xref: /linux/arch/mips/mm/tlbex.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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/pgtable.h>
35 #include <asm/war.h>
36 #include <asm/uasm.h>
37 #include <asm/setup.h>
38 #include <asm/tlbex.h>
39 
40 static int mips_xpa_disabled;
41 
42 static int __init xpa_disable(char *s)
43 {
44 	mips_xpa_disabled = 1;
45 
46 	return 1;
47 }
48 
49 __setup("noxpa", xpa_disable);
50 
51 /*
52  * TLB load/store/modify handlers.
53  *
54  * Only the fastpath gets synthesized at runtime, the slowpath for
55  * do_page_fault remains normal asm.
56  */
57 extern void tlb_do_page_fault_0(void);
58 extern void tlb_do_page_fault_1(void);
59 
60 struct work_registers {
61 	int r1;
62 	int r2;
63 	int r3;
64 };
65 
66 struct tlb_reg_save {
67 	unsigned long a;
68 	unsigned long b;
69 } ____cacheline_aligned_in_smp;
70 
71 static struct tlb_reg_save handler_reg_save[NR_CPUS];
72 
73 static inline int r45k_bvahwbug(void)
74 {
75 	/* XXX: We should probe for the presence of this bug, but we don't. */
76 	return 0;
77 }
78 
79 static inline int r4k_250MHZhwbug(void)
80 {
81 	/* XXX: We should probe for the presence of this bug, but we don't. */
82 	return 0;
83 }
84 
85 static inline int __maybe_unused bcm1250_m3_war(void)
86 {
87 	return BCM1250_M3_WAR;
88 }
89 
90 static inline int __maybe_unused r10000_llsc_war(void)
91 {
92 	return R10000_LLSC_WAR;
93 }
94 
95 static int use_bbit_insns(void)
96 {
97 	switch (current_cpu_type()) {
98 	case CPU_CAVIUM_OCTEON:
99 	case CPU_CAVIUM_OCTEON_PLUS:
100 	case CPU_CAVIUM_OCTEON2:
101 	case CPU_CAVIUM_OCTEON3:
102 		return 1;
103 	default:
104 		return 0;
105 	}
106 }
107 
108 static int use_lwx_insns(void)
109 {
110 	switch (current_cpu_type()) {
111 	case CPU_CAVIUM_OCTEON2:
112 	case CPU_CAVIUM_OCTEON3:
113 		return 1;
114 	default:
115 		return 0;
116 	}
117 }
118 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
119     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
120 static bool scratchpad_available(void)
121 {
122 	return true;
123 }
124 static int scratchpad_offset(int i)
125 {
126 	/*
127 	 * CVMSEG starts at address -32768 and extends for
128 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
129 	 */
130 	i += 1; /* Kernel use starts at the top and works down. */
131 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
132 }
133 #else
134 static bool scratchpad_available(void)
135 {
136 	return false;
137 }
138 static int scratchpad_offset(int i)
139 {
140 	BUG();
141 	/* Really unreachable, but evidently some GCC want this. */
142 	return 0;
143 }
144 #endif
145 /*
146  * Found by experiment: At least some revisions of the 4kc throw under
147  * some circumstances a machine check exception, triggered by invalid
148  * values in the index register.  Delaying the tlbp instruction until
149  * after the next branch,  plus adding an additional nop in front of
150  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
151  * why; it's not an issue caused by the core RTL.
152  *
153  */
154 static int m4kc_tlbp_war(void)
155 {
156 	return (current_cpu_data.processor_id & 0xffff00) ==
157 	       (PRID_COMP_MIPS | PRID_IMP_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 u32 *handler, int count)
258 {
259 	int i;
260 
261 	pr_debug("LEAF(%s)\n", symbol);
262 
263 	pr_debug("\t.set push\n");
264 	pr_debug("\t.set noreorder\n");
265 
266 	for (i = 0; i < count; i++)
267 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
268 
269 	pr_debug("\t.set\tpop\n");
270 
271 	pr_debug("\tEND(%s)\n", symbol);
272 }
273 
274 /* The only general purpose registers allowed in TLB handlers. */
275 #define K0		26
276 #define K1		27
277 
278 /* Some CP0 registers */
279 #define C0_INDEX	0, 0
280 #define C0_ENTRYLO0	2, 0
281 #define C0_TCBIND	2, 2
282 #define C0_ENTRYLO1	3, 0
283 #define C0_CONTEXT	4, 0
284 #define C0_PAGEMASK	5, 0
285 #define C0_PWBASE	5, 5
286 #define C0_PWFIELD	5, 6
287 #define C0_PWSIZE	5, 7
288 #define C0_PWCTL	6, 6
289 #define C0_BADVADDR	8, 0
290 #define C0_PGD		9, 7
291 #define C0_ENTRYHI	10, 0
292 #define C0_EPC		14, 0
293 #define C0_XCONTEXT	20, 0
294 
295 #ifdef CONFIG_64BIT
296 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
297 #else
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
299 #endif
300 
301 /* The worst case length of the handler is around 18 instructions for
302  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
303  * Maximum space available is 32 instructions for R3000 and 64
304  * instructions for R4000.
305  *
306  * We deliberately chose a buffer size of 128, so we won't scribble
307  * over anything important on overflow before we panic.
308  */
309 static u32 tlb_handler[128];
310 
311 /* simply assume worst case size for labels and relocs */
312 static struct uasm_label labels[128];
313 static struct uasm_reloc relocs[128];
314 
315 static int check_for_high_segbits;
316 static bool fill_includes_sw_bits;
317 
318 static unsigned int kscratch_used_mask;
319 
320 static inline int __maybe_unused c0_kscratch(void)
321 {
322 	switch (current_cpu_type()) {
323 	case CPU_XLP:
324 	case CPU_XLR:
325 		return 22;
326 	default:
327 		return 31;
328 	}
329 }
330 
331 static int allocate_kscratch(void)
332 {
333 	int r;
334 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
335 
336 	r = ffs(a);
337 
338 	if (r == 0)
339 		return -1;
340 
341 	r--; /* make it zero based */
342 
343 	kscratch_used_mask |= (1 << r);
344 
345 	return r;
346 }
347 
348 static int scratch_reg;
349 int pgd_reg;
350 EXPORT_SYMBOL_GPL(pgd_reg);
351 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
352 
353 static struct work_registers build_get_work_registers(u32 **p)
354 {
355 	struct work_registers r;
356 
357 	if (scratch_reg >= 0) {
358 		/* Save in CPU local C0_KScratch? */
359 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
360 		r.r1 = K0;
361 		r.r2 = K1;
362 		r.r3 = 1;
363 		return r;
364 	}
365 
366 	if (num_possible_cpus() > 1) {
367 		/* Get smp_processor_id */
368 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
369 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
370 
371 		/* handler_reg_save index in K0 */
372 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
373 
374 		UASM_i_LA(p, K1, (long)&handler_reg_save);
375 		UASM_i_ADDU(p, K0, K0, K1);
376 	} else {
377 		UASM_i_LA(p, K0, (long)&handler_reg_save);
378 	}
379 	/* K0 now points to save area, save $1 and $2  */
380 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
381 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
382 
383 	r.r1 = K1;
384 	r.r2 = 1;
385 	r.r3 = 2;
386 	return r;
387 }
388 
389 static void build_restore_work_registers(u32 **p)
390 {
391 	if (scratch_reg >= 0) {
392 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
393 		return;
394 	}
395 	/* K0 already points to save area, restore $1 and $2  */
396 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
397 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
398 }
399 
400 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
401 
402 /*
403  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
404  * we cannot do r3000 under these circumstances.
405  *
406  * Declare pgd_current here instead of including mmu_context.h to avoid type
407  * conflicts for tlbmiss_handler_setup_pgd
408  */
409 extern unsigned long pgd_current[];
410 
411 /*
412  * The R3000 TLB handler is simple.
413  */
414 static void build_r3000_tlb_refill_handler(void)
415 {
416 	long pgdc = (long)pgd_current;
417 	u32 *p;
418 
419 	memset(tlb_handler, 0, sizeof(tlb_handler));
420 	p = tlb_handler;
421 
422 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
423 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
424 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
425 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
426 	uasm_i_sll(&p, K0, K0, 2);
427 	uasm_i_addu(&p, K1, K1, K0);
428 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
429 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
430 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
431 	uasm_i_addu(&p, K1, K1, K0);
432 	uasm_i_lw(&p, K0, 0, K1);
433 	uasm_i_nop(&p); /* load delay */
434 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
435 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
436 	uasm_i_tlbwr(&p); /* cp0 delay */
437 	uasm_i_jr(&p, K1);
438 	uasm_i_rfe(&p); /* branch delay */
439 
440 	if (p > tlb_handler + 32)
441 		panic("TLB refill handler space exceeded");
442 
443 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
444 		 (unsigned int)(p - tlb_handler));
445 
446 	memcpy((void *)ebase, tlb_handler, 0x80);
447 	local_flush_icache_range(ebase, ebase + 0x80);
448 
449 	dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
450 }
451 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
452 
453 /*
454  * The R4000 TLB handler is much more complicated. We have two
455  * consecutive handler areas with 32 instructions space each.
