xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 41e4b7dc)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008	 Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009	 Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/pgtable.h>
35 #include <asm/war.h>
36 #include <asm/uasm.h>
37 #include <asm/setup.h>
38 #include <asm/tlbex.h>
39 
40 static int mips_xpa_disabled;
41 
42 static int __init xpa_disable(char *s)
43 {
44 	mips_xpa_disabled = 1;
45 
46 	return 1;
47 }
48 
49 __setup("noxpa", xpa_disable);
50 
51 /*
52  * TLB load/store/modify handlers.
53  *
54  * Only the fastpath gets synthesized at runtime, the slowpath for
55  * do_page_fault remains normal asm.
56  */
57 extern void tlb_do_page_fault_0(void);
58 extern void tlb_do_page_fault_1(void);
59 
60 struct work_registers {
61 	int r1;
62 	int r2;
63 	int r3;
64 };
65 
66 struct tlb_reg_save {
67 	unsigned long a;
68 	unsigned long b;
69 } ____cacheline_aligned_in_smp;
70 
71 static struct tlb_reg_save handler_reg_save[NR_CPUS];
72 
73 static inline int r45k_bvahwbug(void)
74 {
75 	/* XXX: We should probe for the presence of this bug, but we don't. */
76 	return 0;
77 }
78 
79 static inline int r4k_250MHZhwbug(void)
80 {
81 	/* XXX: We should probe for the presence of this bug, but we don't. */
82 	return 0;
83 }
84 
85 static inline int __maybe_unused bcm1250_m3_war(void)
86 {
87 	return BCM1250_M3_WAR;
88 }
89 
90 static inline int __maybe_unused r10000_llsc_war(void)
91 {
92 	return R10000_LLSC_WAR;
93 }
94 
95 static int use_bbit_insns(void)
96 {
97 	switch (current_cpu_type()) {
98 	case CPU_CAVIUM_OCTEON:
99 	case CPU_CAVIUM_OCTEON_PLUS:
100 	case CPU_CAVIUM_OCTEON2:
101 	case CPU_CAVIUM_OCTEON3:
102 		return 1;
103 	default:
104 		return 0;
105 	}
106 }
107 
108 static int use_lwx_insns(void)
109 {
110 	switch (current_cpu_type()) {
111 	case CPU_CAVIUM_OCTEON2:
112 	case CPU_CAVIUM_OCTEON3:
113 		return 1;
114 	default:
115 		return 0;
116 	}
117 }
118 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
119     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
120 static bool scratchpad_available(void)
121 {
122 	return true;
123 }
124 static int scratchpad_offset(int i)
125 {
126 	/*
127 	 * CVMSEG starts at address -32768 and extends for
128 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
129 	 */
130 	i += 1; /* Kernel use starts at the top and works down. */
131 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
132 }
133 #else
134 static bool scratchpad_available(void)
135 {
136 	return false;
137 }
138 static int scratchpad_offset(int i)
139 {
140 	BUG();
141 	/* Really unreachable, but evidently some GCC want this. */
142 	return 0;
143 }
144 #endif
145 /*
146  * Found by experiment: At least some revisions of the 4kc throw under
147  * some circumstances a machine check exception, triggered by invalid
148  * values in the index register.  Delaying the tlbp instruction until
149  * after the next branch,  plus adding an additional nop in front of
150  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
151  * why; it's not an issue caused by the core RTL.
152  *
153  */
154 static int m4kc_tlbp_war(void)
155 {
156 	return current_cpu_type() == CPU_4KC;
157 }
158 
159 /* Handle labels (which must be positive integers). */
160 enum label_id {
161 	label_second_part = 1,
162 	label_leave,
163 	label_vmalloc,
164 	label_vmalloc_done,
165 	label_tlbw_hazard_0,
166 	label_split = label_tlbw_hazard_0 + 8,
167 	label_tlbl_goaround1,
168 	label_tlbl_goaround2,
169 	label_nopage_tlbl,
170 	label_nopage_tlbs,
171 	label_nopage_tlbm,
172 	label_smp_pgtable_change,
173 	label_r3000_write_probe_fail,
174 	label_large_segbits_fault,
175 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
176 	label_tlb_huge_update,
177 #endif
178 };
179 
180 UASM_L_LA(_second_part)
181 UASM_L_LA(_leave)
182 UASM_L_LA(_vmalloc)
183 UASM_L_LA(_vmalloc_done)
184 /* _tlbw_hazard_x is handled differently.  */
185 UASM_L_LA(_split)
186 UASM_L_LA(_tlbl_goaround1)
187 UASM_L_LA(_tlbl_goaround2)
188 UASM_L_LA(_nopage_tlbl)
189 UASM_L_LA(_nopage_tlbs)
190 UASM_L_LA(_nopage_tlbm)
191 UASM_L_LA(_smp_pgtable_change)
192 UASM_L_LA(_r3000_write_probe_fail)
193 UASM_L_LA(_large_segbits_fault)
194 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
195 UASM_L_LA(_tlb_huge_update)
196 #endif
197 
198 static int hazard_instance;
199 
200 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
201 {
202 	switch (instance) {
203 	case 0 ... 7:
204 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
205 		return;
206 	default:
207 		BUG();
208 	}
209 }
210 
211 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
212 {
213 	switch (instance) {
214 	case 0 ... 7:
215 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
216 		break;
217 	default:
218 		BUG();
219 	}
220 }
221 
222 /*
223  * pgtable bits are assigned dynamically depending on processor feature
224  * and statically based on kernel configuration.  This spits out the actual
225  * values the kernel is using.	Required to make sense from disassembled
226  * TLB exception handlers.
227  */
228 static void output_pgtable_bits_defines(void)
229 {
230 #define pr_define(fmt, ...)					\
231 	pr_debug("#define " fmt, ##__VA_ARGS__)
232 
233 	pr_debug("#include <asm/asm.h>\n");
234 	pr_debug("#include <asm/regdef.h>\n");
235 	pr_debug("\n");
236 
237 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
238 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
239 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
240 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
241 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
242 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
243 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
244 #endif
245 #ifdef _PAGE_NO_EXEC_SHIFT
246 	if (cpu_has_rixi)
247 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
248 #endif
249 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
250 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
251 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
252 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
253 	pr_debug("\n");
254 }
255 
256 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
257 {
258 	int i;
259 
260 	pr_debug("LEAF(%s)\n", symbol);
261 
262 	pr_debug("\t.set push\n");
263 	pr_debug("\t.set noreorder\n");
264 
265 	for (i = 0; i < count; i++)
266 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
267 
268 	pr_debug("\t.set\tpop\n");
269 
270 	pr_debug("\tEND(%s)\n", symbol);
271 }
272 
273 /* The only general purpose registers allowed in TLB handlers. */
274 #define K0		26
275 #define K1		27
276 
277 /* Some CP0 registers */
278 #define C0_INDEX	0, 0
279 #define C0_ENTRYLO0	2, 0
280 #define C0_TCBIND	2, 2
281 #define C0_ENTRYLO1	3, 0
282 #define C0_CONTEXT	4, 0
283 #define C0_PAGEMASK	5, 0
284 #define C0_PWBASE	5, 5
285 #define C0_PWFIELD	5, 6
286 #define C0_PWSIZE	5, 7
287 #define C0_PWCTL	6, 6
288 #define C0_BADVADDR	8, 0
289 #define C0_PGD		9, 7
290 #define C0_ENTRYHI	10, 0
291 #define C0_EPC		14, 0
292 #define C0_XCONTEXT	20, 0
293 
294 #ifdef CONFIG_64BIT
295 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
296 #else
297 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
298 #endif
299 
300 /* The worst case length of the handler is around 18 instructions for
301  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
302  * Maximum space available is 32 instructions for R3000 and 64
303  * instructions for R4000.
304  *
305  * We deliberately chose a buffer size of 128, so we won't scribble
306  * over anything important on overflow before we panic.
307  */
308 static u32 tlb_handler[128];
309 
310 /* simply assume worst case size for labels and relocs */
311 static struct uasm_label labels[128];
312 static struct uasm_reloc relocs[128];
313 
314 static int check_for_high_segbits;
315 static bool fill_includes_sw_bits;
316 
317 static unsigned int kscratch_used_mask;
318 
319 static inline int __maybe_unused c0_kscratch(void)
320 {
321 	switch (current_cpu_type()) {
322 	case CPU_XLP:
323 	case CPU_XLR:
324 		return 22;
325 	default:
326 		return 31;
327 	}
328 }
329 
330 static int allocate_kscratch(void)
331 {
332 	int r;
333 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
334 
335 	r = ffs(a);
336 
337 	if (r == 0)
338 		return -1;
339 
340 	r--; /* make it zero based */
341 
342 	kscratch_used_mask |= (1 << r);
343 
344 	return r;
345 }
346 
347 static int scratch_reg;
348 int pgd_reg;
349 EXPORT_SYMBOL_GPL(pgd_reg);
350 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
351 
352 static struct work_registers build_get_work_registers(u32 **p)
353 {
354 	struct work_registers r;
355 
356 	if (scratch_reg >= 0) {
357 		/* Save in CPU local C0_KScratch? */
358 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
359 		r.r1 = K0;
360 		r.r2 = K1;
361 		r.r3 = 1;
362 		return r;
363 	}
364 
365 	if (num_possible_cpus() > 1) {
366 		/* Get smp_processor_id */
367 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
368 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
369 
370 		/* handler_reg_save index in K0 */
371 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
372 
373 		UASM_i_LA(p, K1, (long)&handler_reg_save);
374 		UASM_i_ADDU(p, K0, K0, K1);
375 	} else {
376 		UASM_i_LA(p, K0, (long)&handler_reg_save);
377 	}
378 	/* K0 now points to save area, save $1 and $2  */
379 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
380 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
381 
382 	r.r1 = K1;
383 	r.r2 = 1;
384 	r.r3 = 2;
385 	return r;
386 }
387 
388 static void build_restore_work_registers(u32 **p)
389 {
390 	if (scratch_reg >= 0) {
391 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
392 		return;
393 	}
394 	/* K0 already points to save area, restore $1 and $2  */
395 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
396 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
397 }
398 
399 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
400 
401 /*
402  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
403  * we cannot do r3000 under these circumstances.
