xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 0c874100)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008	 Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009	 Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/mmu_context.h>
35 #include <asm/pgtable.h>
36 #include <asm/war.h>
37 #include <asm/uasm.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
40 
41 static int mips_xpa_disabled;
42 
43 static int __init xpa_disable(char *s)
44 {
45 	mips_xpa_disabled = 1;
46 
47 	return 1;
48 }
49 
50 __setup("noxpa", xpa_disable);
51 
52 /*
53  * TLB load/store/modify handlers.
54  *
55  * Only the fastpath gets synthesized at runtime, the slowpath for
56  * do_page_fault remains normal asm.
57  */
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
60 
61 struct work_registers {
62 	int r1;
63 	int r2;
64 	int r3;
65 };
66 
67 struct tlb_reg_save {
68 	unsigned long a;
69 	unsigned long b;
70 } ____cacheline_aligned_in_smp;
71 
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
73 
74 static inline int r45k_bvahwbug(void)
75 {
76 	/* XXX: We should probe for the presence of this bug, but we don't. */
77 	return 0;
78 }
79 
80 static inline int r4k_250MHZhwbug(void)
81 {
82 	/* XXX: We should probe for the presence of this bug, but we don't. */
83 	return 0;
84 }
85 
86 static inline int __maybe_unused bcm1250_m3_war(void)
87 {
88 	return BCM1250_M3_WAR;
89 }
90 
91 static inline int __maybe_unused r10000_llsc_war(void)
92 {
93 	return R10000_LLSC_WAR;
94 }
95 
96 static int use_bbit_insns(void)
97 {
98 	switch (current_cpu_type()) {
99 	case CPU_CAVIUM_OCTEON:
100 	case CPU_CAVIUM_OCTEON_PLUS:
101 	case CPU_CAVIUM_OCTEON2:
102 	case CPU_CAVIUM_OCTEON3:
103 		return 1;
104 	default:
105 		return 0;
106 	}
107 }
108 
109 static int use_lwx_insns(void)
110 {
111 	switch (current_cpu_type()) {
112 	case CPU_CAVIUM_OCTEON2:
113 	case CPU_CAVIUM_OCTEON3:
114 		return 1;
115 	default:
116 		return 0;
117 	}
118 }
119 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
120     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
121 static bool scratchpad_available(void)
122 {
123 	return true;
124 }
125 static int scratchpad_offset(int i)
126 {
127 	/*
128 	 * CVMSEG starts at address -32768 and extends for
129 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
130 	 */
131 	i += 1; /* Kernel use starts at the top and works down. */
132 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
133 }
134 #else
135 static bool scratchpad_available(void)
136 {
137 	return false;
138 }
139 static int scratchpad_offset(int i)
140 {
141 	BUG();
142 	/* Really unreachable, but evidently some GCC want this. */
143 	return 0;
144 }
145 #endif
146 /*
147  * Found by experiment: At least some revisions of the 4kc throw under
148  * some circumstances a machine check exception, triggered by invalid
149  * values in the index register.  Delaying the tlbp instruction until
150  * after the next branch,  plus adding an additional nop in front of
151  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
152  * why; it's not an issue caused by the core RTL.
153  *
154  */
155 static int m4kc_tlbp_war(void)
156 {
157 	return current_cpu_type() == CPU_4KC;
158 }
159 
160 /* Handle labels (which must be positive integers). */
161 enum label_id {
162 	label_second_part = 1,
163 	label_leave,
164 	label_vmalloc,
165 	label_vmalloc_done,
166 	label_tlbw_hazard_0,
167 	label_split = label_tlbw_hazard_0 + 8,
168 	label_tlbl_goaround1,
169 	label_tlbl_goaround2,
170 	label_nopage_tlbl,
171 	label_nopage_tlbs,
172 	label_nopage_tlbm,
173 	label_smp_pgtable_change,
174 	label_r3000_write_probe_fail,
175 	label_large_segbits_fault,
176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
177 	label_tlb_huge_update,
178 #endif
179 };
180 
181 UASM_L_LA(_second_part)
182 UASM_L_LA(_leave)
183 UASM_L_LA(_vmalloc)
184 UASM_L_LA(_vmalloc_done)
185 /* _tlbw_hazard_x is handled differently.  */
186 UASM_L_LA(_split)
187 UASM_L_LA(_tlbl_goaround1)
188 UASM_L_LA(_tlbl_goaround2)
189 UASM_L_LA(_nopage_tlbl)
190 UASM_L_LA(_nopage_tlbs)
191 UASM_L_LA(_nopage_tlbm)
192 UASM_L_LA(_smp_pgtable_change)
193 UASM_L_LA(_r3000_write_probe_fail)
194 UASM_L_LA(_large_segbits_fault)
195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
196 UASM_L_LA(_tlb_huge_update)
197 #endif
198 
199 static int hazard_instance;
200 
201 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
202 {
203 	switch (instance) {
204 	case 0 ... 7:
205 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
206 		return;
207 	default:
208 		BUG();
209 	}
210 }
211 
212 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
213 {
214 	switch (instance) {
215 	case 0 ... 7:
216 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
217 		break;
218 	default:
219 		BUG();
220 	}
221 }
222 
223 /*
224  * pgtable bits are assigned dynamically depending on processor feature
225  * and statically based on kernel configuration.  This spits out the actual
226  * values the kernel is using.	Required to make sense from disassembled
227  * TLB exception handlers.
228  */
229 static void output_pgtable_bits_defines(void)
230 {
231 #define pr_define(fmt, ...)					\
232 	pr_debug("#define " fmt, ##__VA_ARGS__)
233 
234 	pr_debug("#include <asm/asm.h>\n");
235 	pr_debug("#include <asm/regdef.h>\n");
236 	pr_debug("\n");
237 
238 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
239 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
241 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
242 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
243 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
244 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
245 #endif
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 	if (cpu_has_rixi)
248 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 #endif
250 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
251 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
252 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
253 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
254 	pr_debug("\n");
255 }
256 
257 static inline void dump_handler(const char *symbol, const void *start, const void *end)
258 {
259 	unsigned int count = (end - start) / sizeof(u32);
260 	const u32 *handler = start;
261 	int i;
262 
263 	pr_debug("LEAF(%s)\n", symbol);
264 
265 	pr_debug("\t.set push\n");
266 	pr_debug("\t.set noreorder\n");
267 
268 	for (i = 0; i < count; i++)
269 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
270 
271 	pr_debug("\t.set\tpop\n");
272 
273 	pr_debug("\tEND(%s)\n", symbol);
274 }
275 
276 /* The only general purpose registers allowed in TLB handlers. */
277 #define K0		26
278 #define K1		27
279 
280 /* Some CP0 registers */
281 #define C0_INDEX	0, 0
282 #define C0_ENTRYLO0	2, 0
283 #define C0_TCBIND	2, 2
284 #define C0_ENTRYLO1	3, 0
285 #define C0_CONTEXT	4, 0
286 #define C0_PAGEMASK	5, 0
287 #define C0_PWBASE	5, 5
288 #define C0_PWFIELD	5, 6
289 #define C0_PWSIZE	5, 7
290 #define C0_PWCTL	6, 6
291 #define C0_BADVADDR	8, 0
292 #define C0_PGD		9, 7
293 #define C0_ENTRYHI	10, 0
294 #define C0_EPC		14, 0
295 #define C0_XCONTEXT	20, 0
296 
297 #ifdef CONFIG_64BIT
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
299 #else
300 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
301 #endif
302 
303 /* The worst case length of the handler is around 18 instructions for
304  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
305  * Maximum space available is 32 instructions for R3000 and 64
306  * instructions for R4000.
307  *
308  * We deliberately chose a buffer size of 128, so we won't scribble
309  * over anything important on overflow before we panic.
310  */
311 static u32 tlb_handler[128];
312 
313 /* simply assume worst case size for labels and relocs */
314 static struct uasm_label labels[128];
315 static struct uasm_reloc relocs[128];
316 
317 static int check_for_high_segbits;
318 static bool fill_includes_sw_bits;
319 
320 static unsigned int kscratch_used_mask;
321 
322 static inline int __maybe_unused c0_kscratch(void)
323 {
324 	switch (current_cpu_type()) {
325 	case CPU_XLP:
326 	case CPU_XLR:
327 		return 22;
328 	default:
329 		return 31;
330 	}
331 }
332 
333 static int allocate_kscratch(void)
334 {
335 	int r;
336 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
337 
338 	r = ffs(a);
339 
340 	if (r == 0)
341 		return -1;
342 
343 	r--; /* make it zero based */
344 
345 	kscratch_used_mask |= (1 << r);
346 
347 	return r;
348 }
349 
350 static int scratch_reg;
351 int pgd_reg;
352 EXPORT_SYMBOL_GPL(pgd_reg);
353 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
354 
355 static struct work_registers build_get_work_registers(u32 **p)
356 {
357 	struct work_registers r;
358 
359 	if (scratch_reg >= 0) {
360 		/* Save in CPU local C0_KScratch? */
361 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
362 		r.r1 = K0;
363 		r.r2 = K1;
364 		r.r3 = 1;
365 		return r;
366 	}
367 
368 	if (num_possible_cpus() > 1) {
369 		/* Get smp_processor_id */
370 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
371 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
372 
373 		/* handler_reg_save index in K0 */
374 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
375 
376 		UASM_i_LA(p, K1, (long)&handler_reg_save);
377 		UASM_i_ADDU(p, K0, K0, K1);
378 	} else {
379 		UASM_i_LA(p, K0, (long)&handler_reg_save);
380 	}
381 	/* K0 now points to save area, save $1 and $2  */
382 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
384 
385 	r.r1 = K1;
386 	r.r2 = 1;
387 	r.r3 = 2;
388 	return r;
389 }
390 
391 static void build_restore_work_registers(u32 **p)
392 {
393 	if (scratch_reg >= 0) {
394 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
395 		return;
396 	}
397 	/* K0 already points to save area, restore $1 and $2  */
398 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
399 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
400 }
401 
402 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
403 
404 /*
405  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
406  * we cannot do r3000 under these circumstances.
