xref: /linux/arch/mips/mm/tlbex.c (revision f49f4ab95c301dbccad0efe85296d908b8ae7ad4)
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 completly 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/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/init.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/pgtable.h>
34 #include <asm/war.h>
35 #include <asm/uasm.h>
36 #include <asm/setup.h>
37 
38 /*
39  * TLB load/store/modify handlers.
40  *
41  * Only the fastpath gets synthesized at runtime, the slowpath for
42  * do_page_fault remains normal asm.
43  */
44 extern void tlb_do_page_fault_0(void);
45 extern void tlb_do_page_fault_1(void);
46 
47 struct work_registers {
48 	int r1;
49 	int r2;
50 	int r3;
51 };
52 
53 struct tlb_reg_save {
54 	unsigned long a;
55 	unsigned long b;
56 } ____cacheline_aligned_in_smp;
57 
58 static struct tlb_reg_save handler_reg_save[NR_CPUS];
59 
60 static inline int r45k_bvahwbug(void)
61 {
62 	/* XXX: We should probe for the presence of this bug, but we don't. */
63 	return 0;
64 }
65 
66 static inline int r4k_250MHZhwbug(void)
67 {
68 	/* XXX: We should probe for the presence of this bug, but we don't. */
69 	return 0;
70 }
71 
72 static inline int __maybe_unused bcm1250_m3_war(void)
73 {
74 	return BCM1250_M3_WAR;
75 }
76 
77 static inline int __maybe_unused r10000_llsc_war(void)
78 {
79 	return R10000_LLSC_WAR;
80 }
81 
82 static int use_bbit_insns(void)
83 {
84 	switch (current_cpu_type()) {
85 	case CPU_CAVIUM_OCTEON:
86 	case CPU_CAVIUM_OCTEON_PLUS:
87 	case CPU_CAVIUM_OCTEON2:
88 		return 1;
89 	default:
90 		return 0;
91 	}
92 }
93 
94 static int use_lwx_insns(void)
95 {
96 	switch (current_cpu_type()) {
97 	case CPU_CAVIUM_OCTEON2:
98 		return 1;
99 	default:
100 		return 0;
101 	}
102 }
103 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
104     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
105 static bool scratchpad_available(void)
106 {
107 	return true;
108 }
109 static int scratchpad_offset(int i)
110 {
111 	/*
112 	 * CVMSEG starts at address -32768 and extends for
113 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
114 	 */
115 	i += 1; /* Kernel use starts at the top and works down. */
116 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
117 }
118 #else
119 static bool scratchpad_available(void)
120 {
121 	return false;
122 }
123 static int scratchpad_offset(int i)
124 {
125 	BUG();
126 	/* Really unreachable, but evidently some GCC want this. */
127 	return 0;
128 }
129 #endif
130 /*
131  * Found by experiment: At least some revisions of the 4kc throw under
132  * some circumstances a machine check exception, triggered by invalid
133  * values in the index register.  Delaying the tlbp instruction until
134  * after the next branch,  plus adding an additional nop in front of
135  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
136  * why; it's not an issue caused by the core RTL.
137  *
138  */
139 static int __cpuinit m4kc_tlbp_war(void)
140 {
141 	return (current_cpu_data.processor_id & 0xffff00) ==
142 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
143 }
144 
145 /* Handle labels (which must be positive integers). */
146 enum label_id {
147 	label_second_part = 1,
148 	label_leave,
149 	label_vmalloc,
150 	label_vmalloc_done,
151 	label_tlbw_hazard_0,
152 	label_split = label_tlbw_hazard_0 + 8,
153 	label_tlbl_goaround1,
154 	label_tlbl_goaround2,
155 	label_nopage_tlbl,
156 	label_nopage_tlbs,
157 	label_nopage_tlbm,
158 	label_smp_pgtable_change,
159 	label_r3000_write_probe_fail,
160 	label_large_segbits_fault,
161 #ifdef CONFIG_HUGETLB_PAGE
162 	label_tlb_huge_update,
163 #endif
164 };
165 
166 UASM_L_LA(_second_part)
167 UASM_L_LA(_leave)
168 UASM_L_LA(_vmalloc)
169 UASM_L_LA(_vmalloc_done)
170 /* _tlbw_hazard_x is handled differently.  */
171 UASM_L_LA(_split)
172 UASM_L_LA(_tlbl_goaround1)
173 UASM_L_LA(_tlbl_goaround2)
174 UASM_L_LA(_nopage_tlbl)
175 UASM_L_LA(_nopage_tlbs)
176 UASM_L_LA(_nopage_tlbm)
177 UASM_L_LA(_smp_pgtable_change)
178 UASM_L_LA(_r3000_write_probe_fail)
179 UASM_L_LA(_large_segbits_fault)
180 #ifdef CONFIG_HUGETLB_PAGE
181 UASM_L_LA(_tlb_huge_update)
182 #endif
183 
184 static int __cpuinitdata hazard_instance;
185 
186 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
187 {
188 	switch (instance) {
189 	case 0 ... 7:
190 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
191 		return;
192 	default:
193 		BUG();
194 	}
195 }
196 
197 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
198 {
199 	switch (instance) {
200 	case 0 ... 7:
201 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
202 		break;
203 	default:
204 		BUG();
205 	}
206 }
207 
208 /*
209  * For debug purposes.