404  *
405  * Declare pgd_current here instead of including mmu_context.h to avoid type
406  * conflicts for tlbmiss_handler_setup_pgd
407  */
408 extern unsigned long pgd_current[];
409 
410 /*
411  * The R3000 TLB handler is simple.
412  */
413 static void build_r3000_tlb_refill_handler(void)
414 {
415 	long pgdc = (long)pgd_current;
416 	u32 *p;
417 
418 	memset(tlb_handler, 0, sizeof(tlb_handler));
419 	p = tlb_handler;
420 
421 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
422 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
423 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
424 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
425 	uasm_i_sll(&p, K0, K0, 2);
426 	uasm_i_addu(&p, K1, K1, K0);
427 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
428 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
429 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
430 	uasm_i_addu(&p, K1, K1, K0);
431 	uasm_i_lw(&p, K0, 0, K1);
432 	uasm_i_nop(&p); /* load delay */
433 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
434 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
435 	uasm_i_tlbwr(&p); /* cp0 delay */
436 	uasm_i_jr(&p, K1);
437 	uasm_i_rfe(&p); /* branch delay */
438 
439 	if (p > tlb_handler + 32)
440 		panic("TLB refill handler space exceeded");
441 
442 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
443 		 (unsigned int)(p - tlb_handler));
444 
445 	memcpy((void *)ebase, tlb_handler, 0x80);
446 	local_flush_icache_range(ebase, ebase + 0x80);
447 
448 	dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
449 }
450 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
451 
452 /*
453  * The R4000 TLB handler is much more complicated. We have two
454  * consecutive handler areas with 32 instructions space each.
455  * Since they aren't used at the same time, we can overflow in the
456  * other one.To keep things simple, we first assume linear space,
457  * then we relocate it to the final handler layout as needed.
458  */
459 static u32 final_handler[64];
460 
461 /*
462  * Hazards
463  *
464  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
465  * 2. A timing hazard exists for the TLBP instruction.
466  *
467  *	stalling_instruction
468  *	TLBP
469  *
470  * The JTLB is being read for the TLBP throughout the stall generated by the
471  * previous instruction. This is not really correct as the stalling instruction
472  * can modify the address used to access the JTLB.  The failure symptom is that
473  * the TLBP instruction will use an address created for the stalling instruction
474  * and not the address held in C0_ENHI and thus report the wrong results.
475  *
476  * The software work-around is to not allow the instruction preceding the TLBP
477  * to stall - make it an NOP or some other instruction guaranteed not to stall.
478  *
479  * Errata 2 will not be fixed.	This errata is also on the R5000.
480  *
481  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
482  */
483 static void __maybe_unused build_tlb_probe_entry(u32 **p)
484 {
485 	switch (current_cpu_type()) {
486 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
487 	case CPU_R4600:
488 	case CPU_R4700:
489 	case CPU_R5000:
490 	case CPU_NEVADA:
491 		uasm_i_nop(p);
492 		uasm_i_tlbp(p);
493 		break;
494 
495 	default:
496 		uasm_i_tlbp(p);
497 		break;
498 	}
499 }
500 
501 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
502 			   struct uasm_reloc **r,
503 			   enum tlb_write_entry wmode)
504 {
505 	void(*tlbw)(u32 **) = NULL;
506 
507 	switch (wmode) {
508 	case tlb_random: tlbw = uasm_i_tlbwr; break;
509 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
510 	}
511 
512 	if (cpu_has_mips_r2_r6) {
513 		if (cpu_has_mips_r2_exec_hazard)
514 			uasm_i_ehb(p);
515 		tlbw(p);
516 		return;
517 	}
518 
519 	switch (current_cpu_type()) {
520 	case CPU_R4000PC:
521 	case CPU_R4000SC:
522 	case CPU_R4000MC:
523 	case CPU_R4400PC:
524 	case CPU_R4400SC:
525 	case CPU_R4400MC:
526 		/*
527 		 * This branch uses up a mtc0 hazard nop slot and saves
528 		 * two nops after the tlbw instruction.
529 		 */
530 		uasm_bgezl_hazard(p, r, hazard_instance);
531 		tlbw(p);
532 		uasm_bgezl_label(l, p, hazard_instance);
533 		hazard_instance++;
534 		uasm_i_nop(p);
535 		break;
536 
537 	case CPU_R4600:
538 	case CPU_R4700:
539 		uasm_i_nop(p);
540 		tlbw(p);
541 		uasm_i_nop(p);
542 		break;
543 
544 	case CPU_R5000:
545 	case CPU_NEVADA:
546 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
547 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
548 		tlbw(p);
549 		break;
550 
551 	case CPU_R4300:
552 	case CPU_5KC:
553 	case CPU_TX49XX:
554 	case CPU_PR4450:
555 	case CPU_XLR:
556 		uasm_i_nop(p);
557 		tlbw(p);
558 		break;
559 
560 	case CPU_R10000:
561 	case CPU_R12000:
562 	case CPU_R14000:
563 	case CPU_R16000:
564 	case CPU_4KC:
565 	case CPU_4KEC:
566 	case CPU_M14KC:
567 	case CPU_M14KEC:
568 	case CPU_SB1:
569 	case CPU_SB1A:
570 	case CPU_4KSC:
571 	case CPU_20KC:
572 	case CPU_25KF:
573 	case CPU_BMIPS32:
574 	case CPU_BMIPS3300:
575 	case CPU_BMIPS4350:
576 	case CPU_BMIPS4380:
577 	case CPU_BMIPS5000:
578 	case CPU_LOONGSON2:
579 	case CPU_LOONGSON3:
580 	case CPU_R5500:
581 		if (m4kc_tlbp_war())
582 			uasm_i_nop(p);
583 	case CPU_ALCHEMY:
584 		tlbw(p);
585 		break;
586 
587 	case CPU_RM7000:
588 		uasm_i_nop(p);
589 		uasm_i_nop(p);
590 		uasm_i_nop(p);
591 		uasm_i_nop(p);
592 		tlbw(p);
593 		break;
594 
595 	case CPU_VR4111:
596 	case CPU_VR4121:
597 	case CPU_VR4122:
598 	case CPU_VR4181:
599 	case CPU_VR4181A:
600 		uasm_i_nop(p);
601 		uasm_i_nop(p);
602 		tlbw(p);
603 		uasm_i_nop(p);
604 		uasm_i_nop(p);
605 		break;
606 
607 	case CPU_VR4131:
608 	case CPU_VR4133:
609 	case CPU_R5432:
610 		uasm_i_nop(p);
611 		uasm_i_nop(p);
612 		tlbw(p);
613 		break;
614 
615 	case CPU_JZRISC:
616 		tlbw(p);
617 		uasm_i_nop(p);
618 		break;
619 
620 	default:
621 		panic("No TLB refill handler yet (CPU type: %d)",
622 		      current_cpu_type());
623 		break;
624 	}
625 }
626 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
627 
628 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
629 							unsigned int reg)
630 {
631 	if (_PAGE_GLOBAL_SHIFT == 0) {
632 		/* pte_t is already in EntryLo format */
633 		return;
634 	}
635 
636 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
637 		if (fill_includes_sw_bits) {
638 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
639 		} else {
640 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
641 			UASM_i_ROTR(p, reg, reg,
642 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
643 		}
644 	} else {
645 #ifdef CONFIG_PHYS_ADDR_T_64BIT
646 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
647 #else
648 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
649 #endif
650 	}
651 }
652 
653 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
654 
655 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
656 				   unsigned int tmp, enum label_id lid,
657 				   int restore_scratch)
658 {
659 	if (restore_scratch) {
660 		/* Reset default page size */
661 		if (PM_DEFAULT_MASK >> 16) {
662 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
663 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
664 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
665 			uasm_il_b(p, r, lid);
666 		} else if (PM_DEFAULT_MASK) {
667 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
668 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
669 			uasm_il_b(p, r, lid);
670 		} else {
671 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
672 			uasm_il_b(p, r, lid);
673 		}
674 		if (scratch_reg >= 0)
675 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
676 		else
677 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
678 	} else {
679 		/* Reset default page size */
680 		if (PM_DEFAULT_MASK >> 16) {
681 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
682 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
683 			uasm_il_b(p, r, lid);
684 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
685 		} else if (PM_DEFAULT_MASK) {
686 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
687 			uasm_il_b(p, r, lid);
688 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
689 		} else {
690 			uasm_il_b(p, r, lid);
691 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
692 		}
693 	}
694 }
695 
696 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
697 				       struct uasm_reloc **r,
698 				       unsigned int tmp,
699 				       enum tlb_write_entry wmode,
700 				       int restore_scratch)
701 {
702 	/* Set huge page tlb entry size */
703 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
704 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
705 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
706 
707 	build_tlb_write_entry(p, l, r, wmode);
708 
709 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
710 }
711 
712 /*
713  * Check if Huge PTE is present, if so then jump to LABEL.