407  *
408  * The R3000 TLB handler is simple.
409  */
410 static void build_r3000_tlb_refill_handler(void)
411 {
412 	long pgdc = (long)pgd_current;
413 	u32 *p;
414 
415 	memset(tlb_handler, 0, sizeof(tlb_handler));
416 	p = tlb_handler;
417 
418 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
419 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
420 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
421 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
422 	uasm_i_sll(&p, K0, K0, 2);
423 	uasm_i_addu(&p, K1, K1, K0);
424 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
425 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
426 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
427 	uasm_i_addu(&p, K1, K1, K0);
428 	uasm_i_lw(&p, K0, 0, K1);
429 	uasm_i_nop(&p); /* load delay */
430 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
431 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
432 	uasm_i_tlbwr(&p); /* cp0 delay */
433 	uasm_i_jr(&p, K1);
434 	uasm_i_rfe(&p); /* branch delay */
435 
436 	if (p > tlb_handler + 32)
437 		panic("TLB refill handler space exceeded");
438 
439 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
440 		 (unsigned int)(p - tlb_handler));
441 
442 	memcpy((void *)ebase, tlb_handler, 0x80);
443 	local_flush_icache_range(ebase, ebase + 0x80);
444 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
445 }
446 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
447 
448 /*
449  * The R4000 TLB handler is much more complicated. We have two
450  * consecutive handler areas with 32 instructions space each.
451  * Since they aren't used at the same time, we can overflow in the
452  * other one.To keep things simple, we first assume linear space,
453  * then we relocate it to the final handler layout as needed.
454  */
455 static u32 final_handler[64];
456 
457 /*
458  * Hazards
459  *
460  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
461  * 2. A timing hazard exists for the TLBP instruction.
462  *
463  *	stalling_instruction
464  *	TLBP
465  *
466  * The JTLB is being read for the TLBP throughout the stall generated by the
467  * previous instruction. This is not really correct as the stalling instruction
468  * can modify the address used to access the JTLB.  The failure symptom is that
469  * the TLBP instruction will use an address created for the stalling instruction
470  * and not the address held in C0_ENHI and thus report the wrong results.
471  *
472  * The software work-around is to not allow the instruction preceding the TLBP
473  * to stall - make it an NOP or some other instruction guaranteed not to stall.
474  *
475  * Errata 2 will not be fixed.	This errata is also on the R5000.
476  *
477  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
478  */
479 static void __maybe_unused build_tlb_probe_entry(u32 **p)
480 {
481 	switch (current_cpu_type()) {
482 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
483 	case CPU_R4600:
484 	case CPU_R4700:
485 	case CPU_R5000:
486 	case CPU_NEVADA:
487 		uasm_i_nop(p);
488 		uasm_i_tlbp(p);
489 		break;
490 
491 	default:
492 		uasm_i_tlbp(p);
493 		break;
494 	}
495 }
496 
497 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
498 			   struct uasm_reloc **r,
499 			   enum tlb_write_entry wmode)
500 {
501 	void(*tlbw)(u32 **) = NULL;
502 
503 	switch (wmode) {
504 	case tlb_random: tlbw = uasm_i_tlbwr; break;
505 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
506 	}
507 
508 	if (cpu_has_mips_r2_r6) {
509 		if (cpu_has_mips_r2_exec_hazard)
510 			uasm_i_ehb(p);
511 		tlbw(p);
512 		return;
513 	}
514 
515 	switch (current_cpu_type()) {
516 	case CPU_R4000PC:
517 	case CPU_R4000SC:
518 	case CPU_R4000MC:
519 	case CPU_R4400PC:
520 	case CPU_R4400SC:
521 	case CPU_R4400MC:
522 		/*
523 		 * This branch uses up a mtc0 hazard nop slot and saves
524 		 * two nops after the tlbw instruction.
525 		 */
526 		uasm_bgezl_hazard(p, r, hazard_instance);
527 		tlbw(p);
528 		uasm_bgezl_label(l, p, hazard_instance);
529 		hazard_instance++;
530 		uasm_i_nop(p);
531 		break;
532 
533 	case CPU_R4600:
534 	case CPU_R4700:
535 		uasm_i_nop(p);
536 		tlbw(p);
537 		uasm_i_nop(p);
538 		break;
539 
540 	case CPU_R5000:
541 	case CPU_NEVADA:
542 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
543 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
544 		tlbw(p);
545 		break;
546 
547 	case CPU_R4300:
548 	case CPU_5KC:
549 	case CPU_TX49XX:
550 	case CPU_PR4450:
551 	case CPU_XLR:
552 		uasm_i_nop(p);
553 		tlbw(p);
554 		break;
555 
556 	case CPU_R10000:
557 	case CPU_R12000:
558 	case CPU_R14000:
559 	case CPU_R16000:
560 	case CPU_4KC:
561 	case CPU_4KEC:
562 	case CPU_M14KC:
563 	case CPU_M14KEC:
564 	case CPU_SB1:
565 	case CPU_SB1A:
566 	case CPU_4KSC:
567 	case CPU_20KC:
568 	case CPU_25KF:
569 	case CPU_BMIPS32:
570 	case CPU_BMIPS3300:
571 	case CPU_BMIPS4350:
572 	case CPU_BMIPS4380:
573 	case CPU_BMIPS5000:
574 	case CPU_LOONGSON2:
575 	case CPU_LOONGSON3:
576 	case CPU_R5500:
577 		if (m4kc_tlbp_war())
578 			uasm_i_nop(p);
579 	case CPU_ALCHEMY:
580 		tlbw(p);
581 		break;
582 
583 	case CPU_RM7000:
584 		uasm_i_nop(p);
585 		uasm_i_nop(p);
586 		uasm_i_nop(p);
587 		uasm_i_nop(p);
588 		tlbw(p);
589 		break;
590 
591 	case CPU_VR4111:
592 	case CPU_VR4121:
593 	case CPU_VR4122:
594 	case CPU_VR4181:
595 	case CPU_VR4181A:
596 		uasm_i_nop(p);
597 		uasm_i_nop(p);
598 		tlbw(p);
599 		uasm_i_nop(p);
600 		uasm_i_nop(p);
601 		break;
602 
603 	case CPU_VR4131:
604 	case CPU_VR4133:
605 	case CPU_R5432:
606 		uasm_i_nop(p);
607 		uasm_i_nop(p);
608 		tlbw(p);
609 		break;
610 
611 	case CPU_JZRISC:
612 		tlbw(p);
613 		uasm_i_nop(p);
614 		break;
615 
616 	default:
617 		panic("No TLB refill handler yet (CPU type: %d)",
618 		      current_cpu_type());
619 		break;
620 	}
621 }
622 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
623 
624 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
625 							unsigned int reg)
626 {
627 	if (_PAGE_GLOBAL_SHIFT == 0) {
628 		/* pte_t is already in EntryLo format */
629 		return;
630 	}
631 
632 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
633 		if (fill_includes_sw_bits) {
634 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
635 		} else {
636 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
637 			UASM_i_ROTR(p, reg, reg,
638 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
639 		}
640 	} else {
641 #ifdef CONFIG_PHYS_ADDR_T_64BIT
642 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
643 #else
644 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 #endif
646 	}
647 }
648 
649 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
650 
651 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
652 				   unsigned int tmp, enum label_id lid,
653 				   int restore_scratch)
654 {
655 	if (restore_scratch) {
656 		/* Reset default page size */
657 		if (PM_DEFAULT_MASK >> 16) {
658 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
659 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
660 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
661 			uasm_il_b(p, r, lid);
662 		} else if (PM_DEFAULT_MASK) {
663 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
664 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
665 			uasm_il_b(p, r, lid);
666 		} else {
667 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
668 			uasm_il_b(p, r, lid);
669 		}
670 		if (scratch_reg >= 0)
671 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
672 		else
673 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
674 	} else {
675 		/* Reset default page size */
676 		if (PM_DEFAULT_MASK >> 16) {
677 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
678 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
679 			uasm_il_b(p, r, lid);
680 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
681 		} else if (PM_DEFAULT_MASK) {
682 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
683 			uasm_il_b(p, r, lid);
684 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
685 		} else {
686 			uasm_il_b(p, r, lid);
687 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
688 		}
689 	}
690 }
691 
692 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
693 				       struct uasm_reloc **r,
694 				       unsigned int tmp,
695 				       enum tlb_write_entry wmode,
696 				       int restore_scratch)
697 {
698 	/* Set huge page tlb entry size */
699 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
700 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
701 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
702 
703 	build_tlb_write_entry(p, l, r, wmode);
704 
705 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
706 }
707 
708 /*
709  * Check if Huge PTE is present, if so then jump to LABEL.
710  */
711 static void
712 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
713 		  unsigned int pmd, int lid)
714 {
715 	UASM_i_LW(p, tmp, 0, pmd);
716 	if (use_bbit_insns()) {
717 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
718 	} else {
719 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
720 		uasm_il_bnez(p, r, tmp, lid);
721 	}
722 }
723 
724 static void build_huge_update_entries(u32 **p, unsigned int pte,
725 				      unsigned int tmp)
726 {
727 	int small_sequence;
728 
729 	/*
730 	 * A huge PTE describes an area the size of the
731 	 * configured huge page size. This is twice the
732 	 * of the large TLB entry size we intend to use.
