xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision d9fd5a71)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008	 Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009	 Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 #include <linux/pgtable.h>
32 
33 #include <asm/cacheflush.h>
34 #include <asm/cpu-type.h>
35 #include <asm/mmu_context.h>
36 #include <asm/war.h>
37 #include <asm/uasm.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
40 
41 static int mips_xpa_disabled;
42 
43 static int __init xpa_disable(char *s)
44 {
45 	mips_xpa_disabled = 1;
46 
47 	return 1;
48 }
49 
50 __setup("noxpa", xpa_disable);
51 
52 /*
53  * TLB load/store/modify handlers.
54  *
55  * Only the fastpath gets synthesized at runtime, the slowpath for
56  * do_page_fault remains normal asm.
57  */
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
60 
61 struct work_registers {
62 	int r1;
63 	int r2;
64 	int r3;
65 };
66 
67 struct tlb_reg_save {
68 	unsigned long a;
69 	unsigned long b;
70 } ____cacheline_aligned_in_smp;
71 
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
73 
74 static inline int r45k_bvahwbug(void)
75 {
76 	/* XXX: We should probe for the presence of this bug, but we don't. */
77 	return 0;
78 }
79 
80 static inline int r4k_250MHZhwbug(void)
81 {
82 	/* XXX: We should probe for the presence of this bug, but we don't. */
83 	return 0;
84 }
85 
86 extern int sb1250_m3_workaround_needed(void);
87 
88 static inline int __maybe_unused bcm1250_m3_war(void)
89 {
90 	if (IS_ENABLED(CONFIG_SB1_PASS_2_WORKAROUNDS))
91 		return sb1250_m3_workaround_needed();
92 	return 0;
93 }
94 
95 static inline int __maybe_unused r10000_llsc_war(void)
96 {
97 	return IS_ENABLED(CONFIG_WAR_R10000_LLSC);
98 }
99 
100 static int use_bbit_insns(void)
101 {
102 	switch (current_cpu_type()) {
103 	case CPU_CAVIUM_OCTEON:
104 	case CPU_CAVIUM_OCTEON_PLUS:
105 	case CPU_CAVIUM_OCTEON2:
106 	case CPU_CAVIUM_OCTEON3:
107 		return 1;
108 	default:
109 		return 0;
110 	}
111 }
112 
113 static int use_lwx_insns(void)
114 {
115 	switch (current_cpu_type()) {
116 	case CPU_CAVIUM_OCTEON2:
117 	case CPU_CAVIUM_OCTEON3:
118 		return 1;
119 	default:
120 		return 0;
121 	}
122 }
123 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
124     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
125 static bool scratchpad_available(void)
126 {
127 	return true;
128 }
129 static int scratchpad_offset(int i)
130 {
131 	/*
132 	 * CVMSEG starts at address -32768 and extends for
133 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
134 	 */
135 	i += 1; /* Kernel use starts at the top and works down. */
136 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
137 }
138 #else
139 static bool scratchpad_available(void)
140 {
141 	return false;
142 }
143 static int scratchpad_offset(int i)
144 {
145 	BUG();
146 	/* Really unreachable, but evidently some GCC want this. */
147 	return 0;
148 }
149 #endif
150 /*
151  * Found by experiment: At least some revisions of the 4kc throw under
152  * some circumstances a machine check exception, triggered by invalid
153  * values in the index register.  Delaying the tlbp instruction until
154  * after the next branch,  plus adding an additional nop in front of
155  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
156  * why; it's not an issue caused by the core RTL.
157  *
158  */
159 static int m4kc_tlbp_war(void)
160 {
161 	return current_cpu_type() == CPU_4KC;
162 }
163 
164 /* Handle labels (which must be positive integers). */
165 enum label_id {
166 	label_second_part = 1,
167 	label_leave,
168 	label_vmalloc,
169 	label_vmalloc_done,
170 	label_tlbw_hazard_0,
171 	label_split = label_tlbw_hazard_0 + 8,
172 	label_tlbl_goaround1,
173 	label_tlbl_goaround2,
174 	label_nopage_tlbl,
175 	label_nopage_tlbs,
176 	label_nopage_tlbm,
177 	label_smp_pgtable_change,
178 	label_r3000_write_probe_fail,
179 	label_large_segbits_fault,
180 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
181 	label_tlb_huge_update,
182 #endif
183 };
184 
185 UASM_L_LA(_second_part)
186 UASM_L_LA(_leave)
187 UASM_L_LA(_vmalloc)
188 UASM_L_LA(_vmalloc_done)
189 /* _tlbw_hazard_x is handled differently.  */
190 UASM_L_LA(_split)
191 UASM_L_LA(_tlbl_goaround1)
192 UASM_L_LA(_tlbl_goaround2)
193 UASM_L_LA(_nopage_tlbl)
194 UASM_L_LA(_nopage_tlbs)
195 UASM_L_LA(_nopage_tlbm)
196 UASM_L_LA(_smp_pgtable_change)
197 UASM_L_LA(_r3000_write_probe_fail)
198 UASM_L_LA(_large_segbits_fault)
199 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
200 UASM_L_LA(_tlb_huge_update)
201 #endif
202 
203 static int hazard_instance;
204 
205 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
206 {
207 	switch (instance) {
208 	case 0 ... 7:
209 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
210 		return;
211 	default:
212 		BUG();
213 	}
214 }
215 
216 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
217 {
218 	switch (instance) {
219 	case 0 ... 7:
220 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
221 		break;
222 	default:
223 		BUG();
224 	}
225 }
226 
227 /*
228  * pgtable bits are assigned dynamically depending on processor feature
229  * and statically based on kernel configuration.  This spits out the actual
230  * values the kernel is using.	Required to make sense from disassembled
231  * TLB exception handlers.
232  */
233 static void output_pgtable_bits_defines(void)
234 {
235 #define pr_define(fmt, ...)					\
236 	pr_debug("#define " fmt, ##__VA_ARGS__)
237 
238 	pr_debug("#include <asm/asm.h>\n");
239 	pr_debug("#include <asm/regdef.h>\n");
240 	pr_debug("\n");
241 
242 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
243 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
244 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
245 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
246 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
247 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
248 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
249 #endif
250 #ifdef _PAGE_NO_EXEC_SHIFT
251 	if (cpu_has_rixi)
252 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
253 #endif
254 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
255 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
256 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
257 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
258 	pr_debug("\n");
259 }
260 
261 static inline void dump_handler(const char *symbol, const void *start, const void *end)
262 {
263 	unsigned int count = (end - start) / sizeof(u32);
264 	const u32 *handler = start;
265 	int i;
266 
267 	pr_debug("LEAF(%s)\n", symbol);
268 
269 	pr_debug("\t.set push\n");
270 	pr_debug("\t.set noreorder\n");
271 
272 	for (i = 0; i < count; i++)
273 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
274 
275 	pr_debug("\t.set\tpop\n");
276 
277 	pr_debug("\tEND(%s)\n", symbol);
278 }
279 
280 /* The only general purpose registers allowed in TLB handlers. */
281 #define K0		26
282 #define K1		27
283 
284 /* Some CP0 registers */
285 #define C0_INDEX	0, 0
286 #define C0_ENTRYLO0	2, 0
287 #define C0_TCBIND	2, 2
288 #define C0_ENTRYLO1	3, 0
289 #define C0_CONTEXT	4, 0
290 #define C0_PAGEMASK	5, 0
291 #define C0_PWBASE	5, 5
292 #define C0_PWFIELD	5, 6
293 #define C0_PWSIZE	5, 7
294 #define C0_PWCTL	6, 6
295 #define C0_BADVADDR	8, 0
296 #define C0_PGD		9, 7
297 #define C0_ENTRYHI	10, 0
298 #define C0_EPC		14, 0
299 #define C0_XCONTEXT	20, 0
300 
301 #ifdef CONFIG_64BIT
302 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
303 #else
304 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
305 #endif
306 
307 /* The worst case length of the handler is around 18 instructions for
308  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
309  * Maximum space available is 32 instructions for R3000 and 64
310  * instructions for R4000.
311  *
312  * We deliberately chose a buffer size of 128, so we won't scribble
313  * over anything important on overflow before we panic.