714  */
715 static void
716 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
717 		  unsigned int pmd, int lid)
718 {
719 	UASM_i_LW(p, tmp, 0, pmd);
720 	if (use_bbit_insns()) {
721 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
722 	} else {
723 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
724 		uasm_il_bnez(p, r, tmp, lid);
725 	}
726 }
727 
728 static void build_huge_update_entries(u32 **p, unsigned int pte,
729 				      unsigned int tmp)
730 {
731 	int small_sequence;
732 
733 	/*
734 	 * A huge PTE describes an area the size of the
735 	 * configured huge page size. This is twice the
736 	 * of the large TLB entry size we intend to use.
737 	 * A TLB entry half the size of the configured
738 	 * huge page size is configured into entrylo0
739 	 * and entrylo1 to cover the contiguous huge PTE
740 	 * address space.
741 	 */
742 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
743 
744 	/* We can clobber tmp.	It isn't used after this.*/
745 	if (!small_sequence)
746 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
747 
748 	build_convert_pte_to_entrylo(p, pte);
749 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
750 	/* convert to entrylo1 */
751 	if (small_sequence)
752 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
753 	else
754 		UASM_i_ADDU(p, pte, pte, tmp);
755 
756 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
757 }
758 
759 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
760 				    struct uasm_label **l,
761 				    unsigned int pte,
762 				    unsigned int ptr,
763 				    unsigned int flush)
764 {
765 #ifdef CONFIG_SMP
766 	UASM_i_SC(p, pte, 0, ptr);
767 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
768 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
769 #else
770 	UASM_i_SW(p, pte, 0, ptr);
771 #endif
772 	if (cpu_has_ftlb && flush) {
773 		BUG_ON(!cpu_has_tlbinv);
774 
775 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
776 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
777 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
778 		build_tlb_write_entry(p, l, r, tlb_indexed);
779 
780 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
781 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
782 		build_huge_update_entries(p, pte, ptr);
783 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
784 
785 		return;
786 	}
787 
788 	build_huge_update_entries(p, pte, ptr);
789 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
790 }
791 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
792 
793 #ifdef CONFIG_64BIT
794 /*
795  * TMP and PTR are scratch.
796  * TMP will be clobbered, PTR will hold the pmd entry.
797  */
798 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
799 		      unsigned int tmp, unsigned int ptr)
800 {
801 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
802 	long pgdc = (long)pgd_current;
803 #endif
804 	/*
805 	 * The vmalloc handling is not in the hotpath.
806 	 */
807 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
808 
809 	if (check_for_high_segbits) {
810 		/*
811 		 * The kernel currently implicitely assumes that the
812 		 * MIPS SEGBITS parameter for the processor is
813 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
814 		 * allocate virtual addresses outside the maximum
815 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
816 		 * that doesn't prevent user code from accessing the
817 		 * higher xuseg addresses.  Here, we make sure that
818 		 * everything but the lower xuseg addresses goes down
819 		 * the module_alloc/vmalloc path.
820 		 */
821 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
822 		uasm_il_bnez(p, r, ptr, label_vmalloc);
823 	} else {
824 		uasm_il_bltz(p, r, tmp, label_vmalloc);
825 	}
826 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
827 
828 	if (pgd_reg != -1) {
829 		/* pgd is in pgd_reg */
830 		if (cpu_has_ldpte)
831 			UASM_i_MFC0(p, ptr, C0_PWBASE);
832 		else
833 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
834 	} else {
835 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
836 		/*
837 		 * &pgd << 11 stored in CONTEXT [23..63].
838 		 */
839 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
840 
841 		/* Clear lower 23 bits of context. */
842 		uasm_i_dins(p, ptr, 0, 0, 23);
843 
844 		/* 1 0	1 0 1  << 6  xkphys cached */
845 		uasm_i_ori(p, ptr, ptr, 0x540);
846 		uasm_i_drotr(p, ptr, ptr, 11);
847 #elif defined(CONFIG_SMP)
848 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
849 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
850 		UASM_i_LA_mostly(p, tmp, pgdc);
851 		uasm_i_daddu(p, ptr, ptr, tmp);
852 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
853 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
854 #else
855 		UASM_i_LA_mostly(p, ptr, pgdc);
856 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
857 #endif
858 	}
859 
860 	uasm_l_vmalloc_done(l, *p);
861 
862 	/* get pgd offset in bytes */
863 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
864 
865 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
866 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
867 #ifndef __PAGETABLE_PUD_FOLDED
868 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
869 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
870 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
871 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
872 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
873 #endif
874 #ifndef __PAGETABLE_PMD_FOLDED
875 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
876 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
877 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
878 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
879 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
880 #endif
881 }
882 EXPORT_SYMBOL_GPL(build_get_pmde64);
883 
884 /*
885  * BVADDR is the faulting address, PTR is scratch.
886  * PTR will hold the pgd for vmalloc.
887  */
888 static void
889 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
890 			unsigned int bvaddr, unsigned int ptr,
891 			enum vmalloc64_mode mode)
892 {
893 	long swpd = (long)swapper_pg_dir;
894 	int single_insn_swpd;
895 	int did_vmalloc_branch = 0;
896 
897 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
898 
899 	uasm_l_vmalloc(l, *p);
900 
901 	if (mode != not_refill && check_for_high_segbits) {
902 		if (single_insn_swpd) {
903 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
904 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
905 			did_vmalloc_branch = 1;
906 			/* fall through */
907 		} else {
908 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
909 		}
910 	}
911 	if (!did_vmalloc_branch) {
912 		if (single_insn_swpd) {
913 			uasm_il_b(p, r, label_vmalloc_done);
914 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
915 		} else {
916 			UASM_i_LA_mostly(p, ptr, swpd);
917 			uasm_il_b(p, r, label_vmalloc_done);
918 			if (uasm_in_compat_space_p(swpd))
919 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
920 			else
921 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
922 		}
923 	}
924 	if (mode != not_refill && check_for_high_segbits) {
925 		uasm_l_large_segbits_fault(l, *p);
926 		/*
927 		 * We get here if we are an xsseg address, or if we are
928 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
929 		 *
930 		 * Ignoring xsseg (assume disabled so would generate
931 		 * (address errors?), the only remaining possibility
932 		 * is the upper xuseg addresses.  On processors with
933 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
934 		 * addresses would have taken an address error. We try
935 		 * to mimic that here by taking a load/istream page
936 		 * fault.
937 		 */
938 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
939 		uasm_i_jr(p, ptr);
940 
941 		if (mode == refill_scratch) {
942 			if (scratch_reg >= 0)
943 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
944 			else
945 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
946 		} else {
947 			uasm_i_nop(p);
948 		}
949 	}
950 }
951 
952 #else /* !CONFIG_64BIT */
953 
954 /*
955  * TMP and PTR are scratch.
956  * TMP will be clobbered, PTR will hold the pgd entry.
957  */
958 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
959 {
960 	if (pgd_reg != -1) {
961 		/* pgd is in pgd_reg */
962 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
963 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
964 	} else {
965 		long pgdc = (long)pgd_current;
966 
967 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
968 #ifdef CONFIG_SMP
969 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
970 		UASM_i_LA_mostly(p, tmp, pgdc);
971 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
972 		uasm_i_addu(p, ptr, tmp, ptr);
973 #else
974 		UASM_i_LA_mostly(p, ptr, pgdc);
975 #endif
976 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
977 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
978 	}
979 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
980 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
981 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
982 }
983 EXPORT_SYMBOL_GPL(build_get_pgde32);
984 
985 #endif /* !CONFIG_64BIT */
986 
987 static void build_adjust_context(u32 **p, unsigned int ctx)
988 {
989 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
990 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
991 
992 	switch (current_cpu_type()) {
993 	case CPU_VR41XX:
994 	case CPU_VR4111:
995 	case CPU_VR4121:
996 	case CPU_VR4122:
997 	case CPU_VR4131:
998 	case CPU_VR4181:
999 	case CPU_VR4181A:
1000 	case CPU_VR4133:
1001 		shift += 2;
1002 		break;
1003 
1004 	default:
1005 		break;
1006 	}
1007 
1008 	if (shift)
1009 		UASM_i_SRL(p, ctx, ctx, shift);
1010 	uasm_i_andi(p, ctx, ctx, mask);
1011 }
1012 
1013 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1014 {
1015 	/*
1016 	 * Bug workaround for the Nevada. It seems as if under certain
1017 	 * circumstances the move from cp0_context might produce a
1018 	 * bogus result when the mfc0 instruction and its consumer are
1019 	 * in a different cacheline or a load instruction, probably any
1020 	 * memory reference, is between them.
