 					}
1456 				}
1457 
1458 				/* Copy the rest of the handler. */
1459 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1460 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1461 					    (p - split);
1462 			}
1463 		}
1464 		break;
1465 	}
1466 
1467 	uasm_resolve_relocs(relocs, labels);
1468 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1469 		 final_len);
1470 
1471 	memcpy((void *)ebase, final_handler, 0x100);
1472 	local_flush_icache_range(ebase, ebase + 0x100);
1473 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1474 }
1475 
1476 static void setup_pw(void)
1477 {
1478 	unsigned long pgd_i, pgd_w;
1479 #ifndef __PAGETABLE_PMD_FOLDED
1480 	unsigned long pmd_i, pmd_w;
1481 #endif
1482 	unsigned long pt_i, pt_w;
1483 	unsigned long pte_i, pte_w;
1484 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1485 	unsigned long psn;
1486 
1487 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1488 #endif
1489 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1490 #ifndef __PAGETABLE_PMD_FOLDED
1491 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1492 
1493 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1494 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1495 #else
1496 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1497 #endif
1498 
1499 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1500 	pt_w  = PAGE_SHIFT - 3;
1501 
1502 	pte_i = ilog2(_PAGE_GLOBAL);
1503 	pte_w = 0;
1504 
1505 #ifndef __PAGETABLE_PMD_FOLDED
1506 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1507 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1508 #else
1509 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1510 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1511 #endif
1512 
1513 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1514 	write_c0_pwctl(1 << 6 | psn);
1515 #endif
1516 	write_c0_kpgd((long)swapper_pg_dir);
1517 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1518 }
1519 
1520 static void build_loongson3_tlb_refill_handler(void)
1521 {
1522 	u32 *p = tlb_handler;
1523 	struct uasm_label *l = labels;
1524 	struct uasm_reloc *r = relocs;
1525 
1526 	memset(labels, 0, sizeof(labels));
1527 	memset(relocs, 0, sizeof(relocs));
1528 	memset(tlb_handler, 0, sizeof(tlb_handler));
1529 
1530 	if (check_for_high_segbits) {
1531 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1532 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1533 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1534 		uasm_i_nop(&p);
1535 
1536 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1537 		uasm_i_nop(&p);
1538 		uasm_l_vmalloc(&l, p);
1539 	}
1540 
1541 	uasm_i_dmfc0(&p, K1, C0_PGD);
1542 
1543 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1544 #ifndef __PAGETABLE_PMD_FOLDED
1545 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1546 #endif
1547 	uasm_i_ldpte(&p, K1, 0);      /* even */
1548 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1549 	uasm_i_tlbwr(&p);
1550 
1551 	/* restore page mask */
1552 	if (PM_DEFAULT_MASK >> 16) {
1553 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1554 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1555 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1556 	} else if (PM_DEFAULT_MASK) {
1557 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1558 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1559 	} else {
1560 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1561 	}
1562 
1563 	uasm_i_eret(&p);
1564 
1565 	if (check_for_high_segbits) {
1566 		uasm_l_large_segbits_fault(&l, p);
1567 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1568 		uasm_i_jr(&p, K1);
1569 		uasm_i_nop(&p);
1570 	}
1571 
1572 	uasm_resolve_relocs(relocs, labels);
1573 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1574 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1575 	dump_handler("loongson3_tlb_refill",
1576 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1577 }
1578 
1579 static void build_setup_pgd(void)
1580 {
1581 	const int a0 = 4;
1582 	const int __maybe_unused a1 = 5;
1583 	const int __maybe_unused a2 = 6;
1584 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1585 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1586 	long pgdc = (long)pgd_current;
1587 #endif
1588 
1589 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1590 	memset(labels, 0, sizeof(labels));
1591 	memset(relocs, 0, sizeof(relocs));
1592 	pgd_reg = allocate_kscratch();
1593 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1594 	if (pgd_reg == -1) {
1595 		struct uasm_label *l = labels;
1596 		struct uasm_reloc *r = relocs;
1597 
1598 		/* PGD << 11 in c0_Context */
1599 		/*
1600 		 * If it is a ckseg0 address, convert to a physical
1601 		 * address.  Shifting right by 29 and adding 4 will
1602 		 * result in zero for these addresses.