456  * Since they aren't used at the same time, we can overflow in the
457  * other one.To keep things simple, we first assume linear space,
458  * then we relocate it to the final handler layout as needed.
459  */
460 static u32 final_handler[64];
461 
462 /*
463  * Hazards
464  *
465  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
466  * 2. A timing hazard exists for the TLBP instruction.
467  *
468  *	stalling_instruction
469  *	TLBP
470  *
471  * The JTLB is being read for the TLBP throughout the stall generated by the
472  * previous instruction. This is not really correct as the stalling instruction
473  * can modify the address used to access the JTLB.  The failure symptom is that
474  * the TLBP instruction will use an address created for the stalling instruction
475  * and not the address held in C0_ENHI and thus report the wrong results.
476  *
477  * The software work-around is to not allow the instruction preceding the TLBP
478  * to stall - make it an NOP or some other instruction guaranteed not to stall.
479  *
480  * Errata 2 will not be fixed.	This errata is also on the R5000.
481  *
482  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
483  */
484 static void __maybe_unused build_tlb_probe_entry(u32 **p)
485 {
486 	switch (current_cpu_type()) {
487 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
488 	case CPU_R4600:
489 	case CPU_R4700:
490 	case CPU_R5000:
491 	case CPU_NEVADA:
492 		uasm_i_nop(p);
493 		uasm_i_tlbp(p);
494 		break;
495 
496 	default:
497 		uasm_i_tlbp(p);
498 		break;
499 	}
500 }
501 
502 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
503 			   struct uasm_reloc **r,
504 			   enum tlb_write_entry wmode)
505 {
506 	void(*tlbw)(u32 **) = NULL;
507 
508 	switch (wmode) {
509 	case tlb_random: tlbw = uasm_i_tlbwr; break;
510 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
511 	}
512 
513 	if (cpu_has_mips_r2_r6) {
514 		if (cpu_has_mips_r2_exec_hazard)
515 			uasm_i_ehb(p);
516 		tlbw(p);
517 		return;
518 	}
519 
520 	switch (current_cpu_type()) {
521 	case CPU_R4000PC:
522 	case CPU_R4000SC:
523 	case CPU_R4000MC:
524 	case CPU_R4400PC:
525 	case CPU_R4400SC:
526 	case CPU_R4400MC:
527 		/*
528 		 * This branch uses up a mtc0 hazard nop slot and saves
529 		 * two nops after the tlbw instruction.
530 		 */
531 		uasm_bgezl_hazard(p, r, hazard_instance);
532 		tlbw(p);
533 		uasm_bgezl_label(l, p, hazard_instance);
534 		hazard_instance++;
535 		uasm_i_nop(p);
536 		break;
537 
538 	case CPU_R4600:
539 	case CPU_R4700:
540 		uasm_i_nop(p);
541 		tlbw(p);
542 		uasm_i_nop(p);
543 		break;
544 
545 	case CPU_R5000:
546 	case CPU_NEVADA:
547 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
548 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
549 		tlbw(p);
550 		break;
551 
552 	case CPU_R4300:
553 	case CPU_5KC:
554 	case CPU_TX49XX:
555 	case CPU_PR4450:
556 	case CPU_XLR:
557 		uasm_i_nop(p);
558 		tlbw(p);
559 		break;
560 
561 	case CPU_R10000:
562 	case CPU_R12000:
563 	case CPU_R14000:
564 	case CPU_R16000:
565 	case CPU_4KC:
566 	case CPU_4KEC:
567 	case CPU_M14KC:
568 	case CPU_M14KEC:
569 	case CPU_SB1:
570 	case CPU_SB1A:
571 	case CPU_4KSC:
572 	case CPU_20KC:
573 	case CPU_25KF:
574 	case CPU_BMIPS32:
575 	case CPU_BMIPS3300:
576 	case CPU_BMIPS4350:
577 	case CPU_BMIPS4380:
578 	case CPU_BMIPS5000:
579 	case CPU_LOONGSON2:
580 	case CPU_LOONGSON3:
581 	case CPU_R5500:
582 		if (m4kc_tlbp_war())
583 			uasm_i_nop(p);
584 	case CPU_ALCHEMY:
585 		tlbw(p);
586 		break;
587 
588 	case CPU_RM7000:
589 		uasm_i_nop(p);
590 		uasm_i_nop(p);
591 		uasm_i_nop(p);
592 		uasm_i_nop(p);
593 		tlbw(p);
594 		break;
595 
596 	case CPU_VR4111:
597 	case CPU_VR4121:
598 	case CPU_VR4122:
599 	case CPU_VR4181:
600 	case CPU_VR4181A:
601 		uasm_i_nop(p);
602 		uasm_i_nop(p);
603 		tlbw(p);
604 		uasm_i_nop(p);
605 		uasm_i_nop(p);
606 		break;
607 
608 	case CPU_VR4131:
609 	case CPU_VR4133:
610 	case CPU_R5432:
611 		uasm_i_nop(p);
612 		uasm_i_nop(p);
613 		tlbw(p);
614 		break;
615 
616 	case CPU_JZRISC:
617 		tlbw(p);
618 		uasm_i_nop(p);
619 		break;
620 
621 	default:
622 		panic("No TLB refill handler yet (CPU type: %d)",
623 		      current_cpu_type());
624 		break;
625 	}
626 }
627 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
628 
629 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
630 							unsigned int reg)
631 {
632 	if (_PAGE_GLOBAL_SHIFT == 0) {
633 		/* pte_t is already in EntryLo format */
634 		return;
635 	}
636 
637 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
638 		if (fill_includes_sw_bits) {
639 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
640 		} else {
641 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
642 			UASM_i_ROTR(p, reg, reg,
643 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
644 		}
645 	} else {
646 #ifdef CONFIG_PHYS_ADDR_T_64BIT
647 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
648 #else
649 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
650 #endif
651 	}
652 }
653 
654 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
655 
656 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
657 				   unsigned int tmp, enum label_id lid,
658 				   int restore_scratch)
659 {
660 	if (restore_scratch) {
661 		/* Reset default page size */
662 		if (PM_DEFAULT_MASK >> 16) {
663 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
664 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
665 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
666 			uasm_il_b(p, r, lid);
667 		} else if (PM_DEFAULT_MASK) {
668 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
669 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
670 			uasm_il_b(p, r, lid);
671 		} else {
672 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
673 			uasm_il_b(p, r, lid);
674 		}
675 		if (scratch_reg >= 0)
676 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
677 		else
678 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
679 	} else {
680 		/* Reset default page size */
681 		if (PM_DEFAULT_MASK >> 16) {
682 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
683 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
684 			uasm_il_b(p, r, lid);
685 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
686 		} else if (PM_DEFAULT_MASK) {
687 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
688 			uasm_il_b(p, r, lid);
689 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
690 		} else {
691 			uasm_il_b(p, r, lid);
692 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
693 		}
694 	}
695 }
696 
697 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
698 				       struct uasm_reloc **r,
699 				       unsigned int tmp,
700 				       enum tlb_write_entry wmode,
701 				       int restore_scratch)
702 {
703 	/* Set huge page tlb entry size */
704 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
705 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
706 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
707 
708 	build_tlb_write_entry(p, l, r, wmode);
709 
710 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
711 }
712 
713 /*
714  * Check if Huge PTE is present, if so then jump to LABEL.
715  */
716 static void
717 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
718 		  unsigned int pmd, int lid)
719 {
720 	UASM_i_LW(p, tmp, 0, pmd);
721 	if (use_bbit_insns()) {
722 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
723 	} else {
724 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
725 		uasm_il_bnez(p, r, tmp, lid);
726 	}
727 }
728 
729 static void build_huge_update_entries(u32 **p, unsigned int pte,
730 				      unsigned int tmp)
731 {
732 	int small_sequence;
733 
734 	/*
735 	 * A huge PTE describes an area the size of the
736 	 * configured huge page size. This is twice the
737 	 * of the large TLB entry size we intend to use.
738 	 * A TLB entry half the size of the configured
739 	 * huge page size is configured into entrylo0
740 	 * and entrylo1 to cover the contiguous huge PTE
741 	 * address space.
742 	 */
743 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
744 
745 	/* We can clobber tmp.	It isn't used after this.*/
746 	if (!small_sequence)
747 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
748 
749 	build_convert_pte_to_entrylo(p, pte);
750 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
751 	/* convert to entrylo1 */
752 	if (small_sequence)
753 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
754 	else
755 		UASM_i_ADDU(p, pte, pte, tmp);
756 
757 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
758 }
759 
760 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
761 				    struct uasm_label **l,
762 				    unsigned int pte,
763 				    unsigned int ptr,
764 				    unsigned int flush)
765 {
766 #ifdef CONFIG_SMP
767 	UASM_i_SC(p, pte, 0, ptr);
768 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
769 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
770 #else
771 	UASM_i_SW(p, pte, 0, ptr);
772 #endif
773 	if (cpu_has_ftlb && flush) {
774 		BUG_ON(!cpu_has_tlbinv);
775 
776 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
777 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
778 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
779 		build_tlb_write_entry(p, l, r, tlb_indexed);
780 
781 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
782 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
783 		build_huge_update_entries(p, pte, ptr);
784 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
785 
786 		return;
787 	}
788 
789 	build_huge_update_entries(p, pte, ptr);
790 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
791 }
792 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
793 
794 #ifdef CONFIG_64BIT
795 /*
796  * TMP and PTR are scratch.