1021 	 */
1022 	switch (current_cpu_type()) {
1023 	case CPU_NEVADA:
1024 		UASM_i_LW(p, ptr, 0, ptr);
1025 		GET_CONTEXT(p, tmp); /* get context reg */
1026 		break;
1027 
1028 	default:
1029 		GET_CONTEXT(p, tmp); /* get context reg */
1030 		UASM_i_LW(p, ptr, 0, ptr);
1031 		break;
1032 	}
1033 
1034 	build_adjust_context(p, tmp);
1035 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1036 }
1037 EXPORT_SYMBOL_GPL(build_get_ptep);
1038 
1039 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1040 {
1041 	int pte_off_even = 0;
1042 	int pte_off_odd = sizeof(pte_t);
1043 
1044 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1045 	/* The low 32 bits of EntryLo is stored in pte_high */
1046 	pte_off_even += offsetof(pte_t, pte_high);
1047 	pte_off_odd += offsetof(pte_t, pte_high);
1048 #endif
1049 
1050 	if (IS_ENABLED(CONFIG_XPA)) {
1051 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1052 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1053 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1054 
1055 		if (cpu_has_xpa && !mips_xpa_disabled) {
1056 			uasm_i_lw(p, tmp, 0, ptep);
1057 			uasm_i_ext(p, tmp, tmp, 0, 24);
1058 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1059 		}
1060 
1061 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1062 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1063 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1064 
1065 		if (cpu_has_xpa && !mips_xpa_disabled) {
1066 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1067 			uasm_i_ext(p, tmp, tmp, 0, 24);
1068 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1069 		}
1070 		return;
1071 	}
1072 
1073 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1074 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1075 	if (r45k_bvahwbug())
1076 		build_tlb_probe_entry(p);
1077 	build_convert_pte_to_entrylo(p, tmp);
1078 	if (r4k_250MHZhwbug())
1079 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1080 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1081 	build_convert_pte_to_entrylo(p, ptep);
1082 	if (r45k_bvahwbug())
1083 		uasm_i_mfc0(p, tmp, C0_INDEX);
1084 	if (r4k_250MHZhwbug())
1085 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1086 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1087 }
1088 EXPORT_SYMBOL_GPL(build_update_entries);
1089 
1090 struct mips_huge_tlb_info {
1091 	int huge_pte;
1092 	int restore_scratch;
1093 	bool need_reload_pte;
1094 };
1095 
1096 static struct mips_huge_tlb_info
1097 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1098 			       struct uasm_reloc **r, unsigned int tmp,
1099 			       unsigned int ptr, int c0_scratch_reg)
1100 {
1101 	struct mips_huge_tlb_info rv;
1102 	unsigned int even, odd;
1103 	int vmalloc_branch_delay_filled = 0;
1104 	const int scratch = 1; /* Our extra working register */
1105 
1106 	rv.huge_pte = scratch;
1107 	rv.restore_scratch = 0;
1108 	rv.need_reload_pte = false;
1109 
1110 	if (check_for_high_segbits) {
1111 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1112 
1113 		if (pgd_reg != -1)
1114 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1115 		else
1116 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1117 
1118 		if (c0_scratch_reg >= 0)
1119 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1120 		else
1121 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1122 
1123 		uasm_i_dsrl_safe(p, scratch, tmp,
1124 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1125 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1126 
1127 		if (pgd_reg == -1) {
1128 			vmalloc_branch_delay_filled = 1;
1129 			/* Clear lower 23 bits of context. */
1130 			uasm_i_dins(p, ptr, 0, 0, 23);
1131 		}
1132 	} else {
1133 		if (pgd_reg != -1)
1134 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1135 		else
1136 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1137 
1138 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1139 
1140 		if (c0_scratch_reg >= 0)
1141 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1142 		else
1143 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1144 
1145 		if (pgd_reg == -1)
1146 			/* Clear lower 23 bits of context. */
1147 			uasm_i_dins(p, ptr, 0, 0, 23);
1148 
1149 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1150 	}
1151 
1152 	if (pgd_reg == -1) {
1153 		vmalloc_branch_delay_filled = 1;
1154 		/* 1 0	1 0 1  << 6  xkphys cached */
1155 		uasm_i_ori(p, ptr, ptr, 0x540);
1156 		uasm_i_drotr(p, ptr, ptr, 11);
1157 	}
1158 
1159 #ifdef __PAGETABLE_PMD_FOLDED
1160 #define LOC_PTEP scratch
1161 #else
1162 #define LOC_PTEP ptr
1163 #endif
1164 
1165 	if (!vmalloc_branch_delay_filled)
1166 		/* get pgd offset in bytes */
1167 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1168 
1169 	uasm_l_vmalloc_done(l, *p);
1170 
1171 	/*
1172 	 *			   tmp		ptr
1173 	 * fall-through case =	 badvaddr  *pgd_current
1174 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1175 	 */
1176 
1177 	if (vmalloc_branch_delay_filled)
1178 		/* get pgd offset in bytes */
1179 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1180 
1181 #ifdef __PAGETABLE_PMD_FOLDED
1182 	GET_CONTEXT(p, tmp); /* get context reg */
1183 #endif
1184 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1185 
1186 	if (use_lwx_insns()) {
1187 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1188 	} else {
1189 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1190 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1191 	}
1192 
1193 #ifndef __PAGETABLE_PUD_FOLDED
1194 	/* get pud offset in bytes */
1195 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1196 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1197 
1198 	if (use_lwx_insns()) {
1199 		UASM_i_LWX(p, ptr, scratch, ptr);
1200 	} else {
1201 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1202 		UASM_i_LW(p, ptr, 0, ptr);
1203 	}
1204 	/* ptr contains a pointer to PMD entry */
1205 	/* tmp contains the address */
1206 #endif
1207 
1208 #ifndef __PAGETABLE_PMD_FOLDED
1209 	/* get pmd offset in bytes */
1210 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1211 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1212 	GET_CONTEXT(p, tmp); /* get context reg */
1213 
1214 	if (use_lwx_insns()) {
1215 		UASM_i_LWX(p, scratch, scratch, ptr);
1216 	} else {
1217 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1218 		UASM_i_LW(p, scratch, 0, ptr);
1219 	}
1220 #endif
1221 	/* Adjust the context during the load latency. */
1222 	build_adjust_context(p, tmp);
1223 
1224 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1225 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1226 	/*
1227 	 * The in the LWX case we don't want to do the load in the
1228 	 * delay slot.	It cannot issue in the same cycle and may be
1229 	 * speculative and unneeded.
1230 	 */
1231 	if (use_lwx_insns())
1232 		uasm_i_nop(p);
1233 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1234 
1235 
1236 	/* build_update_entries */
1237 	if (use_lwx_insns()) {
1238 		even = ptr;
1239 		odd = tmp;
1240 		UASM_i_LWX(p, even, scratch, tmp);
1241 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1242 		UASM_i_LWX(p, odd, scratch, tmp);
1243 	} else {
1244 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1245 		even = tmp;
1246 		odd = ptr;
1247 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1248 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1249 	}
1250 	if (cpu_has_rixi) {
1251 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1252 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1253 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1254 	} else {
1255 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1256 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1257 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1258 	}
1259 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1260 
1261 	if (c0_scratch_reg >= 0) {
1262 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1263 		build_tlb_write_entry(p, l, r, tlb_random);
1264 		uasm_l_leave(l, *p);
1265 		rv.restore_scratch = 1;
1266 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1267 		build_tlb_write_entry(p, l, r, tlb_random);
1268 		uasm_l_leave(l, *p);
1269 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1270 	} else {
1271 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1272 		build_tlb_write_entry(p, l, r, tlb_random);
1273 		uasm_l_leave(l, *p);
1274 		rv.restore_scratch = 1;
1275 	}
1276 
1277 	uasm_i_eret(p); /* return from trap */
1278 
1279 	return rv;
1280 }
1281 
1282 /*
1283  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1284  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1285  * slots before the XTLB refill exception handler which belong to the
1286  * unused TLB refill exception.
1287  */
1288 #define MIPS64_REFILL_INSNS 32
1289 
1290 static void build_r4000_tlb_refill_handler(void)
1291 {
1292 	u32 *p = tlb_handler;
1293 	struct uasm_label *l = labels;
1294 	struct uasm_reloc *r = relocs;
1295 	u32 *f;
1296 	unsigned int final_len;
1297 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1298 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1299 
1300 	memset(tlb_handler, 0, sizeof(tlb_handler));
1301 	memset(labels, 0, sizeof(labels));
1302 	memset(relocs, 0, sizeof(relocs));
1303 	memset(final_handler, 0, sizeof(final_handler));
1304 
1305 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1306 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1307 							  scratch_reg);
1308 		vmalloc_mode = refill_scratch;
1309 	} else {
1310 		htlb_info.huge_pte = K0;
1311 		htlb_info.restore_scratch = 0;
1312 		htlb_info.need_reload_pte = true;
1313 		vmalloc_mode = refill_noscratch;
1314 		/*
1315 		 * create the plain linear handler
1316 		 */
1317 		if (bcm1250_m3_war()) {
1318 			unsigned int segbits = 44;
1319 
1320 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1321 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1322 			uasm_i_xor(&p, K0, K0, K1);
1323 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1324 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1325 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1326 			uasm_i_or(&p, K0, K0, K1);
1327 			uasm_il_bnez(&p, &r, K0, label_leave);
1328 			/* No need for uasm_i_nop */
1329 		}
1330 
1331 #ifdef CONFIG_64BIT
1332 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1333 #else
1334 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1335 #endif
1336 
1337 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1338 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1339 #endif
1340 
1341 		build_get_ptep(&p, K0, K1);
1342 		build_update_entries(&p, K0, K1);
1343 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1344 		uasm_l_leave(&l, p);
1345 		uasm_i_eret(&p); /* return from trap */
1346 	}
1347 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1348 	uasm_l_tlb_huge_update(&l, p);
1349 	if (htlb_info.need_reload_pte)
1350 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1351 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1352 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1353 				   htlb_info.restore_scratch);
1354 #endif
1355 
1356 #ifdef CONFIG_64BIT
1357 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1358 #endif
1359 
1360 	/*
1361 	 * Overflow check: For the 64bit handler, we need at least one
1362 	 * free instruction slot for the wrap-around branch. In worst
1363 	 * case, if the intended insertion point is a delay slot, we
1364 	 * need three, with the second nop'ed and the third being
1365 	 * unused.
1366 	 */
1367 	switch (boot_cpu_type()) {
1368 	default:
1369 		if (sizeof(long) == 4) {
1370 	case CPU_LOONGSON2:
1371 		/* Loongson2 ebase is different than r4k, we have more space */
1372 			if ((p - tlb_handler) > 64)
1373 				panic("TLB refill handler space exceeded");
1374 			/*
1375 			 * Now fold the handler in the TLB refill handler space.
