xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision ccd51b9f)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008	 Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009	 Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/mmu_context.h>
35 #include <asm/pgtable.h>
36 #include <asm/war.h>
37 #include <asm/uasm.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
40 
41 static int mips_xpa_disabled;
42 
43 static int __init xpa_disable(char *s)
44 {
45 	mips_xpa_disabled = 1;
46 
47 	return 1;
48 }
49 
50 __setup("noxpa", xpa_disable);
51 
52 /*
53  * TLB load/store/modify handlers.
54  *
55  * Only the fastpath gets synthesized at runtime, the slowpath for
56  * do_page_fault remains normal asm.
57  */
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
60 
61 struct work_registers {
62 	int r1;
63 	int r2;
64 	int r3;
65 };
66 
67 struct tlb_reg_save {
68 	unsigned long a;
69 	unsigned long b;
70 } ____cacheline_aligned_in_smp;
71 
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
73 
74 static inline int r45k_bvahwbug(void)
75 {
76 	/* XXX: We should probe for the presence of this bug, but we don't. */
77 	return 0;
78 }
79 
80 static inline int r4k_250MHZhwbug(void)
81 {
82 	/* XXX: We should probe for the presence of this bug, but we don't. */
83 	return 0;
84 }
85 
86 static inline int __maybe_unused bcm1250_m3_war(void)
87 {
88 	return BCM1250_M3_WAR;
89 }
90 
91 static inline int __maybe_unused r10000_llsc_war(void)
92 {
93 	return R10000_LLSC_WAR;
94 }
95 
96 static int use_bbit_insns(void)
97 {
98 	switch (current_cpu_type()) {
99 	case CPU_CAVIUM_OCTEON:
100 	case CPU_CAVIUM_OCTEON_PLUS:
101 	case CPU_CAVIUM_OCTEON2:
102 	case CPU_CAVIUM_OCTEON3:
103 		return 1;
104 	default:
105 		return 0;
106 	}
107 }
108 
109 static int use_lwx_insns(void)
110 {
111 	switch (current_cpu_type()) {
112 	case CPU_CAVIUM_OCTEON2:
113 	case CPU_CAVIUM_OCTEON3:
114 		return 1;
115 	default:
116 		return 0;
117 	}
118 }
119 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
120     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
121 static bool scratchpad_available(void)
122 {
123 	return true;
124 }
125 static int scratchpad_offset(int i)
126 {
127 	/*
128 	 * CVMSEG starts at address -32768 and extends for
129 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
130 	 */
131 	i += 1; /* Kernel use starts at the top and works down. */
132 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
133 }
134 #else
135 static bool scratchpad_available(void)
136 {
137 	return false;
138 }
139 static int scratchpad_offset(int i)
140 {
141 	BUG();
142 	/* Really unreachable, but evidently some GCC want this. */
143 	return 0;
144 }
145 #endif
146 /*
147  * Found by experiment: At least some revisions of the 4kc throw under
148  * some circumstances a machine check exception, triggered by invalid
149  * values in the index register.  Delaying the tlbp instruction until
150  * after the next branch,  plus adding an additional nop in front of
151  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
152  * why; it's not an issue caused by the core RTL.
153  *
154  */
155 static int m4kc_tlbp_war(void)
156 {
157 	return current_cpu_type() == CPU_4KC;
158 }
159 
160 /* Handle labels (which must be positive integers). */
161 enum label_id {
162 	label_second_part = 1,
163 	label_leave,
164 	label_vmalloc,
165 	label_vmalloc_done,
166 	label_tlbw_hazard_0,
167 	label_split = label_tlbw_hazard_0 + 8,
168 	label_tlbl_goaround1,
169 	label_tlbl_goaround2,
170 	label_nopage_tlbl,
171 	label_nopage_tlbs,
172 	label_nopage_tlbm,
173 	label_smp_pgtable_change,
174 	label_r3000_write_probe_fail,
175 	label_large_segbits_fault,
176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
177 	label_tlb_huge_update,
178 #endif
179 };
180 
181 UASM_L_LA(_second_part)
182 UASM_L_LA(_leave)
183 UASM_L_LA(_vmalloc)
184 UASM_L_LA(_vmalloc_done)
185 /* _tlbw_hazard_x is handled differently.  */
186 UASM_L_LA(_split)
187 UASM_L_LA(_tlbl_goaround1)
188 UASM_L_LA(_tlbl_goaround2)
189 UASM_L_LA(_nopage_tlbl)
190 UASM_L_LA(_nopage_tlbs)
191 UASM_L_LA(_nopage_tlbm)
192 UASM_L_LA(_smp_pgtable_change)
193 UASM_L_LA(_r3000_write_probe_fail)
194 UASM_L_LA(_large_segbits_fault)
195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
196 UASM_L_LA(_tlb_huge_update)
197 #endif
198 
199 static int hazard_instance;
200 
201 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
202 {
203 	switch (instance) {
204 	case 0 ... 7:
205 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
206 		return;
207 	default:
208 		BUG();
209 	}
210 }
211 
212 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
213 {
214 	switch (instance) {
215 	case 0 ... 7:
216 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
217 		break;
218 	default:
219 		BUG();
220 	}
221 }
222 
223 /*
224  * pgtable bits are assigned dynamically depending on processor feature
225  * and statically based on kernel configuration.  This spits out the actual
226  * values the kernel is using.	Required to make sense from disassembled
227  * TLB exception handlers.
228  */
229 static void output_pgtable_bits_defines(void)
230 {
231 #define pr_define(fmt, ...)					\
232 	pr_debug("#define " fmt, ##__VA_ARGS__)
233 
234 	pr_debug("#include <asm/asm.h>\n");
235 	pr_debug("#include <asm/regdef.h>\n");
236 	pr_debug("\n");
237 
238 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
239 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
241 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
242 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
243 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
244 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
245 #endif
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 	if (cpu_has_rixi)
248 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 #endif
250 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
251 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
252 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
253 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
254 	pr_debug("\n");
255 }
256 
257 static inline void dump_handler(const char *symbol, const void *start, const void *end)
258 {
259 	unsigned int count = (end - start) / sizeof(u32);
260 	const u32 *handler = start;
261 	int i;
262 
263 	pr_debug("LEAF(%s)\n", symbol);
264 
265 	pr_debug("\t.set push\n");
266 	pr_debug("\t.set noreorder\n");
267 
268 	for (i = 0; i < count; i++)
269 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
270 
271 	pr_debug("\t.set\tpop\n");
272 
273 	pr_debug("\tEND(%s)\n", symbol);
274 }
275 
276 /* The only general purpose registers allowed in TLB handlers. */
277 #define K0		26
278 #define K1		27
279 
280 /* Some CP0 registers */
281 #define C0_INDEX	0, 0
282 #define C0_ENTRYLO0	2, 0
283 #define C0_TCBIND	2, 2
284 #define C0_ENTRYLO1	3, 0
285 #define C0_CONTEXT	4, 0
286 #define C0_PAGEMASK	5, 0
287 #define C0_PWBASE	5, 5
288 #define C0_PWFIELD	5, 6
289 #define C0_PWSIZE	5, 7
290 #define C0_PWCTL	6, 6
291 #define C0_BADVADDR	8, 0
292 #define C0_PGD		9, 7
293 #define C0_ENTRYHI	10, 0
294 #define C0_EPC		14, 0
295 #define C0_XCONTEXT	20, 0
296 
297 #ifdef CONFIG_64BIT
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
299 #else
300 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
301 #endif
302 
303 /* The worst case length of the handler is around 18 instructions for
304  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
305  * Maximum space available is 32 instructions for R3000 and 64
306  * instructions for R4000.
307  *
308  * We deliberately chose a buffer size of 128, so we won't scribble
309  * over anything important on overflow before we panic.
310  */
311 static u32 tlb_handler[128];
312 
313 /* simply assume worst case size for labels and relocs */
314 static struct uasm_label labels[128];
315 static struct uasm_reloc relocs[128];
316 
317 static int check_for_high_segbits;
318 static bool fill_includes_sw_bits;
319 
320 static unsigned int kscratch_used_mask;
321 
322 static inline int __maybe_unused c0_kscratch(void)
323 {
324 	switch (current_cpu_type()) {
325 	case CPU_XLP:
326 	case CPU_XLR:
327 		return 22;
328 	default:
329 		return 31;
330 	}
331 }
332 
333 static int allocate_kscratch(void)
334 {
335 	int r;
336 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
337 
338 	r = ffs(a);
339 
340 	if (r == 0)
341 		return -1;
342 
343 	r--; /* make it zero based */
344 
345 	kscratch_used_mask |= (1 << r);
346 
347 	return r;
348 }
349 
350 static int scratch_reg;
351 int pgd_reg;
352 EXPORT_SYMBOL_GPL(pgd_reg);
353 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
354 
355 static struct work_registers build_get_work_registers(u32 **p)
356 {
357 	struct work_registers r;
358 
359 	if (scratch_reg >= 0) {
360 		/* Save in CPU local C0_KScratch? */
361 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
362 		r.r1 = K0;
363 		r.r2 = K1;
364 		r.r3 = 1;
365 		return r;
366 	}
367 
368 	if (num_possible_cpus() > 1) {
369 		/* Get smp_processor_id */
370 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
371 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
372 
373 		/* handler_reg_save index in K0 */
374 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
375 
376 		UASM_i_LA(p, K1, (long)&handler_reg_save);
377 		UASM_i_ADDU(p, K0, K0, K1);
378 	} else {
379 		UASM_i_LA(p, K0, (long)&handler_reg_save);
380 	}
381 	/* K0 now points to save area, save $1 and $2  */
382 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
384 
385 	r.r1 = K1;
386 	r.r2 = 1;
387 	r.r3 = 2;
388 	return r;
389 }
390 
391 static void build_restore_work_registers(u32 **p)
392 {
393 	if (scratch_reg >= 0) {
394 		uasm_i_ehb(p);
395 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
396 		return;
397 	}
398 	/* K0 already points to save area, restore $1 and $2  */
399 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
400 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
401 }
402 
403 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
404 
405 /*
406  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
407  * we cannot do r3000 under these circumstances.