314  */
315 static u32 tlb_handler[128];
316 
317 /* simply assume worst case size for labels and relocs */
318 static struct uasm_label labels[128];
319 static struct uasm_reloc relocs[128];
320 
321 static int check_for_high_segbits;
322 static bool fill_includes_sw_bits;
323 
324 static unsigned int kscratch_used_mask;
325 
326 static inline int __maybe_unused c0_kscratch(void)
327 {
328 	switch (current_cpu_type()) {
329 	case CPU_XLP:
330 	case CPU_XLR:
331 		return 22;
332 	default:
333 		return 31;
334 	}
335 }
336 
337 static int allocate_kscratch(void)
338 {
339 	int r;
340 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
341 
342 	r = ffs(a);
343 
344 	if (r == 0)
345 		return -1;
346 
347 	r--; /* make it zero based */
348 
349 	kscratch_used_mask |= (1 << r);
350 
351 	return r;
352 }
353 
354 static int scratch_reg;
355 int pgd_reg;
356 EXPORT_SYMBOL_GPL(pgd_reg);
357 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
358 
359 static struct work_registers build_get_work_registers(u32 **p)
360 {
361 	struct work_registers r;
362 
363 	if (scratch_reg >= 0) {
364 		/* Save in CPU local C0_KScratch? */
365 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
366 		r.r1 = K0;
367 		r.r2 = K1;
368 		r.r3 = 1;
369 		return r;
370 	}
371 
372 	if (num_possible_cpus() > 1) {
373 		/* Get smp_processor_id */
374 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
375 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
376 
377 		/* handler_reg_save index in K0 */
378 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
379 
380 		UASM_i_LA(p, K1, (long)&handler_reg_save);
381 		UASM_i_ADDU(p, K0, K0, K1);
382 	} else {
383 		UASM_i_LA(p, K0, (long)&handler_reg_save);
384 	}
385 	/* K0 now points to save area, save $1 and $2  */
386 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
387 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
388 
389 	r.r1 = K1;
390 	r.r2 = 1;
391 	r.r3 = 2;
392 	return r;
393 }
394 
395 static void build_restore_work_registers(u32 **p)
396 {
397 	if (scratch_reg >= 0) {
398 		uasm_i_ehb(p);
399 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
400 		return;
401 	}
402 	/* K0 already points to save area, restore $1 and $2  */
403 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
404 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
405 }
406 
407 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
408 
409 /*
410  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
411  * we cannot do r3000 under these circumstances.
412  *
413  * The R3000 TLB handler is simple.
414  */
415 static void build_r3000_tlb_refill_handler(void)
416 {
417 	long pgdc = (long)pgd_current;
418 	u32 *p;
419 
420 	memset(tlb_handler, 0, sizeof(tlb_handler));
421 	p = tlb_handler;
422 
423 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
424 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
425 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
426 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
427 	uasm_i_sll(&p, K0, K0, 2);
428 	uasm_i_addu(&p, K1, K1, K0);
429 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
430 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
431 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
432 	uasm_i_addu(&p, K1, K1, K0);
433 	uasm_i_lw(&p, K0, 0, K1);
434 	uasm_i_nop(&p); /* load delay */
435 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
436 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
437 	uasm_i_tlbwr(&p); /* cp0 delay */
438 	uasm_i_jr(&p, K1);
439 	uasm_i_rfe(&p); /* branch delay */
440 
441 	if (p > tlb_handler + 32)
442 		panic("TLB refill handler space exceeded");
443 
444 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
445 		 (unsigned int)(p - tlb_handler));
446 
447 	memcpy((void *)ebase, tlb_handler, 0x80);
448 	local_flush_icache_range(ebase, ebase + 0x80);
449 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
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_LOONGSON2EF:
580 	case CPU_LOONGSON64:
581 	case CPU_R5500:
582 		if (m4kc_tlbp_war())
583 			uasm_i_nop(p);
584 		fallthrough;
585 	case CPU_ALCHEMY:
586 		tlbw(p);
587 		break;
588 
589 	case CPU_RM7000:
590 		uasm_i_nop(p);
591 		uasm_i_nop(p);
592 		uasm_i_nop(p);
593 		uasm_i_nop(p);
594 		tlbw(p);
595 		break;
596 
597 	case CPU_VR4111:
598 	case CPU_VR4121:
599 	case CPU_VR4122:
600 	case CPU_VR4181:
601 	case CPU_VR4181A:
602 		uasm_i_nop(p);
603 		uasm_i_nop(p);
604 		tlbw(p);
605 		uasm_i_nop(p);
606 		uasm_i_nop(p);
607 		break;
608 
609 	case CPU_VR4131:
610 	case CPU_VR4133:
611 		uasm_i_nop(p);
612 		uasm_i_nop(p);
613 		tlbw(p);
614 		break;
615 
616 	case CPU_XBURST:
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 		/*
662 		 * Ensure the MFC0 below observes the value written to the
663 		 * KScratch register by the prior MTC0.
664 		 */
665 		if (scratch_reg >= 0)
666 			uasm_i_ehb(p);
667 
668 		/* Reset default page size */
669 		if (PM_DEFAULT_MASK >> 16) {
670 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
671 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
672 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
673 			uasm_il_b(p, r, lid);
674 		} else if (PM_DEFAULT_MASK) {
675 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
676 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
677 			uasm_il_b(p, r, lid);
678 		} else {
679 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
680 			uasm_il_b(p, r, lid);
681 		}
682 		if (scratch_reg >= 0)
683 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
684 		else
685 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
686 	} else {
687 		/* Reset default page size */
688 		if (PM_DEFAULT_MASK >> 16) {
689 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
690 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
691 			uasm_il_b(p, r, lid);
692 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
693 		} else if (PM_DEFAULT_MASK) {
694 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
695 			uasm_il_b(p, r, lid);
696 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
697 		} else {
698 			uasm_il_b(p, r, lid);
699 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
700 		}
701 	}
702 }
703 
704 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
705 				       struct uasm_reloc **r,
706 				       unsigned int tmp,
707 				       enum tlb_write_entry wmode,
708 				       int restore_scratch)
709 {
710 	/* Set huge page tlb entry size */
711 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
712 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
713 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
714 
715 	build_tlb_write_entry(p, l, r, wmode);
716 
717 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
718 }
719 
720 /*
721  * Check if Huge PTE is present, if so then jump to LABEL.
722  */
723 static void
724 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
725 		  unsigned int pmd, int lid)
726 {
727 	UASM_i_LW(p, tmp, 0, pmd);
728 	if (use_bbit_insns()) {
729 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
730 	} else {
731 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
732 		uasm_il_bnez(p, r, tmp, lid);
733 	}
734 }
735 
736 static void build_huge_update_entries(u32 **p, unsigned int pte,
737 				      unsigned int tmp)
738 {
739 	int small_sequence;
740 
741 	/*
742 	 * A huge PTE describes an area the size of the
743 	 * configured huge page size. This is twice the
744 	 * of the large TLB entry size we intend to use.
745 	 * A TLB entry half the size of the configured
746 	 * huge page size is configured into entrylo0
747 	 * and entrylo1 to cover the contiguous huge PTE
748 	 * address space.
749 	 */
750 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
751 
752 	/* We can clobber tmp.	It isn't used after this.*/
753 	if (!small_sequence)
754 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
755 
756 	build_convert_pte_to_entrylo(p, pte);
757 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
758 	/* convert to entrylo1 */
759 	if (small_sequence)
760 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
761 	else
762 		UASM_i_ADDU(p, pte, pte, tmp);
763 
764 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
765 }
766 
767 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
768 				    struct uasm_label **l,
769 				    unsigned int pte,
770 				    unsigned int ptr,
771 				    unsigned int flush)
772 {
773 #ifdef CONFIG_SMP
774 	UASM_i_SC(p, pte, 0, ptr);
775 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
776 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
777 #else
778 	UASM_i_SW(p, pte, 0, ptr);
779 #endif
780 	if (cpu_has_ftlb && flush) {
781 		BUG_ON(!cpu_has_tlbinv);
782 
783 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
784 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
785 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
786 		build_tlb_write_entry(p, l, r, tlb_indexed);
787 
788 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
789 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
790 		build_huge_update_entries(p, pte, ptr);
791 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
792 
793 		return;
794 	}
795 
796 	build_huge_update_entries(p, pte, ptr);
797 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
798 }
799 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
800 
801 #ifdef CONFIG_64BIT
802 /*
803  * TMP and PTR are scratch.
804  * TMP will be clobbered, PTR will hold the pmd entry.
805  */
806 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
807 		      unsigned int tmp, unsigned int ptr)
808 {
809 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
810 	long pgdc = (long)pgd_current;
811 #endif
812 	/*
813 	 * The vmalloc handling is not in the hotpath.
814 	 */
815 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
816 
817 	if (check_for_high_segbits) {
818 		/*
819 		 * The kernel currently implicitely assumes that the
820 		 * MIPS SEGBITS parameter for the processor is
821 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
822 		 * allocate virtual addresses outside the maximum
823 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
824 		 * that doesn't prevent user code from accessing the
825 		 * higher xuseg addresses.  Here, we make sure that
826 		 * everything but the lower xuseg addresses goes down
827 		 * the module_alloc/vmalloc path.
828 		 */
829 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
830 		uasm_il_bnez(p, r, ptr, label_vmalloc);
831 	} else {
832 		uasm_il_bltz(p, r, tmp, label_vmalloc);
833 	}
834 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
835 
836 	if (pgd_reg != -1) {
837 		/* pgd is in pgd_reg */
838 		if (cpu_has_ldpte)
839 			UASM_i_MFC0(p, ptr, C0_PWBASE);
840 		else
841 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
842 	} else {
843 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
844 		/*
845 		 * &pgd << 11 stored in CONTEXT [23..63].