1376 			 */
1377 			f = final_handler;
1378 			/* Simplest case, just copy the handler. */
1379 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1380 			final_len = p - tlb_handler;
1381 			break;
1382 		} else {
1383 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1384 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1385 				&& uasm_insn_has_bdelay(relocs,
1386 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1387 				panic("TLB refill handler space exceeded");
1388 			/*
1389 			 * Now fold the handler in the TLB refill handler space.
1390 			 */
1391 			f = final_handler + MIPS64_REFILL_INSNS;
1392 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1393 				/* Just copy the handler. */
1394 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1395 				final_len = p - tlb_handler;
1396 			} else {
1397 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1398 				const enum label_id ls = label_tlb_huge_update;
1399 #else
1400 				const enum label_id ls = label_vmalloc;
1401 #endif
1402 				u32 *split;
1403 				int ov = 0;
1404 				int i;
1405 
1406 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1407 					;
1408 				BUG_ON(i == ARRAY_SIZE(labels));
1409 				split = labels[i].addr;
1410 
1411 				/*
1412 				 * See if we have overflown one way or the other.
1413 				 */
1414 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1415 				    split < p - MIPS64_REFILL_INSNS)
1416 					ov = 1;
1417 
1418 				if (ov) {
1419 					/*
1420 					 * Split two instructions before the end.  One
1421 					 * for the branch and one for the instruction
1422 					 * in the delay slot.
1423 					 */
1424 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1425 
1426 					/*
1427 					 * If the branch would fall in a delay slot,
1428 					 * we must back up an additional instruction
1429 					 * so that it is no longer in a delay slot.
1430 					 */
1431 					if (uasm_insn_has_bdelay(relocs, split - 1))
1432 						split--;
1433 				}
1434 				/* Copy first part of the handler. */
1435 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1436 				f += split - tlb_handler;
1437 
1438 				if (ov) {
1439 					/* Insert branch. */
1440 					uasm_l_split(&l, final_handler);
1441 					uasm_il_b(&f, &r, label_split);
1442 					if (uasm_insn_has_bdelay(relocs, split))
1443 						uasm_i_nop(&f);
1444 					else {
1445 						uasm_copy_handler(relocs, labels,
1446 								  split, split + 1, f);
1447 						uasm_move_labels(labels, f, f + 1, -1);
1448 						f++;
1449 						split++;
1450 					}
1451 				}
1452 
1453 				/* Copy the rest of the handler. */
1454 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1455 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1456 					    (p - split);
1457 			}
1458 		}
1459 		break;
1460 	}
1461 
1462 	uasm_resolve_relocs(relocs, labels);
1463 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1464 		 final_len);
1465 
1466 	memcpy((void *)ebase, final_handler, 0x100);
1467 	local_flush_icache_range(ebase, ebase + 0x100);
1468 
1469 	dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1470 }
1471 
1472 static void setup_pw(void)
1473 {
1474 	unsigned long pgd_i, pgd_w;
1475 #ifndef __PAGETABLE_PMD_FOLDED
1476 	unsigned long pmd_i, pmd_w;
1477 #endif
1478 	unsigned long pt_i, pt_w;
1479 	unsigned long pte_i, pte_w;
1480 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1481 	unsigned long psn;
1482 
1483 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1484 #endif
1485 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1486 #ifndef __PAGETABLE_PMD_FOLDED
1487 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1488 
1489 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1490 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1491 #else
1492 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1493 #endif
1494 
1495 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1496 	pt_w  = PAGE_SHIFT - 3;
1497 
1498 	pte_i = ilog2(_PAGE_GLOBAL);
1499 	pte_w = 0;
1500 
1501 #ifndef __PAGETABLE_PMD_FOLDED
1502 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1503 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1504 #else
1505 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1506 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1507 #endif
1508 
1509 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1510 	write_c0_pwctl(1 << 6 | psn);
1511 #endif
1512 	write_c0_kpgd(swapper_pg_dir);
1513 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1514 }
1515 
1516 static void build_loongson3_tlb_refill_handler(void)
1517 {
1518 	u32 *p = tlb_handler;
1519 	struct uasm_label *l = labels;
1520 	struct uasm_reloc *r = relocs;
1521 
1522 	memset(labels, 0, sizeof(labels));
1523 	memset(relocs, 0, sizeof(relocs));
1524 	memset(tlb_handler, 0, sizeof(tlb_handler));
1525 
1526 	if (check_for_high_segbits) {
1527 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1528 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1529 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1530 		uasm_i_nop(&p);
1531 
1532 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1533 		uasm_i_nop(&p);
1534 		uasm_l_vmalloc(&l, p);
1535 	}
1536 
1537 	uasm_i_dmfc0(&p, K1, C0_PGD);
1538 
1539 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1540 #ifndef __PAGETABLE_PMD_FOLDED
1541 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1542 #endif
1543 	uasm_i_ldpte(&p, K1, 0);      /* even */
1544 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1545 	uasm_i_tlbwr(&p);
1546 
1547 	/* restore page mask */
1548 	if (PM_DEFAULT_MASK >> 16) {
1549 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1550 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1551 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1552 	} else if (PM_DEFAULT_MASK) {
1553 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1554 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1555 	} else {
1556 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1557 	}
1558 
1559 	uasm_i_eret(&p);
1560 
1561 	if (check_for_high_segbits) {
1562 		uasm_l_large_segbits_fault(&l, p);
1563 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1564 		uasm_i_jr(&p, K1);
1565 		uasm_i_nop(&p);
1566 	}
1567 
1568 	uasm_resolve_relocs(relocs, labels);
1569 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1570 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1571 	dump_handler("loongson3_tlb_refill", (u32 *)(ebase + 0x80), 32);
1572 }
1573 
1574 extern u32 handle_tlbl[], handle_tlbl_end[];
1575 extern u32 handle_tlbs[], handle_tlbs_end[];
1576 extern u32 handle_tlbm[], handle_tlbm_end[];
1577 extern u32 tlbmiss_handler_setup_pgd_start[];
1578 extern u32 tlbmiss_handler_setup_pgd[];
1579 EXPORT_SYMBOL_GPL(tlbmiss_handler_setup_pgd);
1580 extern u32 tlbmiss_handler_setup_pgd_end[];
1581 
1582 static void build_setup_pgd(void)
1583 {
1584 	const int a0 = 4;
1585 	const int __maybe_unused a1 = 5;
1586 	const int __maybe_unused a2 = 6;
1587 	u32 *p = tlbmiss_handler_setup_pgd_start;
1588 	const int tlbmiss_handler_setup_pgd_size =
1589 		tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1590 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1591 	long pgdc = (long)pgd_current;
1592 #endif
1593 
1594 	memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1595 					sizeof(tlbmiss_handler_setup_pgd[0]));
1596 	memset(labels, 0, sizeof(labels));
1597 	memset(relocs, 0, sizeof(relocs));
1598 	pgd_reg = allocate_kscratch();
1599 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1600 	if (pgd_reg == -1) {
1601 		struct uasm_label *l = labels;
1602 		struct uasm_reloc *r = relocs;
1603 
1604 		/* PGD << 11 in c0_Context */
1605 		/*
1606 		 * If it is a ckseg0 address, convert to a physical
1607 		 * address.  Shifting right by 29 and adding 4 will
1608 		 * result in zero for these addresses.
1609 		 *
1610 		 */
1611 		UASM_i_SRA(&p, a1, a0, 29);
1612 		UASM_i_ADDIU(&p, a1, a1, 4);
1613 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1614 		uasm_i_nop(&p);
1615 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1616 		uasm_l_tlbl_goaround1(&l, p);
1617 		UASM_i_SLL(&p, a0, a0, 11);
1618 		uasm_i_jr(&p, 31);
1619 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1620 	} else {
1621 		/* PGD in c0_KScratch */
1622 		uasm_i_jr(&p, 31);
1623 		if (cpu_has_ldpte)
1624 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1625 		else
1626 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1627 	}
1628 #else
1629 #ifdef CONFIG_SMP
1630 	/* Save PGD to pgd_current[smp_processor_id()] */
1631 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1632 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1633 	UASM_i_LA_mostly(&p, a2, pgdc);
1634 	UASM_i_ADDU(&p, a2, a2, a1);
1635 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1636 #else
1637 	UASM_i_LA_mostly(&p, a2, pgdc);
1638 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1639 #endif /* SMP */
1640 	uasm_i_jr(&p, 31);
1641 
1642 	/* if pgd_reg is allocated, save PGD also to scratch register */
1643 	if (pgd_reg != -1)
1644 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1645 	else
1646 		uasm_i_nop(&p);
1647 #endif
1648 	if (p >= tlbmiss_handler_setup_pgd_end)
1649 		panic("tlbmiss_handler_setup_pgd space exceeded");
1650 
1651 	uasm_resolve_relocs(relocs, labels);
1652 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1653 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1654 
1655 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1656 					tlbmiss_handler_setup_pgd_size);
1657 }
1658 
1659 static void
1660 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1661 {
1662 #ifdef CONFIG_SMP
1663 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1664 	if (cpu_has_64bits)
1665 		uasm_i_lld(p, pte, 0, ptr);
1666 	else
1667 # endif
1668 		UASM_i_LL(p, pte, 0, ptr);
1669 #else
1670 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1671 	if (cpu_has_64bits)
1672 		uasm_i_ld(p, pte, 0, ptr);
1673 	else
1674 # endif
1675 		UASM_i_LW(p, pte, 0, ptr);
1676 #endif
1677 }
1678 
1679 static void
1680 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1681 	unsigned int mode, unsigned int scratch)
1682 {
1683 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1684 	unsigned int swmode = mode & ~hwmode;
1685 
1686 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1687 		uasm_i_lui(p, scratch, swmode >> 16);
1688 		uasm_i_or(p, pte, pte, scratch);
1689 		BUG_ON(swmode & 0xffff);
1690 	} else {
1691 		uasm_i_ori(p, pte, pte, mode);
1692 	}
1693 
1694 #ifdef CONFIG_SMP
1695 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1696 	if (cpu_has_64bits)
1697 		uasm_i_scd(p, pte, 0, ptr);
1698 	else
1699 # endif
1700 		UASM_i_SC(p, pte, 0, ptr);
1701 
1702 	if (r10000_llsc_war())
1703 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1704 	else
1705 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1706 
1707 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1708 	if (!cpu_has_64bits) {
1709 		/* no uasm_i_nop needed */
1710 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1711 		uasm_i_ori(p, pte, pte, hwmode);
1712 		BUG_ON(hwmode & ~0xffff);
1713 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1714 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1715 		/* no uasm_i_nop needed */
1716 		uasm_i_lw(p, pte, 0, ptr);
1717 	} else
1718 		uasm_i_nop(p);
1719 # else
1720 	uasm_i_nop(p);
1721 # endif
1722 #else
1723 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1724 	if (cpu_has_64bits)
1725 		uasm_i_sd(p, pte, 0, ptr);
1726 	else
1727 # endif
1728 		UASM_i_SW(p, pte, 0, ptr);
1729 
1730 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1731 	if (!cpu_has_64bits) {
1732 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1733 		uasm_i_ori(p, pte, pte, hwmode);
1734 		BUG_ON(hwmode & ~0xffff);
1735 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1736 		uasm_i_lw(p, pte, 0, ptr);
1737 	}
1738 # endif
1739 #endif
1740 }
1741 
1742 /*
1743  * Check if PTE is present, if not then jump to LABEL. PTR points to
1744  * the page table where this PTE is located, PTE will be re-loaded
1745  * with it's original value.