408  *
409  * The R3000 TLB handler is simple.
410  */
411 static void build_r3000_tlb_refill_handler(void)
412 {
413 	long pgdc = (long)pgd_current;
414 	u32 *p;
415 
416 	memset(tlb_handler, 0, sizeof(tlb_handler));
417 	p = tlb_handler;
418 
419 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
420 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
421 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
422 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
423 	uasm_i_sll(&p, K0, K0, 2);
424 	uasm_i_addu(&p, K1, K1, K0);
425 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
426 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
427 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
428 	uasm_i_addu(&p, K1, K1, K0);
429 	uasm_i_lw(&p, K0, 0, K1);
430 	uasm_i_nop(&p); /* load delay */
431 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
432 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
433 	uasm_i_tlbwr(&p); /* cp0 delay */
434 	uasm_i_jr(&p, K1);
435 	uasm_i_rfe(&p); /* branch delay */
436 
437 	if (p > tlb_handler + 32)
438 		panic("TLB refill handler space exceeded");
439 
440 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
441 		 (unsigned int)(p - tlb_handler));
442 
443 	memcpy((void *)ebase, tlb_handler, 0x80);
444 	local_flush_icache_range(ebase, ebase + 0x80);
445 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
446 }
447 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
448 
449 /*
450  * The R4000 TLB handler is much more complicated. We have two
451  * consecutive handler areas with 32 instructions space each.
452  * Since they aren't used at the same time, we can overflow in the
453  * other one.To keep things simple, we first assume linear space,
454  * then we relocate it to the final handler layout as needed.
455  */
456 static u32 final_handler[64];
457 
458 /*
459  * Hazards
460  *
461  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
462  * 2. A timing hazard exists for the TLBP instruction.
463  *
464  *	stalling_instruction
465  *	TLBP
466  *
467  * The JTLB is being read for the TLBP throughout the stall generated by the
468  * previous instruction. This is not really correct as the stalling instruction
469  * can modify the address used to access the JTLB.  The failure symptom is that
470  * the TLBP instruction will use an address created for the stalling instruction
471  * and not the address held in C0_ENHI and thus report the wrong results.
472  *
473  * The software work-around is to not allow the instruction preceding the TLBP
474  * to stall - make it an NOP or some other instruction guaranteed not to stall.
475  *
476  * Errata 2 will not be fixed.	This errata is also on the R5000.
477  *
478  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
479  */
480 static void __maybe_unused build_tlb_probe_entry(u32 **p)
481 {
482 	switch (current_cpu_type()) {
483 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
484 	case CPU_R4600:
485 	case CPU_R4700:
486 	case CPU_R5000:
487 	case CPU_NEVADA:
488 		uasm_i_nop(p);
489 		uasm_i_tlbp(p);
490 		break;
491 
492 	default:
493 		uasm_i_tlbp(p);
494 		break;
495 	}
496 }
497 
498 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
499 			   struct uasm_reloc **r,
500 			   enum tlb_write_entry wmode)
501 {
502 	void(*tlbw)(u32 **) = NULL;
503 
504 	switch (wmode) {
505 	case tlb_random: tlbw = uasm_i_tlbwr; break;
506 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
507 	}
508 
509 	if (cpu_has_mips_r2_r6) {
510 		if (cpu_has_mips_r2_exec_hazard)
511 			uasm_i_ehb(p);
512 		tlbw(p);
513 		return;
514 	}
515 
516 	switch (current_cpu_type()) {
517 	case CPU_R4000PC:
518 	case CPU_R4000SC:
519 	case CPU_R4000MC:
520 	case CPU_R4400PC:
521 	case CPU_R4400SC:
522 	case CPU_R4400MC:
523 		/*
524 		 * This branch uses up a mtc0 hazard nop slot and saves
525 		 * two nops after the tlbw instruction.
526 		 */
527 		uasm_bgezl_hazard(p, r, hazard_instance);
528 		tlbw(p);
529 		uasm_bgezl_label(l, p, hazard_instance);
530 		hazard_instance++;
531 		uasm_i_nop(p);
532 		break;
533 
534 	case CPU_R4600:
535 	case CPU_R4700:
536 		uasm_i_nop(p);
537 		tlbw(p);
538 		uasm_i_nop(p);
539 		break;
540 
541 	case CPU_R5000:
542 	case CPU_NEVADA:
543 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
544 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
545 		tlbw(p);
546 		break;
547 
548 	case CPU_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 		/* fall through */
580 	case CPU_ALCHEMY:
581 		tlbw(p);
582 		break;
583 
584 	case CPU_RM7000:
585 		uasm_i_nop(p);
586 		uasm_i_nop(p);
587 		uasm_i_nop(p);
588 		uasm_i_nop(p);
589 		tlbw(p);
590 		break;
591 
592 	case CPU_VR4111:
593 	case CPU_VR4121:
594 	case CPU_VR4122:
595 	case CPU_VR4181:
596 	case CPU_VR4181A:
597 		uasm_i_nop(p);
598 		uasm_i_nop(p);
599 		tlbw(p);
600 		uasm_i_nop(p);
601 		uasm_i_nop(p);
602 		break;
603 
604 	case CPU_VR4131:
605 	case CPU_VR4133:
606 		uasm_i_nop(p);
607 		uasm_i_nop(p);
608 		tlbw(p);
609 		break;
610 
611 	case CPU_XBURST:
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_ehb(p);
672 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
673 		} else {
674 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
675 		}
676 	} else {
677 		/* Reset default page size */
678 		if (PM_DEFAULT_MASK >> 16) {
679 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
680 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
681 			uasm_il_b(p, r, lid);
682 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
683 		} else if (PM_DEFAULT_MASK) {
684 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
685 			uasm_il_b(p, r, lid);
686 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
687 		} else {
688 			uasm_il_b(p, r, lid);
689 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
690 		}
691 	}
692 }
693 
694 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
695 				       struct uasm_reloc **r,
696 				       unsigned int tmp,
697 				       enum tlb_write_entry wmode,
698 				       int restore_scratch)
699 {
700 	/* Set huge page tlb entry size */
701 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
702 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
703 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
704 
705 	build_tlb_write_entry(p, l, r, wmode);
706 
707 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
708 }
709 
710 /*
711  * Check if Huge PTE is present, if so then jump to LABEL.
712  */
713 static void
714 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
715 		  unsigned int pmd, int lid)
716 {
717 	UASM_i_LW(p, tmp, 0, pmd);
718 	if (use_bbit_insns()) {
719 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
720 	} else {
721 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
722 		uasm_il_bnez(p, r, tmp, lid);
723 	}
724 }
725 
726 static void build_huge_update_entries(u32 **p, unsigned int pte,
727 				      unsigned int tmp)
728 {
729 	int small_sequence;
730 
731 	/*
732 	 * A huge PTE describes an area the size of the
733 	 * configured huge page size. This is twice the
734 	 * of the large TLB entry size we intend to use.
735 	 * A TLB entry half the size of the configured
736 	 * huge page size is configured into entrylo0
737 	 * and entrylo1 to cover the contiguous huge PTE
738 	 * address space.
739 	 */
740 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
741 
742 	/* We can clobber tmp.	It isn't used after this.*/
743 	if (!small_sequence)
744 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
745 
746 	build_convert_pte_to_entrylo(p, pte);
747 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
748 	/* convert to entrylo1 */
749 	if (small_sequence)
750 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
751 	else
752 		UASM_i_ADDU(p, pte, pte, tmp);
753 
754 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
755 }
756 
757 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
758 				    struct uasm_label **l,
759 				    unsigned int pte,
760 				    unsigned int ptr,
761 				    unsigned int flush)
762 {
763 #ifdef CONFIG_SMP
764 	UASM_i_SC(p, pte, 0, ptr);
765 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
766 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
767 #else
768 	UASM_i_SW(p, pte, 0, ptr);
769 #endif
770 	if (cpu_has_ftlb && flush) {
771 		BUG_ON(!cpu_has_tlbinv);
772 
773 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
774 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
775 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
776 		build_tlb_write_entry(p, l, r, tlb_indexed);
777 
778 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
779 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
780 		build_huge_update_entries(p, pte, ptr);
781 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
782 
783 		return;
784 	}
785 
786 	build_huge_update_entries(p, pte, ptr);
787 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
788 }
789 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
790 
791 #ifdef CONFIG_64BIT
792 /*
793  * TMP and PTR are scratch.
794  * TMP will be clobbered, PTR will hold the pmd entry.
795  */
796 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
797 		      unsigned int tmp, unsigned int ptr)
798 {
799 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
800 	long pgdc = (long)pgd_current;
801 #endif
802 	/*
803 	 * The vmalloc handling is not in the hotpath.