797  * TMP will be clobbered, PTR will hold the pmd entry.
798  */
799 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
800 		      unsigned int tmp, unsigned int ptr)
801 {
802 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
803 	long pgdc = (long)pgd_current;
804 #endif
805 	/*
806 	 * The vmalloc handling is not in the hotpath.
807 	 */
808 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
809 
810 	if (check_for_high_segbits) {
811 		/*
812 		 * The kernel currently implicitely assumes that the
813 		 * MIPS SEGBITS parameter for the processor is
814 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
815 		 * allocate virtual addresses outside the maximum
816 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
817 		 * that doesn't prevent user code from accessing the
818 		 * higher xuseg addresses.  Here, we make sure that
819 		 * everything but the lower xuseg addresses goes down
820 		 * the module_alloc/vmalloc path.
821 		 */
822 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
823 		uasm_il_bnez(p, r, ptr, label_vmalloc);
824 	} else {
825 		uasm_il_bltz(p, r, tmp, label_vmalloc);
826 	}
827 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
828 
829 	if (pgd_reg != -1) {
830 		/* pgd is in pgd_reg */
831 		if (cpu_has_ldpte)
832 			UASM_i_MFC0(p, ptr, C0_PWBASE);
833 		else
834 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
835 	} else {
836 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
837 		/*
838 		 * &pgd << 11 stored in CONTEXT [23..63].
839 		 */
840 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
841 
842 		/* Clear lower 23 bits of context. */
843 		uasm_i_dins(p, ptr, 0, 0, 23);
844 
845 		/* 1 0	1 0 1  << 6  xkphys cached */
846 		uasm_i_ori(p, ptr, ptr, 0x540);
847 		uasm_i_drotr(p, ptr, ptr, 11);
848 #elif defined(CONFIG_SMP)
849 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
850 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
851 		UASM_i_LA_mostly(p, tmp, pgdc);
852 		uasm_i_daddu(p, ptr, ptr, tmp);
853 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
854 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
855 #else
856 		UASM_i_LA_mostly(p, ptr, pgdc);
857 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
858 #endif
859 	}
860 
861 	uasm_l_vmalloc_done(l, *p);
862 
863 	/* get pgd offset in bytes */
864 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
865 
866 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
867 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
868 #ifndef __PAGETABLE_PMD_FOLDED
869 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
870 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
871 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
872 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
873 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
874 #endif
875 }
876 EXPORT_SYMBOL_GPL(build_get_pmde64);
877 
878 /*
879  * BVADDR is the faulting address, PTR is scratch.
880  * PTR will hold the pgd for vmalloc.
881  */
882 static void
883 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
884 			unsigned int bvaddr, unsigned int ptr,
885 			enum vmalloc64_mode mode)
886 {
887 	long swpd = (long)swapper_pg_dir;
888 	int single_insn_swpd;
889 	int did_vmalloc_branch = 0;
890 
891 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
892 
893 	uasm_l_vmalloc(l, *p);
894 
895 	if (mode != not_refill && check_for_high_segbits) {
896 		if (single_insn_swpd) {
897 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
898 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
899 			did_vmalloc_branch = 1;
900 			/* fall through */
901 		} else {
902 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
903 		}
904 	}
905 	if (!did_vmalloc_branch) {
906 		if (single_insn_swpd) {
907 			uasm_il_b(p, r, label_vmalloc_done);
908 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
909 		} else {
910 			UASM_i_LA_mostly(p, ptr, swpd);
911 			uasm_il_b(p, r, label_vmalloc_done);
912 			if (uasm_in_compat_space_p(swpd))
913 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
914 			else
915 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
916 		}
917 	}
918 	if (mode != not_refill && check_for_high_segbits) {
919 		uasm_l_large_segbits_fault(l, *p);
920 		/*
921 		 * We get here if we are an xsseg address, or if we are
922 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
923 		 *
924 		 * Ignoring xsseg (assume disabled so would generate
925 		 * (address errors?), the only remaining possibility
926 		 * is the upper xuseg addresses.  On processors with
927 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
928 		 * addresses would have taken an address error. We try
929 		 * to mimic that here by taking a load/istream page
930 		 * fault.
931 		 */
932 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
933 		uasm_i_jr(p, ptr);
934 
935 		if (mode == refill_scratch) {
936 			if (scratch_reg >= 0)
937 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
938 			else
939 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
940 		} else {
941 			uasm_i_nop(p);
942 		}
943 	}
944 }
945 
946 #else /* !CONFIG_64BIT */
947 
948 /*
949  * TMP and PTR are scratch.
950  * TMP will be clobbered, PTR will hold the pgd entry.
951  */
952 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
953 {
954 	if (pgd_reg != -1) {
955 		/* pgd is in pgd_reg */
956 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
957 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
958 	} else {
959 		long pgdc = (long)pgd_current;
960 
961 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
962 #ifdef CONFIG_SMP
963 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
964 		UASM_i_LA_mostly(p, tmp, pgdc);
965 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
966 		uasm_i_addu(p, ptr, tmp, ptr);
967 #else
968 		UASM_i_LA_mostly(p, ptr, pgdc);
969 #endif
970 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
971 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
972 	}
973 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
974 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
975 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
976 }
977 EXPORT_SYMBOL_GPL(build_get_pgde32);
978 
979 #endif /* !CONFIG_64BIT */
980 
981 static void build_adjust_context(u32 **p, unsigned int ctx)
982 {
983 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
984 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
985 
986 	switch (current_cpu_type()) {
987 	case CPU_VR41XX:
988 	case CPU_VR4111:
989 	case CPU_VR4121:
990 	case CPU_VR4122:
991 	case CPU_VR4131:
992 	case CPU_VR4181:
993 	case CPU_VR4181A:
994 	case CPU_VR4133:
995 		shift += 2;
996 		break;
997 
998 	default:
999 		break;
1000 	}
1001 
1002 	if (shift)
1003 		UASM_i_SRL(p, ctx, ctx, shift);
1004 	uasm_i_andi(p, ctx, ctx, mask);
1005 }
1006 
1007 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1008 {
1009 	/*
1010 	 * Bug workaround for the Nevada. It seems as if under certain
1011 	 * circumstances the move from cp0_context might produce a
1012 	 * bogus result when the mfc0 instruction and its consumer are
1013 	 * in a different cacheline or a load instruction, probably any
1014 	 * memory reference, is between them.