1746  */
1747 static void
1748 build_pte_present(u32 **p, struct uasm_reloc **r,
1749 		  int pte, int ptr, int scratch, enum label_id lid)
1750 {
1751 	int t = scratch >= 0 ? scratch : pte;
1752 	int cur = pte;
1753 
1754 	if (cpu_has_rixi) {
1755 		if (use_bbit_insns()) {
1756 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1757 			uasm_i_nop(p);
1758 		} else {
1759 			if (_PAGE_PRESENT_SHIFT) {
1760 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1761 				cur = t;
1762 			}
1763 			uasm_i_andi(p, t, cur, 1);
1764 			uasm_il_beqz(p, r, t, lid);
1765 			if (pte == t)
1766 				/* You lose the SMP race :-(*/
1767 				iPTE_LW(p, pte, ptr);
1768 		}
1769 	} else {
1770 		if (_PAGE_PRESENT_SHIFT) {
1771 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1772 			cur = t;
1773 		}
1774 		uasm_i_andi(p, t, cur,
1775 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1776 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1777 		uasm_il_bnez(p, r, t, lid);
1778 		if (pte == t)
1779 			/* You lose the SMP race :-(*/
1780 			iPTE_LW(p, pte, ptr);
1781 	}
1782 }
1783 
1784 /* Make PTE valid, store result in PTR. */
1785 static void
1786 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1787 		 unsigned int ptr, unsigned int scratch)
1788 {
1789 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1790 
1791 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1792 }
1793 
1794 /*
1795  * Check if PTE can be written to, if not branch to LABEL. Regardless
1796  * restore PTE with value from PTR when done.
1797  */
1798 static void
1799 build_pte_writable(u32 **p, struct uasm_reloc **r,
1800 		   unsigned int pte, unsigned int ptr, int scratch,
1801 		   enum label_id lid)
1802 {
1803 	int t = scratch >= 0 ? scratch : pte;
1804 	int cur = pte;
1805 
1806 	if (_PAGE_PRESENT_SHIFT) {
1807 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1808 		cur = t;
1809 	}
1810 	uasm_i_andi(p, t, cur,
1811 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1812 	uasm_i_xori(p, t, t,
1813 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1814 	uasm_il_bnez(p, r, t, lid);
1815 	if (pte == t)
1816 		/* You lose the SMP race :-(*/
1817 		iPTE_LW(p, pte, ptr);
1818 	else
1819 		uasm_i_nop(p);
1820 }
1821 
1822 /* Make PTE writable, update software status bits as well, then store
1823  * at PTR.
1824  */
1825 static void
1826 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1827 		 unsigned int ptr, unsigned int scratch)
1828 {
1829 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1830 			     | _PAGE_DIRTY);
1831 
1832 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1833 }
1834 
1835 /*
1836  * Check if PTE can be modified, if not branch to LABEL. Regardless
1837  * restore PTE with value from PTR when done.
1838  */
1839 static void
1840 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1841 		     unsigned int pte, unsigned int ptr, int scratch,
1842 		     enum label_id lid)
1843 {
1844 	if (use_bbit_insns()) {
1845 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1846 		uasm_i_nop(p);
1847 	} else {
1848 		int t = scratch >= 0 ? scratch : pte;
1849 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1850 		uasm_i_andi(p, t, t, 1);
1851 		uasm_il_beqz(p, r, t, lid);
1852 		if (pte == t)
1853 			/* You lose the SMP race :-(*/
1854 			iPTE_LW(p, pte, ptr);
1855 	}
1856 }
1857 
1858 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1859 
1860 
1861 /*
1862  * R3000 style TLB load/store/modify handlers.
1863  */
1864 
1865 /*
1866  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1867  * Then it returns.
1868  */
1869 static void
1870 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1871 {
1872 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1873 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1874 	uasm_i_tlbwi(p);
1875 	uasm_i_jr(p, tmp);
1876 	uasm_i_rfe(p); /* branch delay */
1877 }
1878 
1879 /*
1880  * This places the pte into ENTRYLO0 and writes it with tlbwi
1881  * or tlbwr as appropriate.  This is because the index register
1882  * may have the probe fail bit set as a result of a trap on a
1883  * kseg2 access, i.e. without refill.  Then it returns.
1884  */
1885 static void
1886 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1887 			     struct uasm_reloc **r, unsigned int pte,
1888 			     unsigned int tmp)
1889 {
1890 	uasm_i_mfc0(p, tmp, C0_INDEX);
1891 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1892 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1893 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1894 	uasm_i_tlbwi(p); /* cp0 delay */
1895 	uasm_i_jr(p, tmp);
1896 	uasm_i_rfe(p); /* branch delay */
1897 	uasm_l_r3000_write_probe_fail(l, *p);
1898 	uasm_i_tlbwr(p); /* cp0 delay */
1899 	uasm_i_jr(p, tmp);
1900 	uasm_i_rfe(p); /* branch delay */
1901 }
1902 
1903 static void
1904 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1905 				   unsigned int ptr)
1906 {
1907 	long pgdc = (long)pgd_current;
1908 
1909 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1910 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1911 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1912 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1913 	uasm_i_sll(p, pte, pte, 2);
1914 	uasm_i_addu(p, ptr, ptr, pte);
1915 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1916 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1917 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1918 	uasm_i_addu(p, ptr, ptr, pte);
1919 	uasm_i_lw(p, pte, 0, ptr);
1920 	uasm_i_tlbp(p); /* load delay */
1921 }
1922 
1923 static void build_r3000_tlb_load_handler(void)
1924 {
1925 	u32 *p = handle_tlbl;
1926 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1927 	struct uasm_label *l = labels;
1928 	struct uasm_reloc *r = relocs;
1929 
1930 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1931 	memset(labels, 0, sizeof(labels));
1932 	memset(relocs, 0, sizeof(relocs));
1933 
1934 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1935 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1936 	uasm_i_nop(&p); /* load delay */
1937 	build_make_valid(&p, &r, K0, K1, -1);
1938 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1939 
1940 	uasm_l_nopage_tlbl(&l, p);
1941 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1942 	uasm_i_nop(&p);
1943 
1944 	if (p >= handle_tlbl_end)
1945 		panic("TLB load handler fastpath space exceeded");
1946 
1947 	uasm_resolve_relocs(relocs, labels);
1948 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1949 		 (unsigned int)(p - handle_tlbl));
1950 
1951 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1952 }
1953 
1954 static void build_r3000_tlb_store_handler(void)
1955 {
1956 	u32 *p = handle_tlbs;
1957 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1958 	struct uasm_label *l = labels;
1959 	struct uasm_reloc *r = relocs;
1960 
1961 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1962 	memset(labels, 0, sizeof(labels));
1963 	memset(relocs, 0, sizeof(relocs));
1964 
1965 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1966 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1967 	uasm_i_nop(&p); /* load delay */
1968 	build_make_write(&p, &r, K0, K1, -1);
1969 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1970 
1971 	uasm_l_nopage_tlbs(&l, p);
1972 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1973 	uasm_i_nop(&p);
1974 
1975 	if (p >= handle_tlbs_end)
1976 		panic("TLB store handler fastpath space exceeded");
1977 
1978 	uasm_resolve_relocs(relocs, labels);
1979 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1980 		 (unsigned int)(p - handle_tlbs));
1981 
1982 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1983 }
1984 
1985 static void build_r3000_tlb_modify_handler(void)
1986 {
1987 	u32 *p = handle_tlbm;
1988 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1989 	struct uasm_label *l = labels;
1990 	struct uasm_reloc *r = relocs;
1991 
1992 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1993 	memset(labels, 0, sizeof(labels));
1994 	memset(relocs, 0, sizeof(relocs));
1995 
1996 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1997 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1998 	uasm_i_nop(&p); /* load delay */
1999 	build_make_write(&p, &r, K0, K1, -1);
2000 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
2001 
2002 	uasm_l_nopage_tlbm(&l, p);
2003 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2004 	uasm_i_nop(&p);
2005 
2006 	if (p >= handle_tlbm_end)
2007 		panic("TLB modify handler fastpath space exceeded");
2008 
2009 	uasm_resolve_relocs(relocs, labels);
2010 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2011 		 (unsigned int)(p - handle_tlbm));
2012 
2013 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
2014 }
2015 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2016 
2017 static bool cpu_has_tlbex_tlbp_race(void)
2018 {
2019 	/*
2020 	 * When a Hardware Table Walker is running it can replace TLB entries
2021 	 * at any time, leading to a race between it & the CPU.