733 	 * A TLB entry half the size of the configured
734 	 * huge page size is configured into entrylo0
735 	 * and entrylo1 to cover the contiguous huge PTE
736 	 * address space.
737 	 */
738 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
739 
740 	/* We can clobber tmp.	It isn't used after this.*/
741 	if (!small_sequence)
742 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
743 
744 	build_convert_pte_to_entrylo(p, pte);
745 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
746 	/* convert to entrylo1 */
747 	if (small_sequence)
748 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
749 	else
750 		UASM_i_ADDU(p, pte, pte, tmp);
751 
752 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
753 }
754 
755 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
756 				    struct uasm_label **l,
757 				    unsigned int pte,
758 				    unsigned int ptr,
759 				    unsigned int flush)
760 {
761 #ifdef CONFIG_SMP
762 	UASM_i_SC(p, pte, 0, ptr);
763 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
764 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
765 #else
766 	UASM_i_SW(p, pte, 0, ptr);
767 #endif
768 	if (cpu_has_ftlb && flush) {
769 		BUG_ON(!cpu_has_tlbinv);
770 
771 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
772 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
773 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
774 		build_tlb_write_entry(p, l, r, tlb_indexed);
775 
776 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
777 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
778 		build_huge_update_entries(p, pte, ptr);
779 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
780 
781 		return;
782 	}
783 
784 	build_huge_update_entries(p, pte, ptr);
785 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
786 }
787 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
788 
789 #ifdef CONFIG_64BIT
790 /*
791  * TMP and PTR are scratch.
792  * TMP will be clobbered, PTR will hold the pmd entry.
793  */
794 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
795 		      unsigned int tmp, unsigned int ptr)
796 {
797 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
798 	long pgdc = (long)pgd_current;
799 #endif
800 	/*
801 	 * The vmalloc handling is not in the hotpath.
802 	 */
803 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
804 
805 	if (check_for_high_segbits) {
806 		/*
807 		 * The kernel currently implicitely assumes that the
808 		 * MIPS SEGBITS parameter for the processor is
809 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
810 		 * allocate virtual addresses outside the maximum
811 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
812 		 * that doesn't prevent user code from accessing the
813 		 * higher xuseg addresses.  Here, we make sure that
814 		 * everything but the lower xuseg addresses goes down
815 		 * the module_alloc/vmalloc path.
816 		 */
817 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
818 		uasm_il_bnez(p, r, ptr, label_vmalloc);
819 	} else {
820 		uasm_il_bltz(p, r, tmp, label_vmalloc);
821 	}
822 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
823 
824 	if (pgd_reg != -1) {
825 		/* pgd is in pgd_reg */
826 		if (cpu_has_ldpte)
827 			UASM_i_MFC0(p, ptr, C0_PWBASE);
828 		else
829 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
830 	} else {
831 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
832 		/*
833 		 * &pgd << 11 stored in CONTEXT [23..63].
834 		 */
835 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
836 
837 		/* Clear lower 23 bits of context. */
838 		uasm_i_dins(p, ptr, 0, 0, 23);
839 
840 		/* 1 0	1 0 1  << 6  xkphys cached */
841 		uasm_i_ori(p, ptr, ptr, 0x540);
842 		uasm_i_drotr(p, ptr, ptr, 11);
843 #elif defined(CONFIG_SMP)
844 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
845 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
846 		UASM_i_LA_mostly(p, tmp, pgdc);
847 		uasm_i_daddu(p, ptr, ptr, tmp);
848 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
849 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
850 #else
851 		UASM_i_LA_mostly(p, ptr, pgdc);
852 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
853 #endif
854 	}
855 
856 	uasm_l_vmalloc_done(l, *p);
857 
858 	/* get pgd offset in bytes */
859 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
860 
861 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
862 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
863 #ifndef __PAGETABLE_PUD_FOLDED
864 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
865 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
866 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
867 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
868 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
869 #endif
870 #ifndef __PAGETABLE_PMD_FOLDED
871 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
872 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
873 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
874 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
875 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
876 #endif
877 }
878 EXPORT_SYMBOL_GPL(build_get_pmde64);
879 
880 /*
881  * BVADDR is the faulting address, PTR is scratch.
882  * PTR will hold the pgd for vmalloc.
883  */
884 static void
885 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
886 			unsigned int bvaddr, unsigned int ptr,
887 			enum vmalloc64_mode mode)
888 {
889 	long swpd = (long)swapper_pg_dir;
890 	int single_insn_swpd;
891 	int did_vmalloc_branch = 0;
892 
893 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
894 
895 	uasm_l_vmalloc(l, *p);
896 
897 	if (mode != not_refill && check_for_high_segbits) {
898 		if (single_insn_swpd) {
899 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
900 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
901 			did_vmalloc_branch = 1;
902 			/* fall through */
903 		} else {
904 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
905 		}
906 	}
907 	if (!did_vmalloc_branch) {
908 		if (single_insn_swpd) {
909 			uasm_il_b(p, r, label_vmalloc_done);
910 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
911 		} else {
912 			UASM_i_LA_mostly(p, ptr, swpd);
913 			uasm_il_b(p, r, label_vmalloc_done);
914 			if (uasm_in_compat_space_p(swpd))
915 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
916 			else
917 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
918 		}
919 	}
920 	if (mode != not_refill && check_for_high_segbits) {
921 		uasm_l_large_segbits_fault(l, *p);
922 		/*
923 		 * We get here if we are an xsseg address, or if we are
924 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
925 		 *
926 		 * Ignoring xsseg (assume disabled so would generate
927 		 * (address errors?), the only remaining possibility
928 		 * is the upper xuseg addresses.  On processors with
929 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
930 		 * addresses would have taken an address error. We try
931 		 * to mimic that here by taking a load/istream page
932 		 * fault.
933 		 */
934 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
935 		uasm_i_jr(p, ptr);
936 
937 		if (mode == refill_scratch) {
938 			if (scratch_reg >= 0)
939 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
940 			else
941 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
942 		} else {
943 			uasm_i_nop(p);
944 		}
945 	}
946 }
947 
948 #else /* !CONFIG_64BIT */
949 
950 /*
951  * TMP and PTR are scratch.
952  * TMP will be clobbered, PTR will hold the pgd entry.
953  */
954 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
955 {
956 	if (pgd_reg != -1) {
957 		/* pgd is in pgd_reg */
958 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
959 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
960 	} else {
961 		long pgdc = (long)pgd_current;
962 
963 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
964 #ifdef CONFIG_SMP
965 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
966 		UASM_i_LA_mostly(p, tmp, pgdc);
967 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
968 		uasm_i_addu(p, ptr, tmp, ptr);
969 #else
970 		UASM_i_LA_mostly(p, ptr, pgdc);
971 #endif
972 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
973 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
974 	}
975 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
976 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
977 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
978 }
979 EXPORT_SYMBOL_GPL(build_get_pgde32);
980 
981 #endif /* !CONFIG_64BIT */
982 
983 static void build_adjust_context(u32 **p, unsigned int ctx)
984 {
985 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
986 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
987 
988 	switch (current_cpu_type()) {
989 	case CPU_VR41XX:
990 	case CPU_VR4111:
991 	case CPU_VR4121:
992 	case CPU_VR4122:
993 	case CPU_VR4131:
994 	case CPU_VR4181:
995 	case CPU_VR4181A:
996 	case CPU_VR4133:
997 		shift += 2;
998 		break;
999 
1000 	default:
1001 		break;
1002 	}
1003 
1004 	if (shift)
1005 		UASM_i_SRL(p, ctx, ctx, shift);
1006 	uasm_i_andi(p, ctx, ctx, mask);
1007 }
1008 
1009 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1010 {
1011 	/*
1012 	 * Bug workaround for the Nevada. It seems as if under certain
1013 	 * circumstances the move from cp0_context might produce a
1014 	 * bogus result when the mfc0 instruction and its consumer are
1015 	 * in a different cacheline or a load instruction, probably any
1016 	 * memory reference, is between them.