210  */
211 static inline void dump_handler(const u32 *handler, int count)
212 {
213 	int i;
214 
215 	pr_debug("\t.set push\n");
216 	pr_debug("\t.set noreorder\n");
217 
218 	for (i = 0; i < count; i++)
219 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
220 
221 	pr_debug("\t.set pop\n");
222 }
223 
224 /* The only general purpose registers allowed in TLB handlers. */
225 #define K0		26
226 #define K1		27
227 
228 /* Some CP0 registers */
229 #define C0_INDEX	0, 0
230 #define C0_ENTRYLO0	2, 0
231 #define C0_TCBIND	2, 2
232 #define C0_ENTRYLO1	3, 0
233 #define C0_CONTEXT	4, 0
234 #define C0_PAGEMASK	5, 0
235 #define C0_BADVADDR	8, 0
236 #define C0_ENTRYHI	10, 0
237 #define C0_EPC		14, 0
238 #define C0_XCONTEXT	20, 0
239 
240 #ifdef CONFIG_64BIT
241 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
242 #else
243 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
244 #endif
245 
246 /* The worst case length of the handler is around 18 instructions for
247  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
248  * Maximum space available is 32 instructions for R3000 and 64
249  * instructions for R4000.
250  *
251  * We deliberately chose a buffer size of 128, so we won't scribble
252  * over anything important on overflow before we panic.
253  */
254 static u32 tlb_handler[128] __cpuinitdata;
255 
256 /* simply assume worst case size for labels and relocs */
257 static struct uasm_label labels[128] __cpuinitdata;
258 static struct uasm_reloc relocs[128] __cpuinitdata;
259 
260 #ifdef CONFIG_64BIT
261 static int check_for_high_segbits __cpuinitdata;
262 #endif
263 
264 static int check_for_high_segbits __cpuinitdata;
265 
266 static unsigned int kscratch_used_mask __cpuinitdata;
267 
268 static int __cpuinit allocate_kscratch(void)
269 {
270 	int r;
271 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
272 
273 	r = ffs(a);
274 
275 	if (r == 0)
276 		return -1;
277 
278 	r--; /* make it zero based */
279 
280 	kscratch_used_mask |= (1 << r);
281 
282 	return r;
283 }
284 
285 static int scratch_reg __cpuinitdata;
286 static int pgd_reg __cpuinitdata;
287 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
288 
289 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
290 {
291 	struct work_registers r;
292 
293 	int smp_processor_id_reg;
294 	int smp_processor_id_sel;
295 	int smp_processor_id_shift;
296 
297 	if (scratch_reg > 0) {
298 		/* Save in CPU local C0_KScratch? */
299 		UASM_i_MTC0(p, 1, 31, scratch_reg);
300 		r.r1 = K0;
301 		r.r2 = K1;
302 		r.r3 = 1;
303 		return r;
304 	}
305 
306 	if (num_possible_cpus() > 1) {
307 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
308 		smp_processor_id_shift = 51;
309 		smp_processor_id_reg = 20; /* XContext */
310 		smp_processor_id_sel = 0;
311 #else
312 # ifdef CONFIG_32BIT
313 		smp_processor_id_shift = 25;
314 		smp_processor_id_reg = 4; /* Context */
315 		smp_processor_id_sel = 0;
316 # endif
317 # ifdef CONFIG_64BIT
318 		smp_processor_id_shift = 26;
319 		smp_processor_id_reg = 4; /* Context */
320 		smp_processor_id_sel = 0;
321 # endif
322 #endif
323 		/* Get smp_processor_id */
324 		UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
325 		UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
326 
327 		/* handler_reg_save index in K0 */
328 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
329 
330 		UASM_i_LA(p, K1, (long)&handler_reg_save);
331 		UASM_i_ADDU(p, K0, K0, K1);
332 	} else {
333 		UASM_i_LA(p, K0, (long)&handler_reg_save);
334 	}
335 	/* K0 now points to save area, save $1 and $2  */
336 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
337 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
338 
339 	r.r1 = K1;
340 	r.r2 = 1;
341 	r.r3 = 2;
342 	return r;
343 }
344 
345 static void __cpuinit build_restore_work_registers(u32 **p)
346 {
347 	if (scratch_reg > 0) {
348 		UASM_i_MFC0(p, 1, 31, scratch_reg);
349 		return;
350 	}
351 	/* K0 already points to save area, restore $1 and $2  */
352 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
353 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
354 }
355 
356 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
357 
358 /*
359  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
360  * we cannot do r3000 under these circumstances.