1603 		 *
1604 		 */
1605 		UASM_i_SRA(&p, a1, a0, 29);
1606 		UASM_i_ADDIU(&p, a1, a1, 4);
1607 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1608 		uasm_i_nop(&p);
1609 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1610 		uasm_l_tlbl_goaround1(&l, p);
1611 		UASM_i_SLL(&p, a0, a0, 11);
1612 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1613 		uasm_i_jr(&p, 31);
1614 		uasm_i_ehb(&p);
1615 	} else {
1616 		/* PGD in c0_KScratch */
1617 		if (cpu_has_ldpte)
1618 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1619 		else
1620 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1621 		uasm_i_jr(&p, 31);
1622 		uasm_i_ehb(&p);
1623 	}
1624 #else
1625 #ifdef CONFIG_SMP
1626 	/* Save PGD to pgd_current[smp_processor_id()] */
1627 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1628 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1629 	UASM_i_LA_mostly(&p, a2, pgdc);
1630 	UASM_i_ADDU(&p, a2, a2, a1);
1631 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1632 #else
1633 	UASM_i_LA_mostly(&p, a2, pgdc);
1634 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1635 #endif /* SMP */
1636 
1637 	/* if pgd_reg is allocated, save PGD also to scratch register */
1638 	if (pgd_reg != -1) {
1639 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1640 		uasm_i_jr(&p, 31);
1641 		uasm_i_ehb(&p);
1642 	} else {
1643 		uasm_i_jr(&p, 31);
1644 		uasm_i_nop(&p);
1645 	}
1646 #endif
1647 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1648 		panic("tlbmiss_handler_setup_pgd space exceeded");
1649 
1650 	uasm_resolve_relocs(relocs, labels);
1651 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1652 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1653 
1654 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1655 					tlbmiss_handler_setup_pgd_end);
1656 }
1657 
1658 static void
1659 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1660 {
1661 #ifdef CONFIG_SMP
1662 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1663 		uasm_i_sync(p, 0);
1664 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1665 	if (cpu_has_64bits)
1666 		uasm_i_lld(p, pte, 0, ptr);
1667 	else
1668 # endif
1669 		UASM_i_LL(p, pte, 0, ptr);
1670 #else
1671 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1672 	if (cpu_has_64bits)
1673 		uasm_i_ld(p, pte, 0, ptr);
1674 	else
1675 # endif
1676 		UASM_i_LW(p, pte, 0, ptr);
1677 #endif
1678 }
1679 
1680 static void
1681 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1682 	unsigned int mode, unsigned int scratch)
1683 {
1684 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1685 	unsigned int swmode = mode & ~hwmode;
1686 
1687 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1688 		uasm_i_lui(p, scratch, swmode >> 16);
1689 		uasm_i_or(p, pte, pte, scratch);
1690 		BUG_ON(swmode & 0xffff);
1691 	} else {
1692 		uasm_i_ori(p, pte, pte, mode);
1693 	}
1694 
1695 #ifdef CONFIG_SMP
1696 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1697 	if (cpu_has_64bits)
1698 		uasm_i_scd(p, pte, 0, ptr);
1699 	else
1700 # endif
1701 		UASM_i_SC(p, pte, 0, ptr);
1702 
1703 	if (r10000_llsc_war())
1704 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1705 	else
1706 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1707 
1708 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1709 	if (!cpu_has_64bits) {
1710 		/* no uasm_i_nop needed */
1711 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1712 		uasm_i_ori(p, pte, pte, hwmode);
1713 		BUG_ON(hwmode & ~0xffff);
1714 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1715 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1716 		/* no uasm_i_nop needed */
1717 		uasm_i_lw(p, pte, 0, ptr);
1718 	} else
1719 		uasm_i_nop(p);
1720 # else
1721 	uasm_i_nop(p);
1722 # endif
1723 #else
1724 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1725 	if (cpu_has_64bits)
1726 		uasm_i_sd(p, pte, 0, ptr);
1727 	else
1728 # endif
1729 		UASM_i_SW(p, pte, 0, ptr);
1730 
1731 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1732 	if (!cpu_has_64bits) {
1733 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1734 		uasm_i_ori(p, pte, pte, hwmode);
1735 		BUG_ON(hwmode & ~0xffff);
1736 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1737 		uasm_i_lw(p, pte, 0, ptr);
1738 	}
1739 # endif
1740 #endif
1741 }
1742 
1743 /*
1744  * Check if PTE is present, if not then jump to LABEL. PTR points to
1745  * the page table where this PTE is located, PTE will be re-loaded
1746  * with it's original value.