2022 	 */
2023 	if (cpu_has_htw)
2024 		return true;
2025 
2026 	/*
2027 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2028 	 * replaced at any time by a sibling performing a write to the FTLB.
2029 	 */
2030 	if (cpu_has_shared_ftlb_ram)
2031 		return true;
2032 
2033 	/* In all other cases there ought to be no race condition to handle */
2034 	return false;
2035 }
2036 
2037 /*
2038  * R4000 style TLB load/store/modify handlers.
2039  */
2040 static struct work_registers
2041 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2042 				   struct uasm_reloc **r)
2043 {
2044 	struct work_registers wr = build_get_work_registers(p);
2045 
2046 #ifdef CONFIG_64BIT
2047 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2048 #else
2049 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2050 #endif
2051 
2052 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2053 	/*
2054 	 * For huge tlb entries, pmd doesn't contain an address but
2055 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2056 	 * see if we need to jump to huge tlb processing.
2057 	 */
2058 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2059 #endif
2060 
2061 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2062 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2063 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2064 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2065 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2066 
2067 #ifdef CONFIG_SMP
2068 	uasm_l_smp_pgtable_change(l, *p);
2069 #endif
2070 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2071 	if (!m4kc_tlbp_war()) {
2072 		build_tlb_probe_entry(p);
2073 		if (cpu_has_tlbex_tlbp_race()) {
2074 			/* race condition happens, leaving */
2075 			uasm_i_ehb(p);
2076 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2077 			uasm_il_bltz(p, r, wr.r3, label_leave);
2078 			uasm_i_nop(p);
2079 		}
2080 	}
2081 	return wr;
2082 }
2083 
2084 static void
2085 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2086 				   struct uasm_reloc **r, unsigned int tmp,
2087 				   unsigned int ptr)
2088 {
2089 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2090 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2091 	build_update_entries(p, tmp, ptr);
2092 	build_tlb_write_entry(p, l, r, tlb_indexed);
2093 	uasm_l_leave(l, *p);
2094 	build_restore_work_registers(p);
2095 	uasm_i_eret(p); /* return from trap */
2096 
2097 #ifdef CONFIG_64BIT
2098 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2099 #endif
2100 }
2101 
2102 static void build_r4000_tlb_load_handler(void)
2103 {
2104 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2105 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
2106 	struct uasm_label *l = labels;
2107 	struct uasm_reloc *r = relocs;
2108 	struct work_registers wr;
2109 
2110 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
2111 	memset(labels, 0, sizeof(labels));
2112 	memset(relocs, 0, sizeof(relocs));
2113 
2114 	if (bcm1250_m3_war()) {
2115 		unsigned int segbits = 44;
2116 
2117 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2118 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2119 		uasm_i_xor(&p, K0, K0, K1);
2120 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2121 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2122 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2123 		uasm_i_or(&p, K0, K0, K1);
2124 		uasm_il_bnez(&p, &r, K0, label_leave);
2125 		/* No need for uasm_i_nop */
2126 	}
2127 
2128 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2129 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2130 	if (m4kc_tlbp_war())
2131 		build_tlb_probe_entry(&p);
2132 
2133 	if (cpu_has_rixi && !cpu_has_rixiex) {
2134 		/*
2135 		 * If the page is not _PAGE_VALID, RI or XI could not
2136 		 * have triggered it.  Skip the expensive test..
2137 		 */
2138 		if (use_bbit_insns()) {
2139 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2140 				      label_tlbl_goaround1);
2141 		} else {
2142 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2143 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2144 		}
2145 		uasm_i_nop(&p);
2146 
2147 		/*
2148 		 * Warn if something may race with us & replace the TLB entry
2149 		 * before we read it here. Everything with such races should
2150 		 * also have dedicated RiXi exception handlers, so this
2151 		 * shouldn't be hit.
2152 		 */
2153 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2154 
2155 		uasm_i_tlbr(&p);
2156 
2157 		switch (current_cpu_type()) {
2158 		default:
2159 			if (cpu_has_mips_r2_exec_hazard) {
2160 				uasm_i_ehb(&p);
2161 
2162 		case CPU_CAVIUM_OCTEON:
2163 		case CPU_CAVIUM_OCTEON_PLUS:
2164 		case CPU_CAVIUM_OCTEON2:
2165 				break;
2166 			}
2167 		}
2168 
2169 		/* Examine  entrylo 0 or 1 based on ptr. */
2170 		if (use_bbit_insns()) {
2171 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2172 		} else {
2173 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2174 			uasm_i_beqz(&p, wr.r3, 8);
2175 		}
2176 		/* load it in the delay slot*/
2177 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2178 		/* load it if ptr is odd */
2179 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2180 		/*
2181 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2182 		 * XI must have triggered it.
2183 		 */
2184 		if (use_bbit_insns()) {
2185 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2186 			uasm_i_nop(&p);
2187 			uasm_l_tlbl_goaround1(&l, p);
2188 		} else {
2189 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2190 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2191 			uasm_i_nop(&p);
2192 		}
2193 		uasm_l_tlbl_goaround1(&l, p);
2194 	}
2195 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2196 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2197 
2198 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2199 	/*
2200 	 * This is the entry point when build_r4000_tlbchange_handler_head
2201 	 * spots a huge page.
2202 	 */
2203 	uasm_l_tlb_huge_update(&l, p);
2204 	iPTE_LW(&p, wr.r1, wr.r2);
2205 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2206 	build_tlb_probe_entry(&p);
2207 
2208 	if (cpu_has_rixi && !cpu_has_rixiex) {
2209 		/*
2210 		 * If the page is not _PAGE_VALID, RI or XI could not
2211 		 * have triggered it.  Skip the expensive test..
2212 		 */
2213 		if (use_bbit_insns()) {
2214 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2215 				      label_tlbl_goaround2);
2216 		} else {
2217 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2218 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2219 		}
2220 		uasm_i_nop(&p);
2221 
2222 		/*
2223 		 * Warn if something may race with us & replace the TLB entry
2224 		 * before we read it here. Everything with such races should
2225 		 * also have dedicated RiXi exception handlers, so this
2226 		 * shouldn't be hit.
2227 		 */
2228 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2229 
2230 		uasm_i_tlbr(&p);
2231 
2232 		switch (current_cpu_type()) {
2233 		default:
2234 			if (cpu_has_mips_r2_exec_hazard) {
2235 				uasm_i_ehb(&p);
2236 
2237 		case CPU_CAVIUM_OCTEON:
2238 		case CPU_CAVIUM_OCTEON_PLUS:
2239 		case CPU_CAVIUM_OCTEON2:
2240 				break;
2241 			}
2242 		}
2243 
2244 		/* Examine  entrylo 0 or 1 based on ptr. */
2245 		if (use_bbit_insns()) {
2246 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2247 		} else {
2248 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2249 			uasm_i_beqz(&p, wr.r3, 8);
2250 		}
2251 		/* load it in the delay slot*/
2252 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2253 		/* load it if ptr is odd */
2254 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2255 		/*
2256 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2257 		 * XI must have triggered it.
2258 		 */
2259 		if (use_bbit_insns()) {
2260 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2261 		} else {
2262 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2263 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2264 		}
2265 		if (PM_DEFAULT_MASK == 0)
2266 			uasm_i_nop(&p);
2267 		/*
2268 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2269 		 * it is restored in build_huge_tlb_write_entry.
2270 		 */
2271 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2272 
2273 		uasm_l_tlbl_goaround2(&l, p);
2274 	}
2275 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2276 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2277 #endif
2278 
2279 	uasm_l_nopage_tlbl(&l, p);
2280 	build_restore_work_registers(&p);
2281 #ifdef CONFIG_CPU_MICROMIPS
2282 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2283 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2284 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2285 		uasm_i_jr(&p, K0);
2286 	} else
2287 #endif
2288 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2289 	uasm_i_nop(&p);
2290 
2291 	if (p >= handle_tlbl_end)
2292 		panic("TLB load handler fastpath space exceeded");
2293 
2294 	uasm_resolve_relocs(relocs, labels);
2295 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2296 		 (unsigned int)(p - handle_tlbl));
2297 
2298 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2299 }
2300 
2301 static void build_r4000_tlb_store_handler(void)
2302 {
2303 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2304 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2305 	struct uasm_label *l = labels;
2306 	struct uasm_reloc *r = relocs;
2307 	struct work_registers wr;
2308 
2309 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2310 	memset(labels, 0, sizeof(labels));
2311 	memset(relocs, 0, sizeof(relocs));
2312 
2313 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2314 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2315 	if (m4kc_tlbp_war())
2316 		build_tlb_probe_entry(&p);
2317 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2318 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2319 
2320 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2321 	/*
2322 	 * This is the entry point when
2323 	 * build_r4000_tlbchange_handler_head spots a huge page.