1017 	 */
1018 	switch (current_cpu_type()) {
1019 	case CPU_NEVADA:
1020 		UASM_i_LW(p, ptr, 0, ptr);
1021 		GET_CONTEXT(p, tmp); /* get context reg */
1022 		break;
1023 
1024 	default:
1025 		GET_CONTEXT(p, tmp); /* get context reg */
1026 		UASM_i_LW(p, ptr, 0, ptr);
1027 		break;
1028 	}
1029 
1030 	build_adjust_context(p, tmp);
1031 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1032 }
1033 EXPORT_SYMBOL_GPL(build_get_ptep);
1034 
1035 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1036 {
1037 	int pte_off_even = 0;
1038 	int pte_off_odd = sizeof(pte_t);
1039 
1040 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1041 	/* The low 32 bits of EntryLo is stored in pte_high */
1042 	pte_off_even += offsetof(pte_t, pte_high);
1043 	pte_off_odd += offsetof(pte_t, pte_high);
1044 #endif
1045 
1046 	if (IS_ENABLED(CONFIG_XPA)) {
1047 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1048 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1049 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1050 
1051 		if (cpu_has_xpa && !mips_xpa_disabled) {
1052 			uasm_i_lw(p, tmp, 0, ptep);
1053 			uasm_i_ext(p, tmp, tmp, 0, 24);
1054 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1055 		}
1056 
1057 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1058 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1059 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1060 
1061 		if (cpu_has_xpa && !mips_xpa_disabled) {
1062 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1063 			uasm_i_ext(p, tmp, tmp, 0, 24);
1064 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1065 		}
1066 		return;
1067 	}
1068 
1069 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1070 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1071 	if (r45k_bvahwbug())
1072 		build_tlb_probe_entry(p);
1073 	build_convert_pte_to_entrylo(p, tmp);
1074 	if (r4k_250MHZhwbug())
1075 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1076 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1077 	build_convert_pte_to_entrylo(p, ptep);
1078 	if (r45k_bvahwbug())
1079 		uasm_i_mfc0(p, tmp, C0_INDEX);
1080 	if (r4k_250MHZhwbug())
1081 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1082 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1083 }
1084 EXPORT_SYMBOL_GPL(build_update_entries);
1085 
1086 struct mips_huge_tlb_info {
1087 	int huge_pte;
1088 	int restore_scratch;
1089 	bool need_reload_pte;
1090 };
1091 
1092 static struct mips_huge_tlb_info
1093 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1094 			       struct uasm_reloc **r, unsigned int tmp,
1095 			       unsigned int ptr, int c0_scratch_reg)
1096 {
1097 	struct mips_huge_tlb_info rv;
1098 	unsigned int even, odd;
1099 	int vmalloc_branch_delay_filled = 0;
1100 	const int scratch = 1; /* Our extra working register */
1101 
1102 	rv.huge_pte = scratch;
1103 	rv.restore_scratch = 0;
1104 	rv.need_reload_pte = false;
1105 
1106 	if (check_for_high_segbits) {
1107 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1108 
1109 		if (pgd_reg != -1)
1110 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1111 		else
1112 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1113 
1114 		if (c0_scratch_reg >= 0)
1115 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1116 		else
1117 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1118 
1119 		uasm_i_dsrl_safe(p, scratch, tmp,
1120 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1121 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1122 
1123 		if (pgd_reg == -1) {
1124 			vmalloc_branch_delay_filled = 1;
1125 			/* Clear lower 23 bits of context. */
1126 			uasm_i_dins(p, ptr, 0, 0, 23);
1127 		}
1128 	} else {
1129 		if (pgd_reg != -1)
1130 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1131 		else
1132 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1133 
1134 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1135 
1136 		if (c0_scratch_reg >= 0)
1137 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1138 		else
1139 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1140 
1141 		if (pgd_reg == -1)
1142 			/* Clear lower 23 bits of context. */
1143 			uasm_i_dins(p, ptr, 0, 0, 23);
1144 
1145 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1146 	}
1147 
1148 	if (pgd_reg == -1) {
1149 		vmalloc_branch_delay_filled = 1;
1150 		/* 1 0	1 0 1  << 6  xkphys cached */
1151 		uasm_i_ori(p, ptr, ptr, 0x540);
1152 		uasm_i_drotr(p, ptr, ptr, 11);
1153 	}
1154 
1155 #ifdef __PAGETABLE_PMD_FOLDED
1156 #define LOC_PTEP scratch
1157 #else
1158 #define LOC_PTEP ptr
1159 #endif
1160 
1161 	if (!vmalloc_branch_delay_filled)
1162 		/* get pgd offset in bytes */
1163 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1164 
1165 	uasm_l_vmalloc_done(l, *p);
1166 
1167 	/*
1168 	 *			   tmp		ptr
1169 	 * fall-through case =	 badvaddr  *pgd_current
1170 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1171 	 */
1172 
1173 	if (vmalloc_branch_delay_filled)
1174 		/* get pgd offset in bytes */
1175 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1176 
1177 #ifdef __PAGETABLE_PMD_FOLDED
1178 	GET_CONTEXT(p, tmp); /* get context reg */
1179 #endif
1180 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1181 
1182 	if (use_lwx_insns()) {
1183 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1184 	} else {
1185 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1186 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1187 	}
1188 
1189 #ifndef __PAGETABLE_PUD_FOLDED
1190 	/* get pud offset in bytes */
1191 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1192 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1193 
1194 	if (use_lwx_insns()) {
1195 		UASM_i_LWX(p, ptr, scratch, ptr);
1196 	} else {
1197 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1198 		UASM_i_LW(p, ptr, 0, ptr);
1199 	}
1200 	/* ptr contains a pointer to PMD entry */
1201 	/* tmp contains the address */
1202 #endif
1203 
1204 #ifndef __PAGETABLE_PMD_FOLDED
1205 	/* get pmd offset in bytes */
1206 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1207 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1208 	GET_CONTEXT(p, tmp); /* get context reg */
1209 
1210 	if (use_lwx_insns()) {
1211 		UASM_i_LWX(p, scratch, scratch, ptr);
1212 	} else {
1213 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1214 		UASM_i_LW(p, scratch, 0, ptr);
1215 	}
1216 #endif
1217 	/* Adjust the context during the load latency. */
1218 	build_adjust_context(p, tmp);
1219 
1220 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1221 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1222 	/*
1223 	 * The in the LWX case we don't want to do the load in the
1224 	 * delay slot.	It cannot issue in the same cycle and may be
1225 	 * speculative and unneeded.
1226 	 */
1227 	if (use_lwx_insns())
1228 		uasm_i_nop(p);
1229 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1230 
1231 
1232 	/* build_update_entries */
1233 	if (use_lwx_insns()) {
1234 		even = ptr;
1235 		odd = tmp;
1236 		UASM_i_LWX(p, even, scratch, tmp);
1237 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1238 		UASM_i_LWX(p, odd, scratch, tmp);
1239 	} else {
1240 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1241 		even = tmp;
1242 		odd = ptr;
1243 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1244 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1245 	}
1246 	if (cpu_has_rixi) {
1247 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1248 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1249 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1250 	} else {
1251 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1252 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1253 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1254 	}
1255 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1256 
1257 	if (c0_scratch_reg >= 0) {
1258 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1259 		build_tlb_write_entry(p, l, r, tlb_random);
1260 		uasm_l_leave(l, *p);
1261 		rv.restore_scratch = 1;
1262 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1263 		build_tlb_write_entry(p, l, r, tlb_random);
1264 		uasm_l_leave(l, *p);
1265 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1266 	} else {
1267 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1268 		build_tlb_write_entry(p, l, r, tlb_random);
1269 		uasm_l_leave(l, *p);
1270 		rv.restore_scratch = 1;
1271 	}
1272 
1273 	uasm_i_eret(p); /* return from trap */
1274 
1275 	return rv;
1276 }
1277 
1278 /*
1279  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1280  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1281  * slots before the XTLB refill exception handler which belong to the
1282  * unused TLB refill exception.
1283  */
1284 #define MIPS64_REFILL_INSNS 32
1285 
1286 static void build_r4000_tlb_refill_handler(void)
1287 {
1288 	u32 *p = tlb_handler;
1289 	struct uasm_label *l = labels;
1290 	struct uasm_reloc *r = relocs;
1291 	u32 *f;
1292 	unsigned int final_len;
1293 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1294 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1295 
1296 	memset(tlb_handler, 0, sizeof(tlb_handler));
1297 	memset(labels, 0, sizeof(labels));
1298 	memset(relocs, 0, sizeof(relocs));
1299 	memset(final_handler, 0, sizeof(final_handler));
1300 
1301 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1302 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1303 							  scratch_reg);
1304 		vmalloc_mode = refill_scratch;
1305 	} else {
1306 		htlb_info.huge_pte = K0;
1307 		htlb_info.restore_scratch = 0;
1308 		htlb_info.need_reload_pte = true;
1309 		vmalloc_mode = refill_noscratch;
1310 		/*
1311 		 * create the plain linear handler
1312 		 */
1313 		if (bcm1250_m3_war()) {
1314 			unsigned int segbits = 44;
1315 
1316 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1317 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1318 			uasm_i_xor(&p, K0, K0, K1);
1319 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1320 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1321 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1322 			uasm_i_or(&p, K0, K0, K1);
1323 			uasm_il_bnez(&p, &r, K0, label_leave);
1324 			/* No need for uasm_i_nop */
1325 		}
1326 
1327 #ifdef CONFIG_64BIT
1328 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1329 #else
1330 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1331 #endif
1332 
1333 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1334 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1335 #endif
1336 
1337 		build_get_ptep(&p, K0, K1);
1338 		build_update_entries(&p, K0, K1);
1339 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1340 		uasm_l_leave(&l, p);
1341 		uasm_i_eret(&p); /* return from trap */
1342 	}
1343 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1344 	uasm_l_tlb_huge_update(&l, p);
1345 	if (htlb_info.need_reload_pte)
1346 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1347 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1348 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1349 				   htlb_info.restore_scratch);
1350 #endif
1351 
1352 #ifdef CONFIG_64BIT
1353 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1354 #endif
1355 
1356 	/*
1357 	 * Overflow check: For the 64bit handler, we need at least one
1358 	 * free instruction slot for the wrap-around branch. In worst
1359 	 * case, if the intended insertion point is a delay slot, we
1360 	 * need three, with the second nop'ed and the third being
1361 	 * unused.
1362 	 */
1363 	switch (boot_cpu_type()) {
1364 	default:
1365 		if (sizeof(long) == 4) {
1366 	case CPU_LOONGSON2:
1367 		/* Loongson2 ebase is different than r4k, we have more space */
1368 			if ((p - tlb_handler) > 64)
1369 				panic("TLB refill handler space exceeded");
1370 			/*
1371 			 * Now fold the handler in the TLB refill handler space.