846 		 */
847 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
848 
849 		/* Clear lower 23 bits of context. */
850 		uasm_i_dins(p, ptr, 0, 0, 23);
851 
852 		/* 1 0	1 0 1  << 6  xkphys cached */
853 		uasm_i_ori(p, ptr, ptr, 0x540);
854 		uasm_i_drotr(p, ptr, ptr, 11);
855 #elif defined(CONFIG_SMP)
856 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
857 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
858 		UASM_i_LA_mostly(p, tmp, pgdc);
859 		uasm_i_daddu(p, ptr, ptr, tmp);
860 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
861 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
862 #else
863 		UASM_i_LA_mostly(p, ptr, pgdc);
864 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
865 #endif
866 	}
867 
868 	uasm_l_vmalloc_done(l, *p);
869 
870 	/* get pgd offset in bytes */
871 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
872 
873 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
874 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
875 #ifndef __PAGETABLE_PUD_FOLDED
876 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
877 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
878 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
879 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
880 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
881 #endif
882 #ifndef __PAGETABLE_PMD_FOLDED
883 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
884 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
885 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
886 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
887 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
888 #endif
889 }
890 EXPORT_SYMBOL_GPL(build_get_pmde64);
891 
892 /*
893  * BVADDR is the faulting address, PTR is scratch.
894  * PTR will hold the pgd for vmalloc.
895  */
896 static void
897 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
898 			unsigned int bvaddr, unsigned int ptr,
899 			enum vmalloc64_mode mode)
900 {
901 	long swpd = (long)swapper_pg_dir;
902 	int single_insn_swpd;
903 	int did_vmalloc_branch = 0;
904 
905 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
906 
907 	uasm_l_vmalloc(l, *p);
908 
909 	if (mode != not_refill && check_for_high_segbits) {
910 		if (single_insn_swpd) {
911 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
912 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
913 			did_vmalloc_branch = 1;
914 			/* fall through */
915 		} else {
916 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
917 		}
918 	}
919 	if (!did_vmalloc_branch) {
920 		if (single_insn_swpd) {
921 			uasm_il_b(p, r, label_vmalloc_done);
922 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
923 		} else {
924 			UASM_i_LA_mostly(p, ptr, swpd);
925 			uasm_il_b(p, r, label_vmalloc_done);
926 			if (uasm_in_compat_space_p(swpd))
927 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
928 			else
929 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
930 		}
931 	}
932 	if (mode != not_refill && check_for_high_segbits) {
933 		uasm_l_large_segbits_fault(l, *p);
934 
935 		if (mode == refill_scratch && scratch_reg >= 0)
936 			uasm_i_ehb(p);
937 
938 		/*
939 		 * We get here if we are an xsseg address, or if we are
940 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
941 		 *
942 		 * Ignoring xsseg (assume disabled so would generate
943 		 * (address errors?), the only remaining possibility
944 		 * is the upper xuseg addresses.  On processors with
945 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
946 		 * addresses would have taken an address error. We try
947 		 * to mimic that here by taking a load/istream page
948 		 * fault.
949 		 */
950 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
951 			uasm_i_sync(p, 0);
952 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
953 		uasm_i_jr(p, ptr);
954 
955 		if (mode == refill_scratch) {
956 			if (scratch_reg >= 0)
957 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
958 			else
959 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
960 		} else {
961 			uasm_i_nop(p);
962 		}
963 	}
964 }
965 
966 #else /* !CONFIG_64BIT */
967 
968 /*
969  * TMP and PTR are scratch.
970  * TMP will be clobbered, PTR will hold the pgd entry.
971  */
972 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
973 {
974 	if (pgd_reg != -1) {
975 		/* pgd is in pgd_reg */
976 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
977 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
978 	} else {
979 		long pgdc = (long)pgd_current;
980 
981 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
982 #ifdef CONFIG_SMP
983 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
984 		UASM_i_LA_mostly(p, tmp, pgdc);
985 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
986 		uasm_i_addu(p, ptr, tmp, ptr);
987 #else
988 		UASM_i_LA_mostly(p, ptr, pgdc);
989 #endif
990 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
991 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
992 	}
993 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
994 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
995 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
996 }
997 EXPORT_SYMBOL_GPL(build_get_pgde32);
998 
999 #endif /* !CONFIG_64BIT */
1000 
1001 static void build_adjust_context(u32 **p, unsigned int ctx)
1002 {
1003 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
1004 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1005 
1006 	switch (current_cpu_type()) {
1007 	case CPU_VR41XX:
1008 	case CPU_VR4111:
1009 	case CPU_VR4121:
1010 	case CPU_VR4122:
1011 	case CPU_VR4131:
1012 	case CPU_VR4181:
1013 	case CPU_VR4181A:
1014 	case CPU_VR4133:
1015 		shift += 2;
1016 		break;
1017 
1018 	default:
1019 		break;
1020 	}
1021 
1022 	if (shift)
1023 		UASM_i_SRL(p, ctx, ctx, shift);
1024 	uasm_i_andi(p, ctx, ctx, mask);
1025 }
1026 
1027 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1028 {
1029 	/*
1030 	 * Bug workaround for the Nevada. It seems as if under certain
1031 	 * circumstances the move from cp0_context might produce a
1032 	 * bogus result when the mfc0 instruction and its consumer are
1033 	 * in a different cacheline or a load instruction, probably any
1034 	 * memory reference, is between them.
1035 	 */
1036 	switch (current_cpu_type()) {
1037 	case CPU_NEVADA:
1038 		UASM_i_LW(p, ptr, 0, ptr);
1039 		GET_CONTEXT(p, tmp); /* get context reg */
1040 		break;
1041 
1042 	default:
1043 		GET_CONTEXT(p, tmp); /* get context reg */
1044 		UASM_i_LW(p, ptr, 0, ptr);
1045 		break;
1046 	}
1047 
1048 	build_adjust_context(p, tmp);
1049 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1050 }
1051 EXPORT_SYMBOL_GPL(build_get_ptep);
1052 
1053 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1054 {
1055 	int pte_off_even = 0;
1056 	int pte_off_odd = sizeof(pte_t);
1057 
1058 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1059 	/* The low 32 bits of EntryLo is stored in pte_high */
1060 	pte_off_even += offsetof(pte_t, pte_high);
1061 	pte_off_odd += offsetof(pte_t, pte_high);
1062 #endif
1063 
1064 	if (IS_ENABLED(CONFIG_XPA)) {
1065 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1066 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1067 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1068 
1069 		if (cpu_has_xpa && !mips_xpa_disabled) {
1070 			uasm_i_lw(p, tmp, 0, ptep);
1071 			uasm_i_ext(p, tmp, tmp, 0, 24);
1072 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1073 		}
1074 
1075 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1076 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1077 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1078 
1079 		if (cpu_has_xpa && !mips_xpa_disabled) {
1080 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1081 			uasm_i_ext(p, tmp, tmp, 0, 24);
1082 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1083 		}
1084 		return;
1085 	}
1086 
1087 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1088 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1089 	if (r45k_bvahwbug())
1090 		build_tlb_probe_entry(p);
1091 	build_convert_pte_to_entrylo(p, tmp);
1092 	if (r4k_250MHZhwbug())
1093 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1094 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1095 	build_convert_pte_to_entrylo(p, ptep);
1096 	if (r45k_bvahwbug())
1097 		uasm_i_mfc0(p, tmp, C0_INDEX);
1098 	if (r4k_250MHZhwbug())
1099 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1100 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1101 }
1102 EXPORT_SYMBOL_GPL(build_update_entries);
1103 
1104 struct mips_huge_tlb_info {
1105 	int huge_pte;
1106 	int restore_scratch;
1107 	bool need_reload_pte;
1108 };
1109 
1110 static struct mips_huge_tlb_info
1111 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1112 			       struct uasm_reloc **r, unsigned int tmp,
1113 			       unsigned int ptr, int c0_scratch_reg)
1114 {
1115 	struct mips_huge_tlb_info rv;
1116 	unsigned int even, odd;
1117 	int vmalloc_branch_delay_filled = 0;
1118 	const int scratch = 1; /* Our extra working register */
1119 
1120 	rv.huge_pte = scratch;
1121 	rv.restore_scratch = 0;
1122 	rv.need_reload_pte = false;
1123 
1124 	if (check_for_high_segbits) {
1125 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1126 
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 		if (c0_scratch_reg >= 0)
1133 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1134 		else
1135 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1136 
1137 		uasm_i_dsrl_safe(p, scratch, tmp,
1138 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1139 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1140 
1141 		if (pgd_reg == -1) {
1142 			vmalloc_branch_delay_filled = 1;
1143 			/* Clear lower 23 bits of context. */
1144 			uasm_i_dins(p, ptr, 0, 0, 23);
1145 		}
1146 	} else {
1147 		if (pgd_reg != -1)
1148 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1149 		else
1150 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1151 
1152 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1153 
1154 		if (c0_scratch_reg >= 0)
1155 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1156 		else
1157 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1158 
1159 		if (pgd_reg == -1)
1160 			/* Clear lower 23 bits of context. */
1161 			uasm_i_dins(p, ptr, 0, 0, 23);
1162 
1163 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1164 	}
1165 
1166 	if (pgd_reg == -1) {
1167 		vmalloc_branch_delay_filled = 1;
1168 		/* 1 0	1 0 1  << 6  xkphys cached */
1169 		uasm_i_ori(p, ptr, ptr, 0x540);
1170 		uasm_i_drotr(p, ptr, ptr, 11);
1171 	}
1172 
1173 #ifdef __PAGETABLE_PMD_FOLDED
1174 #define LOC_PTEP scratch
1175 #else
1176 #define LOC_PTEP ptr
1177 #endif
1178 
1179 	if (!vmalloc_branch_delay_filled)
1180 		/* get pgd offset in bytes */
1181 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1182 
1183 	uasm_l_vmalloc_done(l, *p);
1184 
1185 	/*
1186 	 *			   tmp		ptr
1187 	 * fall-through case =	 badvaddr  *pgd_current
1188 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1189 	 */
1190 
1191 	if (vmalloc_branch_delay_filled)
1192 		/* get pgd offset in bytes */
1193 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1194 
1195 #ifdef __PAGETABLE_PMD_FOLDED
1196 	GET_CONTEXT(p, tmp); /* get context reg */
1197 #endif
1198 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1199 
1200 	if (use_lwx_insns()) {
1201 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1202 	} else {
1203 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1204 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1205 	}
1206 
1207 #ifndef __PAGETABLE_PUD_FOLDED
1208 	/* get pud offset in bytes */
1209 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1210 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1211 
1212 	if (use_lwx_insns()) {
1213 		UASM_i_LWX(p, ptr, scratch, ptr);
1214 	} else {
1215 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1216 		UASM_i_LW(p, ptr, 0, ptr);
1217 	}
1218 	/* ptr contains a pointer to PMD entry */
1219 	/* tmp contains the address */
1220 #endif
1221 
1222 #ifndef __PAGETABLE_PMD_FOLDED
1223 	/* get pmd offset in bytes */
1224 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1225 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1226 	GET_CONTEXT(p, tmp); /* get context reg */
1227 
1228 	if (use_lwx_insns()) {
1229 		UASM_i_LWX(p, scratch, scratch, ptr);
1230 	} else {
1231 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1232 		UASM_i_LW(p, scratch, 0, ptr);
1233 	}
1234 #endif
1235 	/* Adjust the context during the load latency. */
1236 	build_adjust_context(p, tmp);
1237 
1238 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1239 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1240 	/*
1241 	 * The in the LWX case we don't want to do the load in the
1242 	 * delay slot.	It cannot issue in the same cycle and may be
1243 	 * speculative and unneeded.