1747  */
1748 static void
1749 build_pte_present(u32 **p, struct uasm_reloc **r,
1750 		  int pte, int ptr, int scratch, enum label_id lid)
1751 {
1752 	int t = scratch >= 0 ? scratch : pte;
1753 	int cur = pte;
1754 
1755 	if (cpu_has_rixi) {
1756 		if (use_bbit_insns()) {
1757 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1758 			uasm_i_nop(p);
1759 		} else {
1760 			if (_PAGE_PRESENT_SHIFT) {
1761 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1762 				cur = t;
1763 			}
1764 			uasm_i_andi(p, t, cur, 1);
1765 			uasm_il_beqz(p, r, t, lid);
1766 			if (pte == t)
1767 				/* You lose the SMP race :-(*/
1768 				iPTE_LW(p, pte, ptr);
1769 		}
1770 	} else {
1771 		if (_PAGE_PRESENT_SHIFT) {
1772 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1773 			cur = t;
1774 		}
1775 		uasm_i_andi(p, t, cur,
1776 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1777 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1778 		uasm_il_bnez(p, r, t, lid);
1779 		if (pte == t)
1780 			/* You lose the SMP race :-(*/
1781 			iPTE_LW(p, pte, ptr);
1782 	}
1783 }
1784 
1785 /* Make PTE valid, store result in PTR. */
1786 static void
1787 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1788 		 unsigned int ptr, unsigned int scratch)
1789 {
1790 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1791 
1792 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1793 }
1794 
1795 /*
1796  * Check if PTE can be written to, if not branch to LABEL. Regardless
1797  * restore PTE with value from PTR when done.
1798  */
1799 static void
1800 build_pte_writable(u32 **p, struct uasm_reloc **r,
1801 		   unsigned int pte, unsigned int ptr, int scratch,
1802 		   enum label_id lid)
1803 {
1804 	int t = scratch >= 0 ? scratch : pte;
1805 	int cur = pte;
1806 
1807 	if (_PAGE_PRESENT_SHIFT) {
1808 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1809 		cur = t;
1810 	}
1811 	uasm_i_andi(p, t, cur,
1812 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1813 	uasm_i_xori(p, t, t,
1814 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1815 	uasm_il_bnez(p, r, t, lid);
1816 	if (pte == t)
1817 		/* You lose the SMP race :-(*/
1818 		iPTE_LW(p, pte, ptr);
1819 	else
1820 		uasm_i_nop(p);
1821 }
1822 
1823 /* Make PTE writable, update software status bits as well, then store
1824  * at PTR.
1825  */
1826 static void
1827 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1828 		 unsigned int ptr, unsigned int scratch)
1829 {
1830 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1831 			     | _PAGE_DIRTY);
1832 
1833 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1834 }
1835 
1836 /*
1837  * Check if PTE can be modified, if not branch to LABEL. Regardless
1838  * restore PTE with value from PTR when done.
1839  */
1840 static void
1841 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1842 		     unsigned int pte, unsigned int ptr, int scratch,
1843 		     enum label_id lid)
1844 {
1845 	if (use_bbit_insns()) {
1846 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1847 		uasm_i_nop(p);
1848 	} else {
1849 		int t = scratch >= 0 ? scratch : pte;
1850 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1851 		uasm_i_andi(p, t, t, 1);
1852 		uasm_il_beqz(p, r, t, lid);
1853 		if (pte == t)
1854 			/* You lose the SMP race :-(*/
1855 			iPTE_LW(p, pte, ptr);
1856 	}
1857 }
1858 
1859 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1860 
1861 
1862 /*
1863  * R3000 style TLB load/store/modify handlers.
1864  */
1865 
1866 /*
1867  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1868  * Then it returns.
1869  */
1870 static void
1871 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1872 {
1873 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1874 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1875 	uasm_i_tlbwi(p);
1876 	uasm_i_jr(p, tmp);
1877 	uasm_i_rfe(p); /* branch delay */
1878 }
1879 
1880 /*
1881  * This places the pte into ENTRYLO0 and writes it with tlbwi
1882  * or tlbwr as appropriate.  This is because the index register
1883  * may have the probe fail bit set as a result of a trap on a
1884  * kseg2 access, i.e. without refill.  Then it returns.