1372 			 */
1373 			f = final_handler;
1374 			/* Simplest case, just copy the handler. */
1375 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1376 			final_len = p - tlb_handler;
1377 			break;
1378 		} else {
1379 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1380 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1381 				&& uasm_insn_has_bdelay(relocs,
1382 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1383 				panic("TLB refill handler space exceeded");
1384 			/*
1385 			 * Now fold the handler in the TLB refill handler space.
1386 			 */
1387 			f = final_handler + MIPS64_REFILL_INSNS;
1388 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1389 				/* Just copy the handler. */
1390 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1391 				final_len = p - tlb_handler;
1392 			} else {
1393 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1394 				const enum label_id ls = label_tlb_huge_update;
1395 #else
1396 				const enum label_id ls = label_vmalloc;
1397 #endif
1398 				u32 *split;
1399 				int ov = 0;
1400 				int i;
1401 
1402 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1403 					;
1404 				BUG_ON(i == ARRAY_SIZE(labels));
1405 				split = labels[i].addr;
1406 
1407 				/*
1408 				 * See if we have overflown one way or the other.
1409 				 */
1410 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1411 				    split < p - MIPS64_REFILL_INSNS)
1412 					ov = 1;
1413 
1414 				if (ov) {
1415 					/*
1416 					 * Split two instructions before the end.  One
1417 					 * for the branch and one for the instruction
1418 					 * in the delay slot.
1419 					 */
1420 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1421 
1422 					/*
1423 					 * If the branch would fall in a delay slot,
1424 					 * we must back up an additional instruction
1425 					 * so that it is no longer in a delay slot.
1426 					 */
1427 					if (uasm_insn_has_bdelay(relocs, split - 1))
1428 						split--;
1429 				}
1430 				/* Copy first part of the handler. */
1431 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1432 				f += split - tlb_handler;
1433 
1434 				if (ov) {
1435 					/* Insert branch. */
1436 					uasm_l_split(&l, final_handler);
1437 					uasm_il_b(&f, &r, label_split);
1438 					if (uasm_insn_has_bdelay(relocs, split))
1439 						uasm_i_nop(&f);
1440 					else {
1441 						uasm_copy_handler(relocs, labels,
1442 								  split, split + 1, f);
1443 						uasm_move_labels(labels, f, f + 1, -1);
1444 						f++;
1445 						split++;
1446 					}
1447 				}
1448 
1449 				/* Copy the rest of the handler. */
1450 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1451 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1452 					    (p - split);
1453 			}
1454 		}
1455 		break;
1456 	}
1457 
1458 	uasm_resolve_relocs(relocs, labels);
1459 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1460 		 final_len);
1461 
1462 	memcpy((void *)ebase, final_handler, 0x100);
1463 	local_flush_icache_range(ebase, ebase + 0x100);
1464 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1465 }
1466 
1467 static void setup_pw(void)
1468 {
1469 	unsigned long pgd_i, pgd_w;
1470 #ifndef __PAGETABLE_PMD_FOLDED
1471 	unsigned long pmd_i, pmd_w;
1472 #endif
1473 	unsigned long pt_i, pt_w;
1474 	unsigned long pte_i, pte_w;
1475 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1476 	unsigned long psn;
1477 
1478 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1479 #endif
1480 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1481 #ifndef __PAGETABLE_PMD_FOLDED
1482 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1483 
1484 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1485 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1486 #else
1487 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1488 #endif
1489 
1490 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1491 	pt_w  = PAGE_SHIFT - 3;
1492 
1493 	pte_i = ilog2(_PAGE_GLOBAL);
1494 	pte_w = 0;
1495 
1496 #ifndef __PAGETABLE_PMD_FOLDED
1497 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1498 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1499 #else
1500 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1501 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1502 #endif
1503 
1504 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1505 	write_c0_pwctl(1 << 6 | psn);
1506 #endif
1507 	write_c0_kpgd((long)swapper_pg_dir);
1508 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1509 }
1510 
1511 static void build_loongson3_tlb_refill_handler(void)
1512 {
1513 	u32 *p = tlb_handler;
1514 	struct uasm_label *l = labels;
1515 	struct uasm_reloc *r = relocs;
1516 
1517 	memset(labels, 0, sizeof(labels));
1518 	memset(relocs, 0, sizeof(relocs));
1519 	memset(tlb_handler, 0, sizeof(tlb_handler));
1520 
1521 	if (check_for_high_segbits) {
1522 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1523 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1524 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1525 		uasm_i_nop(&p);
1526 
1527 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1528 		uasm_i_nop(&p);
1529 		uasm_l_vmalloc(&l, p);
1530 	}
1531 
1532 	uasm_i_dmfc0(&p, K1, C0_PGD);
1533 
1534 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1535 #ifndef __PAGETABLE_PMD_FOLDED
1536 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1537 #endif
1538 	uasm_i_ldpte(&p, K1, 0);      /* even */
1539 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1540 	uasm_i_tlbwr(&p);
1541 
1542 	/* restore page mask */
1543 	if (PM_DEFAULT_MASK >> 16) {
1544 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1545 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1546 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1547 	} else if (PM_DEFAULT_MASK) {
1548 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1549 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1550 	} else {
1551 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1552 	}
1553 
1554 	uasm_i_eret(&p);
1555 
1556 	if (check_for_high_segbits) {
1557 		uasm_l_large_segbits_fault(&l, p);
1558 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1559 		uasm_i_jr(&p, K1);
1560 		uasm_i_nop(&p);
1561 	}
1562 
1563 	uasm_resolve_relocs(relocs, labels);
1564 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1565 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1566 	dump_handler("loongson3_tlb_refill",
1567 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1568 }
1569 
1570 static void build_setup_pgd(void)
1571 {
1572 	const int a0 = 4;
1573 	const int __maybe_unused a1 = 5;
1574 	const int __maybe_unused a2 = 6;
1575 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1576 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1577 	long pgdc = (long)pgd_current;
1578 #endif
1579 
1580 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1581 	memset(labels, 0, sizeof(labels));
1582 	memset(relocs, 0, sizeof(relocs));
1583 	pgd_reg = allocate_kscratch();
1584 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1585 	if (pgd_reg == -1) {
1586 		struct uasm_label *l = labels;
1587 		struct uasm_reloc *r = relocs;
1588 
1589 		/* PGD << 11 in c0_Context */
1590 		/*
1591 		 * If it is a ckseg0 address, convert to a physical
1592 		 * address.  Shifting right by 29 and adding 4 will
1593 		 * result in zero for these addresses.
1594 		 *
1595 		 */
1596 		UASM_i_SRA(&p, a1, a0, 29);
1597 		UASM_i_ADDIU(&p, a1, a1, 4);
1598 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1599 		uasm_i_nop(&p);
1600 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1601 		uasm_l_tlbl_goaround1(&l, p);
1602 		UASM_i_SLL(&p, a0, a0, 11);
1603 		uasm_i_jr(&p, 31);
1604 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1605 	} else {
1606 		/* PGD in c0_KScratch */
1607 		uasm_i_jr(&p, 31);
1608 		if (cpu_has_ldpte)
1609 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1610 		else
1611 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1612 	}
1613 #else
1614 #ifdef CONFIG_SMP
1615 	/* Save PGD to pgd_current[smp_processor_id()] */
1616 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1617 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1618 	UASM_i_LA_mostly(&p, a2, pgdc);
1619 	UASM_i_ADDU(&p, a2, a2, a1);
1620 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1621 #else
1622 	UASM_i_LA_mostly(&p, a2, pgdc);
1623 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1624 #endif /* SMP */
1625 	uasm_i_jr(&p, 31);
1626 
1627 	/* if pgd_reg is allocated, save PGD also to scratch register */
1628 	if (pgd_reg != -1)
1629 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1630 	else
1631 		uasm_i_nop(&p);
1632 #endif
1633 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1634 		panic("tlbmiss_handler_setup_pgd space exceeded");
1635 
1636 	uasm_resolve_relocs(relocs, labels);
1637 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1638 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1639 
1640 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1641 					tlbmiss_handler_setup_pgd_end);
1642 }
1643 
1644 static void
1645 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1646 {
1647 #ifdef CONFIG_SMP
1648 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1649 	if (cpu_has_64bits)
1650 		uasm_i_lld(p, pte, 0, ptr);
1651 	else
1652 # endif
1653 		UASM_i_LL(p, pte, 0, ptr);
1654 #else
1655 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1656 	if (cpu_has_64bits)
1657 		uasm_i_ld(p, pte, 0, ptr);
1658 	else
1659 # endif
1660 		UASM_i_LW(p, pte, 0, ptr);
1661 #endif
1662 }
1663 
1664 static void
1665 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1666 	unsigned int mode, unsigned int scratch)
1667 {
1668 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1669 	unsigned int swmode = mode & ~hwmode;
1670 
1671 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1672 		uasm_i_lui(p, scratch, swmode >> 16);
1673 		uasm_i_or(p, pte, pte, scratch);
1674 		BUG_ON(swmode & 0xffff);
1675 	} else {
1676 		uasm_i_ori(p, pte, pte, mode);
1677 	}
1678 
1679 #ifdef CONFIG_SMP
1680 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1681 	if (cpu_has_64bits)
1682 		uasm_i_scd(p, pte, 0, ptr);
1683 	else
1684 # endif
1685 		UASM_i_SC(p, pte, 0, ptr);
1686 
1687 	if (r10000_llsc_war())
1688 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1689 	else
1690 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1691 
1692 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1693 	if (!cpu_has_64bits) {
1694 		/* no uasm_i_nop needed */
1695 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1696 		uasm_i_ori(p, pte, pte, hwmode);
1697 		BUG_ON(hwmode & ~0xffff);
1698 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1699 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1700 		/* no uasm_i_nop needed */
1701 		uasm_i_lw(p, pte, 0, ptr);
1702 	} else
1703 		uasm_i_nop(p);
1704 # else
1705 	uasm_i_nop(p);
1706 # endif
1707 #else
1708 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1709 	if (cpu_has_64bits)
1710 		uasm_i_sd(p, pte, 0, ptr);
1711 	else
1712 # endif
1713 		UASM_i_SW(p, pte, 0, ptr);
1714 
1715 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1716 	if (!cpu_has_64bits) {
1717 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1718 		uasm_i_ori(p, pte, pte, hwmode);
1719 		BUG_ON(hwmode & ~0xffff);
1720 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1721 		uasm_i_lw(p, pte, 0, ptr);
1722 	}
1723 # endif
1724 #endif
1725 }
1726 
1727 /*
1728  * Check if PTE is present, if not then jump to LABEL. PTR points to
1729  * the page table where this PTE is located, PTE will be re-loaded
1730  * with it's original value.