1015 	 */
1016 	switch (current_cpu_type()) {
1017 	case CPU_NEVADA:
1018 		UASM_i_LW(p, ptr, 0, ptr);
1019 		GET_CONTEXT(p, tmp); /* get context reg */
1020 		break;
1021 
1022 	default:
1023 		GET_CONTEXT(p, tmp); /* get context reg */
1024 		UASM_i_LW(p, ptr, 0, ptr);
1025 		break;
1026 	}
1027 
1028 	build_adjust_context(p, tmp);
1029 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1030 }
1031 EXPORT_SYMBOL_GPL(build_get_ptep);
1032 
1033 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1034 {
1035 	int pte_off_even = 0;
1036 	int pte_off_odd = sizeof(pte_t);
1037 
1038 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1039 	/* The low 32 bits of EntryLo is stored in pte_high */
1040 	pte_off_even += offsetof(pte_t, pte_high);
1041 	pte_off_odd += offsetof(pte_t, pte_high);
1042 #endif
1043 
1044 	if (IS_ENABLED(CONFIG_XPA)) {
1045 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1046 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1047 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1048 
1049 		if (cpu_has_xpa && !mips_xpa_disabled) {
1050 			uasm_i_lw(p, tmp, 0, ptep);
1051 			uasm_i_ext(p, tmp, tmp, 0, 24);
1052 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1053 		}
1054 
1055 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1056 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1057 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1058 
1059 		if (cpu_has_xpa && !mips_xpa_disabled) {
1060 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1061 			uasm_i_ext(p, tmp, tmp, 0, 24);
1062 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1063 		}
1064 		return;
1065 	}
1066 
1067 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1068 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1069 	if (r45k_bvahwbug())
1070 		build_tlb_probe_entry(p);
1071 	build_convert_pte_to_entrylo(p, tmp);
1072 	if (r4k_250MHZhwbug())
1073 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1074 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1075 	build_convert_pte_to_entrylo(p, ptep);
1076 	if (r45k_bvahwbug())
1077 		uasm_i_mfc0(p, tmp, C0_INDEX);
1078 	if (r4k_250MHZhwbug())
1079 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1080 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1081 }
1082 EXPORT_SYMBOL_GPL(build_update_entries);
1083 
1084 struct mips_huge_tlb_info {
1085 	int huge_pte;
1086 	int restore_scratch;
1087 	bool need_reload_pte;
1088 };
1089 
1090 static struct mips_huge_tlb_info
1091 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1092 			       struct uasm_reloc **r, unsigned int tmp,
1093 			       unsigned int ptr, int c0_scratch_reg)
1094 {
1095 	struct mips_huge_tlb_info rv;
1096 	unsigned int even, odd;
1097 	int vmalloc_branch_delay_filled = 0;
1098 	const int scratch = 1; /* Our extra working register */
1099 
1100 	rv.huge_pte = scratch;
1101 	rv.restore_scratch = 0;
1102 	rv.need_reload_pte = false;
1103 
1104 	if (check_for_high_segbits) {
1105 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1106 
1107 		if (pgd_reg != -1)
1108 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1109 		else
1110 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1111 
1112 		if (c0_scratch_reg >= 0)
1113 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1114 		else
1115 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1116 
1117 		uasm_i_dsrl_safe(p, scratch, tmp,
1118 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1119 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1120 
1121 		if (pgd_reg == -1) {
1122 			vmalloc_branch_delay_filled = 1;
1123 			/* Clear lower 23 bits of context. */
1124 			uasm_i_dins(p, ptr, 0, 0, 23);
1125 		}
1126 	} else {
1127 		if (pgd_reg != -1)
1128 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1129 		else
1130 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1131 
1132 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1133 
1134 		if (c0_scratch_reg >= 0)
1135 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1136 		else
1137 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1138 
1139 		if (pgd_reg == -1)
1140 			/* Clear lower 23 bits of context. */
1141 			uasm_i_dins(p, ptr, 0, 0, 23);
1142 
1143 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1144 	}
1145 
1146 	if (pgd_reg == -1) {
1147 		vmalloc_branch_delay_filled = 1;
1148 		/* 1 0	1 0 1  << 6  xkphys cached */
1149 		uasm_i_ori(p, ptr, ptr, 0x540);
1150 		uasm_i_drotr(p, ptr, ptr, 11);
1151 	}
1152 
1153 #ifdef __PAGETABLE_PMD_FOLDED
1154 #define LOC_PTEP scratch
1155 #else
1156 #define LOC_PTEP ptr
1157 #endif
1158 
1159 	if (!vmalloc_branch_delay_filled)
1160 		/* get pgd offset in bytes */
1161 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1162 
1163 	uasm_l_vmalloc_done(l, *p);
1164 
1165 	/*
1166 	 *			   tmp		ptr
1167 	 * fall-through case =	 badvaddr  *pgd_current
1168 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1169 	 */
1170 
1171 	if (vmalloc_branch_delay_filled)
1172 		/* get pgd offset in bytes */
1173 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1174 
1175 #ifdef __PAGETABLE_PMD_FOLDED
1176 	GET_CONTEXT(p, tmp); /* get context reg */
1177 #endif
1178 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1179 
1180 	if (use_lwx_insns()) {
1181 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1182 	} else {
1183 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1184 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1185 	}
1186 
1187 #ifndef __PAGETABLE_PMD_FOLDED
1188 	/* get pmd offset in bytes */
1189 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1190 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1191 	GET_CONTEXT(p, tmp); /* get context reg */
1192 
1193 	if (use_lwx_insns()) {
1194 		UASM_i_LWX(p, scratch, scratch, ptr);
1195 	} else {
1196 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1197 		UASM_i_LW(p, scratch, 0, ptr);
1198 	}
1199 #endif
1200 	/* Adjust the context during the load latency. */
1201 	build_adjust_context(p, tmp);
1202 
1203 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1204 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1205 	/*
1206 	 * The in the LWX case we don't want to do the load in the
1207 	 * delay slot.	It cannot issue in the same cycle and may be
1208 	 * speculative and unneeded.
1209 	 */
1210 	if (use_lwx_insns())
1211 		uasm_i_nop(p);
1212 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1213 
1214 
1215 	/* build_update_entries */
1216 	if (use_lwx_insns()) {
1217 		even = ptr;
1218 		odd = tmp;
1219 		UASM_i_LWX(p, even, scratch, tmp);
1220 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1221 		UASM_i_LWX(p, odd, scratch, tmp);
1222 	} else {
1223 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1224 		even = tmp;
1225 		odd = ptr;
1226 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1227 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1228 	}
1229 	if (cpu_has_rixi) {
1230 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1231 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1232 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1233 	} else {
1234 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1235 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1236 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1237 	}
1238 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1239 
1240 	if (c0_scratch_reg >= 0) {
1241 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1242 		build_tlb_write_entry(p, l, r, tlb_random);
1243 		uasm_l_leave(l, *p);
1244 		rv.restore_scratch = 1;
1245 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1246 		build_tlb_write_entry(p, l, r, tlb_random);
1247 		uasm_l_leave(l, *p);
1248 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1249 	} else {
1250 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1251 		build_tlb_write_entry(p, l, r, tlb_random);
1252 		uasm_l_leave(l, *p);
1253 		rv.restore_scratch = 1;
1254 	}
1255 
1256 	uasm_i_eret(p); /* return from trap */
1257 
1258 	return rv;
1259 }
1260 
1261 /*
1262  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1263  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1264  * slots before the XTLB refill exception handler which belong to the
1265  * unused TLB refill exception.
1266  */
1267 #define MIPS64_REFILL_INSNS 32
1268 
1269 static void build_r4000_tlb_refill_handler(void)
1270 {
1271 	u32 *p = tlb_handler;
1272 	struct uasm_label *l = labels;
1273 	struct uasm_reloc *r = relocs;
1274 	u32 *f;
1275 	unsigned int final_len;
1276 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1277 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1278 
1279 	memset(tlb_handler, 0, sizeof(tlb_handler));
1280 	memset(labels, 0, sizeof(labels));
1281 	memset(relocs, 0, sizeof(relocs));
1282 	memset(final_handler, 0, sizeof(final_handler));
1283 
1284 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1285 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1286 							  scratch_reg);
1287 		vmalloc_mode = refill_scratch;
1288 	} else {
1289 		htlb_info.huge_pte = K0;
1290 		htlb_info.restore_scratch = 0;
1291 		htlb_info.need_reload_pte = true;
1292 		vmalloc_mode = refill_noscratch;
1293 		/*
1294 		 * create the plain linear handler
1295 		 */
1296 		if (bcm1250_m3_war()) {
1297 			unsigned int segbits = 44;
1298 
1299 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1300 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1301 			uasm_i_xor(&p, K0, K0, K1);
1302 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1303 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1304 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1305 			uasm_i_or(&p, K0, K0, K1);
1306 			uasm_il_bnez(&p, &r, K0, label_leave);
1307 			/* No need for uasm_i_nop */
1308 		}
1309 
1310 #ifdef CONFIG_64BIT
1311 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1312 #else
1313 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1314 #endif
1315 
1316 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1317 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1318 #endif
1319 
1320 		build_get_ptep(&p, K0, K1);
1321 		build_update_entries(&p, K0, K1);
1322 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1323 		uasm_l_leave(&l, p);
1324 		uasm_i_eret(&p); /* return from trap */
1325 	}
1326 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1327 	uasm_l_tlb_huge_update(&l, p);
1328 	if (htlb_info.need_reload_pte)
1329 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1330 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1331 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1332 				   htlb_info.restore_scratch);
1333 #endif
1334 
1335 #ifdef CONFIG_64BIT
1336 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1337 #endif
1338 
1339 	/*
1340 	 * Overflow check: For the 64bit handler, we need at least one
1341 	 * free instruction slot for the wrap-around branch. In worst
1342 	 * case, if the intended insertion point is a delay slot, we
1343 	 * need three, with the second nop'ed and the third being
1344 	 * unused.
1345 	 */
1346 	switch (boot_cpu_type()) {
1347 	default:
1348 		if (sizeof(long) == 4) {
1349 	case CPU_LOONGSON2:
1350 		/* Loongson2 ebase is different than r4k, we have more space */
1351 			if ((p - tlb_handler) > 64)
1352 				panic("TLB refill handler space exceeded");
1353 			/*
1354 			 * Now fold the handler in the TLB refill handler space.
1355 			 */
1356 			f = final_handler;
1357 			/* Simplest case, just copy the handler. */
1358 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1359 			final_len = p - tlb_handler;
1360 			break;
1361 		} else {
1362 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1363 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1364 				&& uasm_insn_has_bdelay(relocs,
1365 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1366 				panic("TLB refill handler space exceeded");
1367 			/*
1368 			 * Now fold the handler in the TLB refill handler space.
1369 			 */
1370 			f = final_handler + MIPS64_REFILL_INSNS;
1371 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1372 				/* Just copy the handler. */
1373 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1374 				final_len = p - tlb_handler;
1375 			} else {
1376 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1377 				const enum label_id ls = label_tlb_huge_update;
1378 #else
1379 				const enum label_id ls = label_vmalloc;
1380 #endif
1381 				u32 *split;
1382 				int ov = 0;
1383 				int i;
1384 
1385 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1386 					;
1387 				BUG_ON(i == ARRAY_SIZE(labels));
1388 				split = labels[i].addr;
1389 
1390 				/*
1391 				 * See if we have overflown one way or the other.
1392 				 */
1393 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1394 				    split < p - MIPS64_REFILL_INSNS)
1395 					ov = 1;
1396 
1397 				if (ov) {
1398 					/*
1399 					 * Split two instructions before the end.  One
1400 					 * for the branch and one for the instruction
1401 					 * in the delay slot.