361  *
362  * Declare pgd_current here instead of including mmu_context.h to avoid type
363  * conflicts for tlbmiss_handler_setup_pgd
364  */
365 extern unsigned long pgd_current[];
366 
367 /*
368  * The R3000 TLB handler is simple.
369  */
370 static void __cpuinit build_r3000_tlb_refill_handler(void)
371 {
372 	long pgdc = (long)pgd_current;
373 	u32 *p;
374 
375 	memset(tlb_handler, 0, sizeof(tlb_handler));
376 	p = tlb_handler;
377 
378 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
379 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
380 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
381 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
382 	uasm_i_sll(&p, K0, K0, 2);
383 	uasm_i_addu(&p, K1, K1, K0);
384 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
385 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
386 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
387 	uasm_i_addu(&p, K1, K1, K0);
388 	uasm_i_lw(&p, K0, 0, K1);
389 	uasm_i_nop(&p); /* load delay */
390 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
391 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
392 	uasm_i_tlbwr(&p); /* cp0 delay */
393 	uasm_i_jr(&p, K1);
394 	uasm_i_rfe(&p); /* branch delay */
395 
396 	if (p > tlb_handler + 32)
397 		panic("TLB refill handler space exceeded");
398 
399 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
400 		 (unsigned int)(p - tlb_handler));
401 
402 	memcpy((void *)ebase, tlb_handler, 0x80);
403 
404 	dump_handler((u32 *)ebase, 32);
405 }
406 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
407 
408 /*
409  * The R4000 TLB handler is much more complicated. We have two
410  * consecutive handler areas with 32 instructions space each.
411  * Since they aren't used at the same time, we can overflow in the
412  * other one.To keep things simple, we first assume linear space,
413  * then we relocate it to the final handler layout as needed.
414  */
415 static u32 final_handler[64] __cpuinitdata;
416 
417 /*
418  * Hazards
419  *
420  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
421  * 2. A timing hazard exists for the TLBP instruction.
422  *
423  *      stalling_instruction
424  *      TLBP
425  *
426  * The JTLB is being read for the TLBP throughout the stall generated by the
427  * previous instruction. This is not really correct as the stalling instruction
428  * can modify the address used to access the JTLB.  The failure symptom is that
429  * the TLBP instruction will use an address created for the stalling instruction
430  * and not the address held in C0_ENHI and thus report the wrong results.
431  *
432  * The software work-around is to not allow the instruction preceding the TLBP
433  * to stall - make it an NOP or some other instruction guaranteed not to stall.
434  *
435  * Errata 2 will not be fixed.  This errata is also on the R5000.
436  *
437  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
438  */
439 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
440 {
441 	switch (current_cpu_type()) {
442 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
443 	case CPU_R4600:
444 	case CPU_R4700:
445 	case CPU_R5000:
446 	case CPU_R5000A:
447 	case CPU_NEVADA:
448 		uasm_i_nop(p);
449 		uasm_i_tlbp(p);
450 		break;
451 
452 	default:
453 		uasm_i_tlbp(p);
454 		break;
455 	}
456 }
457 
458 /*
459  * Write random or indexed TLB entry, and care about the hazards from
460  * the preceding mtc0 and for the following eret.