1731  */
1732 static void
1733 build_pte_present(u32 **p, struct uasm_reloc **r,
1734 		  int pte, int ptr, int scratch, enum label_id lid)
1735 {
1736 	int t = scratch >= 0 ? scratch : pte;
1737 	int cur = pte;
1738 
1739 	if (cpu_has_rixi) {
1740 		if (use_bbit_insns()) {
1741 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1742 			uasm_i_nop(p);
1743 		} else {
1744 			if (_PAGE_PRESENT_SHIFT) {
1745 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1746 				cur = t;
1747 			}
1748 			uasm_i_andi(p, t, cur, 1);
1749 			uasm_il_beqz(p, r, t, lid);
1750 			if (pte == t)
1751 				/* You lose the SMP race :-(*/
1752 				iPTE_LW(p, pte, ptr);
1753 		}
1754 	} else {
1755 		if (_PAGE_PRESENT_SHIFT) {
1756 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1757 			cur = t;
1758 		}
1759 		uasm_i_andi(p, t, cur,
1760 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1761 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1762 		uasm_il_bnez(p, r, t, lid);
1763 		if (pte == t)
1764 			/* You lose the SMP race :-(*/
1765 			iPTE_LW(p, pte, ptr);
1766 	}
1767 }
1768 
1769 /* Make PTE valid, store result in PTR. */
1770 static void
1771 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1772 		 unsigned int ptr, unsigned int scratch)
1773 {
1774 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1775 
1776 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1777 }
1778 
1779 /*
1780  * Check if PTE can be written to, if not branch to LABEL. Regardless
1781  * restore PTE with value from PTR when done.
1782  */
1783 static void
1784 build_pte_writable(u32 **p, struct uasm_reloc **r,
1785 		   unsigned int pte, unsigned int ptr, int scratch,
1786 		   enum label_id lid)
1787 {
1788 	int t = scratch >= 0 ? scratch : pte;
1789 	int cur = pte;
1790 
1791 	if (_PAGE_PRESENT_SHIFT) {
1792 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1793 		cur = t;
1794 	}
1795 	uasm_i_andi(p, t, cur,
1796 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1797 	uasm_i_xori(p, t, t,
1798 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1799 	uasm_il_bnez(p, r, t, lid);
1800 	if (pte == t)
1801 		/* You lose the SMP race :-(*/
1802 		iPTE_LW(p, pte, ptr);
1803 	else
1804 		uasm_i_nop(p);
1805 }
1806 
1807 /* Make PTE writable, update software status bits as well, then store
1808  * at PTR.
1809  */
1810 static void
1811 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1812 		 unsigned int ptr, unsigned int scratch)
1813 {
1814 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1815 			     | _PAGE_DIRTY);
1816 
1817 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1818 }
1819 
1820 /*
1821  * Check if PTE can be modified, if not branch to LABEL. Regardless
1822  * restore PTE with value from PTR when done.
1823  */
1824 static void
1825 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1826 		     unsigned int pte, unsigned int ptr, int scratch,
1827 		     enum label_id lid)
1828 {
1829 	if (use_bbit_insns()) {
1830 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1831 		uasm_i_nop(p);
1832 	} else {
1833 		int t = scratch >= 0 ? scratch : pte;
1834 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1835 		uasm_i_andi(p, t, t, 1);
1836 		uasm_il_beqz(p, r, t, lid);
1837 		if (pte == t)
1838 			/* You lose the SMP race :-(*/
1839 			iPTE_LW(p, pte, ptr);
1840 	}
1841 }
1842 
1843 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1844 
1845 
1846 /*
1847  * R3000 style TLB load/store/modify handlers.
1848  */
1849 
1850 /*
1851  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1852  * Then it returns.
1853  */
1854 static void
1855 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1856 {
1857 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1858 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1859 	uasm_i_tlbwi(p);
1860 	uasm_i_jr(p, tmp);
1861 	uasm_i_rfe(p); /* branch delay */
1862 }
1863 
1864 /*
1865  * This places the pte into ENTRYLO0 and writes it with tlbwi
1866  * or tlbwr as appropriate.  This is because the index register
1867  * may have the probe fail bit set as a result of a trap on a
1868  * kseg2 access, i.e. without refill.  Then it returns.
1869  */
1870 static void
1871 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1872 			     struct uasm_reloc **r, unsigned int pte,
1873 			     unsigned int tmp)
1874 {
1875 	uasm_i_mfc0(p, tmp, C0_INDEX);
1876 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1877 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1878 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1879 	uasm_i_tlbwi(p); /* cp0 delay */
1880 	uasm_i_jr(p, tmp);
1881 	uasm_i_rfe(p); /* branch delay */
1882 	uasm_l_r3000_write_probe_fail(l, *p);
1883 	uasm_i_tlbwr(p); /* cp0 delay */
1884 	uasm_i_jr(p, tmp);
1885 	uasm_i_rfe(p); /* branch delay */
1886 }
1887 
1888 static void
1889 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1890 				   unsigned int ptr)
1891 {
1892 	long pgdc = (long)pgd_current;
1893 
1894 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1895 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1896 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1897 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1898 	uasm_i_sll(p, pte, pte, 2);
1899 	uasm_i_addu(p, ptr, ptr, pte);
1900 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1901 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1902 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1903 	uasm_i_addu(p, ptr, ptr, pte);
1904 	uasm_i_lw(p, pte, 0, ptr);
1905 	uasm_i_tlbp(p); /* load delay */
1906 }
1907 
1908 static void build_r3000_tlb_load_handler(void)
1909 {
1910 	u32 *p = (u32 *)handle_tlbl;
1911 	struct uasm_label *l = labels;
1912 	struct uasm_reloc *r = relocs;
1913 
1914 	memset(p, 0, handle_tlbl_end - (char *)p);
1915 	memset(labels, 0, sizeof(labels));
1916 	memset(relocs, 0, sizeof(relocs));
1917 
1918 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1919 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1920 	uasm_i_nop(&p); /* load delay */
1921 	build_make_valid(&p, &r, K0, K1, -1);
1922 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1923 
1924 	uasm_l_nopage_tlbl(&l, p);
1925 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1926 	uasm_i_nop(&p);
1927 
1928 	if (p >= (u32 *)handle_tlbl_end)
1929 		panic("TLB load handler fastpath space exceeded");
1930 
1931 	uasm_resolve_relocs(relocs, labels);
1932 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1933 		 (unsigned int)(p - (u32 *)handle_tlbl));
1934 
1935 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1936 }
1937 
1938 static void build_r3000_tlb_store_handler(void)
1939 {
1940 	u32 *p = (u32 *)handle_tlbs;
1941 	struct uasm_label *l = labels;
1942 	struct uasm_reloc *r = relocs;
1943 
1944 	memset(p, 0, handle_tlbs_end - (char *)p);
1945 	memset(labels, 0, sizeof(labels));
1946 	memset(relocs, 0, sizeof(relocs));
1947 
1948 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1949 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1950 	uasm_i_nop(&p); /* load delay */
1951 	build_make_write(&p, &r, K0, K1, -1);
1952 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1953 
1954 	uasm_l_nopage_tlbs(&l, p);
1955 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1956 	uasm_i_nop(&p);
1957 
1958 	if (p >= (u32 *)handle_tlbs_end)
1959 		panic("TLB store handler fastpath space exceeded");
1960 
1961 	uasm_resolve_relocs(relocs, labels);
1962 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1963 		 (unsigned int)(p - (u32 *)handle_tlbs));
1964 
1965 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1966 }
1967 
1968 static void build_r3000_tlb_modify_handler(void)
1969 {
1970 	u32 *p = (u32 *)handle_tlbm;
1971 	struct uasm_label *l = labels;
1972 	struct uasm_reloc *r = relocs;
1973 
1974 	memset(p, 0, handle_tlbm_end - (char *)p);
1975 	memset(labels, 0, sizeof(labels));
1976 	memset(relocs, 0, sizeof(relocs));
1977 
1978 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1979 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1980 	uasm_i_nop(&p); /* load delay */
1981 	build_make_write(&p, &r, K0, K1, -1);
1982 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1983 
1984 	uasm_l_nopage_tlbm(&l, p);
1985 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1986 	uasm_i_nop(&p);
1987 
1988 	if (p >= (u32 *)handle_tlbm_end)
1989 		panic("TLB modify handler fastpath space exceeded");
1990 
1991 	uasm_resolve_relocs(relocs, labels);
1992 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1993 		 (unsigned int)(p - (u32 *)handle_tlbm));
1994 
1995 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
1996 }
1997 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1998 
1999 static bool cpu_has_tlbex_tlbp_race(void)
2000 {
2001 	/*
2002 	 * When a Hardware Table Walker is running it can replace TLB entries
2003 	 * at any time, leading to a race between it & the CPU.