1244 	 */
1245 	if (use_lwx_insns())
1246 		uasm_i_nop(p);
1247 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1248 
1249 
1250 	/* build_update_entries */
1251 	if (use_lwx_insns()) {
1252 		even = ptr;
1253 		odd = tmp;
1254 		UASM_i_LWX(p, even, scratch, tmp);
1255 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1256 		UASM_i_LWX(p, odd, scratch, tmp);
1257 	} else {
1258 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1259 		even = tmp;
1260 		odd = ptr;
1261 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1262 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1263 	}
1264 	if (cpu_has_rixi) {
1265 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1266 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1267 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1268 	} else {
1269 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1270 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1271 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1272 	}
1273 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1274 
1275 	if (c0_scratch_reg >= 0) {
1276 		uasm_i_ehb(p);
1277 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1278 		build_tlb_write_entry(p, l, r, tlb_random);
1279 		uasm_l_leave(l, *p);
1280 		rv.restore_scratch = 1;
1281 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1282 		build_tlb_write_entry(p, l, r, tlb_random);
1283 		uasm_l_leave(l, *p);
1284 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1285 	} else {
1286 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1287 		build_tlb_write_entry(p, l, r, tlb_random);
1288 		uasm_l_leave(l, *p);
1289 		rv.restore_scratch = 1;
1290 	}
1291 
1292 	uasm_i_eret(p); /* return from trap */
1293 
1294 	return rv;
1295 }
1296 
1297 /*
1298  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1299  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1300  * slots before the XTLB refill exception handler which belong to the
1301  * unused TLB refill exception.
1302  */
1303 #define MIPS64_REFILL_INSNS 32
1304 
1305 static void build_r4000_tlb_refill_handler(void)
1306 {
1307 	u32 *p = tlb_handler;
1308 	struct uasm_label *l = labels;
1309 	struct uasm_reloc *r = relocs;
1310 	u32 *f;
1311 	unsigned int final_len;
1312 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1313 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1314 
1315 	memset(tlb_handler, 0, sizeof(tlb_handler));
1316 	memset(labels, 0, sizeof(labels));
1317 	memset(relocs, 0, sizeof(relocs));
1318 	memset(final_handler, 0, sizeof(final_handler));
1319 
1320 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1321 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1322 							  scratch_reg);
1323 		vmalloc_mode = refill_scratch;
1324 	} else {
1325 		htlb_info.huge_pte = K0;
1326 		htlb_info.restore_scratch = 0;
1327 		htlb_info.need_reload_pte = true;
1328 		vmalloc_mode = refill_noscratch;
1329 		/*
1330 		 * create the plain linear handler
1331 		 */
1332 		if (bcm1250_m3_war()) {
1333 			unsigned int segbits = 44;
1334 
1335 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1336 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1337 			uasm_i_xor(&p, K0, K0, K1);
1338 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1339 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1340 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1341 			uasm_i_or(&p, K0, K0, K1);
1342 			uasm_il_bnez(&p, &r, K0, label_leave);
1343 			/* No need for uasm_i_nop */
1344 		}
1345 
1346 #ifdef CONFIG_64BIT
1347 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1348 #else
1349 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1350 #endif
1351 
1352 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1353 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1354 #endif
1355 
1356 		build_get_ptep(&p, K0, K1);
1357 		build_update_entries(&p, K0, K1);
1358 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1359 		uasm_l_leave(&l, p);
1360 		uasm_i_eret(&p); /* return from trap */
1361 	}
1362 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1363 	uasm_l_tlb_huge_update(&l, p);
1364 	if (htlb_info.need_reload_pte)
1365 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1366 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1367 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1368 				   htlb_info.restore_scratch);
1369 #endif
1370 
1371 #ifdef CONFIG_64BIT
1372 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1373 #endif
1374 
1375 	/*
1376 	 * Overflow check: For the 64bit handler, we need at least one
1377 	 * free instruction slot for the wrap-around branch. In worst
1378 	 * case, if the intended insertion point is a delay slot, we
1379 	 * need three, with the second nop'ed and the third being
1380 	 * unused.
1381 	 */
1382 	switch (boot_cpu_type()) {
1383 	default:
1384 		if (sizeof(long) == 4) {
1385 	case CPU_LOONGSON2EF:
1386 		/* Loongson2 ebase is different than r4k, we have more space */
1387 			if ((p - tlb_handler) > 64)
1388 				panic("TLB refill handler space exceeded");
1389 			/*
1390 			 * Now fold the handler in the TLB refill handler space.
1391 			 */
1392 			f = final_handler;
1393 			/* Simplest case, just copy the handler. */
1394 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1395 			final_len = p - tlb_handler;
1396 			break;
1397 		} else {
1398 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1399 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1400 				&& uasm_insn_has_bdelay(relocs,
1401 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1402 				panic("TLB refill handler space exceeded");
1403 			/*
1404 			 * Now fold the handler in the TLB refill handler space.
1405 			 */
1406 			f = final_handler + MIPS64_REFILL_INSNS;
1407 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1408 				/* Just copy the handler. */
1409 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1410 				final_len = p - tlb_handler;
1411 			} else {
1412 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1413 				const enum label_id ls = label_tlb_huge_update;
1414 #else
1415 				const enum label_id ls = label_vmalloc;
1416 #endif
1417 				u32 *split;
1418 				int ov = 0;
1419 				int i;
1420 
1421 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1422 					;
1423 				BUG_ON(i == ARRAY_SIZE(labels));
1424 				split = labels[i].addr;
1425 
1426 				/*
1427 				 * See if we have overflown one way or the other.
1428 				 */
1429 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1430 				    split < p - MIPS64_REFILL_INSNS)
1431 					ov = 1;
1432 
1433 				if (ov) {
1434 					/*
1435 					 * Split two instructions before the end.  One
1436 					 * for the branch and one for the instruction
1437 					 * in the delay slot.
1438 					 */
1439 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1440 
1441 					/*
1442 					 * If the branch would fall in a delay slot,
1443 					 * we must back up an additional instruction
1444 					 * so that it is no longer in a delay slot.