2324 	 */
2325 	uasm_l_tlb_huge_update(&l, p);
2326 	iPTE_LW(&p, wr.r1, wr.r2);
2327 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2328 	build_tlb_probe_entry(&p);
2329 	uasm_i_ori(&p, wr.r1, wr.r1,
2330 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2331 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2332 #endif
2333 
2334 	uasm_l_nopage_tlbs(&l, p);
2335 	build_restore_work_registers(&p);
2336 #ifdef CONFIG_CPU_MICROMIPS
2337 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2338 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2339 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2340 		uasm_i_jr(&p, K0);
2341 	} else
2342 #endif
2343 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2344 	uasm_i_nop(&p);
2345 
2346 	if (p >= handle_tlbs_end)
2347 		panic("TLB store handler fastpath space exceeded");
2348 
2349 	uasm_resolve_relocs(relocs, labels);
2350 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2351 		 (unsigned int)(p - handle_tlbs));
2352 
2353 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2354 }
2355 
2356 static void build_r4000_tlb_modify_handler(void)
2357 {
2358 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2359 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2360 	struct uasm_label *l = labels;
2361 	struct uasm_reloc *r = relocs;
2362 	struct work_registers wr;
2363 
2364 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2365 	memset(labels, 0, sizeof(labels));
2366 	memset(relocs, 0, sizeof(relocs));
2367 
2368 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2369 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2370 	if (m4kc_tlbp_war())
2371 		build_tlb_probe_entry(&p);
2372 	/* Present and writable bits set, set accessed and dirty bits. */
2373 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2374 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2375 
2376 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2377 	/*
2378 	 * This is the entry point when
2379 	 * build_r4000_tlbchange_handler_head spots a huge page.
2380 	 */
2381 	uasm_l_tlb_huge_update(&l, p);
2382 	iPTE_LW(&p, wr.r1, wr.r2);
2383 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2384 	build_tlb_probe_entry(&p);
2385 	uasm_i_ori(&p, wr.r1, wr.r1,
2386 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2387 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2388 #endif
2389 
2390 	uasm_l_nopage_tlbm(&l, p);
2391 	build_restore_work_registers(&p);
2392 #ifdef CONFIG_CPU_MICROMIPS
2393 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2394 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2395 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2396 		uasm_i_jr(&p, K0);
2397 	} else
2398 #endif
2399 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2400 	uasm_i_nop(&p);
2401 
2402 	if (p >= handle_tlbm_end)
2403 		panic("TLB modify handler fastpath space exceeded");
2404 
2405 	uasm_resolve_relocs(relocs, labels);
2406 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2407 		 (unsigned int)(p - handle_tlbm));
2408 
2409 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2410 }
2411 
2412 static void flush_tlb_handlers(void)
2413 {
2414 	local_flush_icache_range((unsigned long)handle_tlbl,
2415 			   (unsigned long)handle_tlbl_end);
2416 	local_flush_icache_range((unsigned long)handle_tlbs,
2417 			   (unsigned long)handle_tlbs_end);
2418 	local_flush_icache_range((unsigned long)handle_tlbm,
2419 			   (unsigned long)handle_tlbm_end);
2420 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2421 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2422 }
2423 
2424 static void print_htw_config(void)
2425 {
2426 	unsigned long config;
2427 	unsigned int pwctl;
2428 	const int field = 2 * sizeof(unsigned long);
2429 
2430 	config = read_c0_pwfield();
2431 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2432 		field, config,
2433 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2434 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2435 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2436 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2437 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2438 
2439 	config = read_c0_pwsize();
2440 	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",
2441 		field, config,
2442 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2443 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2444 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2445 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2446 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2447 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2448 
2449 	pwctl = read_c0_pwctl();
2450 	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",
2451 		pwctl,
2452 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2453 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2454 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2455 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2456 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2457 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2458 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2459 }
2460 
2461 static void config_htw_params(void)
2462 {
2463 	unsigned long pwfield, pwsize, ptei;
2464 	unsigned int config;
2465 
2466 	/*
2467 	 * We are using 2-level page tables, so we only need to
2468 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2469 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2470 	 * write values less than 0xc in these fields because the entire
2471 	 * write will be dropped. As a result of which, we must preserve
2472 	 * the original reset values and overwrite only what we really want.
2473 	 */
2474 
2475 	pwfield = read_c0_pwfield();
2476 	/* re-initialize the GDI field */
2477 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2478 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2479 	/* re-initialize the PTI field including the even/odd bit */
2480 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2481 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2482 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2483 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2484 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2485 	}
2486 	/* Set the PTEI right shift */
2487 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2488 	pwfield |= ptei;
2489 	write_c0_pwfield(pwfield);
2490 	/* Check whether the PTEI value is supported */
2491 	back_to_back_c0_hazard();
2492 	pwfield = read_c0_pwfield();
2493 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2494 		!= ptei) {
2495 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2496 			ptei);
2497 		/*
2498 		 * Drop option to avoid HTW being enabled via another path
2499 		 * (eg htw_reset())
2500 		 */
2501 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2502 		return;
2503 	}
2504 
2505 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2506 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2507 	if (CONFIG_PGTABLE_LEVELS >= 3)
2508 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2509 
2510 	/* Set pointer size to size of directory pointers */
2511 	if (IS_ENABLED(CONFIG_64BIT))
2512 		pwsize |= MIPS_PWSIZE_PS_MASK;
2513 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2514 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2515 			& MIPS_PWSIZE_PTEW_MASK;
2516 
2517 	write_c0_pwsize(pwsize);
2518 
2519 	/* Make sure everything is set before we enable the HTW */
2520 	back_to_back_c0_hazard();
2521 
2522 	/*
2523 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2524 	 * the pwctl fields.
2525 	 */
2526 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2527 	if (IS_ENABLED(CONFIG_64BIT))
2528 		config |= MIPS_PWCTL_XU_MASK;
2529 	write_c0_pwctl(config);
2530 	pr_info("Hardware Page Table Walker enabled\n");
2531 
2532 	print_htw_config();
2533 }
2534 
2535 static void config_xpa_params(void)
2536 {
2537 #ifdef CONFIG_XPA
2538 	unsigned int pagegrain;
2539 
2540 	if (mips_xpa_disabled) {
2541 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2542 		return;
2543 	}
2544 
2545 	pagegrain = read_c0_pagegrain();
2546 	write_c0_pagegrain(pagegrain | PG_ELPA);
2547 	back_to_back_c0_hazard();
2548 	pagegrain = read_c0_pagegrain();
2549 
2550 	if (pagegrain & PG_ELPA)
2551 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2552 	else
2553 		panic("Extended Physical Addressing (XPA) disabled");
2554 #endif
2555 }
2556 
2557 static void check_pabits(void)
2558 {
2559 	unsigned long entry;
2560 	unsigned pabits, fillbits;
2561 
2562 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2563 		/*
2564 		 * We'll only be making use of the fact that we can rotate bits
2565 		 * into the fill if the CPU supports RIXI, so don't bother
2566 		 * probing this for CPUs which don't.
2567 		 */
2568 		return;
2569 	}
2570 
2571 	write_c0_entrylo0(~0ul);
2572 	back_to_back_c0_hazard();
2573 	entry = read_c0_entrylo0();
2574 
2575 	/* clear all non-PFN bits */
2576 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2577 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2578 
2579 	/* find a lower bound on PABITS, and upper bound on fill bits */
2580 	pabits = fls_long(entry) + 6;
2581 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2582 
2583 	/* minus the RI & XI bits */
2584 	fillbits -= min_t(unsigned, fillbits, 2);
2585 
2586 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2587 		fill_includes_sw_bits = true;
2588 
2589 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2590 }
2591 
2592 void build_tlb_refill_handler(void)
2593 {
2594 	/*
2595 	 * The refill handler is generated per-CPU, multi-node systems
2596 	 * may have local storage for it. The other handlers are only
2597 	 * needed once.
2598 	 */
2599 	static int run_once = 0;
2600 
2601 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2602 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2603 
2604 	output_pgtable_bits_defines();
2605 	check_pabits();
2606 
2607 #ifdef CONFIG_64BIT
2608 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2609 #endif
2610 
2611 	switch (current_cpu_type()) {
2612 	case CPU_R2000:
2613 	case CPU_R3000:
2614 	case CPU_R3000A:
2615 	case CPU_R3081E:
2616 	case CPU_TX3912:
2617 	case CPU_TX3922:
2618 	case CPU_TX3927:
2619 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2620 		if (cpu_has_local_ebase)
2621 			build_r3000_tlb_refill_handler();
2622 		if (!run_once) {
2623 			if (!cpu_has_local_ebase)
2624 				build_r3000_tlb_refill_handler();
2625 			build_setup_pgd();
2626 			build_r3000_tlb_load_handler();
2627 			build_r3000_tlb_store_handler();
2628 			build_r3000_tlb_modify_handler();
2629 			flush_tlb_handlers();
2630 			run_once++;
2631 		}
2632 #else
2633 		panic("No R3000 TLB refill handler");
2634 #endif
2635 		break;
2636 
2637 	case CPU_R8000:
2638 		panic("No R8000 TLB refill handler yet");
2639 		break;
2640 
2641 	default:
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 if (!cpu_has_local_ebase)
2654 				build_r4000_tlb_refill_handler();
2655 			flush_tlb_handlers();
2656 			run_once++;
2657 		}
2658 		if (cpu_has_local_ebase)
2659 			build_r4000_tlb_refill_handler();
2660 		if (cpu_has_xpa)
2661 			config_xpa_params();
2662 		if (cpu_has_htw)
2663 			config_htw_params();
2664 	}
2665 }
2666