2004 	 */
2005 	if (cpu_has_htw)
2006 		return true;
2007 
2008 	/*
2009 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2010 	 * replaced at any time by a sibling performing a write to the FTLB.
2011 	 */
2012 	if (cpu_has_shared_ftlb_ram)
2013 		return true;
2014 
2015 	/* In all other cases there ought to be no race condition to handle */
2016 	return false;
2017 }
2018 
2019 /*
2020  * R4000 style TLB load/store/modify handlers.
2021  */
2022 static struct work_registers
2023 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2024 				   struct uasm_reloc **r)
2025 {
2026 	struct work_registers wr = build_get_work_registers(p);
2027 
2028 #ifdef CONFIG_64BIT
2029 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2030 #else
2031 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2032 #endif
2033 
2034 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2035 	/*
2036 	 * For huge tlb entries, pmd doesn't contain an address but
2037 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2038 	 * see if we need to jump to huge tlb processing.
2039 	 */
2040 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2041 #endif
2042 
2043 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2044 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2045 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2046 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2047 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2048 
2049 #ifdef CONFIG_SMP
2050 	uasm_l_smp_pgtable_change(l, *p);
2051 #endif
2052 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2053 	if (!m4kc_tlbp_war()) {
2054 		build_tlb_probe_entry(p);
2055 		if (cpu_has_tlbex_tlbp_race()) {
2056 			/* race condition happens, leaving */
2057 			uasm_i_ehb(p);
2058 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2059 			uasm_il_bltz(p, r, wr.r3, label_leave);
2060 			uasm_i_nop(p);
2061 		}
2062 	}
2063 	return wr;
2064 }
2065 
2066 static void
2067 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2068 				   struct uasm_reloc **r, unsigned int tmp,
2069 				   unsigned int ptr)
2070 {
2071 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2072 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2073 	build_update_entries(p, tmp, ptr);
2074 	build_tlb_write_entry(p, l, r, tlb_indexed);
2075 	uasm_l_leave(l, *p);
2076 	build_restore_work_registers(p);
2077 	uasm_i_eret(p); /* return from trap */
2078 
2079 #ifdef CONFIG_64BIT
2080 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2081 #endif
2082 }
2083 
2084 static void build_r4000_tlb_load_handler(void)
2085 {
2086 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2087 	struct uasm_label *l = labels;
2088 	struct uasm_reloc *r = relocs;
2089 	struct work_registers wr;
2090 
2091 	memset(p, 0, handle_tlbl_end - (char *)p);
2092 	memset(labels, 0, sizeof(labels));
2093 	memset(relocs, 0, sizeof(relocs));
2094 
2095 	if (bcm1250_m3_war()) {
2096 		unsigned int segbits = 44;
2097 
2098 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2099 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2100 		uasm_i_xor(&p, K0, K0, K1);
2101 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2102 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2103 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2104 		uasm_i_or(&p, K0, K0, K1);
2105 		uasm_il_bnez(&p, &r, K0, label_leave);
2106 		/* No need for uasm_i_nop */
2107 	}
2108 
2109 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2110 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2111 	if (m4kc_tlbp_war())
2112 		build_tlb_probe_entry(&p);
2113 
2114 	if (cpu_has_rixi && !cpu_has_rixiex) {
2115 		/*
2116 		 * If the page is not _PAGE_VALID, RI or XI could not
2117 		 * have triggered it.  Skip the expensive test..
2118 		 */
2119 		if (use_bbit_insns()) {
2120 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2121 				      label_tlbl_goaround1);
2122 		} else {
2123 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2124 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2125 		}
2126 		uasm_i_nop(&p);
2127 
2128 		/*
2129 		 * Warn if something may race with us & replace the TLB entry
2130 		 * before we read it here. Everything with such races should
2131 		 * also have dedicated RiXi exception handlers, so this
2132 		 * shouldn't be hit.
2133 		 */
2134 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2135 
2136 		uasm_i_tlbr(&p);
2137 
2138 		switch (current_cpu_type()) {
2139 		default:
2140 			if (cpu_has_mips_r2_exec_hazard) {
2141 				uasm_i_ehb(&p);
2142 
2143 		case CPU_CAVIUM_OCTEON:
2144 		case CPU_CAVIUM_OCTEON_PLUS:
2145 		case CPU_CAVIUM_OCTEON2:
2146 				break;
2147 			}
2148 		}
2149 
2150 		/* Examine  entrylo 0 or 1 based on ptr. */
2151 		if (use_bbit_insns()) {
2152 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2153 		} else {
2154 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2155 			uasm_i_beqz(&p, wr.r3, 8);
2156 		}
2157 		/* load it in the delay slot*/
2158 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2159 		/* load it if ptr is odd */
2160 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2161 		/*
2162 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2163 		 * XI must have triggered it.
2164 		 */
2165 		if (use_bbit_insns()) {
2166 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2167 			uasm_i_nop(&p);
2168 			uasm_l_tlbl_goaround1(&l, p);
2169 		} else {
2170 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2171 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2172 			uasm_i_nop(&p);
2173 		}
2174 		uasm_l_tlbl_goaround1(&l, p);
2175 	}
2176 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2177 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2178 
2179 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2180 	/*
2181 	 * This is the entry point when build_r4000_tlbchange_handler_head
2182 	 * spots a huge page.
2183 	 */
2184 	uasm_l_tlb_huge_update(&l, p);
2185 	iPTE_LW(&p, wr.r1, wr.r2);
2186 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2187 	build_tlb_probe_entry(&p);
2188 
2189 	if (cpu_has_rixi && !cpu_has_rixiex) {
2190 		/*
2191 		 * If the page is not _PAGE_VALID, RI or XI could not
2192 		 * have triggered it.  Skip the expensive test..
2193 		 */
2194 		if (use_bbit_insns()) {
2195 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2196 				      label_tlbl_goaround2);
2197 		} else {
2198 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2199 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2200 		}
2201 		uasm_i_nop(&p);
2202 
2203 		/*
2204 		 * Warn if something may race with us & replace the TLB entry
2205 		 * before we read it here. Everything with such races should
2206 		 * also have dedicated RiXi exception handlers, so this
2207 		 * shouldn't be hit.
2208 		 */
2209 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2210 
2211 		uasm_i_tlbr(&p);
2212 
2213 		switch (current_cpu_type()) {
2214 		default:
2215 			if (cpu_has_mips_r2_exec_hazard) {
2216 				uasm_i_ehb(&p);
2217 
2218 		case CPU_CAVIUM_OCTEON:
2219 		case CPU_CAVIUM_OCTEON_PLUS:
2220 		case CPU_CAVIUM_OCTEON2:
2221 				break;
2222 			}
2223 		}
2224 
2225 		/* Examine  entrylo 0 or 1 based on ptr. */
2226 		if (use_bbit_insns()) {
2227 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2228 		} else {
2229 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2230 			uasm_i_beqz(&p, wr.r3, 8);
2231 		}
2232 		/* load it in the delay slot*/
2233 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2234 		/* load it if ptr is odd */
2235 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2236 		/*
2237 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2238 		 * XI must have triggered it.
2239 		 */
2240 		if (use_bbit_insns()) {
2241 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2242 		} else {
2243 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2244 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2245 		}
2246 		if (PM_DEFAULT_MASK == 0)
2247 			uasm_i_nop(&p);
2248 		/*
2249 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2250 		 * it is restored in build_huge_tlb_write_entry.
2251 		 */
2252 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2253 
2254 		uasm_l_tlbl_goaround2(&l, p);
2255 	}
2256 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2257 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2258 #endif
2259 
2260 	uasm_l_nopage_tlbl(&l, p);
2261 	build_restore_work_registers(&p);
2262 #ifdef CONFIG_CPU_MICROMIPS
2263 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2264 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2265 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2266 		uasm_i_jr(&p, K0);
2267 	} else
2268 #endif
2269 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2270 	uasm_i_nop(&p);
2271 
2272 	if (p >= (u32 *)handle_tlbl_end)
2273 		panic("TLB load handler fastpath space exceeded");
2274 
2275 	uasm_resolve_relocs(relocs, labels);
2276 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2277 		 (unsigned int)(p - (u32 *)handle_tlbl));
2278 
2279 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2280 }
2281 
2282 static void build_r4000_tlb_store_handler(void)
2283 {
2284 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2285 	struct uasm_label *l = labels;
2286 	struct uasm_reloc *r = relocs;
2287 	struct work_registers wr;
2288 
2289 	memset(p, 0, handle_tlbs_end - (char *)p);
2290 	memset(labels, 0, sizeof(labels));
2291 	memset(relocs, 0, sizeof(relocs));
2292 
2293 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2294 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2295 	if (m4kc_tlbp_war())
2296 		build_tlb_probe_entry(&p);
2297 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2298 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2299 
2300 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2301 	/*
2302 	 * This is the entry point when
2303 	 * build_r4000_tlbchange_handler_head spots a huge page.