1885  */
1886 static void
1887 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1888 			     struct uasm_reloc **r, unsigned int pte,
1889 			     unsigned int tmp)
1890 {
1891 	uasm_i_mfc0(p, tmp, C0_INDEX);
1892 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1893 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1894 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1895 	uasm_i_tlbwi(p); /* cp0 delay */
1896 	uasm_i_jr(p, tmp);
1897 	uasm_i_rfe(p); /* branch delay */
1898 	uasm_l_r3000_write_probe_fail(l, *p);
1899 	uasm_i_tlbwr(p); /* cp0 delay */
1900 	uasm_i_jr(p, tmp);
1901 	uasm_i_rfe(p); /* branch delay */
1902 }
1903 
1904 static void
1905 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1906 				   unsigned int ptr)
1907 {
1908 	long pgdc = (long)pgd_current;
1909 
1910 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1911 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1912 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1913 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1914 	uasm_i_sll(p, pte, pte, 2);
1915 	uasm_i_addu(p, ptr, ptr, pte);
1916 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1917 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1918 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1919 	uasm_i_addu(p, ptr, ptr, pte);
1920 	uasm_i_lw(p, pte, 0, ptr);
1921 	uasm_i_tlbp(p); /* load delay */
1922 }
1923 
1924 static void build_r3000_tlb_load_handler(void)
1925 {
1926 	u32 *p = (u32 *)handle_tlbl;
1927 	struct uasm_label *l = labels;
1928 	struct uasm_reloc *r = relocs;
1929 
1930 	memset(p, 0, handle_tlbl_end - (char *)p);
1931 	memset(labels, 0, sizeof(labels));
1932 	memset(relocs, 0, sizeof(relocs));
1933 
1934 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1935 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1936 	uasm_i_nop(&p); /* load delay */
1937 	build_make_valid(&p, &r, K0, K1, -1);
1938 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1939 
1940 	uasm_l_nopage_tlbl(&l, p);
1941 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1942 	uasm_i_nop(&p);
1943 
1944 	if (p >= (u32 *)handle_tlbl_end)
1945 		panic("TLB load handler fastpath space exceeded");
1946 
1947 	uasm_resolve_relocs(relocs, labels);
1948 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1949 		 (unsigned int)(p - (u32 *)handle_tlbl));
1950 
1951 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1952 }
1953 
1954 static void build_r3000_tlb_store_handler(void)
1955 {
1956 	u32 *p = (u32 *)handle_tlbs;
1957 	struct uasm_label *l = labels;
1958 	struct uasm_reloc *r = relocs;
1959 
1960 	memset(p, 0, handle_tlbs_end - (char *)p);
1961 	memset(labels, 0, sizeof(labels));
1962 	memset(relocs, 0, sizeof(relocs));
1963 
1964 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1965 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1966 	uasm_i_nop(&p); /* load delay */
1967 	build_make_write(&p, &r, K0, K1, -1);
1968 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1969 
1970 	uasm_l_nopage_tlbs(&l, p);
1971 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1972 	uasm_i_nop(&p);
1973 
1974 	if (p >= (u32 *)handle_tlbs_end)
1975 		panic("TLB store handler fastpath space exceeded");
1976 
1977 	uasm_resolve_relocs(relocs, labels);
1978 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1979 		 (unsigned int)(p - (u32 *)handle_tlbs));
1980 
1981 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1982 }
1983 
1984 static void build_r3000_tlb_modify_handler(void)
1985 {
1986 	u32 *p = (u32 *)handle_tlbm;
1987 	struct uasm_label *l = labels;
1988 	struct uasm_reloc *r = relocs;
1989 
1990 	memset(p, 0, handle_tlbm_end - (char *)p);
1991 	memset(labels, 0, sizeof(labels));
1992 	memset(relocs, 0, sizeof(relocs));
1993 
1994 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1995 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1996 	uasm_i_nop(&p); /* load delay */
1997 	build_make_write(&p, &r, K0, K1, -1);
1998 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1999 
2000 	uasm_l_nopage_tlbm(&l, p);
2001 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2002 	uasm_i_nop(&p);
2003 
2004 	if (p >= (u32 *)handle_tlbm_end)
2005 		panic("TLB modify handler fastpath space exceeded");
2006 
2007 	uasm_resolve_relocs(relocs, labels);
2008 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2009 		 (unsigned int)(p - (u32 *)handle_tlbm));
2010 
2011 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2012 }
2013 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2014 
2015 static bool cpu_has_tlbex_tlbp_race(void)
2016 {
2017 	/*
2018 	 * When a Hardware Table Walker is running it can replace TLB entries
2019 	 * at any time, leading to a race between it & the CPU.
2020 	 */
2021 	if (cpu_has_htw)
2022 		return true;
2023 
2024 	/*
2025 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2026 	 * replaced at any time by a sibling performing a write to the FTLB.
2027 	 */
2028 	if (cpu_has_shared_ftlb_ram)
2029 		return true;
2030 
2031 	/* In all other cases there ought to be no race condition to handle */
2032 	return false;
2033 }
2034 
2035 /*
2036  * R4000 style TLB load/store/modify handlers.
2037  */
2038 static struct work_registers
2039 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2040 				   struct uasm_reloc **r)
2041 {
2042 	struct work_registers wr = build_get_work_registers(p);
2043 
2044 #ifdef CONFIG_64BIT
2045 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2046 #else
2047 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2048 #endif
2049 
2050 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2051 	/*
2052 	 * For huge tlb entries, pmd doesn't contain an address but
2053 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2054 	 * see if we need to jump to huge tlb processing.