1402 					 */
1403 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1404 
1405 					/*
1406 					 * If the branch would fall in a delay slot,
1407 					 * we must back up an additional instruction
1408 					 * so that it is no longer in a delay slot.
1409 					 */
1410 					if (uasm_insn_has_bdelay(relocs, split - 1))
1411 						split--;
1412 				}
1413 				/* Copy first part of the handler. */
1414 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1415 				f += split - tlb_handler;
1416 
1417 				if (ov) {
1418 					/* Insert branch. */
1419 					uasm_l_split(&l, final_handler);
1420 					uasm_il_b(&f, &r, label_split);
1421 					if (uasm_insn_has_bdelay(relocs, split))
1422 						uasm_i_nop(&f);
1423 					else {
1424 						uasm_copy_handler(relocs, labels,
1425 								  split, split + 1, f);
1426 						uasm_move_labels(labels, f, f + 1, -1);
1427 						f++;
1428 						split++;
1429 					}
1430 				}
1431 
1432 				/* Copy the rest of the handler. */
1433 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1434 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1435 					    (p - split);
1436 			}
1437 		}
1438 		break;
1439 	}
1440 
1441 	uasm_resolve_relocs(relocs, labels);
1442 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1443 		 final_len);
1444 
1445 	memcpy((void *)ebase, final_handler, 0x100);
1446 	local_flush_icache_range(ebase, ebase + 0x100);
1447 
1448 	dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1449 }
1450 
1451 static void setup_pw(void)
1452 {
1453 	unsigned long pgd_i, pgd_w;
1454 #ifndef __PAGETABLE_PMD_FOLDED
1455 	unsigned long pmd_i, pmd_w;
1456 #endif
1457 	unsigned long pt_i, pt_w;
1458 	unsigned long pte_i, pte_w;
1459 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1460 	unsigned long psn;
1461 
1462 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1463 #endif
1464 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1465 #ifndef __PAGETABLE_PMD_FOLDED
1466 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1467 
1468 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1469 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1470 #else
1471 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1472 #endif
1473 
1474 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1475 	pt_w  = PAGE_SHIFT - 3;
1476 
1477 	pte_i = ilog2(_PAGE_GLOBAL);
1478 	pte_w = 0;
1479 
1480 #ifndef __PAGETABLE_PMD_FOLDED
1481 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1482 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1483 #else
1484 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1485 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1486 #endif
1487 
1488 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1489 	write_c0_pwctl(1 << 6 | psn);
1490 #endif
1491 	write_c0_kpgd(swapper_pg_dir);
1492 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1493 }
1494 
1495 static void build_loongson3_tlb_refill_handler(void)
1496 {
1497 	u32 *p = tlb_handler;
1498 	struct uasm_label *l = labels;
1499 	struct uasm_reloc *r = relocs;
1500 
1501 	memset(labels, 0, sizeof(labels));
1502 	memset(relocs, 0, sizeof(relocs));
1503 	memset(tlb_handler, 0, sizeof(tlb_handler));
1504 
1505 	if (check_for_high_segbits) {
1506 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1507 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1508 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1509 		uasm_i_nop(&p);
1510 
1511 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1512 		uasm_i_nop(&p);
1513 		uasm_l_vmalloc(&l, p);
1514 	}
1515 
1516 	uasm_i_dmfc0(&p, K1, C0_PGD);
1517 
1518 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1519 #ifndef __PAGETABLE_PMD_FOLDED
1520 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1521 #endif
1522 	uasm_i_ldpte(&p, K1, 0);      /* even */
1523 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1524 	uasm_i_tlbwr(&p);
1525 
1526 	/* restore page mask */
1527 	if (PM_DEFAULT_MASK >> 16) {
1528 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1529 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1530 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1531 	} else if (PM_DEFAULT_MASK) {
1532 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1533 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1534 	} else {
1535 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1536 	}
1537 
1538 	uasm_i_eret(&p);
1539 
1540 	if (check_for_high_segbits) {
1541 		uasm_l_large_segbits_fault(&l, p);
1542 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1543 		uasm_i_jr(&p, K1);
1544 		uasm_i_nop(&p);
1545 	}
1546 
1547 	uasm_resolve_relocs(relocs, labels);
1548 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1549 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1550 	dump_handler("loongson3_tlb_refill", (u32 *)(ebase + 0x80), 32);
1551 }
1552 
1553 extern u32 handle_tlbl[], handle_tlbl_end[];
1554 extern u32 handle_tlbs[], handle_tlbs_end[];
1555 extern u32 handle_tlbm[], handle_tlbm_end[];
1556 extern u32 tlbmiss_handler_setup_pgd_start[];
1557 extern u32 tlbmiss_handler_setup_pgd[];
1558 EXPORT_SYMBOL_GPL(tlbmiss_handler_setup_pgd);
1559 extern u32 tlbmiss_handler_setup_pgd_end[];
1560 
1561 static void build_setup_pgd(void)
1562 {
1563 	const int a0 = 4;
1564 	const int __maybe_unused a1 = 5;
1565 	const int __maybe_unused a2 = 6;
1566 	u32 *p = tlbmiss_handler_setup_pgd_start;
1567 	const int tlbmiss_handler_setup_pgd_size =
1568 		tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1569 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1570 	long pgdc = (long)pgd_current;
1571 #endif
1572 
1573 	memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1574 					sizeof(tlbmiss_handler_setup_pgd[0]));
1575 	memset(labels, 0, sizeof(labels));
1576 	memset(relocs, 0, sizeof(relocs));
1577 	pgd_reg = allocate_kscratch();
1578 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1579 	if (pgd_reg == -1) {
1580 		struct uasm_label *l = labels;
1581 		struct uasm_reloc *r = relocs;
1582 
1583 		/* PGD << 11 in c0_Context */
1584 		/*
1585 		 * If it is a ckseg0 address, convert to a physical
1586 		 * address.  Shifting right by 29 and adding 4 will
1587 		 * result in zero for these addresses.
1588 		 *
1589 		 */
1590 		UASM_i_SRA(&p, a1, a0, 29);
1591 		UASM_i_ADDIU(&p, a1, a1, 4);
1592 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1593 		uasm_i_nop(&p);
1594 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1595 		uasm_l_tlbl_goaround1(&l, p);
1596 		UASM_i_SLL(&p, a0, a0, 11);
1597 		uasm_i_jr(&p, 31);
1598 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1599 	} else {
1600 		/* PGD in c0_KScratch */
1601 		uasm_i_jr(&p, 31);
1602 		if (cpu_has_ldpte)
1603 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1604 		else
1605 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1606 	}
1607 #else
1608 #ifdef CONFIG_SMP
1609 	/* Save PGD to pgd_current[smp_processor_id()] */
1610 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1611 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1612 	UASM_i_LA_mostly(&p, a2, pgdc);
1613 	UASM_i_ADDU(&p, a2, a2, a1);
1614 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1615 #else
1616 	UASM_i_LA_mostly(&p, a2, pgdc);
1617 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1618 #endif /* SMP */
1619 	uasm_i_jr(&p, 31);
1620 
1621 	/* if pgd_reg is allocated, save PGD also to scratch register */
1622 	if (pgd_reg != -1)
1623 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1624 	else
1625 		uasm_i_nop(&p);
1626 #endif
1627 	if (p >= tlbmiss_handler_setup_pgd_end)
1628 		panic("tlbmiss_handler_setup_pgd space exceeded");
1629 
1630 	uasm_resolve_relocs(relocs, labels);
1631 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1632 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1633 
1634 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1635 					tlbmiss_handler_setup_pgd_size);
1636 }
1637 
1638 static void
1639 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1640 {
1641 #ifdef CONFIG_SMP
1642 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1643 	if (cpu_has_64bits)
1644 		uasm_i_lld(p, pte, 0, ptr);
1645 	else
1646 # endif
1647 		UASM_i_LL(p, pte, 0, ptr);
1648 #else
1649 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1650 	if (cpu_has_64bits)
1651 		uasm_i_ld(p, pte, 0, ptr);
1652 	else
1653 # endif
1654 		UASM_i_LW(p, pte, 0, ptr);
1655 #endif
1656 }
1657 
1658 static void
1659 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1660 	unsigned int mode, unsigned int scratch)
1661 {
1662 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1663 	unsigned int swmode = mode & ~hwmode;
1664 
1665 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1666 		uasm_i_lui(p, scratch, swmode >> 16);
1667 		uasm_i_or(p, pte, pte, scratch);
1668 		BUG_ON(swmode & 0xffff);
1669 	} else {
1670 		uasm_i_ori(p, pte, pte, mode);
1671 	}
1672 
1673 #ifdef CONFIG_SMP
1674 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1675 	if (cpu_has_64bits)
1676 		uasm_i_scd(p, pte, 0, ptr);
1677 	else
1678 # endif
1679 		UASM_i_SC(p, pte, 0, ptr);
1680 
1681 	if (r10000_llsc_war())
1682 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1683 	else
1684 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1685 
1686 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1687 	if (!cpu_has_64bits) {
1688 		/* no uasm_i_nop needed */
1689 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1690 		uasm_i_ori(p, pte, pte, hwmode);
1691 		BUG_ON(hwmode & ~0xffff);
1692 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1693 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1694 		/* no uasm_i_nop needed */
1695 		uasm_i_lw(p, pte, 0, ptr);
1696 	} else
1697 		uasm_i_nop(p);
1698 # else
1699 	uasm_i_nop(p);
1700 # endif
1701 #else
1702 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1703 	if (cpu_has_64bits)
1704 		uasm_i_sd(p, pte, 0, ptr);
1705 	else
1706 # endif
1707 		UASM_i_SW(p, pte, 0, ptr);
1708 
1709 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1710 	if (!cpu_has_64bits) {
1711 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1712 		uasm_i_ori(p, pte, pte, hwmode);
1713 		BUG_ON(hwmode & ~0xffff);
1714 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1715 		uasm_i_lw(p, pte, 0, ptr);
1716 	}
1717 # endif
1718 #endif
1719 }
1720 
1721 /*
1722  * Check if PTE is present, if not then jump to LABEL. PTR points to
1723  * the page table where this PTE is located, PTE will be re-loaded
1724  * with it's original value.