1445 					 */
1446 					if (uasm_insn_has_bdelay(relocs, split - 1))
1447 						split--;
1448 				}
1449 				/* Copy first part of the handler. */
1450 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1451 				f += split - tlb_handler;
1452 
1453 				if (ov) {
1454 					/* Insert branch. */
1455 					uasm_l_split(&l, final_handler);
1456 					uasm_il_b(&f, &r, label_split);
1457 					if (uasm_insn_has_bdelay(relocs, split))
1458 						uasm_i_nop(&f);
1459 					else {
1460 						uasm_copy_handler(relocs, labels,
1461 								  split, split + 1, f);
1462 						uasm_move_labels(labels, f, f + 1, -1);
1463 						f++;
1464 						split++;
1465 					}
1466 				}
1467 
1468 				/* Copy the rest of the handler. */
1469 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1470 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1471 					    (p - split);
1472 			}
1473 		}
1474 		break;
1475 	}
1476 
1477 	uasm_resolve_relocs(relocs, labels);
1478 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1479 		 final_len);
1480 
1481 	memcpy((void *)ebase, final_handler, 0x100);
1482 	local_flush_icache_range(ebase, ebase + 0x100);
1483 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1484 }
1485 
1486 static void setup_pw(void)
1487 {
1488 	unsigned int pwctl;
1489 	unsigned long pgd_i, pgd_w;
1490 #ifndef __PAGETABLE_PMD_FOLDED
1491 	unsigned long pmd_i, pmd_w;
1492 #endif
1493 	unsigned long pt_i, pt_w;
1494 	unsigned long pte_i, pte_w;
1495 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1496 	unsigned long psn;
1497 
1498 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1499 #endif
1500 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1501 #ifndef __PAGETABLE_PMD_FOLDED
1502 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1503 
1504 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1505 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1506 #else
1507 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1508 #endif
1509 
1510 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1511 	pt_w  = PAGE_SHIFT - 3;
1512 
1513 	pte_i = ilog2(_PAGE_GLOBAL);
1514 	pte_w = 0;
1515 	pwctl = 1 << 30; /* Set PWDirExt */
1516 
1517 #ifndef __PAGETABLE_PMD_FOLDED
1518 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1519 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1520 #else
1521 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1522 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1523 #endif
1524 
1525 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1526 	pwctl |= (1 << 6 | psn);
1527 #endif
1528 	write_c0_pwctl(pwctl);
1529 	write_c0_kpgd((long)swapper_pg_dir);
1530 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1531 }
1532 
1533 static void build_loongson3_tlb_refill_handler(void)
1534 {
1535 	u32 *p = tlb_handler;
1536 	struct uasm_label *l = labels;
1537 	struct uasm_reloc *r = relocs;
1538 
1539 	memset(labels, 0, sizeof(labels));
1540 	memset(relocs, 0, sizeof(relocs));
1541 	memset(tlb_handler, 0, sizeof(tlb_handler));
1542 
1543 	if (check_for_high_segbits) {
1544 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1545 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1546 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1547 		uasm_i_nop(&p);
1548 
1549 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1550 		uasm_i_nop(&p);
1551 		uasm_l_vmalloc(&l, p);
1552 	}
1553 
1554 	uasm_i_dmfc0(&p, K1, C0_PGD);
1555 
1556 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1557 #ifndef __PAGETABLE_PMD_FOLDED
1558 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1559 #endif
1560 	uasm_i_ldpte(&p, K1, 0);      /* even */
1561 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1562 	uasm_i_tlbwr(&p);
1563 
1564 	/* restore page mask */
1565 	if (PM_DEFAULT_MASK >> 16) {
1566 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1567 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1568 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1569 	} else if (PM_DEFAULT_MASK) {
1570 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1571 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1572 	} else {
1573 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1574 	}
1575 
1576 	uasm_i_eret(&p);
1577 
1578 	if (check_for_high_segbits) {
1579 		uasm_l_large_segbits_fault(&l, p);
1580 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1581 		uasm_i_jr(&p, K1);
1582 		uasm_i_nop(&p);
1583 	}
1584 
1585 	uasm_resolve_relocs(relocs, labels);
1586 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1587 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1588 	dump_handler("loongson3_tlb_refill",
1589 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1590 }
1591 
1592 static void build_setup_pgd(void)
1593 {
1594 	const int a0 = 4;
1595 	const int __maybe_unused a1 = 5;
1596 	const int __maybe_unused a2 = 6;
1597 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1598 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1599 	long pgdc = (long)pgd_current;
1600 #endif
1601 
1602 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1603 	memset(labels, 0, sizeof(labels));
1604 	memset(relocs, 0, sizeof(relocs));
1605 	pgd_reg = allocate_kscratch();
1606 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1607 	if (pgd_reg == -1) {
1608 		struct uasm_label *l = labels;
1609 		struct uasm_reloc *r = relocs;
1610 
1611 		/* PGD << 11 in c0_Context */
1612 		/*
1613 		 * If it is a ckseg0 address, convert to a physical
1614 		 * address.  Shifting right by 29 and adding 4 will
1615 		 * result in zero for these addresses.
1616 		 *
1617 		 */
1618 		UASM_i_SRA(&p, a1, a0, 29);
1619 		UASM_i_ADDIU(&p, a1, a1, 4);
1620 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1621 		uasm_i_nop(&p);
1622 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1623 		uasm_l_tlbl_goaround1(&l, p);
1624 		UASM_i_SLL(&p, a0, a0, 11);
1625 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1626 		uasm_i_jr(&p, 31);
1627 		uasm_i_ehb(&p);
1628 	} else {
1629 		/* PGD in c0_KScratch */
1630 		if (cpu_has_ldpte)
1631 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1632 		else
1633 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1634 		uasm_i_jr(&p, 31);
1635 		uasm_i_ehb(&p);
1636 	}
1637 #else
1638 #ifdef CONFIG_SMP
1639 	/* Save PGD to pgd_current[smp_processor_id()] */
1640 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1641 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1642 	UASM_i_LA_mostly(&p, a2, pgdc);
1643 	UASM_i_ADDU(&p, a2, a2, a1);
1644 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1645 #else
1646 	UASM_i_LA_mostly(&p, a2, pgdc);
1647 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1648 #endif /* SMP */
1649 
1650 	/* if pgd_reg is allocated, save PGD also to scratch register */
1651 	if (pgd_reg != -1) {
1652 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1653 		uasm_i_jr(&p, 31);
1654 		uasm_i_ehb(&p);
1655 	} else {
1656 		uasm_i_jr(&p, 31);
1657 		uasm_i_nop(&p);
1658 	}
1659 #endif
1660 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1661 		panic("tlbmiss_handler_setup_pgd space exceeded");
1662 
1663 	uasm_resolve_relocs(relocs, labels);
1664 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1665 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1666 
1667 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1668 					tlbmiss_handler_setup_pgd_end);
1669 }
1670 
1671 static void
1672 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1673 {
1674 #ifdef CONFIG_SMP
1675 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1676 		uasm_i_sync(p, 0);
1677 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1678 	if (cpu_has_64bits)
1679 		uasm_i_lld(p, pte, 0, ptr);
1680 	else
1681 # endif
1682 		UASM_i_LL(p, pte, 0, ptr);
1683 #else
1684 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1685 	if (cpu_has_64bits)
1686 		uasm_i_ld(p, pte, 0, ptr);
1687 	else
1688 # endif
1689 		UASM_i_LW(p, pte, 0, ptr);
1690 #endif
1691 }
1692 
1693 static void
1694 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1695 	unsigned int mode, unsigned int scratch)
1696 {
1697 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1698 	unsigned int swmode = mode & ~hwmode;
1699 
1700 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1701 		uasm_i_lui(p, scratch, swmode >> 16);
1702 		uasm_i_or(p, pte, pte, scratch);
1703 		BUG_ON(swmode & 0xffff);
1704 	} else {
1705 		uasm_i_ori(p, pte, pte, mode);
1706 	}
1707 
1708 #ifdef CONFIG_SMP
1709 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1710 	if (cpu_has_64bits)
1711 		uasm_i_scd(p, pte, 0, ptr);
1712 	else
1713 # endif
1714 		UASM_i_SC(p, pte, 0, ptr);
1715 
1716 	if (r10000_llsc_war())
1717 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1718 	else
1719 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1720 
1721 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1722 	if (!cpu_has_64bits) {
1723 		/* no uasm_i_nop needed */
1724 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1725 		uasm_i_ori(p, pte, pte, hwmode);
1726 		BUG_ON(hwmode & ~0xffff);
1727 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1728 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1729 		/* no uasm_i_nop needed */
1730 		uasm_i_lw(p, pte, 0, ptr);
1731 	} else
1732 		uasm_i_nop(p);
1733 # else
1734 	uasm_i_nop(p);
1735 # endif
1736 #else
1737 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1738 	if (cpu_has_64bits)
1739 		uasm_i_sd(p, pte, 0, ptr);
1740 	else
1741 # endif
1742 		UASM_i_SW(p, pte, 0, ptr);
1743 
1744 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1745 	if (!cpu_has_64bits) {
1746 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1747 		uasm_i_ori(p, pte, pte, hwmode);
1748 		BUG_ON(hwmode & ~0xffff);
1749 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1750 		uasm_i_lw(p, pte, 0, ptr);
1751 	}
1752 # endif
1753 #endif
1754 }
1755 
1756 /*
1757  * Check if PTE is present, if not then jump to LABEL. PTR points to
1758  * the page table where this PTE is located, PTE will be re-loaded
1759  * with it's original value.