804 	 */
805 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
806 
807 	if (check_for_high_segbits) {
808 		/*
809 		 * The kernel currently implicitely assumes that the
810 		 * MIPS SEGBITS parameter for the processor is
811 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
812 		 * allocate virtual addresses outside the maximum
813 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
814 		 * that doesn't prevent user code from accessing the
815 		 * higher xuseg addresses.  Here, we make sure that
816 		 * everything but the lower xuseg addresses goes down
817 		 * the module_alloc/vmalloc path.
818 		 */
819 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
820 		uasm_il_bnez(p, r, ptr, label_vmalloc);
821 	} else {
822 		uasm_il_bltz(p, r, tmp, label_vmalloc);
823 	}
824 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
825 
826 	if (pgd_reg != -1) {
827 		/* pgd is in pgd_reg */
828 		if (cpu_has_ldpte)
829 			UASM_i_MFC0(p, ptr, C0_PWBASE);
830 		else
831 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
832 	} else {
833 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
834 		/*
835 		 * &pgd << 11 stored in CONTEXT [23..63].
836 		 */
837 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
838 
839 		/* Clear lower 23 bits of context. */
840 		uasm_i_dins(p, ptr, 0, 0, 23);
841 
842 		/* 1 0	1 0 1  << 6  xkphys cached */
843 		uasm_i_ori(p, ptr, ptr, 0x540);
844 		uasm_i_drotr(p, ptr, ptr, 11);
845 #elif defined(CONFIG_SMP)
846 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
847 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
848 		UASM_i_LA_mostly(p, tmp, pgdc);
849 		uasm_i_daddu(p, ptr, ptr, tmp);
850 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
851 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
852 #else
853 		UASM_i_LA_mostly(p, ptr, pgdc);
854 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
855 #endif
856 	}
857 
858 	uasm_l_vmalloc_done(l, *p);
859 
860 	/* get pgd offset in bytes */
861 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
862 
863 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
864 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
865 #ifndef __PAGETABLE_PUD_FOLDED
866 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
867 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
868 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
869 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
870 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
871 #endif
872 #ifndef __PAGETABLE_PMD_FOLDED
873 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
874 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
875 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
876 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
877 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
878 #endif
879 }
880 EXPORT_SYMBOL_GPL(build_get_pmde64);
881 
882 /*
883  * BVADDR is the faulting address, PTR is scratch.
884  * PTR will hold the pgd for vmalloc.
885  */
886 static void
887 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
888 			unsigned int bvaddr, unsigned int ptr,
889 			enum vmalloc64_mode mode)
890 {
891 	long swpd = (long)swapper_pg_dir;
892 	int single_insn_swpd;
893 	int did_vmalloc_branch = 0;
894 
895 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
896 
897 	uasm_l_vmalloc(l, *p);
898 
899 	if (mode != not_refill && check_for_high_segbits) {
900 		if (single_insn_swpd) {
901 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
902 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
903 			did_vmalloc_branch = 1;
904 			/* fall through */
905 		} else {
906 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
907 		}
908 	}
909 	if (!did_vmalloc_branch) {
910 		if (single_insn_swpd) {
911 			uasm_il_b(p, r, label_vmalloc_done);
912 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
913 		} else {
914 			UASM_i_LA_mostly(p, ptr, swpd);
915 			uasm_il_b(p, r, label_vmalloc_done);
916 			if (uasm_in_compat_space_p(swpd))
917 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
918 			else
919 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
920 		}
921 	}
922 	if (mode != not_refill && check_for_high_segbits) {
923 		uasm_l_large_segbits_fault(l, *p);
924 		/*
925 		 * We get here if we are an xsseg address, or if we are
926 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
927 		 *
928 		 * Ignoring xsseg (assume disabled so would generate
929 		 * (address errors?), the only remaining possibility
930 		 * is the upper xuseg addresses.  On processors with
931 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
932 		 * addresses would have taken an address error. We try
933 		 * to mimic that here by taking a load/istream page
934 		 * fault.
935 		 */
936 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
937 			uasm_i_sync(p, 0);
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_ehb(p);
944 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
945 			} else {
946 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
947 			}
948 		} else {
949 			uasm_i_nop(p);
950 		}
951 	}
952 }
953 
954 #else /* !CONFIG_64BIT */
955 
956 /*
957  * TMP and PTR are scratch.
958  * TMP will be clobbered, PTR will hold the pgd entry.
959  */
960 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
961 {
962 	if (pgd_reg != -1) {
963 		/* pgd is in pgd_reg */
964 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
965 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
966 	} else {
967 		long pgdc = (long)pgd_current;
968 
969 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
970 #ifdef CONFIG_SMP
971 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
972 		UASM_i_LA_mostly(p, tmp, pgdc);
973 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
974 		uasm_i_addu(p, ptr, tmp, ptr);
975 #else
976 		UASM_i_LA_mostly(p, ptr, pgdc);
977 #endif
978 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
979 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
980 	}
981 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
982 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
983 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
984 }
985 EXPORT_SYMBOL_GPL(build_get_pgde32);
986 
987 #endif /* !CONFIG_64BIT */
988 
989 static void build_adjust_context(u32 **p, unsigned int ctx)
990 {
991 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
992 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
993 
994 	switch (current_cpu_type()) {
995 	case CPU_VR41XX:
996 	case CPU_VR4111:
997 	case CPU_VR4121:
998 	case CPU_VR4122:
999 	case CPU_VR4131:
1000 	case CPU_VR4181:
1001 	case CPU_VR4181A:
1002 	case CPU_VR4133:
1003 		shift += 2;
1004 		break;
1005 
1006 	default:
1007 		break;
1008 	}
1009 
1010 	if (shift)
1011 		UASM_i_SRL(p, ctx, ctx, shift);
1012 	uasm_i_andi(p, ctx, ctx, mask);
1013 }
1014 
1015 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1016 {
1017 	/*
1018 	 * Bug workaround for the Nevada. It seems as if under certain
1019 	 * circumstances the move from cp0_context might produce a
1020 	 * bogus result when the mfc0 instruction and its consumer are
1021 	 * in a different cacheline or a load instruction, probably any
1022 	 * memory reference, is between them.
1023 	 */
1024 	switch (current_cpu_type()) {
1025 	case CPU_NEVADA:
1026 		UASM_i_LW(p, ptr, 0, ptr);
1027 		GET_CONTEXT(p, tmp); /* get context reg */
1028 		break;
1029 
1030 	default:
1031 		GET_CONTEXT(p, tmp); /* get context reg */
1032 		UASM_i_LW(p, ptr, 0, ptr);
1033 		break;
1034 	}
1035 
1036 	build_adjust_context(p, tmp);
1037 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1038 }
1039 EXPORT_SYMBOL_GPL(build_get_ptep);
1040 
1041 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1042 {
1043 	int pte_off_even = 0;
1044 	int pte_off_odd = sizeof(pte_t);
1045 
1046 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1047 	/* The low 32 bits of EntryLo is stored in pte_high */
1048 	pte_off_even += offsetof(pte_t, pte_high);
1049 	pte_off_odd += offsetof(pte_t, pte_high);
1050 #endif
1051 
1052 	if (IS_ENABLED(CONFIG_XPA)) {
1053 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1054 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1055 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1056 
1057 		if (cpu_has_xpa && !mips_xpa_disabled) {
1058 			uasm_i_lw(p, tmp, 0, ptep);
1059 			uasm_i_ext(p, tmp, tmp, 0, 24);
1060 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1061 		}
1062 
1063 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1064 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1065 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1066 
1067 		if (cpu_has_xpa && !mips_xpa_disabled) {
1068 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1069 			uasm_i_ext(p, tmp, tmp, 0, 24);
1070 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1071 		}
1072 		return;
1073 	}
1074 
1075 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1076 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1077 	if (r45k_bvahwbug())
1078 		build_tlb_probe_entry(p);
1079 	build_convert_pte_to_entrylo(p, tmp);
1080 	if (r4k_250MHZhwbug())
1081 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1082 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1083 	build_convert_pte_to_entrylo(p, ptep);
1084 	if (r45k_bvahwbug())
1085 		uasm_i_mfc0(p, tmp, C0_INDEX);
1086 	if (r4k_250MHZhwbug())
1087 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1088 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1089 }
1090 EXPORT_SYMBOL_GPL(build_update_entries);
1091 
1092 struct mips_huge_tlb_info {
1093 	int huge_pte;
1094 	int restore_scratch;
1095 	bool need_reload_pte;
1096 };
1097 
1098 static struct mips_huge_tlb_info
1099 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1100 			       struct uasm_reloc **r, unsigned int tmp,
1101 			       unsigned int ptr, int c0_scratch_reg)
1102 {
1103 	struct mips_huge_tlb_info rv;
1104 	unsigned int even, odd;
1105 	int vmalloc_branch_delay_filled = 0;
1106 	const int scratch = 1; /* Our extra working register */
1107 
1108 	rv.huge_pte = scratch;
1109 	rv.restore_scratch = 0;
1110 	rv.need_reload_pte = false;
1111 
1112 	if (check_for_high_segbits) {
1113 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1114 
1115 		if (pgd_reg != -1)
1116 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1117 		else
1118 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1119 
1120 		if (c0_scratch_reg >= 0)
1121 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1122 		else