1725  */
1726 static void
1727 build_pte_present(u32 **p, struct uasm_reloc **r,
1728 		  int pte, int ptr, int scratch, enum label_id lid)
1729 {
1730 	int t = scratch >= 0 ? scratch : pte;
1731 	int cur = pte;
1732 
1733 	if (cpu_has_rixi) {
1734 		if (use_bbit_insns()) {
1735 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1736 			uasm_i_nop(p);
1737 		} else {
1738 			if (_PAGE_PRESENT_SHIFT) {
1739 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1740 				cur = t;
1741 			}
1742 			uasm_i_andi(p, t, cur, 1);
1743 			uasm_il_beqz(p, r, t, lid);
1744 			if (pte == t)
1745 				/* You lose the SMP race :-(*/
1746 				iPTE_LW(p, pte, ptr);
1747 		}
1748 	} else {
1749 		if (_PAGE_PRESENT_SHIFT) {
1750 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1751 			cur = t;
1752 		}
1753 		uasm_i_andi(p, t, cur,
1754 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1755 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1756 		uasm_il_bnez(p, r, t, lid);
1757 		if (pte == t)
1758 			/* You lose the SMP race :-(*/
1759 			iPTE_LW(p, pte, ptr);
1760 	}
1761 }
1762 
1763 /* Make PTE valid, store result in PTR. */
1764 static void
1765 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1766 		 unsigned int ptr, unsigned int scratch)
1767 {
1768 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1769 
1770 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1771 }
1772 
1773 /*
1774  * Check if PTE can be written to, if not branch to LABEL. Regardless
1775  * restore PTE with value from PTR when done.
1776  */
1777 static void
1778 build_pte_writable(u32 **p, struct uasm_reloc **r,
1779 		   unsigned int pte, unsigned int ptr, int scratch,
1780 		   enum label_id lid)
1781 {
1782 	int t = scratch >= 0 ? scratch : pte;
1783 	int cur = pte;
1784 
1785 	if (_PAGE_PRESENT_SHIFT) {
1786 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1787 		cur = t;
1788 	}
1789 	uasm_i_andi(p, t, cur,
1790 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1791 	uasm_i_xori(p, t, t,
1792 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1793 	uasm_il_bnez(p, r, t, lid);
1794 	if (pte == t)
1795 		/* You lose the SMP race :-(*/
1796 		iPTE_LW(p, pte, ptr);
1797 	else
1798 		uasm_i_nop(p);
1799 }
1800 
1801 /* Make PTE writable, update software status bits as well, then store
1802  * at PTR.
1803  */
1804 static void
1805 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1806 		 unsigned int ptr, unsigned int scratch)
1807 {
1808 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1809 			     | _PAGE_DIRTY);
1810 
1811 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1812 }
1813 
1814 /*
1815  * Check if PTE can be modified, if not branch to LABEL. Regardless
1816  * restore PTE with value from PTR when done.
1817  */
1818 static void
1819 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1820 		     unsigned int pte, unsigned int ptr, int scratch,
1821 		     enum label_id lid)
1822 {
1823 	if (use_bbit_insns()) {
1824 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1825 		uasm_i_nop(p);
1826 	} else {
1827 		int t = scratch >= 0 ? scratch : pte;
1828 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1829 		uasm_i_andi(p, t, t, 1);
1830 		uasm_il_beqz(p, r, t, lid);
1831 		if (pte == t)
1832 			/* You lose the SMP race :-(*/
1833 			iPTE_LW(p, pte, ptr);
1834 	}
1835 }
1836 
1837 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1838 
1839 
1840 /*
1841  * R3000 style TLB load/store/modify handlers.
1842  */
1843 
1844 /*
1845  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1846  * Then it returns.
1847  */
1848 static void
1849 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1850 {
1851 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1852 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1853 	uasm_i_tlbwi(p);
1854 	uasm_i_jr(p, tmp);
1855 	uasm_i_rfe(p); /* branch delay */
1856 }
1857 
1858 /*
1859  * This places the pte into ENTRYLO0 and writes it with tlbwi
1860  * or tlbwr as appropriate.  This is because the index register
1861  * may have the probe fail bit set as a result of a trap on a
1862  * kseg2 access, i.e. without refill.  Then it returns.
1863  */
1864 static void
1865 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1866 			     struct uasm_reloc **r, unsigned int pte,
1867 			     unsigned int tmp)
1868 {
1869 	uasm_i_mfc0(p, tmp, C0_INDEX);
1870 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1871 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1872 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1873 	uasm_i_tlbwi(p); /* cp0 delay */
1874 	uasm_i_jr(p, tmp);
1875 	uasm_i_rfe(p); /* branch delay */
1876 	uasm_l_r3000_write_probe_fail(l, *p);
1877 	uasm_i_tlbwr(p); /* cp0 delay */
1878 	uasm_i_jr(p, tmp);
1879 	uasm_i_rfe(p); /* branch delay */
1880 }
1881 
1882 static void
1883 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1884 				   unsigned int ptr)
1885 {
1886 	long pgdc = (long)pgd_current;
1887 
1888 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1889 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1890 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1891 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1892 	uasm_i_sll(p, pte, pte, 2);
1893 	uasm_i_addu(p, ptr, ptr, pte);
1894 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1895 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1896 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1897 	uasm_i_addu(p, ptr, ptr, pte);
1898 	uasm_i_lw(p, pte, 0, ptr);
1899 	uasm_i_tlbp(p); /* load delay */
1900 }
1901 
1902 static void build_r3000_tlb_load_handler(void)
1903 {
1904 	u32 *p = handle_tlbl;
1905 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1906 	struct uasm_label *l = labels;
1907 	struct uasm_reloc *r = relocs;
1908 
1909 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1910 	memset(labels, 0, sizeof(labels));
1911 	memset(relocs, 0, sizeof(relocs));
1912 
1913 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1914 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1915 	uasm_i_nop(&p); /* load delay */
1916 	build_make_valid(&p, &r, K0, K1, -1);
1917 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1918 
1919 	uasm_l_nopage_tlbl(&l, p);
1920 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1921 	uasm_i_nop(&p);
1922 
1923 	if (p >= handle_tlbl_end)
1924 		panic("TLB load handler fastpath space exceeded");
1925 
1926 	uasm_resolve_relocs(relocs, labels);
1927 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1928 		 (unsigned int)(p - handle_tlbl));
1929 
1930 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1931 }
1932 
1933 static void build_r3000_tlb_store_handler(void)
1934 {
1935 	u32 *p = handle_tlbs;
1936 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1937 	struct uasm_label *l = labels;
1938 	struct uasm_reloc *r = relocs;
1939 
1940 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1941 	memset(labels, 0, sizeof(labels));
1942 	memset(relocs, 0, sizeof(relocs));
1943 
1944 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1945 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1946 	uasm_i_nop(&p); /* load delay */
1947 	build_make_write(&p, &r, K0, K1, -1);
1948 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1949 
1950 	uasm_l_nopage_tlbs(&l, p);
1951 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1952 	uasm_i_nop(&p);
1953 
1954 	if (p >= handle_tlbs_end)
1955 		panic("TLB store handler fastpath space exceeded");
1956 
1957 	uasm_resolve_relocs(relocs, labels);
1958 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1959 		 (unsigned int)(p - handle_tlbs));
1960 
1961 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1962 }
1963 
1964 static void build_r3000_tlb_modify_handler(void)
1965 {
1966 	u32 *p = handle_tlbm;
1967 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1968 	struct uasm_label *l = labels;
1969 	struct uasm_reloc *r = relocs;
1970 
1971 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1972 	memset(labels, 0, sizeof(labels));
1973 	memset(relocs, 0, sizeof(relocs));
1974 
1975 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1976 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1977 	uasm_i_nop(&p); /* load delay */
1978 	build_make_write(&p, &r, K0, K1, -1);
1979 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1980 
1981 	uasm_l_nopage_tlbm(&l, p);
1982 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1983 	uasm_i_nop(&p);
1984 
1985 	if (p >= handle_tlbm_end)
1986 		panic("TLB modify handler fastpath space exceeded");
1987 
1988 	uasm_resolve_relocs(relocs, labels);
1989 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1990 		 (unsigned int)(p - handle_tlbm));
1991 
1992 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1993 }
1994 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1995 
1996 /*
1997  * R4000 style TLB load/store/modify handlers.