1760  */
1761 static void
1762 build_pte_present(u32 **p, struct uasm_reloc **r,
1763 		  int pte, int ptr, int scratch, enum label_id lid)
1764 {
1765 	int t = scratch >= 0 ? scratch : pte;
1766 	int cur = pte;
1767 
1768 	if (cpu_has_rixi) {
1769 		if (use_bbit_insns()) {
1770 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1771 			uasm_i_nop(p);
1772 		} else {
1773 			if (_PAGE_PRESENT_SHIFT) {
1774 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1775 				cur = t;
1776 			}
1777 			uasm_i_andi(p, t, cur, 1);
1778 			uasm_il_beqz(p, r, t, lid);
1779 			if (pte == t)
1780 				/* You lose the SMP race :-(*/
1781 				iPTE_LW(p, pte, ptr);
1782 		}
1783 	} else {
1784 		if (_PAGE_PRESENT_SHIFT) {
1785 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1786 			cur = t;
1787 		}
1788 		uasm_i_andi(p, t, cur,
1789 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1790 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1791 		uasm_il_bnez(p, r, t, lid);
1792 		if (pte == t)
1793 			/* You lose the SMP race :-(*/
1794 			iPTE_LW(p, pte, ptr);
1795 	}
1796 }
1797 
1798 /* Make PTE valid, store result in PTR. */
1799 static void
1800 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1801 		 unsigned int ptr, unsigned int scratch)
1802 {
1803 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1804 
1805 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1806 }
1807 
1808 /*
1809  * Check if PTE can be written to, if not branch to LABEL. Regardless
1810  * restore PTE with value from PTR when done.
1811  */
1812 static void
1813 build_pte_writable(u32 **p, struct uasm_reloc **r,
1814 		   unsigned int pte, unsigned int ptr, int scratch,
1815 		   enum label_id lid)
1816 {
1817 	int t = scratch >= 0 ? scratch : pte;
1818 	int cur = pte;
1819 
1820 	if (_PAGE_PRESENT_SHIFT) {
1821 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1822 		cur = t;
1823 	}
1824 	uasm_i_andi(p, t, cur,
1825 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1826 	uasm_i_xori(p, t, t,
1827 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1828 	uasm_il_bnez(p, r, t, lid);
1829 	if (pte == t)
1830 		/* You lose the SMP race :-(*/
1831 		iPTE_LW(p, pte, ptr);
1832 	else
1833 		uasm_i_nop(p);
1834 }
1835 
1836 /* Make PTE writable, update software status bits as well, then store
1837  * at PTR.
1838  */
1839 static void
1840 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1841 		 unsigned int ptr, unsigned int scratch)
1842 {
1843 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1844 			     | _PAGE_DIRTY);
1845 
1846 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1847 }
1848 
1849 /*
1850  * Check if PTE can be modified, if not branch to LABEL. Regardless
1851  * restore PTE with value from PTR when done.
1852  */
1853 static void
1854 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1855 		     unsigned int pte, unsigned int ptr, int scratch,
1856 		     enum label_id lid)
1857 {
1858 	if (use_bbit_insns()) {
1859 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1860 		uasm_i_nop(p);
1861 	} else {
1862 		int t = scratch >= 0 ? scratch : pte;
1863 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1864 		uasm_i_andi(p, t, t, 1);
1865 		uasm_il_beqz(p, r, t, lid);
1866 		if (pte == t)
1867 			/* You lose the SMP race :-(*/
1868 			iPTE_LW(p, pte, ptr);
1869 	}
1870 }
1871 
1872 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1873 
1874 
1875 /*
1876  * R3000 style TLB load/store/modify handlers.
1877  */
1878 
1879 /*
1880  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1881  * Then it returns.
1882  */
1883 static void
1884 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1885 {
1886 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1887 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1888 	uasm_i_tlbwi(p);
1889 	uasm_i_jr(p, tmp);
1890 	uasm_i_rfe(p); /* branch delay */
1891 }
1892 
1893 /*
1894  * This places the pte into ENTRYLO0 and writes it with tlbwi
1895  * or tlbwr as appropriate.  This is because the index register
1896  * may have the probe fail bit set as a result of a trap on a
1897  * kseg2 access, i.e. without refill.  Then it returns.
1898  */
1899 static void
1900 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1901 			     struct uasm_reloc **r, unsigned int pte,
1902 			     unsigned int tmp)
1903 {
1904 	uasm_i_mfc0(p, tmp, C0_INDEX);
1905 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1906 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1907 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1908 	uasm_i_tlbwi(p); /* cp0 delay */
1909 	uasm_i_jr(p, tmp);
1910 	uasm_i_rfe(p); /* branch delay */
1911 	uasm_l_r3000_write_probe_fail(l, *p);
1912 	uasm_i_tlbwr(p); /* cp0 delay */
1913 	uasm_i_jr(p, tmp);
1914 	uasm_i_rfe(p); /* branch delay */
1915 }
1916 
1917 static void
1918 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1919 				   unsigned int ptr)
1920 {
1921 	long pgdc = (long)pgd_current;
1922 
1923 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1924 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1925 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1926 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1927 	uasm_i_sll(p, pte, pte, 2);
1928 	uasm_i_addu(p, ptr, ptr, pte);
1929 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1930 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1931 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1932 	uasm_i_addu(p, ptr, ptr, pte);
1933 	uasm_i_lw(p, pte, 0, ptr);
1934 	uasm_i_tlbp(p); /* load delay */
1935 }
1936 
1937 static void build_r3000_tlb_load_handler(void)
1938 {
1939 	u32 *p = (u32 *)handle_tlbl;
1940 	struct uasm_label *l = labels;
1941 	struct uasm_reloc *r = relocs;
1942 
1943 	memset(p, 0, handle_tlbl_end - (char *)p);
1944 	memset(labels, 0, sizeof(labels));
1945 	memset(relocs, 0, sizeof(relocs));
1946 
1947 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1948 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1949 	uasm_i_nop(&p); /* load delay */
1950 	build_make_valid(&p, &r, K0, K1, -1);
1951 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1952 
1953 	uasm_l_nopage_tlbl(&l, p);
1954 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1955 	uasm_i_nop(&p);
1956 
1957 	if (p >= (u32 *)handle_tlbl_end)
1958 		panic("TLB load handler fastpath space exceeded");
1959 
1960 	uasm_resolve_relocs(relocs, labels);
1961 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1962 		 (unsigned int)(p - (u32 *)handle_tlbl));
1963 
1964 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1965 }
1966 
1967 static void build_r3000_tlb_store_handler(void)
1968 {
1969 	u32 *p = (u32 *)handle_tlbs;
1970 	struct uasm_label *l = labels;
1971 	struct uasm_reloc *r = relocs;
1972 
1973 	memset(p, 0, handle_tlbs_end - (char *)p);
1974 	memset(labels, 0, sizeof(labels));
1975 	memset(relocs, 0, sizeof(relocs));
1976 
1977 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1978 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1979 	uasm_i_nop(&p); /* load delay */
1980 	build_make_write(&p, &r, K0, K1, -1);
1981 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1982 
1983 	uasm_l_nopage_tlbs(&l, p);
1984 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1985 	uasm_i_nop(&p);
1986 
1987 	if (p >= (u32 *)handle_tlbs_end)
1988 		panic("TLB store handler fastpath space exceeded");
1989 
1990 	uasm_resolve_relocs(relocs, labels);
1991 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1992 		 (unsigned int)(p - (u32 *)handle_tlbs));
1993 
1994 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1995 }
1996 
1997 static void build_r3000_tlb_modify_handler(void)
1998 {
1999 	u32 *p = (u32 *)handle_tlbm;
2000 	struct uasm_label *l = labels;
2001 	struct uasm_reloc *r = relocs;
2002 
2003 	memset(p, 0, handle_tlbm_end - (char *)p);
2004 	memset(labels, 0, sizeof(labels));
2005 	memset(relocs, 0, sizeof(relocs));
2006 
2007 	build_r3000_tlbchange_handler_head(&p, K0, K1);
2008 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
2009 	uasm_i_nop(&p); /* load delay */
2010 	build_make_write(&p, &r, K0, K1, -1);
2011 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
2012 
2013 	uasm_l_nopage_tlbm(&l, p);
2014 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2015 	uasm_i_nop(&p);
2016 
2017 	if (p >= (u32 *)handle_tlbm_end)
2018 		panic("TLB modify handler fastpath space exceeded");
2019 
2020 	uasm_resolve_relocs(relocs, labels);
2021 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2022 		 (unsigned int)(p - (u32 *)handle_tlbm));
2023 
2024 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2025 }
2026 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2027 
2028 static bool cpu_has_tlbex_tlbp_race(void)
2029 {
2030 	/*
2031 	 * When a Hardware Table Walker is running it can replace TLB entries
2032 	 * at any time, leading to a race between it & the CPU.
2033 	 */
2034 	if (cpu_has_htw)
2035 		return true;
2036 
2037 	/*
2038 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2039 	 * replaced at any time by a sibling performing a write to the FTLB.
2040 	 */
2041 	if (cpu_has_shared_ftlb_ram)
2042 		return true;
2043 
2044 	/* In all other cases there ought to be no race condition to handle */
2045 	return false;
2046 }
2047 
2048 /*
2049  * R4000 style TLB load/store/modify handlers.