2055 	 */
2056 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2057 #endif
2058 
2059 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2060 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2061 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2062 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2063 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2064 
2065 #ifdef CONFIG_SMP
2066 	uasm_l_smp_pgtable_change(l, *p);
2067 #endif
2068 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2069 	if (!m4kc_tlbp_war()) {
2070 		build_tlb_probe_entry(p);
2071 		if (cpu_has_tlbex_tlbp_race()) {
2072 			/* race condition happens, leaving */
2073 			uasm_i_ehb(p);
2074 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2075 			uasm_il_bltz(p, r, wr.r3, label_leave);
2076 			uasm_i_nop(p);
2077 		}
2078 	}
2079 	return wr;
2080 }
2081 
2082 static void
2083 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2084 				   struct uasm_reloc **r, unsigned int tmp,
2085 				   unsigned int ptr)
2086 {
2087 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2088 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2089 	build_update_entries(p, tmp, ptr);
2090 	build_tlb_write_entry(p, l, r, tlb_indexed);
2091 	uasm_l_leave(l, *p);
2092 	build_restore_work_registers(p);
2093 	uasm_i_eret(p); /* return from trap */
2094 
2095 #ifdef CONFIG_64BIT
2096 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2097 #endif
2098 }
2099 
2100 static void build_r4000_tlb_load_handler(void)
2101 {
2102 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2103 	struct uasm_label *l = labels;
2104 	struct uasm_reloc *r = relocs;
2105 	struct work_registers wr;
2106 
2107 	memset(p, 0, handle_tlbl_end - (char *)p);
2108 	memset(labels, 0, sizeof(labels));
2109 	memset(relocs, 0, sizeof(relocs));
2110 
2111 	if (bcm1250_m3_war()) {
2112 		unsigned int segbits = 44;
2113 
2114 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2115 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2116 		uasm_i_xor(&p, K0, K0, K1);
2117 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2118 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2119 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2120 		uasm_i_or(&p, K0, K0, K1);
2121 		uasm_il_bnez(&p, &r, K0, label_leave);
2122 		/* No need for uasm_i_nop */
2123 	}
2124 
2125 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2126 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2127 	if (m4kc_tlbp_war())
2128 		build_tlb_probe_entry(&p);
2129 
2130 	if (cpu_has_rixi && !cpu_has_rixiex) {
2131 		/*
2132 		 * If the page is not _PAGE_VALID, RI or XI could not
2133 		 * have triggered it.  Skip the expensive test..
2134 		 */
2135 		if (use_bbit_insns()) {
2136 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2137 				      label_tlbl_goaround1);
2138 		} else {
2139 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2140 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2141 		}
2142 		uasm_i_nop(&p);
2143 
2144 		/*
2145 		 * Warn if something may race with us & replace the TLB entry
2146 		 * before we read it here. Everything with such races should
2147 		 * also have dedicated RiXi exception handlers, so this
2148 		 * shouldn't be hit.
2149 		 */
2150 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2151 
2152 		uasm_i_tlbr(&p);
2153 
2154 		switch (current_cpu_type()) {
2155 		default:
2156 			if (cpu_has_mips_r2_exec_hazard) {
2157 				uasm_i_ehb(&p);
2158 
2159 		case CPU_CAVIUM_OCTEON:
2160 		case CPU_CAVIUM_OCTEON_PLUS:
2161 		case CPU_CAVIUM_OCTEON2:
2162 				break;
2163 			}
2164 		}
2165 
2166 		/* Examine  entrylo 0 or 1 based on ptr. */
2167 		if (use_bbit_insns()) {
2168 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2169 		} else {
2170 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2171 			uasm_i_beqz(&p, wr.r3, 8);
2172 		}
2173 		/* load it in the delay slot*/
2174 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2175 		/* load it if ptr is odd */
2176 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2177 		/*
2178 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2179 		 * XI must have triggered it.
2180 		 */
2181 		if (use_bbit_insns()) {
2182 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2183 			uasm_i_nop(&p);
2184 			uasm_l_tlbl_goaround1(&l, p);
2185 		} else {
2186 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2187 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2188 			uasm_i_nop(&p);
2189 		}
2190 		uasm_l_tlbl_goaround1(&l, p);
2191 	}
2192 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2193 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2194 
2195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2196 	/*
2197 	 * This is the entry point when build_r4000_tlbchange_handler_head
2198 	 * spots a huge page.
2199 	 */
2200 	uasm_l_tlb_huge_update(&l, p);
2201 	iPTE_LW(&p, wr.r1, wr.r2);
2202 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2203 	build_tlb_probe_entry(&p);
2204 
2205 	if (cpu_has_rixi && !cpu_has_rixiex) {
2206 		/*
2207 		 * If the page is not _PAGE_VALID, RI or XI could not
2208 		 * have triggered it.  Skip the expensive test..