1998  */
1999 static struct work_registers
2000 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2001 				   struct uasm_reloc **r)
2002 {
2003 	struct work_registers wr = build_get_work_registers(p);
2004 
2005 #ifdef CONFIG_64BIT
2006 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2007 #else
2008 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2009 #endif
2010 
2011 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2012 	/*
2013 	 * For huge tlb entries, pmd doesn't contain an address but
2014 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2015 	 * see if we need to jump to huge tlb processing.
2016 	 */
2017 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2018 #endif
2019 
2020 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2021 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2022 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2023 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2024 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2025 
2026 #ifdef CONFIG_SMP
2027 	uasm_l_smp_pgtable_change(l, *p);
2028 #endif
2029 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2030 	if (!m4kc_tlbp_war()) {
2031 		build_tlb_probe_entry(p);
2032 		if (cpu_has_htw) {
2033 			/* race condition happens, leaving */
2034 			uasm_i_ehb(p);
2035 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2036 			uasm_il_bltz(p, r, wr.r3, label_leave);
2037 			uasm_i_nop(p);
2038 		}
2039 	}
2040 	return wr;
2041 }
2042 
2043 static void
2044 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2045 				   struct uasm_reloc **r, unsigned int tmp,
2046 				   unsigned int ptr)
2047 {
2048 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2049 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2050 	build_update_entries(p, tmp, ptr);
2051 	build_tlb_write_entry(p, l, r, tlb_indexed);
2052 	uasm_l_leave(l, *p);
2053 	build_restore_work_registers(p);
2054 	uasm_i_eret(p); /* return from trap */
2055 
2056 #ifdef CONFIG_64BIT
2057 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2058 #endif
2059 }
2060 
2061 static void build_r4000_tlb_load_handler(void)
2062 {
2063 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2064 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
2065 	struct uasm_label *l = labels;
2066 	struct uasm_reloc *r = relocs;
2067 	struct work_registers wr;
2068 
2069 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
2070 	memset(labels, 0, sizeof(labels));
2071 	memset(relocs, 0, sizeof(relocs));
2072 
2073 	if (bcm1250_m3_war()) {
2074 		unsigned int segbits = 44;
2075 
2076 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2077 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2078 		uasm_i_xor(&p, K0, K0, K1);
2079 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2080 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2081 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2082 		uasm_i_or(&p, K0, K0, K1);
2083 		uasm_il_bnez(&p, &r, K0, label_leave);
2084 		/* No need for uasm_i_nop */
2085 	}
2086 
2087 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2088 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2089 	if (m4kc_tlbp_war())
2090 		build_tlb_probe_entry(&p);
2091 
2092 	if (cpu_has_rixi && !cpu_has_rixiex) {
2093 		/*
2094 		 * If the page is not _PAGE_VALID, RI or XI could not
2095 		 * have triggered it.  Skip the expensive test..
2096 		 */
2097 		if (use_bbit_insns()) {
2098 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2099 				      label_tlbl_goaround1);
2100 		} else {
2101 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2102 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2103 		}
2104 		uasm_i_nop(&p);
2105 
2106 		uasm_i_tlbr(&p);
2107 
2108 		switch (current_cpu_type()) {
2109 		default:
2110 			if (cpu_has_mips_r2_exec_hazard) {
2111 				uasm_i_ehb(&p);
2112 
2113 		case CPU_CAVIUM_OCTEON:
2114 		case CPU_CAVIUM_OCTEON_PLUS:
2115 		case CPU_CAVIUM_OCTEON2:
2116 				break;
2117 			}
2118 		}
2119 
2120 		/* Examine  entrylo 0 or 1 based on ptr. */
2121 		if (use_bbit_insns()) {
2122 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2123 		} else {
2124 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2125 			uasm_i_beqz(&p, wr.r3, 8);
2126 		}
2127 		/* load it in the delay slot*/
2128 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2129 		/* load it if ptr is odd */
2130 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2131 		/*
2132 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2133 		 * XI must have triggered it.
2134 		 */
2135 		if (use_bbit_insns()) {
2136 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2137 			uasm_i_nop(&p);
2138 			uasm_l_tlbl_goaround1(&l, p);
2139 		} else {
2140 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2141 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2142 			uasm_i_nop(&p);
2143 		}
2144 		uasm_l_tlbl_goaround1(&l, p);
2145 	}
2146 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2147 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2148 
2149 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2150 	/*
2151 	 * This is the entry point when build_r4000_tlbchange_handler_head
2152 	 * spots a huge page.
2153 	 */
2154 	uasm_l_tlb_huge_update(&l, p);
2155 	iPTE_LW(&p, wr.r1, wr.r2);
2156 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2157 	build_tlb_probe_entry(&p);
2158 
2159 	if (cpu_has_rixi && !cpu_has_rixiex) {
2160 		/*
2161 		 * If the page is not _PAGE_VALID, RI or XI could not
2162 		 * have triggered it.  Skip the expensive test..
2163 		 */
2164 		if (use_bbit_insns()) {
2165 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2166 				      label_tlbl_goaround2);
2167 		} else {
2168 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2169 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2170 		}
2171 		uasm_i_nop(&p);
2172 
2173 		uasm_i_tlbr(&p);
2174 
2175 		switch (current_cpu_type()) {
2176 		default:
2177 			if (cpu_has_mips_r2_exec_hazard) {
2178 				uasm_i_ehb(&p);
2179 
2180 		case CPU_CAVIUM_OCTEON:
2181 		case CPU_CAVIUM_OCTEON_PLUS:
2182 		case CPU_CAVIUM_OCTEON2:
2183 				break;
2184 			}
2185 		}
2186 
2187 		/* Examine  entrylo 0 or 1 based on ptr. */
2188 		if (use_bbit_insns()) {
2189 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2190 		} else {
2191 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2192 			uasm_i_beqz(&p, wr.r3, 8);
2193 		}
2194 		/* load it in the delay slot*/
2195 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2196 		/* load it if ptr is odd */
2197 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2198 		/*
2199 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2200 		 * XI must have triggered it.
2201 		 */
2202 		if (use_bbit_insns()) {
2203 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2204 		} else {
2205 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2206 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2207 		}
2208 		if (PM_DEFAULT_MASK == 0)
2209 			uasm_i_nop(&p);
2210 		/*
2211 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2212 		 * it is restored in build_huge_tlb_write_entry.
2213 		 */
2214 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2215 
2216 		uasm_l_tlbl_goaround2(&l, p);
2217 	}
2218 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2219 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2220 #endif
2221 
2222 	uasm_l_nopage_tlbl(&l, p);
2223 	build_restore_work_registers(&p);
2224 #ifdef CONFIG_CPU_MICROMIPS
2225 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2226 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2227 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2228 		uasm_i_jr(&p, K0);
2229 	} else
2230 #endif
2231 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2232 	uasm_i_nop(&p);
2233 
2234 	if (p >= handle_tlbl_end)
2235 		panic("TLB load handler fastpath space exceeded");
2236 
2237 	uasm_resolve_relocs(relocs, labels);
2238 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2239 		 (unsigned int)(p - handle_tlbl));
2240 
2241 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2242 }
2243 
2244 static void build_r4000_tlb_store_handler(void)
2245 {
2246 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2247 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2248 	struct uasm_label *l = labels;
2249 	struct uasm_reloc *r = relocs;
2250 	struct work_registers wr;
2251 
2252 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2253 	memset(labels, 0, sizeof(labels));
2254 	memset(relocs, 0, sizeof(relocs));
2255 
2256 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2257 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2258 	if (m4kc_tlbp_war())
2259 		build_tlb_probe_entry(&p);
2260 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2261 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2262 
2263 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2264 	/*
2265 	 * This is the entry point when
2266 	 * build_r4000_tlbchange_handler_head spots a huge page.
2267 	 */
2268 	uasm_l_tlb_huge_update(&l, p);
2269 	iPTE_LW(&p, wr.r1, wr.r2);
2270 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2271 	build_tlb_probe_entry(&p);
2272 	uasm_i_ori(&p, wr.r1, wr.r1,
2273 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2274 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2275 #endif
2276 
2277 	uasm_l_nopage_tlbs(&l, p);
2278 	build_restore_work_registers(&p);
2279 #ifdef CONFIG_CPU_MICROMIPS
2280 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2281 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2282 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2283 		uasm_i_jr(&p, K0);
2284 	} else
2285 #endif
2286 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2287 	uasm_i_nop(&p);
2288 
2289 	if (p >= handle_tlbs_end)
2290 		panic("TLB store handler fastpath space exceeded");
2291 
2292 	uasm_resolve_relocs(relocs, labels);
2293 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2294 		 (unsigned int)(p - handle_tlbs));
2295 
2296 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2297 }
2298 
2299 static void build_r4000_tlb_modify_handler(void)
2300 {
2301 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2302 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2303 	struct uasm_label *l = labels;
2304 	struct uasm_reloc *r = relocs;
2305 	struct work_registers wr;
2306 
2307 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
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_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2313 	if (m4kc_tlbp_war())
2314 		build_tlb_probe_entry(&p);
2315 	/* Present and writable bits set, set accessed and dirty bits. */
2316 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2317 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2318 
2319 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2320 	/*
2321 	 * This is the entry point when
2322 	 * build_r4000_tlbchange_handler_head spots a huge page.