2304 	 */
2305 	uasm_l_tlb_huge_update(&l, p);
2306 	iPTE_LW(&p, wr.r1, wr.r2);
2307 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2308 	build_tlb_probe_entry(&p);
2309 	uasm_i_ori(&p, wr.r1, wr.r1,
2310 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2311 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2312 #endif
2313 
2314 	uasm_l_nopage_tlbs(&l, p);
2315 	build_restore_work_registers(&p);
2316 #ifdef CONFIG_CPU_MICROMIPS
2317 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2318 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2319 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2320 		uasm_i_jr(&p, K0);
2321 	} else
2322 #endif
2323 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2324 	uasm_i_nop(&p);
2325 
2326 	if (p >= (u32 *)handle_tlbs_end)
2327 		panic("TLB store handler fastpath space exceeded");
2328 
2329 	uasm_resolve_relocs(relocs, labels);
2330 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2331 		 (unsigned int)(p - (u32 *)handle_tlbs));
2332 
2333 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2334 }
2335 
2336 static void build_r4000_tlb_modify_handler(void)
2337 {
2338 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2339 	struct uasm_label *l = labels;
2340 	struct uasm_reloc *r = relocs;
2341 	struct work_registers wr;
2342 
2343 	memset(p, 0, handle_tlbm_end - (char *)p);
2344 	memset(labels, 0, sizeof(labels));
2345 	memset(relocs, 0, sizeof(relocs));
2346 
2347 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2348 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2349 	if (m4kc_tlbp_war())
2350 		build_tlb_probe_entry(&p);
2351 	/* Present and writable bits set, set accessed and dirty bits. */
2352 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2353 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2354 
2355 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2356 	/*
2357 	 * This is the entry point when
2358 	 * build_r4000_tlbchange_handler_head spots a huge page.
2359 	 */
2360 	uasm_l_tlb_huge_update(&l, p);
2361 	iPTE_LW(&p, wr.r1, wr.r2);
2362 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2363 	build_tlb_probe_entry(&p);
2364 	uasm_i_ori(&p, wr.r1, wr.r1,
2365 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2366 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2367 #endif
2368 
2369 	uasm_l_nopage_tlbm(&l, p);
2370 	build_restore_work_registers(&p);
2371 #ifdef CONFIG_CPU_MICROMIPS
2372 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2373 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2374 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2375 		uasm_i_jr(&p, K0);
2376 	} else
2377 #endif
2378 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2379 	uasm_i_nop(&p);
2380 
2381 	if (p >= (u32 *)handle_tlbm_end)
2382 		panic("TLB modify handler fastpath space exceeded");
2383 
2384 	uasm_resolve_relocs(relocs, labels);
2385 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2386 		 (unsigned int)(p - (u32 *)handle_tlbm));
2387 
2388 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2389 }
2390 
2391 static void flush_tlb_handlers(void)
2392 {
2393 	local_flush_icache_range((unsigned long)handle_tlbl,
2394 			   (unsigned long)handle_tlbl_end);
2395 	local_flush_icache_range((unsigned long)handle_tlbs,
2396 			   (unsigned long)handle_tlbs_end);
2397 	local_flush_icache_range((unsigned long)handle_tlbm,
2398 			   (unsigned long)handle_tlbm_end);
2399 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2400 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2401 }
2402 
2403 static void print_htw_config(void)
2404 {
2405 	unsigned long config;
2406 	unsigned int pwctl;
2407 	const int field = 2 * sizeof(unsigned long);
2408 
2409 	config = read_c0_pwfield();
2410 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2411 		field, config,
2412 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2413 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2414 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2415 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2416 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2417 
2418 	config = read_c0_pwsize();
2419 	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",
2420 		field, config,
2421 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2422 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2423 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2424 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2425 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2426 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2427 
2428 	pwctl = read_c0_pwctl();
2429 	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",
2430 		pwctl,
2431 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2432 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2433 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2434 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2435 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2436 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2437 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2438 }
2439 
2440 static void config_htw_params(void)
2441 {
2442 	unsigned long pwfield, pwsize, ptei;
2443 	unsigned int config;
2444 
2445 	/*
2446 	 * We are using 2-level page tables, so we only need to
2447 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2448 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2449 	 * write values less than 0xc in these fields because the entire
2450 	 * write will be dropped. As a result of which, we must preserve
2451 	 * the original reset values and overwrite only what we really want.
2452 	 */
2453 
2454 	pwfield = read_c0_pwfield();
2455 	/* re-initialize the GDI field */
2456 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2457 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2458 	/* re-initialize the PTI field including the even/odd bit */
2459 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2460 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2461 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2462 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2463 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2464 	}
2465 	/* Set the PTEI right shift */
2466 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2467 	pwfield |= ptei;
2468 	write_c0_pwfield(pwfield);
2469 	/* Check whether the PTEI value is supported */
2470 	back_to_back_c0_hazard();
2471 	pwfield = read_c0_pwfield();
2472 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2473 		!= ptei) {
2474 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2475 			ptei);
2476 		/*
2477 		 * Drop option to avoid HTW being enabled via another path
2478 		 * (eg htw_reset())
2479 		 */
2480 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2481 		return;
2482 	}
2483 
2484 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2485 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2486 	if (CONFIG_PGTABLE_LEVELS >= 3)
2487 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2488 
2489 	/* Set pointer size to size of directory pointers */
2490 	if (IS_ENABLED(CONFIG_64BIT))
2491 		pwsize |= MIPS_PWSIZE_PS_MASK;
2492 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2493 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2494 			& MIPS_PWSIZE_PTEW_MASK;
2495 
2496 	write_c0_pwsize(pwsize);
2497 
2498 	/* Make sure everything is set before we enable the HTW */
2499 	back_to_back_c0_hazard();
2500 
2501 	/*
2502 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2503 	 * the pwctl fields.
2504 	 */
2505 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2506 	if (IS_ENABLED(CONFIG_64BIT))
2507 		config |= MIPS_PWCTL_XU_MASK;
2508 	write_c0_pwctl(config);
2509 	pr_info("Hardware Page Table Walker enabled\n");
2510 
2511 	print_htw_config();
2512 }
2513 
2514 static void config_xpa_params(void)
2515 {
2516 #ifdef CONFIG_XPA
2517 	unsigned int pagegrain;
2518 
2519 	if (mips_xpa_disabled) {
2520 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2521 		return;
2522 	}
2523 
2524 	pagegrain = read_c0_pagegrain();
2525 	write_c0_pagegrain(pagegrain | PG_ELPA);
2526 	back_to_back_c0_hazard();
2527 	pagegrain = read_c0_pagegrain();
2528 
2529 	if (pagegrain & PG_ELPA)
2530 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2531 	else
2532 		panic("Extended Physical Addressing (XPA) disabled");
2533 #endif
2534 }
2535 
2536 static void check_pabits(void)
2537 {
2538 	unsigned long entry;
2539 	unsigned pabits, fillbits;
2540 
2541 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2542 		/*
2543 		 * We'll only be making use of the fact that we can rotate bits
2544 		 * into the fill if the CPU supports RIXI, so don't bother
2545 		 * probing this for CPUs which don't.
2546 		 */
2547 		return;
2548 	}
2549 
2550 	write_c0_entrylo0(~0ul);
2551 	back_to_back_c0_hazard();
2552 	entry = read_c0_entrylo0();
2553 
2554 	/* clear all non-PFN bits */
2555 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2556 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2557 
2558 	/* find a lower bound on PABITS, and upper bound on fill bits */
2559 	pabits = fls_long(entry) + 6;
2560 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2561 
2562 	/* minus the RI & XI bits */
2563 	fillbits -= min_t(unsigned, fillbits, 2);
2564 
2565 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2566 		fill_includes_sw_bits = true;
2567 
2568 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2569 }
2570 
2571 void build_tlb_refill_handler(void)
2572 {
2573 	/*
2574 	 * The refill handler is generated per-CPU, multi-node systems
2575 	 * may have local storage for it. The other handlers are only
2576 	 * needed once.
2577 	 */
2578 	static int run_once = 0;
2579 
2580 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2581 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2582 
2583 	output_pgtable_bits_defines();
2584 	check_pabits();
2585 
2586 #ifdef CONFIG_64BIT
2587 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2588 #endif
2589 
2590 	switch (current_cpu_type()) {
2591 	case CPU_R2000:
2592 	case CPU_R3000:
2593 	case CPU_R3000A:
2594 	case CPU_R3081E:
2595 	case CPU_TX3912:
2596 	case CPU_TX3922:
2597 	case CPU_TX3927:
2598 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2599 		if (cpu_has_local_ebase)
2600 			build_r3000_tlb_refill_handler();
2601 		if (!run_once) {
2602 			if (!cpu_has_local_ebase)
2603 				build_r3000_tlb_refill_handler();
2604 			build_setup_pgd();
2605 			build_r3000_tlb_load_handler();
2606 			build_r3000_tlb_store_handler();
2607 			build_r3000_tlb_modify_handler();
2608 			flush_tlb_handlers();
2609 			run_once++;
2610 		}
2611 #else
2612 		panic("No R3000 TLB refill handler");
2613 #endif
2614 		break;
2615 
2616 	case CPU_R8000:
2617 		panic("No R8000 TLB refill handler yet");
2618 		break;
2619 
2620 	default:
2621 		if (cpu_has_ldpte)
2622 			setup_pw();
2623 
2624 		if (!run_once) {
2625 			scratch_reg = allocate_kscratch();
2626 			build_setup_pgd();
2627 			build_r4000_tlb_load_handler();
2628 			build_r4000_tlb_store_handler();
2629 			build_r4000_tlb_modify_handler();
2630 			if (cpu_has_ldpte)
2631 				build_loongson3_tlb_refill_handler();
2632 			else if (!cpu_has_local_ebase)
2633 				build_r4000_tlb_refill_handler();
2634 			flush_tlb_handlers();
2635 			run_once++;
2636 		}
2637 		if (cpu_has_local_ebase)
2638 			build_r4000_tlb_refill_handler();
2639 		if (cpu_has_xpa)
2640 			config_xpa_params();
2641 		if (cpu_has_htw)
2642 			config_htw_params();
2643 	}
2644 }
2645