1123 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1124 
1125 		uasm_i_dsrl_safe(p, scratch, tmp,
1126 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1127 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1128 
1129 		if (pgd_reg == -1) {
1130 			vmalloc_branch_delay_filled = 1;
1131 			/* Clear lower 23 bits of context. */
1132 			uasm_i_dins(p, ptr, 0, 0, 23);
1133 		}
1134 	} else {
1135 		if (pgd_reg != -1)
1136 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1137 		else
1138 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1139 
1140 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1141 
1142 		if (c0_scratch_reg >= 0)
1143 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1144 		else
1145 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1146 
1147 		if (pgd_reg == -1)
1148 			/* Clear lower 23 bits of context. */
1149 			uasm_i_dins(p, ptr, 0, 0, 23);
1150 
1151 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1152 	}
1153 
1154 	if (pgd_reg == -1) {
1155 		vmalloc_branch_delay_filled = 1;
1156 		/* 1 0	1 0 1  << 6  xkphys cached */
1157 		uasm_i_ori(p, ptr, ptr, 0x540);
1158 		uasm_i_drotr(p, ptr, ptr, 11);
1159 	}
1160 
1161 #ifdef __PAGETABLE_PMD_FOLDED
1162 #define LOC_PTEP scratch
1163 #else
1164 #define LOC_PTEP ptr
1165 #endif
1166 
1167 	if (!vmalloc_branch_delay_filled)
1168 		/* get pgd offset in bytes */
1169 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1170 
1171 	uasm_l_vmalloc_done(l, *p);
1172 
1173 	/*
1174 	 *			   tmp		ptr
1175 	 * fall-through case =	 badvaddr  *pgd_current
1176 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1177 	 */
1178 
1179 	if (vmalloc_branch_delay_filled)
1180 		/* get pgd offset in bytes */
1181 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1182 
1183 #ifdef __PAGETABLE_PMD_FOLDED
1184 	GET_CONTEXT(p, tmp); /* get context reg */
1185 #endif
1186 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1187 
1188 	if (use_lwx_insns()) {
1189 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1190 	} else {
1191 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1192 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1193 	}
1194 
1195 #ifndef __PAGETABLE_PUD_FOLDED
1196 	/* get pud offset in bytes */
1197 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1198 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1199 
1200 	if (use_lwx_insns()) {
1201 		UASM_i_LWX(p, ptr, scratch, ptr);
1202 	} else {
1203 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1204 		UASM_i_LW(p, ptr, 0, ptr);
1205 	}
1206 	/* ptr contains a pointer to PMD entry */
1207 	/* tmp contains the address */
1208 #endif
1209 
1210 #ifndef __PAGETABLE_PMD_FOLDED
1211 	/* get pmd offset in bytes */
1212 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1213 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1214 	GET_CONTEXT(p, tmp); /* get context reg */
1215 
1216 	if (use_lwx_insns()) {
1217 		UASM_i_LWX(p, scratch, scratch, ptr);
1218 	} else {
1219 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1220 		UASM_i_LW(p, scratch, 0, ptr);
1221 	}
1222 #endif
1223 	/* Adjust the context during the load latency. */
1224 	build_adjust_context(p, tmp);
1225 
1226 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1227 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1228 	/*
1229 	 * The in the LWX case we don't want to do the load in the
1230 	 * delay slot.	It cannot issue in the same cycle and may be
1231 	 * speculative and unneeded.
1232 	 */
1233 	if (use_lwx_insns())
1234 		uasm_i_nop(p);
1235 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1236 
1237 
1238 	/* build_update_entries */
1239 	if (use_lwx_insns()) {
1240 		even = ptr;
1241 		odd = tmp;
1242 		UASM_i_LWX(p, even, scratch, tmp);
1243 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1244 		UASM_i_LWX(p, odd, scratch, tmp);
1245 	} else {
1246 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1247 		even = tmp;
1248 		odd = ptr;
1249 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1250 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1251 	}
1252 	if (cpu_has_rixi) {
1253 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1254 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1255 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1256 	} else {
1257 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1258 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1259 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1260 	}
1261 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1262 
1263 	if (c0_scratch_reg >= 0) {
1264 		uasm_i_ehb(p);
1265 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1266 		build_tlb_write_entry(p, l, r, tlb_random);
1267 		uasm_l_leave(l, *p);
1268 		rv.restore_scratch = 1;
1269 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1270 		build_tlb_write_entry(p, l, r, tlb_random);
1271 		uasm_l_leave(l, *p);
1272 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1273 	} else {
1274 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1275 		build_tlb_write_entry(p, l, r, tlb_random);
1276 		uasm_l_leave(l, *p);
1277 		rv.restore_scratch = 1;
1278 	}
1279 
1280 	uasm_i_eret(p); /* return from trap */
1281 
1282 	return rv;
1283 }
1284 
1285 /*
1286  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1287  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1288  * slots before the XTLB refill exception handler which belong to the
1289  * unused TLB refill exception.
1290  */
1291 #define MIPS64_REFILL_INSNS 32
1292 
1293 static void build_r4000_tlb_refill_handler(void)
1294 {
1295 	u32 *p = tlb_handler;
1296 	struct uasm_label *l = labels;
1297 	struct uasm_reloc *r = relocs;
1298 	u32 *f;
1299 	unsigned int final_len;
1300 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1301 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1302 
1303 	memset(tlb_handler, 0, sizeof(tlb_handler));
1304 	memset(labels, 0, sizeof(labels));
1305 	memset(relocs, 0, sizeof(relocs));
1306 	memset(final_handler, 0, sizeof(final_handler));
1307 
1308 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1309 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1310 							  scratch_reg);
1311 		vmalloc_mode = refill_scratch;
1312 	} else {
1313 		htlb_info.huge_pte = K0;
1314 		htlb_info.restore_scratch = 0;
1315 		htlb_info.need_reload_pte = true;
1316 		vmalloc_mode = refill_noscratch;
1317 		/*
1318 		 * create the plain linear handler
1319 		 */
1320 		if (bcm1250_m3_war()) {
1321 			unsigned int segbits = 44;
1322 
1323 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1324 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1325 			uasm_i_xor(&p, K0, K0, K1);
1326 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1327 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1328 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1329 			uasm_i_or(&p, K0, K0, K1);
1330 			uasm_il_bnez(&p, &r, K0, label_leave);
1331 			/* No need for uasm_i_nop */
1332 		}
1333 
1334 #ifdef CONFIG_64BIT
1335 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1336 #else
1337 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1338 #endif
1339 
1340 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1341 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1342 #endif
1343 
1344 		build_get_ptep(&p, K0, K1);
1345 		build_update_entries(&p, K0, K1);
1346 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1347 		uasm_l_leave(&l, p);
1348 		uasm_i_eret(&p); /* return from trap */
1349 	}
1350 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1351 	uasm_l_tlb_huge_update(&l, p);
1352 	if (htlb_info.need_reload_pte)
1353 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1354 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1355 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1356 				   htlb_info.restore_scratch);
1357 #endif
1358 
1359 #ifdef CONFIG_64BIT
1360 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1361 #endif
1362 
1363 	/*
1364 	 * Overflow check: For the 64bit handler, we need at least one
1365 	 * free instruction slot for the wrap-around branch. In worst
1366 	 * case, if the intended insertion point is a delay slot, we
1367 	 * need three, with the second nop'ed and the third being
1368 	 * unused.
1369 	 */
1370 	switch (boot_cpu_type()) {
1371 	default:
1372 		if (sizeof(long) == 4) {
1373 	case CPU_LOONGSON2:
1374 		/* Loongson2 ebase is different than r4k, we have more space */
1375 			if ((p - tlb_handler) > 64)
1376 				panic("TLB refill handler space exceeded");
1377 			/*
1378 			 * Now fold the handler in the TLB refill handler space.
1379 			 */
1380 			f = final_handler;
1381 			/* Simplest case, just copy the handler. */
1382 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1383 			final_len = p - tlb_handler;
1384 			break;
1385 		} else {
1386 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1387 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1388 				&& uasm_insn_has_bdelay(relocs,
1389 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1390 				panic("TLB refill handler space exceeded");
1391 			/*
1392 			 * Now fold the handler in the TLB refill handler space.
1393 			 */
1394 			f = final_handler + MIPS64_REFILL_INSNS;
1395 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1396 				/* Just copy the handler. */
1397 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1398 				final_len = p - tlb_handler;
1399 			} else {
1400 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1401 				const enum label_id ls = label_tlb_huge_update;
1402 #else
1403 				const enum label_id ls = label_vmalloc;
1404 #endif
1405 				u32 *split;
1406 				int ov = 0;
1407 				int i;
1408 
1409 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1410 					;
1411 				BUG_ON(i == ARRAY_SIZE(labels));
1412 				split = labels[i].addr;
1413 
1414 				/*
1415 				 * See if we have overflown one way or the other.
1416 				 */
1417 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1418 				    split < p - MIPS64_REFILL_INSNS)
1419 					ov = 1;
1420 
1421 				if (ov) {
1422 					/*
1423 					 * Split two instructions before the end.  One
1424 					 * for the branch and one for the instruction
1425 					 * in the delay slot.
1426 					 */
1427 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1428 
1429 					/*
1430 					 * If the branch would fall in a delay slot,
1431 					 * we must back up an additional instruction
1432 					 * so that it is no longer in a delay slot.