2323 	 */
2324 	uasm_l_tlb_huge_update(&l, p);
2325 	iPTE_LW(&p, wr.r1, wr.r2);
2326 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2327 	build_tlb_probe_entry(&p);
2328 	uasm_i_ori(&p, wr.r1, wr.r1,
2329 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2330 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2331 #endif
2332 
2333 	uasm_l_nopage_tlbm(&l, p);
2334 	build_restore_work_registers(&p);
2335 #ifdef CONFIG_CPU_MICROMIPS
2336 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2337 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2338 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2339 		uasm_i_jr(&p, K0);
2340 	} else
2341 #endif
2342 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2343 	uasm_i_nop(&p);
2344 
2345 	if (p >= handle_tlbm_end)
2346 		panic("TLB modify handler fastpath space exceeded");
2347 
2348 	uasm_resolve_relocs(relocs, labels);
2349 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2350 		 (unsigned int)(p - handle_tlbm));
2351 
2352 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2353 }
2354 
2355 static void flush_tlb_handlers(void)
2356 {
2357 	local_flush_icache_range((unsigned long)handle_tlbl,
2358 			   (unsigned long)handle_tlbl_end);
2359 	local_flush_icache_range((unsigned long)handle_tlbs,
2360 			   (unsigned long)handle_tlbs_end);
2361 	local_flush_icache_range((unsigned long)handle_tlbm,
2362 			   (unsigned long)handle_tlbm_end);
2363 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2364 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2365 }
2366 
2367 static void print_htw_config(void)
2368 {
2369 	unsigned long config;
2370 	unsigned int pwctl;
2371 	const int field = 2 * sizeof(unsigned long);
2372 
2373 	config = read_c0_pwfield();
2374 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2375 		field, config,
2376 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2377 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2378 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2379 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2380 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2381 
2382 	config = read_c0_pwsize();
2383 	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",
2384 		field, config,
2385 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2386 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2387 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2388 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2389 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2390 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2391 
2392 	pwctl = read_c0_pwctl();
2393 	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",
2394 		pwctl,
2395 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2396 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2397 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2398 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2399 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2400 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2401 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2402 }
2403 
2404 static void config_htw_params(void)
2405 {
2406 	unsigned long pwfield, pwsize, ptei;
2407 	unsigned int config;
2408 
2409 	/*
2410 	 * We are using 2-level page tables, so we only need to
2411 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2412 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2413 	 * write values less than 0xc in these fields because the entire
2414 	 * write will be dropped. As a result of which, we must preserve
2415 	 * the original reset values and overwrite only what we really want.
2416 	 */
2417 
2418 	pwfield = read_c0_pwfield();
2419 	/* re-initialize the GDI field */
2420 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2421 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2422 	/* re-initialize the PTI field including the even/odd bit */
2423 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2424 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2425 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2426 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2427 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2428 	}
2429 	/* Set the PTEI right shift */
2430 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2431 	pwfield |= ptei;
2432 	write_c0_pwfield(pwfield);
2433 	/* Check whether the PTEI value is supported */
2434 	back_to_back_c0_hazard();
2435 	pwfield = read_c0_pwfield();
2436 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2437 		!= ptei) {
2438 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2439 			ptei);
2440 		/*
2441 		 * Drop option to avoid HTW being enabled via another path
2442 		 * (eg htw_reset())
2443 		 */
2444 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2445 		return;
2446 	}
2447 
2448 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2449 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2450 	if (CONFIG_PGTABLE_LEVELS >= 3)
2451 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2452 
2453 	/* Set pointer size to size of directory pointers */
2454 	if (IS_ENABLED(CONFIG_64BIT))
2455 		pwsize |= MIPS_PWSIZE_PS_MASK;
2456 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2457 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2458 			& MIPS_PWSIZE_PTEW_MASK;
2459 
2460 	write_c0_pwsize(pwsize);
2461 
2462 	/* Make sure everything is set before we enable the HTW */
2463 	back_to_back_c0_hazard();
2464 
2465 	/*
2466 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2467 	 * the pwctl fields.
2468 	 */
2469 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2470 	if (IS_ENABLED(CONFIG_64BIT))
2471 		config |= MIPS_PWCTL_XU_MASK;
2472 	write_c0_pwctl(config);
2473 	pr_info("Hardware Page Table Walker enabled\n");
2474 
2475 	print_htw_config();
2476 }
2477 
2478 static void config_xpa_params(void)
2479 {
2480 #ifdef CONFIG_XPA
2481 	unsigned int pagegrain;
2482 
2483 	if (mips_xpa_disabled) {
2484 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2485 		return;
2486 	}
2487 
2488 	pagegrain = read_c0_pagegrain();
2489 	write_c0_pagegrain(pagegrain | PG_ELPA);
2490 	back_to_back_c0_hazard();
2491 	pagegrain = read_c0_pagegrain();
2492 
2493 	if (pagegrain & PG_ELPA)
2494 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2495 	else
2496 		panic("Extended Physical Addressing (XPA) disabled");
2497 #endif
2498 }
2499 
2500 static void check_pabits(void)
2501 {
2502 	unsigned long entry;
2503 	unsigned pabits, fillbits;
2504 
2505 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2506 		/*
2507 		 * We'll only be making use of the fact that we can rotate bits
2508 		 * into the fill if the CPU supports RIXI, so don't bother
2509 		 * probing this for CPUs which don't.
2510 		 */
2511 		return;
2512 	}
2513 
2514 	write_c0_entrylo0(~0ul);
2515 	back_to_back_c0_hazard();
2516 	entry = read_c0_entrylo0();
2517 
2518 	/* clear all non-PFN bits */
2519 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2520 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2521 
2522 	/* find a lower bound on PABITS, and upper bound on fill bits */
2523 	pabits = fls_long(entry) + 6;
2524 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2525 
2526 	/* minus the RI & XI bits */
2527 	fillbits -= min_t(unsigned, fillbits, 2);
2528 
2529 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2530 		fill_includes_sw_bits = true;
2531 
2532 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2533 }
2534 
2535 void build_tlb_refill_handler(void)
2536 {
2537 	/*
2538 	 * The refill handler is generated per-CPU, multi-node systems
2539 	 * may have local storage for it. The other handlers are only
2540 	 * needed once.
2541 	 */
2542 	static int run_once = 0;
2543 
2544 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2545 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2546 
2547 	output_pgtable_bits_defines();
2548 	check_pabits();
2549 
2550 #ifdef CONFIG_64BIT
2551 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2552 #endif
2553 
2554 	switch (current_cpu_type()) {
2555 	case CPU_R2000:
2556 	case CPU_R3000:
2557 	case CPU_R3000A:
2558 	case CPU_R3081E:
2559 	case CPU_TX3912:
2560 	case CPU_TX3922:
2561 	case CPU_TX3927:
2562 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2563 		if (cpu_has_local_ebase)
2564 			build_r3000_tlb_refill_handler();
2565 		if (!run_once) {
2566 			if (!cpu_has_local_ebase)
2567 				build_r3000_tlb_refill_handler();
2568 			build_setup_pgd();
2569 			build_r3000_tlb_load_handler();
2570 			build_r3000_tlb_store_handler();
2571 			build_r3000_tlb_modify_handler();
2572 			flush_tlb_handlers();
2573 			run_once++;
2574 		}
2575 #else
2576 		panic("No R3000 TLB refill handler");
2577 #endif
2578 		break;
2579 
2580 	case CPU_R6000:
2581 	case CPU_R6000A:
2582 		panic("No R6000 TLB refill handler yet");
2583 		break;
2584 
2585 	case CPU_R8000:
2586 		panic("No R8000 TLB refill handler yet");
2587 		break;
2588 
2589 	default:
2590 		if (cpu_has_ldpte)
2591 			setup_pw();
2592 
2593 		if (!run_once) {
2594 			scratch_reg = allocate_kscratch();
2595 			build_setup_pgd();
2596 			build_r4000_tlb_load_handler();
2597 			build_r4000_tlb_store_handler();
2598 			build_r4000_tlb_modify_handler();
2599 			if (cpu_has_ldpte)
2600 				build_loongson3_tlb_refill_handler();
2601 			else if (!cpu_has_local_ebase)
2602 				build_r4000_tlb_refill_handler();
2603 			flush_tlb_handlers();
2604 			run_once++;
2605 		}
2606 		if (cpu_has_local_ebase)
2607 			build_r4000_tlb_refill_handler();
2608 		if (cpu_has_xpa)
2609 			config_xpa_params();
2610 		if (cpu_has_htw)
2611 			config_htw_params();
2612 	}
2613 }
2614