2209 		 */
2210 		if (use_bbit_insns()) {
2211 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2212 				      label_tlbl_goaround2);
2213 		} else {
2214 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2215 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2216 		}
2217 		uasm_i_nop(&p);
2218 
2219 		/*
2220 		 * Warn if something may race with us & replace the TLB entry
2221 		 * before we read it here. Everything with such races should
2222 		 * also have dedicated RiXi exception handlers, so this
2223 		 * shouldn't be hit.
2224 		 */
2225 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2226 
2227 		uasm_i_tlbr(&p);
2228 
2229 		switch (current_cpu_type()) {
2230 		default:
2231 			if (cpu_has_mips_r2_exec_hazard) {
2232 				uasm_i_ehb(&p);
2233 
2234 		case CPU_CAVIUM_OCTEON:
2235 		case CPU_CAVIUM_OCTEON_PLUS:
2236 		case CPU_CAVIUM_OCTEON2:
2237 				break;
2238 			}
2239 		}
2240 
2241 		/* Examine  entrylo 0 or 1 based on ptr. */
2242 		if (use_bbit_insns()) {
2243 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2244 		} else {
2245 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2246 			uasm_i_beqz(&p, wr.r3, 8);
2247 		}
2248 		/* load it in the delay slot*/
2249 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2250 		/* load it if ptr is odd */
2251 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2252 		/*
2253 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2254 		 * XI must have triggered it.
2255 		 */
2256 		if (use_bbit_insns()) {
2257 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2258 		} else {
2259 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2260 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2261 		}
2262 		if (PM_DEFAULT_MASK == 0)
2263 			uasm_i_nop(&p);
2264 		/*
2265 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2266 		 * it is restored in build_huge_tlb_write_entry.
2267 		 */
2268 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2269 
2270 		uasm_l_tlbl_goaround2(&l, p);
2271 	}
2272 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2273 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2274 #endif
2275 
2276 	uasm_l_nopage_tlbl(&l, p);
2277 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2278 		uasm_i_sync(&p, 0);
2279 	build_restore_work_registers(&p);
2280 #ifdef CONFIG_CPU_MICROMIPS
2281 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2282 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2283 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2284 		uasm_i_jr(&p, K0);
2285 	} else
2286 #endif
2287 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2288 	uasm_i_nop(&p);
2289 
2290 	if (p >= (u32 *)handle_tlbl_end)
2291 		panic("TLB load handler fastpath space exceeded");
2292 
2293 	uasm_resolve_relocs(relocs, labels);
2294 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2295 		 (unsigned int)(p - (u32 *)handle_tlbl));
2296 
2297 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2298 }
2299 
2300 static void build_r4000_tlb_store_handler(void)
2301 {
2302 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2303 	struct uasm_label *l = labels;
2304 	struct uasm_reloc *r = relocs;
2305 	struct work_registers wr;
2306 
2307 	memset(p, 0, handle_tlbs_end - (char *)p);
2308 	memset(labels, 0, sizeof(labels));
2309 	memset(relocs, 0, sizeof(relocs));
2310 
2311 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2312 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2313 	if (m4kc_tlbp_war())
2314 		build_tlb_probe_entry(&p);
2315 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2316 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2317 
2318 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2319 	/*
2320 	 * This is the entry point when
2321 	 * build_r4000_tlbchange_handler_head spots a huge page.
2322 	 */
2323 	uasm_l_tlb_huge_update(&l, p);
2324 	iPTE_LW(&p, wr.r1, wr.r2);
2325 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2326 	build_tlb_probe_entry(&p);
2327 	uasm_i_ori(&p, wr.r1, wr.r1,
2328 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2329 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2330 #endif
2331 
2332 	uasm_l_nopage_tlbs(&l, p);
2333 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2334 		uasm_i_sync(&p, 0);
2335 	build_restore_work_registers(&p);
2336 #ifdef CONFIG_CPU_MICROMIPS
2337 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2338 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2339 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2340 		uasm_i_jr(&p, K0);
2341 	} else
2342 #endif
2343 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2344 	uasm_i_nop(&p);
2345 
2346 	if (p >= (u32 *)handle_tlbs_end)
2347 		panic("TLB store handler fastpath space exceeded");
2348 
2349 	uasm_resolve_relocs(relocs, labels);
2350 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2351 		 (unsigned int)(p - (u32 *)handle_tlbs));
2352 
2353 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2354 }
2355 
2356 static void build_r4000_tlb_modify_handler(void)
2357 {
2358 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2359 	struct uasm_label *l = labels;
2360 	struct uasm_reloc *r = relocs;
2361 	struct work_registers wr;
2362 
2363 	memset(p, 0, handle_tlbm_end - (char *)p);
2364 	memset(labels, 0, sizeof(labels));
2365 	memset(relocs, 0, sizeof(relocs));
2366 
2367 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2368 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2369 	if (m4kc_tlbp_war())
2370 		build_tlb_probe_entry(&p);
2371 	/* Present and writable bits set, set accessed and dirty bits. */
2372 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2373 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2374 
2375 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2376 	/*
2377 	 * This is the entry point when
2378 	 * build_r4000_tlbchange_handler_head spots a huge page.