1433 					 */
1434 					if (uasm_insn_has_bdelay(relocs, split - 1))
1435 						split--;
1436 				}
1437 				/* Copy first part of the handler. */
1438 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1439 				f += split - tlb_handler;
1440 
1441 				if (ov) {
1442 					/* Insert branch. */
1443 					uasm_l_split(&l, final_handler);
1444 					uasm_il_b(&f, &r, label_split);
1445 					if (uasm_insn_has_bdelay(relocs, split))
1446 						uasm_i_nop(&f);
1447 					else {
1448 						uasm_copy_handler(relocs, labels,
1449 								  split, split + 1, f);
1450 						uasm_move_labels(labels, f, f + 1, -1);
1451 						f++;
1452 						split++;
1453 					}
1454 				}
1455 
1456 				/* Copy the rest of the handler. */
1457 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1458 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1459 					    (p - split);
1460 			}
1461 		}
1462 		break;
1463 	}
1464 
1465 	uasm_resolve_relocs(relocs, labels);
1466 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1467 		 final_len);
1468 
1469 	memcpy((void *)ebase, final_handler, 0x100);
1470 	local_flush_icache_range(ebase, ebase + 0x100);
1471 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1472 }
1473 
1474 static void setup_pw(void)
1475 {
1476 	unsigned long pgd_i, pgd_w;
1477 #ifndef __PAGETABLE_PMD_FOLDED
1478 	unsigned long pmd_i, pmd_w;
1479 #endif
1480 	unsigned long pt_i, pt_w;
1481 	unsigned long pte_i, pte_w;
1482 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1483 	unsigned long psn;
1484 
1485 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1486 #endif
1487 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1488 #ifndef __PAGETABLE_PMD_FOLDED
1489 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1490 
1491 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1492 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1493 #else
1494 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1495 #endif
1496 
1497 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1498 	pt_w  = PAGE_SHIFT - 3;
1499 
1500 	pte_i = ilog2(_PAGE_GLOBAL);
1501 	pte_w = 0;
1502 
1503 #ifndef __PAGETABLE_PMD_FOLDED
1504 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1505 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1506 #else
1507 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1508 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1509 #endif
1510 
1511 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1512 	write_c0_pwctl(1 << 6 | psn);
1513 #endif
1514 	write_c0_kpgd((long)swapper_pg_dir);
1515 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1516 }
1517 
1518 static void build_loongson3_tlb_refill_handler(void)
1519 {
1520 	u32 *p = tlb_handler;
1521 	struct uasm_label *l = labels;
1522 	struct uasm_reloc *r = relocs;
1523 
1524 	memset(labels, 0, sizeof(labels));
1525 	memset(relocs, 0, sizeof(relocs));
1526 	memset(tlb_handler, 0, sizeof(tlb_handler));
1527 
1528 	if (check_for_high_segbits) {
1529 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1530 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1531 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1532 		uasm_i_nop(&p);
1533 
1534 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1535 		uasm_i_nop(&p);
1536 		uasm_l_vmalloc(&l, p);
1537 	}
1538 
1539 	uasm_i_dmfc0(&p, K1, C0_PGD);
1540 
1541 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1542 #ifndef __PAGETABLE_PMD_FOLDED
1543 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1544 #endif
1545 	uasm_i_ldpte(&p, K1, 0);      /* even */
1546 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1547 	uasm_i_tlbwr(&p);
1548 
1549 	/* restore page mask */
1550 	if (PM_DEFAULT_MASK >> 16) {
1551 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1552 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1553 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1554 	} else if (PM_DEFAULT_MASK) {
1555 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1556 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1557 	} else {
1558 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1559 	}
1560 
1561 	uasm_i_eret(&p);
1562 
1563 	if (check_for_high_segbits) {
1564 		uasm_l_large_segbits_fault(&l, p);
1565 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1566 		uasm_i_jr(&p, K1);
1567 		uasm_i_nop(&p);
1568 	}
1569 
1570 	uasm_resolve_relocs(relocs, labels);
1571 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1572 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1573 	dump_handler("loongson3_tlb_refill",
1574 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1575 }
1576 
1577 static void build_setup_pgd(void)
1578 {
1579 	const int a0 = 4;
1580 	const int __maybe_unused a1 = 5;
1581 	const int __maybe_unused a2 = 6;
1582 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1583 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1584 	long pgdc = (long)pgd_current;
1585 #endif
1586 
1587 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1588 	memset(labels, 0, sizeof(labels));
1589 	memset(relocs, 0, sizeof(relocs));
1590 	pgd_reg = allocate_kscratch();
1591 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1592 	if (pgd_reg == -1) {
1593 		struct uasm_label *l = labels;
1594 		struct uasm_reloc *r = relocs;
1595 
1596 		/* PGD << 11 in c0_Context */
1597 		/*
1598 		 * If it is a ckseg0 address, convert to a physical
1599 		 * address.  Shifting right by 29 and adding 4 will
1600 		 * result in zero for these addresses.
1601 		 *
1602 		 */
1603 		UASM_i_SRA(&p, a1, a0, 29);
1604 		UASM_i_ADDIU(&p, a1, a1, 4);
1605 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1606 		uasm_i_nop(&p);
1607 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1608 		uasm_l_tlbl_goaround1(&l, p);
1609 		UASM_i_SLL(&p, a0, a0, 11);
1610 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1611 		uasm_i_jr(&p, 31);
1612 		uasm_i_ehb(&p);
1613 	} else {
1614 		/* PGD in c0_KScratch */
1615 		if (cpu_has_ldpte)
1616 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1617 		else
1618 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1619 		uasm_i_jr(&p, 31);
1620 		uasm_i_ehb(&p);
1621 	}
1622 #else
1623 #ifdef CONFIG_SMP
1624 	/* Save PGD to pgd_current[smp_processor_id()] */
1625 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1626 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1627 	UASM_i_LA_mostly(&p, a2, pgdc);
1628 	UASM_i_ADDU(&p, a2, a2, a1);
1629 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1630 #else
1631 	UASM_i_LA_mostly(&p, a2, pgdc);
1632 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1633 #endif /* SMP */
1634 
1635 	/* if pgd_reg is allocated, save PGD also to scratch register */
1636 	if (pgd_reg != -1) {
1637 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1638 		uasm_i_jr(&p, 31);
1639 		uasm_i_ehb(&p);
1640 	} else {
1641 		uasm_i_jr(&p, 31);
1642 		uasm_i_nop(&p);
1643 	}
1644 #endif
1645 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1646 		panic("tlbmiss_handler_setup_pgd space exceeded");
1647 
1648 	uasm_resolve_relocs(relocs, labels);
1649 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1650 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1651 
1652 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1653 					tlbmiss_handler_setup_pgd_end);
1654 }
1655 
1656 static void
1657 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1658 {
1659 #ifdef CONFIG_SMP
1660 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1661 		uasm_i_sync(p, 0);
1662 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1663 	if (cpu_has_64bits)
1664 		uasm_i_lld(p, pte, 0, ptr);
1665 	else
1666 # endif
1667 		UASM_i_LL(p, pte, 0, ptr);
1668 #else
1669 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1670 	if (cpu_has_64bits)
1671 		uasm_i_ld(p, pte, 0, ptr);
1672 	else
1673 # endif
1674 		UASM_i_LW(p, pte, 0, ptr);
1675 #endif
1676 }
1677 
1678 static void
1679 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1680 	unsigned int mode, unsigned int scratch)
1681 {
1682 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1683 	unsigned int swmode = mode & ~hwmode;
1684 
1685 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1686 		uasm_i_lui(p, scratch, swmode >> 16);
1687 		uasm_i_or(p, pte, pte, scratch);
1688 		BUG_ON(swmode & 0xffff);
1689 	} else {
1690 		uasm_i_ori(p, pte, pte, mode);
1691 	}
1692 
1693 #ifdef CONFIG_SMP
1694 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1695 	if (cpu_has_64bits)
1696 		uasm_i_scd(p, pte, 0, ptr);
1697 	else
1698 # endif
1699 		UASM_i_SC(p, pte, 0, ptr);
1700 
1701 	if (r10000_llsc_war())
1702 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1703 	else
1704 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1705 
1706 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1707 	if (!cpu_has_64bits) {
1708 		/* no uasm_i_nop needed */
1709 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1710 		uasm_i_ori(p, pte, pte, hwmode);
1711 		BUG_ON(hwmode & ~0xffff);
1712 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1713 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1714 		/* no uasm_i_nop needed */
1715 		uasm_i_lw(p, pte, 0, ptr);
1716 	} else
1717 		uasm_i_nop(p);
1718 # else
1719 	uasm_i_nop(p);
1720 # endif
1721 #else
1722 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1723 	if (cpu_has_64bits)
1724 		uasm_i_sd(p, pte, 0, ptr);
1725 	else
1726 # endif
1727 		UASM_i_SW(p, pte, 0, ptr);
1728 
1729 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1730 	if (!cpu_has_64bits) {
1731 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1732 		uasm_i_ori(p, pte, pte, hwmode);
1733 		BUG_ON(hwmode & ~0xffff);
1734 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1735 		uasm_i_lw(p, pte, 0, ptr);
1736 	}
1737 # endif
1738 #endif
1739 }
1740 
1741 /*
1742  * Check if PTE is present, if not then jump to LABEL. PTR points to
1743  * the page table where this PTE is located, PTE will be re-loaded
1744  * with it's original value.