2050  */
2051 static struct work_registers
2052 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2053 				   struct uasm_reloc **r)
2054 {
2055 	struct work_registers wr = build_get_work_registers(p);
2056 
2057 #ifdef CONFIG_64BIT
2058 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2059 #else
2060 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2061 #endif
2062 
2063 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2064 	/*
2065 	 * For huge tlb entries, pmd doesn't contain an address but
2066 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2067 	 * see if we need to jump to huge tlb processing.
2068 	 */
2069 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2070 #endif
2071 
2072 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2073 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2074 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2075 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2076 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2077 
2078 #ifdef CONFIG_SMP
2079 	uasm_l_smp_pgtable_change(l, *p);
2080 #endif
2081 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2082 	if (!m4kc_tlbp_war()) {
2083 		build_tlb_probe_entry(p);
2084 		if (cpu_has_tlbex_tlbp_race()) {
2085 			/* race condition happens, leaving */
2086 			uasm_i_ehb(p);
2087 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2088 			uasm_il_bltz(p, r, wr.r3, label_leave);
2089 			uasm_i_nop(p);
2090 		}
2091 	}
2092 	return wr;
2093 }
2094 
2095 static void
2096 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2097 				   struct uasm_reloc **r, unsigned int tmp,
2098 				   unsigned int ptr)
2099 {
2100 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2101 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2102 	build_update_entries(p, tmp, ptr);
2103 	build_tlb_write_entry(p, l, r, tlb_indexed);
2104 	uasm_l_leave(l, *p);
2105 	build_restore_work_registers(p);
2106 	uasm_i_eret(p); /* return from trap */
2107 
2108 #ifdef CONFIG_64BIT
2109 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2110 #endif
2111 }
2112 
2113 static void build_r4000_tlb_load_handler(void)
2114 {
2115 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2116 	struct uasm_label *l = labels;
2117 	struct uasm_reloc *r = relocs;
2118 	struct work_registers wr;
2119 
2120 	memset(p, 0, handle_tlbl_end - (char *)p);
2121 	memset(labels, 0, sizeof(labels));
2122 	memset(relocs, 0, sizeof(relocs));
2123 
2124 	if (bcm1250_m3_war()) {
2125 		unsigned int segbits = 44;
2126 
2127 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2128 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2129 		uasm_i_xor(&p, K0, K0, K1);
2130 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2131 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2132 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2133 		uasm_i_or(&p, K0, K0, K1);
2134 		uasm_il_bnez(&p, &r, K0, label_leave);
2135 		/* No need for uasm_i_nop */
2136 	}
2137 
2138 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2139 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2140 	if (m4kc_tlbp_war())
2141 		build_tlb_probe_entry(&p);
2142 
2143 	if (cpu_has_rixi && !cpu_has_rixiex) {
2144 		/*
2145 		 * If the page is not _PAGE_VALID, RI or XI could not
2146 		 * have triggered it.  Skip the expensive test..
2147 		 */
2148 		if (use_bbit_insns()) {
2149 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2150 				      label_tlbl_goaround1);
2151 		} else {
2152 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2153 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2154 		}
2155 		uasm_i_nop(&p);
2156 
2157 		/*
2158 		 * Warn if something may race with us & replace the TLB entry
2159 		 * before we read it here. Everything with such races should
2160 		 * also have dedicated RiXi exception handlers, so this
2161 		 * shouldn't be hit.
2162 		 */
2163 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2164 
2165 		uasm_i_tlbr(&p);
2166 
2167 		switch (current_cpu_type()) {
2168 		default:
2169 			if (cpu_has_mips_r2_exec_hazard) {
2170 				uasm_i_ehb(&p);
2171 
2172 		case CPU_CAVIUM_OCTEON:
2173 		case CPU_CAVIUM_OCTEON_PLUS:
2174 		case CPU_CAVIUM_OCTEON2:
2175 				break;
2176 			}
2177 		}
2178 
2179 		/* Examine  entrylo 0 or 1 based on ptr. */
2180 		if (use_bbit_insns()) {
2181 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2182 		} else {
2183 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2184 			uasm_i_beqz(&p, wr.r3, 8);
2185 		}
2186 		/* load it in the delay slot*/
2187 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2188 		/* load it if ptr is odd */
2189 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2190 		/*
2191 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2192 		 * XI must have triggered it.
2193 		 */
2194 		if (use_bbit_insns()) {
2195 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2196 			uasm_i_nop(&p);
2197 			uasm_l_tlbl_goaround1(&l, p);
2198 		} else {
2199 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2200 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2201 			uasm_i_nop(&p);
2202 		}
2203 		uasm_l_tlbl_goaround1(&l, p);
2204 	}
2205 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2206 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2207 
2208 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2209 	/*
2210 	 * This is the entry point when build_r4000_tlbchange_handler_head
2211 	 * spots a huge page.
2212 	 */
2213 	uasm_l_tlb_huge_update(&l, p);
2214 	iPTE_LW(&p, wr.r1, wr.r2);
2215 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2216 	build_tlb_probe_entry(&p);
2217 
2218 	if (cpu_has_rixi && !cpu_has_rixiex) {
2219 		/*
2220 		 * If the page is not _PAGE_VALID, RI or XI could not
2221 		 * have triggered it.  Skip the expensive test..
2222 		 */
2223 		if (use_bbit_insns()) {
2224 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2225 				      label_tlbl_goaround2);
2226 		} else {
2227 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2228 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2229 		}
2230 		uasm_i_nop(&p);
2231 
2232 		/*
2233 		 * Warn if something may race with us & replace the TLB entry
2234 		 * before we read it here. Everything with such races should
2235 		 * also have dedicated RiXi exception handlers, so this
2236 		 * shouldn't be hit.
2237 		 */
2238 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2239 
2240 		uasm_i_tlbr(&p);
2241 
2242 		switch (current_cpu_type()) {
2243 		default:
2244 			if (cpu_has_mips_r2_exec_hazard) {
2245 				uasm_i_ehb(&p);
2246 
2247 		case CPU_CAVIUM_OCTEON:
2248 		case CPU_CAVIUM_OCTEON_PLUS:
2249 		case CPU_CAVIUM_OCTEON2:
2250 				break;
2251 			}
2252 		}
2253 
2254 		/* Examine  entrylo 0 or 1 based on ptr. */
2255 		if (use_bbit_insns()) {
2256 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2257 		} else {
2258 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2259 			uasm_i_beqz(&p, wr.r3, 8);
2260 		}
2261 		/* load it in the delay slot*/
2262 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2263 		/* load it if ptr is odd */
2264 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2265 		/*
2266 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2267 		 * XI must have triggered it.
2268 		 */
2269 		if (use_bbit_insns()) {
2270 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2271 		} else {
2272 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2273 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2274 		}
2275 		if (PM_DEFAULT_MASK == 0)
2276 			uasm_i_nop(&p);
2277 		/*
2278 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2279 		 * it is restored in build_huge_tlb_write_entry.
2280 		 */
2281 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2282 
2283 		uasm_l_tlbl_goaround2(&l, p);
2284 	}
2285 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2286 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2287 #endif
2288 
2289 	uasm_l_nopage_tlbl(&l, p);
2290 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2291 		uasm_i_sync(&p, 0);
2292 	build_restore_work_registers(&p);
2293 #ifdef CONFIG_CPU_MICROMIPS
2294 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2295 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2296 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2297 		uasm_i_jr(&p, K0);
2298 	} else
2299 #endif
2300 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2301 	uasm_i_nop(&p);
2302 
2303 	if (p >= (u32 *)handle_tlbl_end)
2304 		panic("TLB load handler fastpath space exceeded");
2305 
2306 	uasm_resolve_relocs(relocs, labels);
2307 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2308 		 (unsigned int)(p - (u32 *)handle_tlbl));
2309 
2310 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2311 }
2312 
2313 static void build_r4000_tlb_store_handler(void)
2314 {
2315 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2316 	struct uasm_label *l = labels;
2317 	struct uasm_reloc *r = relocs;
2318 	struct work_registers wr;
2319 
2320 	memset(p, 0, handle_tlbs_end - (char *)p);
2321 	memset(labels, 0, sizeof(labels));
2322 	memset(relocs, 0, sizeof(relocs));
2323 
2324 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2325 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2326 	if (m4kc_tlbp_war())
2327 		build_tlb_probe_entry(&p);
2328 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2329 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2330 
2331 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2332 	/*
2333 	 * This is the entry point when
2334 	 * build_r4000_tlbchange_handler_head spots a huge page.