2379 	 */
2380 	uasm_l_tlb_huge_update(&l, p);
2381 	iPTE_LW(&p, wr.r1, wr.r2);
2382 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2383 	build_tlb_probe_entry(&p);
2384 	uasm_i_ori(&p, wr.r1, wr.r1,
2385 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2386 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2387 #endif
2388 
2389 	uasm_l_nopage_tlbm(&l, p);
2390 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2391 		uasm_i_sync(&p, 0);
2392 	build_restore_work_registers(&p);
2393 #ifdef CONFIG_CPU_MICROMIPS
2394 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2395 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2396 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2397 		uasm_i_jr(&p, K0);
2398 	} else
2399 #endif
2400 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2401 	uasm_i_nop(&p);
2402 
2403 	if (p >= (u32 *)handle_tlbm_end)
2404 		panic("TLB modify handler fastpath space exceeded");
2405 
2406 	uasm_resolve_relocs(relocs, labels);
2407 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2408 		 (unsigned int)(p - (u32 *)handle_tlbm));
2409 
2410 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2411 }
2412 
2413 static void flush_tlb_handlers(void)
2414 {
2415 	local_flush_icache_range((unsigned long)handle_tlbl,
2416 			   (unsigned long)handle_tlbl_end);
2417 	local_flush_icache_range((unsigned long)handle_tlbs,
2418 			   (unsigned long)handle_tlbs_end);
2419 	local_flush_icache_range((unsigned long)handle_tlbm,
2420 			   (unsigned long)handle_tlbm_end);
2421 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2422 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2423 }
2424 
2425 static void print_htw_config(void)
2426 {
2427 	unsigned long config;
2428 	unsigned int pwctl;
2429 	const int field = 2 * sizeof(unsigned long);
2430 
2431 	config = read_c0_pwfield();
2432 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2433 		field, config,
2434 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2435 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2436 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2437 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2438 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2439 
2440 	config = read_c0_pwsize();
2441 	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",
2442 		field, config,
2443 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2444 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2445 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2446 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2447 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2448 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2449 
2450 	pwctl = read_c0_pwctl();
2451 	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",
2452 		pwctl,
2453 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2454 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2455 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2456 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2457 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2458 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2459 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2460 }
2461 
2462 static void config_htw_params(void)
2463 {
2464 	unsigned long pwfield, pwsize, ptei;
2465 	unsigned int config;
2466 
2467 	/*
2468 	 * We are using 2-level page tables, so we only need to
2469 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2470 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2471 	 * write values less than 0xc in these fields because the entire
2472 	 * write will be dropped. As a result of which, we must preserve
2473 	 * the original reset values and overwrite only what we really want.
2474 	 */
2475 
2476 	pwfield = read_c0_pwfield();
2477 	/* re-initialize the GDI field */
2478 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2479 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2480 	/* re-initialize the PTI field including the even/odd bit */
2481 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2482 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2483 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2484 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2485 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2486 	}
2487 	/* Set the PTEI right shift */
2488 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2489 	pwfield |= ptei;
2490 	write_c0_pwfield(pwfield);
2491 	/* Check whether the PTEI value is supported */
2492 	back_to_back_c0_hazard();
2493 	pwfield = read_c0_pwfield();
2494 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2495 		!= ptei) {
2496 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2497 			ptei);
2498 		/*
2499 		 * Drop option to avoid HTW being enabled via another path
2500 		 * (eg htw_reset())
2501 		 */
2502 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2503 		return;
2504 	}
2505 
2506 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2507 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2508 	if (CONFIG_PGTABLE_LEVELS >= 3)
2509 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2510 
2511 	/* Set pointer size to size of directory pointers */
2512 	if (IS_ENABLED(CONFIG_64BIT))
2513 		pwsize |= MIPS_PWSIZE_PS_MASK;
2514 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2515 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2516 			& MIPS_PWSIZE_PTEW_MASK;
2517 
2518 	write_c0_pwsize(pwsize);
2519 
2520 	/* Make sure everything is set before we enable the HTW */
2521 	back_to_back_c0_hazard();
2522 
2523 	/*
2524 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2525 	 * the pwctl fields.