1745  */
1746 static void
1747 build_pte_present(u32 **p, struct uasm_reloc **r,
1748 		  int pte, int ptr, int scratch, enum label_id lid)
1749 {
1750 	int t = scratch >= 0 ? scratch : pte;
1751 	int cur = pte;
1752 
1753 	if (cpu_has_rixi) {
1754 		if (use_bbit_insns()) {
1755 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1756 			uasm_i_nop(p);
1757 		} else {
1758 			if (_PAGE_PRESENT_SHIFT) {
1759 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1760 				cur = t;
1761 			}
1762 			uasm_i_andi(p, t, cur, 1);
1763 			uasm_il_beqz(p, r, t, lid);
1764 			if (pte == t)
1765 				/* You lose the SMP race :-(*/
1766 				iPTE_LW(p, pte, ptr);
1767 		}
1768 	} else {
1769 		if (_PAGE_PRESENT_SHIFT) {
1770 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1771 			cur = t;
1772 		}
1773 		uasm_i_andi(p, t, cur,
1774 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1775 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1776 		uasm_il_bnez(p, r, t, lid);
1777 		if (pte == t)
1778 			/* You lose the SMP race :-(*/
1779 			iPTE_LW(p, pte, ptr);
1780 	}
1781 }
1782 
1783 /* Make PTE valid, store result in PTR. */
1784 static void
1785 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1786 		 unsigned int ptr, unsigned int scratch)
1787 {
1788 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1789 
1790 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1791 }
1792 
1793 /*
1794  * Check if PTE can be written to, if not branch to LABEL. Regardless
1795  * restore PTE with value from PTR when done.
1796  */
1797 static void
1798 build_pte_writable(u32 **p, struct uasm_reloc **r,
1799 		   unsigned int pte, unsigned int ptr, int scratch,
1800 		   enum label_id lid)
1801 {
1802 	int t = scratch >= 0 ? scratch : pte;
1803 	int cur = pte;
1804 
1805 	if (_PAGE_PRESENT_SHIFT) {
1806 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1807 		cur = t;
1808 	}
1809 	uasm_i_andi(p, t, cur,
1810 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1811 	uasm_i_xori(p, t, t,
1812 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1813 	uasm_il_bnez(p, r, t, lid);
1814 	if (pte == t)
1815 		/* You lose the SMP race :-(*/
1816 		iPTE_LW(p, pte, ptr);
1817 	else
1818 		uasm_i_nop(p);
1819 }
1820 
1821 /* Make PTE writable, update software status bits as well, then store
1822  * at PTR.
1823  */
1824 static void
1825 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1826 		 unsigned int ptr, unsigned int scratch)
1827 {
1828 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1829 			     | _PAGE_DIRTY);
1830 
1831 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1832 }
1833 
1834 /*
1835  * Check if PTE can be modified, if not branch to LABEL. Regardless
1836  * restore PTE with value from PTR when done.
1837  */
1838 static void
1839 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1840 		     unsigned int pte, unsigned int ptr, int scratch,
1841 		     enum label_id lid)
1842 {
1843 	if (use_bbit_insns()) {
1844 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1845 		uasm_i_nop(p);
1846 	} else {
1847 		int t = scratch >= 0 ? scratch : pte;
1848 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1849 		uasm_i_andi(p, t, t, 1);
1850 		uasm_il_beqz(p, r, t, lid);
1851 		if (pte == t)
1852 			/* You lose the SMP race :-(*/
1853 			iPTE_LW(p, pte, ptr);
1854 	}
1855 }
1856 
1857 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1858 
1859 
1860 /*
1861  * R3000 style TLB load/store/modify handlers.
1862  */
1863 
1864 /*
1865  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1866  * Then it returns.
1867  */
1868 static void
1869 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1870 {
1871 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1872 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1873 	uasm_i_tlbwi(p);
1874 	uasm_i_jr(p, tmp);
1875 	uasm_i_rfe(p); /* branch delay */
1876 }
1877 
1878 /*
1879  * This places the pte into ENTRYLO0 and writes it with tlbwi
1880  * or tlbwr as appropriate.  This is because the index register
1881  * may have the probe fail bit set as a result of a trap on a
1882  * kseg2 access, i.e. without refill.  Then it returns.
1883  */
1884 static void
1885 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1886 			     struct uasm_reloc **r, unsigned int pte,
1887 			     unsigned int tmp)
1888 {
1889 	uasm_i_mfc0(p, tmp, C0_INDEX);
1890 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1891 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1892 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1893 	uasm_i_tlbwi(p); /* cp0 delay */
1894 	uasm_i_jr(p, tmp);
1895 	uasm_i_rfe(p); /* branch delay */
1896 	uasm_l_r3000_write_probe_fail(l, *p);
1897 	uasm_i_tlbwr(p); /* cp0 delay */
1898 	uasm_i_jr(p, tmp);
1899 	uasm_i_rfe(p); /* branch delay */
1900 }
1901 
1902 static void
1903 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1904 				   unsigned int ptr)
1905 {
1906 	long pgdc = (long)pgd_current;
1907 
1908 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1909 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1910 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1911 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1912 	uasm_i_sll(p, pte, pte, 2);
1913 	uasm_i_addu(p, ptr, ptr, pte);
1914 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1915 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1916 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1917 	uasm_i_addu(p, ptr, ptr, pte);
1918 	uasm_i_lw(p, pte, 0, ptr);
1919 	uasm_i_tlbp(p); /* load delay */
1920 }
1921 
1922 static void build_r3000_tlb_load_handler(void)
1923 {
1924 	u32 *p = (u32 *)handle_tlbl;
1925 	struct uasm_label *l = labels;
1926 	struct uasm_reloc *r = relocs;
1927 
1928 	memset(p, 0, handle_tlbl_end - (char *)p);
1929 	memset(labels, 0, sizeof(labels));
1930 	memset(relocs, 0, sizeof(relocs));
1931 
1932 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1933 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1934 	uasm_i_nop(&p); /* load delay */
1935 	build_make_valid(&p, &r, K0, K1, -1);
1936 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1937 
1938 	uasm_l_nopage_tlbl(&l, p);
1939 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1940 	uasm_i_nop(&p);
1941 
1942 	if (p >= (u32 *)handle_tlbl_end)
1943 		panic("TLB load handler fastpath space exceeded");
1944 
1945 	uasm_resolve_relocs(relocs, labels);
1946 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1947 		 (unsigned int)(p - (u32 *)handle_tlbl));
1948 
1949 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1950 }
1951 
1952 static void build_r3000_tlb_store_handler(void)
1953 {
1954 	u32 *p = (u32 *)handle_tlbs;
1955 	struct uasm_label *l = labels;
1956 	struct uasm_reloc *r = relocs;
1957 
1958 	memset(p, 0, handle_tlbs_end - (char *)p);
1959 	memset(labels, 0, sizeof(labels));
1960 	memset(relocs, 0, sizeof(relocs));
1961 
1962 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1963 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1964 	uasm_i_nop(&p); /* load delay */
1965 	build_make_write(&p, &r, K0, K1, -1);
1966 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1967 
1968 	uasm_l_nopage_tlbs(&l, p);
1969 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1970 	uasm_i_nop(&p);
1971 
1972 	if (p >= (u32 *)handle_tlbs_end)
1973 		panic("TLB store handler fastpath space exceeded");
1974 
1975 	uasm_resolve_relocs(relocs, labels);
1976 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1977 		 (unsigned int)(p - (u32 *)handle_tlbs));
1978 
1979 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1980 }
1981 
1982 static void build_r3000_tlb_modify_handler(void)
1983 {
1984 	u32 *p = (u32 *)handle_tlbm;
1985 	struct uasm_label *l = labels;
1986 	struct uasm_reloc *r = relocs;
1987 
1988 	memset(p, 0, handle_tlbm_end - (char *)p);
1989 	memset(labels, 0, sizeof(labels));
1990 	memset(relocs, 0, sizeof(relocs));
1991 
1992 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1993 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1994 	uasm_i_nop(&p); /* load delay */
1995 	build_make_write(&p, &r, K0, K1, -1);
1996 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1997 
1998 	uasm_l_nopage_tlbm(&l, p);
1999 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2000 	uasm_i_nop(&p);
2001 
2002 	if (p >= (u32 *)handle_tlbm_end)
2003 		panic("TLB modify handler fastpath space exceeded");
2004 
2005 	uasm_resolve_relocs(relocs, labels);
2006 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2007 		 (unsigned int)(p - (u32 *)handle_tlbm));
2008 
2009 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2010 }
2011 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2012 
2013 static bool cpu_has_tlbex_tlbp_race(void)
2014 {
2015 	/*
2016 	 * When a Hardware Table Walker is running it can replace TLB entries
2017 	 * at any time, leading to a race between it & the CPU.
2018 	 */
2019 	if (cpu_has_htw)
2020 		return true;
2021 
2022 	/*
2023 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2024 	 * replaced at any time by a sibling performing a write to the FTLB.
2025 	 */
2026 	if (cpu_has_shared_ftlb_ram)
2027 		return true;
2028 
2029 	/* In all other cases there ought to be no race condition to handle */
2030 	return false;
2031 }
2032 
2033 /*
2034  * R4000 style TLB load/store/modify handlers.