461  */
462 enum tlb_write_entry { tlb_random, tlb_indexed };
463 
464 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
465 					 struct uasm_reloc **r,
466 					 enum tlb_write_entry wmode)
467 {
468 	void(*tlbw)(u32 **) = NULL;
469 
470 	switch (wmode) {
471 	case tlb_random: tlbw = uasm_i_tlbwr; break;
472 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
473 	}
474 
475 	if (cpu_has_mips_r2) {
476 		/*
477 		 * The architecture spec says an ehb is required here,
478 		 * but a number of cores do not have the hazard and
479 		 * using an ehb causes an expensive pipeline stall.
480 		 */
481 		switch (current_cpu_type()) {
482 		case CPU_M14KC:
483 		case CPU_74K:
484 			break;
485 
486 		default:
487 			uasm_i_ehb(p);
488 			break;
489 		}
490 		tlbw(p);
491 		return;
492 	}
493 
494 	switch (current_cpu_type()) {
495 	case CPU_R4000PC:
496 	case CPU_R4000SC:
497 	case CPU_R4000MC:
498 	case CPU_R4400PC:
499 	case CPU_R4400SC:
500 	case CPU_R4400MC:
501 		/*
502 		 * This branch uses up a mtc0 hazard nop slot and saves
503 		 * two nops after the tlbw instruction.
504 		 */
505 		uasm_bgezl_hazard(p, r, hazard_instance);
506 		tlbw(p);
507 		uasm_bgezl_label(l, p, hazard_instance);
508 		hazard_instance++;
509 		uasm_i_nop(p);
510 		break;
511 
512 	case CPU_R4600:
513 	case CPU_R4700:
514 		uasm_i_nop(p);
515 		tlbw(p);
516 		uasm_i_nop(p);
517 		break;
518 
519 	case CPU_R5000:
520 	case CPU_R5000A:
521 	case CPU_NEVADA:
522 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
523 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
524 		tlbw(p);
525 		break;
526 
527 	case CPU_R4300:
528 	case CPU_5KC:
529 	case CPU_TX49XX:
530 	case CPU_PR4450:
531 	case CPU_XLR:
532 		uasm_i_nop(p);
533 		tlbw(p);
534 		break;
535 
536 	case CPU_R10000:
537 	case CPU_R12000:
538 	case CPU_R14000:
539 	case CPU_4KC:
540 	case CPU_4KEC:
541 	case CPU_M14KC:
542 	case CPU_SB1:
543 	case CPU_SB1A:
544 	case CPU_4KSC:
545 	case CPU_20KC:
546 	case CPU_25KF:
547 	case CPU_BMIPS32:
548 	case CPU_BMIPS3300:
549 	case CPU_BMIPS4350:
550 	case CPU_BMIPS4380:
551 	case CPU_BMIPS5000:
552 	case CPU_LOONGSON2:
553 	case CPU_R5500:
554 		if (m4kc_tlbp_war())
555 			uasm_i_nop(p);
556 	case CPU_ALCHEMY:
557 		tlbw(p);
558 		break;
559 
560 	case CPU_RM7000:
561 		uasm_i_nop(p);
562 		uasm_i_nop(p);
563 		uasm_i_nop(p);
564 		uasm_i_nop(p);
565 		tlbw(p);
566 		break;
567 
568 	case CPU_RM9000:
569 		/*
570 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
571 		 * use of the JTLB for instructions should not occur for 4
572 		 * cpu cycles and use for data translations should not occur
573 		 * for 3 cpu cycles.