2526 	 */
2527 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2528 	if (IS_ENABLED(CONFIG_64BIT))
2529 		config |= MIPS_PWCTL_XU_MASK;
2530 	write_c0_pwctl(config);
2531 	pr_info("Hardware Page Table Walker enabled\n");
2532 
2533 	print_htw_config();
2534 }
2535 
2536 static void config_xpa_params(void)
2537 {
2538 #ifdef CONFIG_XPA
2539 	unsigned int pagegrain;
2540 
2541 	if (mips_xpa_disabled) {
2542 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2543 		return;
2544 	}
2545 
2546 	pagegrain = read_c0_pagegrain();
2547 	write_c0_pagegrain(pagegrain | PG_ELPA);
2548 	back_to_back_c0_hazard();
2549 	pagegrain = read_c0_pagegrain();
2550 
2551 	if (pagegrain & PG_ELPA)
2552 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2553 	else
2554 		panic("Extended Physical Addressing (XPA) disabled");
2555 #endif
2556 }
2557 
2558 static void check_pabits(void)
2559 {
2560 	unsigned long entry;
2561 	unsigned pabits, fillbits;
2562 
2563 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2564 		/*
2565 		 * We'll only be making use of the fact that we can rotate bits
2566 		 * into the fill if the CPU supports RIXI, so don't bother
2567 		 * probing this for CPUs which don't.
2568 		 */
2569 		return;
2570 	}
2571 
2572 	write_c0_entrylo0(~0ul);
2573 	back_to_back_c0_hazard();
2574 	entry = read_c0_entrylo0();
2575 
2576 	/* clear all non-PFN bits */
2577 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2578 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2579 
2580 	/* find a lower bound on PABITS, and upper bound on fill bits */
2581 	pabits = fls_long(entry) + 6;
2582 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2583 
2584 	/* minus the RI & XI bits */
2585 	fillbits -= min_t(unsigned, fillbits, 2);
2586 
2587 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2588 		fill_includes_sw_bits = true;
2589 
2590 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2591 }
2592 
2593 void build_tlb_refill_handler(void)
2594 {
2595 	/*
2596 	 * The refill handler is generated per-CPU, multi-node systems
2597 	 * may have local storage for it. The other handlers are only
2598 	 * needed once.
2599 	 */
2600 	static int run_once = 0;
2601 
2602 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2603 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2604 
2605 	output_pgtable_bits_defines();
2606 	check_pabits();
2607 
2608 #ifdef CONFIG_64BIT
2609 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2610 #endif
2611 
2612 	switch (current_cpu_type()) {
2613 	case CPU_R2000:
2614 	case CPU_R3000:
2615 	case CPU_R3000A:
2616 	case CPU_R3081E:
2617 	case CPU_TX3912:
2618 	case CPU_TX3922:
2619 	case CPU_TX3927:
2620 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2621 		if (cpu_has_local_ebase)
2622 			build_r3000_tlb_refill_handler();
2623 		if (!run_once) {
2624 			if (!cpu_has_local_ebase)
2625 				build_r3000_tlb_refill_handler();
2626 			build_setup_pgd();
2627 			build_r3000_tlb_load_handler();
2628 			build_r3000_tlb_store_handler();
2629 			build_r3000_tlb_modify_handler();
2630 			flush_tlb_handlers();
2631 			run_once++;
2632 		}
2633 #else
2634 		panic("No R3000 TLB refill handler");
2635 #endif
2636 		break;
2637 
2638 	case CPU_R8000:
2639 		panic("No R8000 TLB refill handler yet");
2640 		break;
2641 
2642 	default:
2643 		if (cpu_has_ldpte)
2644 			setup_pw();
2645 
2646 		if (!run_once) {
2647 			scratch_reg = allocate_kscratch();
2648 			build_setup_pgd();
2649 			build_r4000_tlb_load_handler();
2650 			build_r4000_tlb_store_handler();
2651 			build_r4000_tlb_modify_handler();
2652 			if (cpu_has_ldpte)
2653 				build_loongson3_tlb_refill_handler();
2654 			else if (!cpu_has_local_ebase)
2655 				build_r4000_tlb_refill_handler();
2656 			flush_tlb_handlers();
2657 			run_once++;
2658 		}
2659 		if (cpu_has_local_ebase)
2660 			build_r4000_tlb_refill_handler();
2661 		if (cpu_has_xpa)
2662 			config_xpa_params();
2663 		if (cpu_has_htw)
2664 			config_htw_params();
2665 	}
2666 }
2667