2035  */
2036 static struct work_registers
2037 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2038 				   struct uasm_reloc **r)
2039 {
2040 	struct work_registers wr = build_get_work_registers(p);
2041 
2042 #ifdef CONFIG_64BIT
2043 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2044 #else
2045 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2046 #endif
2047 
2048 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2049 	/*
2050 	 * For huge tlb entries, pmd doesn't contain an address but
2051 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2052 	 * see if we need to jump to huge tlb processing.
2053 	 */
2054 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2055 #endif
2056 
2057 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2058 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2059 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2060 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2061 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2062 
2063 #ifdef CONFIG_SMP
2064 	uasm_l_smp_pgtable_change(l, *p);
2065 #endif
2066 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2067 	if (!m4kc_tlbp_war()) {
2068 		build_tlb_probe_entry(p);
2069 		if (cpu_has_tlbex_tlbp_race()) {
2070 			/* race condition happens, leaving */
2071 			uasm_i_ehb(p);
2072 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2073 			uasm_il_bltz(p, r, wr.r3, label_leave);
2074 			uasm_i_nop(p);
2075 		}
2076 	}
2077 	return wr;
2078 }
2079 
2080 static void
2081 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2082 				   struct uasm_reloc **r, unsigned int tmp,
2083 				   unsigned int ptr)
2084 {
2085 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2086 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2087 	build_update_entries(p, tmp, ptr);
2088 	build_tlb_write_entry(p, l, r, tlb_indexed);
2089 	uasm_l_leave(l, *p);
2090 	build_restore_work_registers(p);
2091 	uasm_i_eret(p); /* return from trap */
2092 
2093 #ifdef CONFIG_64BIT
2094 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2095 #endif
2096 }
2097 
2098 static void build_r4000_tlb_load_handler(void)
2099 {
2100 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2101 	struct uasm_label *l = labels;
2102 	struct uasm_reloc *r = relocs;
2103 	struct work_registers wr;
2104 
2105 	memset(p, 0, handle_tlbl_end - (char *)p);
2106 	memset(labels, 0, sizeof(labels));
2107 	memset(relocs, 0, sizeof(relocs));
2108 
2109 	if (bcm1250_m3_war()) {
2110 		unsigned int segbits = 44;
2111 
2112 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2113 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2114 		uasm_i_xor(&p, K0, K0, K1);
2115 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2116 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2117 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2118 		uasm_i_or(&p, K0, K0, K1);
2119 		uasm_il_bnez(&p, &r, K0, label_leave);
2120 		/* No need for uasm_i_nop */
2121 	}
2122 
2123 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2124 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2125 	if (m4kc_tlbp_war())
2126 		build_tlb_probe_entry(&p);
2127 
2128 	if (cpu_has_rixi && !cpu_has_rixiex) {
2129 		/*
2130 		 * If the page is not _PAGE_VALID, RI or XI could not
2131 		 * have triggered it.  Skip the expensive test..
2132 		 */
2133 		if (use_bbit_insns()) {
2134 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2135 				      label_tlbl_goaround1);
2136 		} else {
2137 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2138 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2139 		}
2140 		uasm_i_nop(&p);
2141 
2142 		/*
2143 		 * Warn if something may race with us & replace the TLB entry
2144 		 * before we read it here. Everything with such races should
2145 		 * also have dedicated RiXi exception handlers, so this
2146 		 * shouldn't be hit.
2147 		 */
2148 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2149 
2150 		uasm_i_tlbr(&p);
2151 
2152 		switch (current_cpu_type()) {
2153 		default:
2154 			if (cpu_has_mips_r2_exec_hazard) {
2155 				uasm_i_ehb(&p);
2156 
2157 		case CPU_CAVIUM_OCTEON:
2158 		case CPU_CAVIUM_OCTEON_PLUS:
2159 		case CPU_CAVIUM_OCTEON2:
2160 				break;
2161 			}
2162 		}
2163 
2164 		/* Examine  entrylo 0 or 1 based on ptr. */
2165 		if (use_bbit_insns()) {
2166 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2167 		} else {
2168 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2169 			uasm_i_beqz(&p, wr.r3, 8);
2170 		}
2171 		/* load it in the delay slot*/
2172 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2173 		/* load it if ptr is odd */
2174 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2175 		/*
2176 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2177 		 * XI must have triggered it.
2178 		 */
2179 		if (use_bbit_insns()) {
2180 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2181 			uasm_i_nop(&p);
2182 			uasm_l_tlbl_goaround1(&l, p);
2183 		} else {
2184 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2185 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2186 			uasm_i_nop(&p);
2187 		}
2188 		uasm_l_tlbl_goaround1(&l, p);
2189 	}
2190 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2191 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2192 
2193 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2194 	/*
2195 	 * This is the entry point when build_r4000_tlbchange_handler_head
2196 	 * spots a huge page.
2197 	 */
2198 	uasm_l_tlb_huge_update(&l, p);
2199 	iPTE_LW(&p, wr.r1, wr.r2);
2200 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2201 	build_tlb_probe_entry(&p);
2202 
2203 	if (cpu_has_rixi && !cpu_has_rixiex) {
2204 		/*
2205 		 * If the page is not _PAGE_VALID, RI or XI could not
2206 		 * have triggered it.  Skip the expensive test..
2207 		 */
2208 		if (use_bbit_insns()) {
2209 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2210 				      label_tlbl_goaround2);
2211 		} else {
2212 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2213 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2214 		}
2215 		uasm_i_nop(&p);
2216 
2217 		/*
2218 		 * Warn if something may race with us & replace the TLB entry
2219 		 * before we read it here. Everything with such races should
2220 		 * also have dedicated RiXi exception handlers, so this
2221 		 * shouldn't be hit.
2222 		 */
2223 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2224 
2225 		uasm_i_tlbr(&p);
2226 
2227 		switch (current_cpu_type()) {
2228 		default:
2229 			if (cpu_has_mips_r2_exec_hazard) {
2230 				uasm_i_ehb(&p);
2231 
2232 		case CPU_CAVIUM_OCTEON:
2233 		case CPU_CAVIUM_OCTEON_PLUS:
2234 		case CPU_CAVIUM_OCTEON2:
2235 				break;
2236 			}
2237 		}
2238 
2239 		/* Examine  entrylo 0 or 1 based on ptr. */
2240 		if (use_bbit_insns()) {
2241 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2242 		} else {
2243 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2244 			uasm_i_beqz(&p, wr.r3, 8);
2245 		}
2246 		/* load it in the delay slot*/
2247 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2248 		/* load it if ptr is odd */
2249 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2250 		/*
2251 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2252 		 * XI must have triggered it.
2253 		 */
2254 		if (use_bbit_insns()) {
2255 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2256 		} else {
2257 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2258 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2259 		}
2260 		if (PM_DEFAULT_MASK == 0)
2261 			uasm_i_nop(&p);
2262 		/*
2263 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2264 		 * it is restored in build_huge_tlb_write_entry.
2265 		 */
2266 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2267 
2268 		uasm_l_tlbl_goaround2(&l, p);
2269 	}
2270 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2271 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2272 #endif
2273 
2274 	uasm_l_nopage_tlbl(&l, p);
2275 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2276 		uasm_i_sync(&p, 0);
2277 	build_restore_work_registers(&p);
2278 #ifdef CONFIG_CPU_MICROMIPS
2279 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2280 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2281 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2282 		uasm_i_jr(&p, K0);
2283 	} else
2284 #endif
2285 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2286 	uasm_i_nop(&p);
2287 
2288 	if (p >= (u32 *)handle_tlbl_end)
2289 		panic("TLB load handler fastpath space exceeded");
2290 
2291 	uasm_resolve_relocs(relocs, labels);
2292 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2293 		 (unsigned int)(p - (u32 *)handle_tlbl));
2294 
2295 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2296 }
2297 
2298 static void build_r4000_tlb_store_handler(void)
2299 {
2300 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2301 	struct uasm_label *l = labels;
2302 	struct uasm_reloc *r = relocs;
2303 	struct work_registers wr;
2304 
2305 	memset(p, 0, handle_tlbs_end - (char *)p);
2306 	memset(labels, 0, sizeof(labels));
2307 	memset(relocs, 0, sizeof(relocs));
2308 
2309 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2310 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2311 	if (m4kc_tlbp_war())
2312 		build_tlb_probe_entry(&p);
2313 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2314 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2315 
2316 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2317 	/*
2318 	 * This is the entry point when
2319 	 * build_r4000_tlbchange_handler_head spots a huge page.
2320 	 */
2321 	uasm_l_tlb_huge_update(&l, p);
2322 	iPTE_LW(&p, wr.r1, wr.r2);
2323 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2324 	build_tlb_probe_entry(&p);
2325 	uasm_i_ori(&p, wr.r1, wr.r1,
2326 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2327 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2328 #endif
2329 
2330 	uasm_l_nopage_tlbs(&l, p);
2331 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2332 		uasm_i_sync(&p, 0);
2333 	build_restore_work_registers(&p);
2334 #ifdef CONFIG_CPU_MICROMIPS
2335 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2336 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2337 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2338 		uasm_i_jr(&p, K0);
2339 	} else
2340 #endif
2341 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2342 	uasm_i_nop(&p);
2343 
2344 	if (p >= (u32 *)handle_tlbs_end)
2345 		panic("TLB store handler fastpath space exceeded");
2346 
2347 	uasm_resolve_relocs(relocs, labels);
2348 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2349 		 (unsigned int)(p - (u32 *)handle_tlbs));
2350 
2351 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2352 }
2353 
2354 static void build_r4000_tlb_modify_handler(void)
2355 {
2356 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2357 	struct uasm_label *l = labels;
2358 	struct uasm_reloc *r = relocs;
2359 	struct work_registers wr;
2360 
2361 	memset(p, 0, handle_tlbm_end - (char *)p);
2362 	memset(labels, 0, sizeof(labels));
2363 	memset(relocs, 0, sizeof(relocs));
2364 
2365 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2366 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2367 	if (m4kc_tlbp_war())
2368 		build_tlb_probe_entry(&p);
2369 	/* Present and writable bits set, set accessed and dirty bits. */
2370 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2371 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2372 
2373 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2374 	/*
2375 	 * This is the entry point when
2376 	 * build_r4000_tlbchange_handler_head spots a huge page.