2335 	 */
2336 	uasm_l_tlb_huge_update(&l, p);
2337 	iPTE_LW(&p, wr.r1, wr.r2);
2338 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2339 	build_tlb_probe_entry(&p);
2340 	uasm_i_ori(&p, wr.r1, wr.r1,
2341 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2342 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2343 #endif
2344 
2345 	uasm_l_nopage_tlbs(&l, p);
2346 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2347 		uasm_i_sync(&p, 0);
2348 	build_restore_work_registers(&p);
2349 #ifdef CONFIG_CPU_MICROMIPS
2350 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2351 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2352 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2353 		uasm_i_jr(&p, K0);
2354 	} else
2355 #endif
2356 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2357 	uasm_i_nop(&p);
2358 
2359 	if (p >= (u32 *)handle_tlbs_end)
2360 		panic("TLB store handler fastpath space exceeded");
2361 
2362 	uasm_resolve_relocs(relocs, labels);
2363 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2364 		 (unsigned int)(p - (u32 *)handle_tlbs));
2365 
2366 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2367 }
2368 
2369 static void build_r4000_tlb_modify_handler(void)
2370 {
2371 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2372 	struct uasm_label *l = labels;
2373 	struct uasm_reloc *r = relocs;
2374 	struct work_registers wr;
2375 
2376 	memset(p, 0, handle_tlbm_end - (char *)p);
2377 	memset(labels, 0, sizeof(labels));
2378 	memset(relocs, 0, sizeof(relocs));
2379 
2380 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2381 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2382 	if (m4kc_tlbp_war())
2383 		build_tlb_probe_entry(&p);
2384 	/* Present and writable bits set, set accessed and dirty bits. */
2385 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2386 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2387 
2388 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2389 	/*
2390 	 * This is the entry point when
2391 	 * build_r4000_tlbchange_handler_head spots a huge page.
2392 	 */
2393 	uasm_l_tlb_huge_update(&l, p);
2394 	iPTE_LW(&p, wr.r1, wr.r2);
2395 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2396 	build_tlb_probe_entry(&p);
2397 	uasm_i_ori(&p, wr.r1, wr.r1,
2398 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2399 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2400 #endif
2401 
2402 	uasm_l_nopage_tlbm(&l, p);
2403 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2404 		uasm_i_sync(&p, 0);
2405 	build_restore_work_registers(&p);
2406 #ifdef CONFIG_CPU_MICROMIPS
2407 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2408 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2409 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2410 		uasm_i_jr(&p, K0);
2411 	} else
2412 #endif
2413 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2414 	uasm_i_nop(&p);
2415 
2416 	if (p >= (u32 *)handle_tlbm_end)
2417 		panic("TLB modify handler fastpath space exceeded");
2418 
2419 	uasm_resolve_relocs(relocs, labels);
2420 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2421 		 (unsigned int)(p - (u32 *)handle_tlbm));
2422 
2423 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2424 }
2425 
2426 static void flush_tlb_handlers(void)
2427 {
2428 	local_flush_icache_range((unsigned long)handle_tlbl,
2429 			   (unsigned long)handle_tlbl_end);
2430 	local_flush_icache_range((unsigned long)handle_tlbs,
2431 			   (unsigned long)handle_tlbs_end);
2432 	local_flush_icache_range((unsigned long)handle_tlbm,
2433 			   (unsigned long)handle_tlbm_end);
2434 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2435 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2436 }
2437 
2438 static void print_htw_config(void)
2439 {
2440 	unsigned long config;
2441 	unsigned int pwctl;
2442 	const int field = 2 * sizeof(unsigned long);
2443 
2444 	config = read_c0_pwfield();
2445 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2446 		field, config,
2447 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2448 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2449 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2450 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2451 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2452 
2453 	config = read_c0_pwsize();
2454 	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",
2455 		field, config,
2456 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2457 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2458 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2459 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2460 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2461 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2462 
2463 	pwctl = read_c0_pwctl();
2464 	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",
2465 		pwctl,
2466 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2467 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2468 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2469 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2470 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2471 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2472 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2473 }
2474 
2475 static void config_htw_params(void)
2476 {
2477 	unsigned long pwfield, pwsize, ptei;
2478 	unsigned int config;
2479 
2480 	/*
2481 	 * We are using 2-level page tables, so we only need to
2482 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2483 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2484 	 * write values less than 0xc in these fields because the entire
2485 	 * write will be dropped. As a result of which, we must preserve
2486 	 * the original reset values and overwrite only what we really want.
2487 	 */
2488 
2489 	pwfield = read_c0_pwfield();
2490 	/* re-initialize the GDI field */
2491 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2492 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2493 	/* re-initialize the PTI field including the even/odd bit */
2494 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2495 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2496 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2497 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2498 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2499 	}
2500 	/* Set the PTEI right shift */
2501 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2502 	pwfield |= ptei;
2503 	write_c0_pwfield(pwfield);
2504 	/* Check whether the PTEI value is supported */
2505 	back_to_back_c0_hazard();
2506 	pwfield = read_c0_pwfield();
2507 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2508 		!= ptei) {
2509 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2510 			ptei);
2511 		/*
2512 		 * Drop option to avoid HTW being enabled via another path
2513 		 * (eg htw_reset())
2514 		 */
2515 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2516 		return;
2517 	}
2518 
2519 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2520 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2521 	if (CONFIG_PGTABLE_LEVELS >= 3)
2522 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2523 
2524 	/* Set pointer size to size of directory pointers */
2525 	if (IS_ENABLED(CONFIG_64BIT))
2526 		pwsize |= MIPS_PWSIZE_PS_MASK;
2527 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2528 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2529 			& MIPS_PWSIZE_PTEW_MASK;
2530 
2531 	write_c0_pwsize(pwsize);
2532 
2533 	/* Make sure everything is set before we enable the HTW */
2534 	back_to_back_c0_hazard();
2535 
2536 	/*
2537 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2538 	 * the pwctl fields.
2539 	 */
2540 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2541 	if (IS_ENABLED(CONFIG_64BIT))
2542 		config |= MIPS_PWCTL_XU_MASK;
2543 	write_c0_pwctl(config);
2544 	pr_info("Hardware Page Table Walker enabled\n");
2545 
2546 	print_htw_config();
2547 }
2548 
2549 static void config_xpa_params(void)
2550 {
2551 #ifdef CONFIG_XPA
2552 	unsigned int pagegrain;
2553 
2554 	if (mips_xpa_disabled) {
2555 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2556 		return;
2557 	}
2558 
2559 	pagegrain = read_c0_pagegrain();
2560 	write_c0_pagegrain(pagegrain | PG_ELPA);
2561 	back_to_back_c0_hazard();
2562 	pagegrain = read_c0_pagegrain();
2563 
2564 	if (pagegrain & PG_ELPA)
2565 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2566 	else
2567 		panic("Extended Physical Addressing (XPA) disabled");
2568 #endif
2569 }
2570 
2571 static void check_pabits(void)
2572 {
2573 	unsigned long entry;
2574 	unsigned pabits, fillbits;
2575 
2576 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2577 		/*
2578 		 * We'll only be making use of the fact that we can rotate bits
2579 		 * into the fill if the CPU supports RIXI, so don't bother
2580 		 * probing this for CPUs which don't.
2581 		 */
2582 		return;
2583 	}
2584 
2585 	write_c0_entrylo0(~0ul);
2586 	back_to_back_c0_hazard();
2587 	entry = read_c0_entrylo0();
2588 
2589 	/* clear all non-PFN bits */
2590 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2591 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2592 
2593 	/* find a lower bound on PABITS, and upper bound on fill bits */
2594 	pabits = fls_long(entry) + 6;
2595 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2596 
2597 	/* minus the RI & XI bits */
2598 	fillbits -= min_t(unsigned, fillbits, 2);
2599 
2600 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2601 		fill_includes_sw_bits = true;
2602 
2603 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2604 }
2605 
2606 void build_tlb_refill_handler(void)
2607 {
2608 	/*
2609 	 * The refill handler is generated per-CPU, multi-node systems
2610 	 * may have local storage for it. The other handlers are only
2611 	 * needed once.
2612 	 */
2613 	static int run_once = 0;
2614 
2615 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2616 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2617 
2618 	output_pgtable_bits_defines();
2619 	check_pabits();
2620 
2621 #ifdef CONFIG_64BIT
2622 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2623 #endif
2624 
2625 	if (cpu_has_3kex) {
2626 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2627 		if (!run_once) {
2628 			build_setup_pgd();
2629 			build_r3000_tlb_refill_handler();
2630 			build_r3000_tlb_load_handler();
2631 			build_r3000_tlb_store_handler();
2632 			build_r3000_tlb_modify_handler();
2633 			flush_tlb_handlers();
2634 			run_once++;
2635 		}
2636 #else
2637 		panic("No R3000 TLB refill handler");
2638 #endif
2639 		return;
2640 	}
2641 
2642 	if (cpu_has_ldpte)
2643 		setup_pw();
2644 
2645 	if (!run_once) {
2646 		scratch_reg = allocate_kscratch();
2647 		build_setup_pgd();
2648 		build_r4000_tlb_load_handler();
2649 		build_r4000_tlb_store_handler();
2650 		build_r4000_tlb_modify_handler();
2651 		if (cpu_has_ldpte)
2652 			build_loongson3_tlb_refill_handler();
2653 		else
2654 			build_r4000_tlb_refill_handler();
2655 		flush_tlb_handlers();
2656 		run_once++;
2657 	}
2658 	if (cpu_has_xpa)
2659 		config_xpa_params();
2660 	if (cpu_has_htw)
2661 		config_htw_params();
2662 }
2663