2377 	 */
2378 	uasm_l_tlb_huge_update(&l, p);
2379 	iPTE_LW(&p, wr.r1, wr.r2);
2380 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2381 	build_tlb_probe_entry(&p);
2382 	uasm_i_ori(&p, wr.r1, wr.r1,
2383 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2384 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2385 #endif
2386 
2387 	uasm_l_nopage_tlbm(&l, p);
2388 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2389 		uasm_i_sync(&p, 0);
2390 	build_restore_work_registers(&p);
2391 #ifdef CONFIG_CPU_MICROMIPS
2392 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2393 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2394 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2395 		uasm_i_jr(&p, K0);
2396 	} else
2397 #endif
2398 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2399 	uasm_i_nop(&p);
2400 
2401 	if (p >= (u32 *)handle_tlbm_end)
2402 		panic("TLB modify handler fastpath space exceeded");
2403 
2404 	uasm_resolve_relocs(relocs, labels);
2405 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2406 		 (unsigned int)(p - (u32 *)handle_tlbm));
2407 
2408 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2409 }
2410 
2411 static void flush_tlb_handlers(void)
2412 {
2413 	local_flush_icache_range((unsigned long)handle_tlbl,
2414 			   (unsigned long)handle_tlbl_end);
2415 	local_flush_icache_range((unsigned long)handle_tlbs,
2416 			   (unsigned long)handle_tlbs_end);
2417 	local_flush_icache_range((unsigned long)handle_tlbm,
2418 			   (unsigned long)handle_tlbm_end);
2419 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2420 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2421 }
2422 
2423 static void print_htw_config(void)
2424 {
2425 	unsigned long config;
2426 	unsigned int pwctl;
2427 	const int field = 2 * sizeof(unsigned long);
2428 
2429 	config = read_c0_pwfield();
2430 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2431 		field, config,
2432 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2433 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2434 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2435 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2436 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2437 
2438 	config = read_c0_pwsize();
2439 	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",
2440 		field, config,
2441 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2442 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2443 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2444 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2445 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2446 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2447 
2448 	pwctl = read_c0_pwctl();
2449 	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",
2450 		pwctl,
2451 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2452 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2453 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2454 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2455 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2456 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2457 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2458 }
2459 
2460 static void config_htw_params(void)
2461 {
2462 	unsigned long pwfield, pwsize, ptei;
2463 	unsigned int config;
2464 
2465 	/*
2466 	 * We are using 2-level page tables, so we only need to
2467 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2468 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2469 	 * write values less than 0xc in these fields because the entire
2470 	 * write will be dropped. As a result of which, we must preserve
2471 	 * the original reset values and overwrite only what we really want.
2472 	 */
2473 
2474 	pwfield = read_c0_pwfield();
2475 	/* re-initialize the GDI field */
2476 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2477 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2478 	/* re-initialize the PTI field including the even/odd bit */
2479 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2480 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2481 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2482 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2483 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2484 	}
2485 	/* Set the PTEI right shift */
2486 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2487 	pwfield |= ptei;
2488 	write_c0_pwfield(pwfield);
2489 	/* Check whether the PTEI value is supported */
2490 	back_to_back_c0_hazard();
2491 	pwfield = read_c0_pwfield();
2492 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2493 		!= ptei) {
2494 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2495 			ptei);
2496 		/*
2497 		 * Drop option to avoid HTW being enabled via another path
2498 		 * (eg htw_reset())
2499 		 */
2500 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2501 		return;
2502 	}
2503 
2504 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2505 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2506 	if (CONFIG_PGTABLE_LEVELS >= 3)
2507 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2508 
2509 	/* Set pointer size to size of directory pointers */
2510 	if (IS_ENABLED(CONFIG_64BIT))
2511 		pwsize |= MIPS_PWSIZE_PS_MASK;
2512 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2513 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2514 			& MIPS_PWSIZE_PTEW_MASK;
2515 
2516 	write_c0_pwsize(pwsize);
2517 
2518 	/* Make sure everything is set before we enable the HTW */
2519 	back_to_back_c0_hazard();
2520 
2521 	/*
2522 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2523 	 * the pwctl fields.
2524 	 */
2525 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2526 	if (IS_ENABLED(CONFIG_64BIT))
2527 		config |= MIPS_PWCTL_XU_MASK;
2528 	write_c0_pwctl(config);
2529 	pr_info("Hardware Page Table Walker enabled\n");
2530 
2531 	print_htw_config();
2532 }
2533 
2534 static void config_xpa_params(void)
2535 {
2536 #ifdef CONFIG_XPA
2537 	unsigned int pagegrain;
2538 
2539 	if (mips_xpa_disabled) {
2540 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2541 		return;
2542 	}
2543 
2544 	pagegrain = read_c0_pagegrain();
2545 	write_c0_pagegrain(pagegrain | PG_ELPA);
2546 	back_to_back_c0_hazard();
2547 	pagegrain = read_c0_pagegrain();
2548 
2549 	if (pagegrain & PG_ELPA)
2550 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2551 	else
2552 		panic("Extended Physical Addressing (XPA) disabled");
2553 #endif
2554 }
2555 
2556 static void check_pabits(void)
2557 {
2558 	unsigned long entry;
2559 	unsigned pabits, fillbits;
2560 
2561 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2562 		/*
2563 		 * We'll only be making use of the fact that we can rotate bits
2564 		 * into the fill if the CPU supports RIXI, so don't bother
2565 		 * probing this for CPUs which don't.
2566 		 */
2567 		return;
2568 	}
2569 
2570 	write_c0_entrylo0(~0ul);
2571 	back_to_back_c0_hazard();
2572 	entry = read_c0_entrylo0();
2573 
2574 	/* clear all non-PFN bits */
2575 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2576 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2577 
2578 	/* find a lower bound on PABITS, and upper bound on fill bits */
2579 	pabits = fls_long(entry) + 6;
2580 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2581 
2582 	/* minus the RI & XI bits */
2583 	fillbits -= min_t(unsigned, fillbits, 2);
2584 
2585 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2586 		fill_includes_sw_bits = true;
2587 
2588 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2589 }
2590 
2591 void build_tlb_refill_handler(void)
2592 {
2593 	/*
2594 	 * The refill handler is generated per-CPU, multi-node systems
2595 	 * may have local storage for it. The other handlers are only
2596 	 * needed once.
2597 	 */
2598 	static int run_once = 0;
2599 
2600 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2601 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2602 
2603 	output_pgtable_bits_defines();
2604 	check_pabits();
2605 
2606 #ifdef CONFIG_64BIT
2607 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2608 #endif
2609 
2610 	switch (current_cpu_type()) {
2611 	case CPU_R2000:
2612 	case CPU_R3000:
2613 	case CPU_R3000A:
2614 	case CPU_R3081E:
2615 	case CPU_TX3912:
2616 	case CPU_TX3922:
2617 	case CPU_TX3927:
2618 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2619 		if (cpu_has_local_ebase)
2620 			build_r3000_tlb_refill_handler();
2621 		if (!run_once) {
2622 			if (!cpu_has_local_ebase)
2623 				build_r3000_tlb_refill_handler();
2624 			build_setup_pgd();
2625 			build_r3000_tlb_load_handler();
2626 			build_r3000_tlb_store_handler();
2627 			build_r3000_tlb_modify_handler();
2628 			flush_tlb_handlers();
2629 			run_once++;
2630 		}
2631 #else
2632 		panic("No R3000 TLB refill handler");
2633 #endif
2634 		break;
2635 
2636 	case CPU_R8000:
2637 		panic("No R8000 TLB refill handler yet");
2638 		break;
2639 
2640 	default:
2641 		if (cpu_has_ldpte)
2642 			setup_pw();
2643 
2644 		if (!run_once) {
2645 			scratch_reg = allocate_kscratch();
2646 			build_setup_pgd();
2647 			build_r4000_tlb_load_handler();
2648 			build_r4000_tlb_store_handler();
2649 			build_r4000_tlb_modify_handler();
2650 			if (cpu_has_ldpte)
2651 				build_loongson3_tlb_refill_handler();
2652 			else if (!cpu_has_local_ebase)
2653 				build_r4000_tlb_refill_handler();
2654 			flush_tlb_handlers();
2655 			run_once++;
2656 		}
2657 		if (cpu_has_local_ebase)
2658 			build_r4000_tlb_refill_handler();
2659 		if (cpu_has_xpa)
2660 			config_xpa_params();
2661 		if (cpu_has_htw)
2662 			config_htw_params();
2663 	}
2664 }
2665