574 		 */
575 		uasm_i_ssnop(p);
576 		uasm_i_ssnop(p);
577 		uasm_i_ssnop(p);
578 		uasm_i_ssnop(p);
579 		tlbw(p);
580 		uasm_i_ssnop(p);
581 		uasm_i_ssnop(p);
582 		uasm_i_ssnop(p);
583 		uasm_i_ssnop(p);
584 		break;
585 
586 	case CPU_VR4111:
587 	case CPU_VR4121:
588 	case CPU_VR4122:
589 	case CPU_VR4181:
590 	case CPU_VR4181A:
591 		uasm_i_nop(p);
592 		uasm_i_nop(p);
593 		tlbw(p);
594 		uasm_i_nop(p);
595 		uasm_i_nop(p);
596 		break;
597 
598 	case CPU_VR4131:
599 	case CPU_VR4133:
600 	case CPU_R5432:
601 		uasm_i_nop(p);
602 		uasm_i_nop(p);
603 		tlbw(p);
604 		break;
605 
606 	case CPU_JZRISC:
607 		tlbw(p);
608 		uasm_i_nop(p);
609 		break;
610 
611 	default:
612 		panic("No TLB refill handler yet (CPU type: %d)",
613 		      current_cpu_data.cputype);
614 		break;
615 	}
616 }
617 
618 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
619 								  unsigned int reg)
620 {
621 	if (cpu_has_rixi) {
622 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
623 	} else {
624 #ifdef CONFIG_64BIT_PHYS_ADDR
625 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
626 #else
627 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
628 #endif
629 	}
630 }
631 
632 #ifdef CONFIG_HUGETLB_PAGE
633 
634 static __cpuinit void build_restore_pagemask(u32 **p,
635 					     struct uasm_reloc **r,
636 					     unsigned int tmp,
637 					     enum label_id lid,
638 					     int restore_scratch)
639 {
640 	if (restore_scratch) {
641 		/* Reset default page size */
642 		if (PM_DEFAULT_MASK >> 16) {
643 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
644 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
645 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
646 			uasm_il_b(p, r, lid);
647 		} else if (PM_DEFAULT_MASK) {
648 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
649 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
650 			uasm_il_b(p, r, lid);
651 		} else {
652 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
653 			uasm_il_b(p, r, lid);
654 		}
655 		if (scratch_reg > 0)
656 			UASM_i_MFC0(p, 1, 31, scratch_reg);
657 		else
658 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
659 	} else {
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_il_b(p, r, lid);
665 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
666 		} else if (PM_DEFAULT_MASK) {
667 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
668 			uasm_il_b(p, r, lid);
669 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
670 		} else {
671 			uasm_il_b(p, r, lid);
672 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
673 		}
674 	}
675 }
676 
677 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
678 						 struct uasm_label **l,
679 						 struct uasm_reloc **r,
680 						 unsigned int tmp,
681 						 enum tlb_write_entry wmode,
682 						 int restore_scratch)
683 {
684 	/* Set huge page tlb entry size */
685 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
686 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
687 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
688 
689 	build_tlb_write_entry(p, l, r, wmode);
690 
691 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
692 }
693 
694 /*
695  * Check if Huge PTE is present, if so then jump to LABEL.
696  */
697 static void __cpuinit
698 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
699 		unsigned int pmd, int lid)
700 {
701 	UASM_i_LW(p, tmp, 0, pmd);
702 	if (use_bbit_insns()) {
703 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
704 	} else {
705 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
706 		uasm_il_bnez(p, r, tmp, lid);
707 	}
708 }
709 
710 static __cpuinit void build_huge_update_entries(u32 **p,
711 						unsigned int pte,
712 						unsigned int tmp)
713 {
714 	int small_sequence;
715 
716 	/*
717 	 * A huge PTE describes an area the size of the
718 	 * configured huge page size. This is twice the
719 	 * of the large TLB entry size we intend to use.
720 	 * A TLB entry half the size of the configured
721 	 * huge page size is configured into entrylo0
722 	 * and entrylo1 to cover the contiguous huge PTE
723 	 * address space.
724 	 */
725 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
726 
727 	/* We can clobber tmp.  It isn't used after this.*/
728 	if (!small_sequence)
729 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
730 
731 	build_convert_pte_to_entrylo(p, pte);
732 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
733 	/* convert to entrylo1 */
734 	if (small_sequence)
735 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
736 	else
737 		UASM_i_ADDU(p, pte, pte, tmp);
738 
739 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
740 }
741 
742 static __cpuinit void build_huge_handler_tail(u32 **p,
743 					      struct uasm_reloc **r,
744 					      struct uasm_label **l,
745 					      unsigned int pte,
746 					      unsigned int ptr)
747 {
748 #ifdef CONFIG_SMP
749 	UASM_i_SC(p, pte, 0, ptr);
750 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
751 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
752 #else
753 	UASM_i_SW(p, pte, 0, ptr);
754 #endif
755 	build_huge_update_entries(p, pte, ptr);
756 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
757 }
758 #endif /* CONFIG_HUGETLB_PAGE */
759 
760 #ifdef CONFIG_64BIT
761 /*
762  * TMP and PTR are scratch.
763  * TMP will be clobbered, PTR will hold the pmd entry.
764  */
765 static void __cpuinit
766 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
767 		 unsigned int tmp, unsigned int ptr)
768 {
769 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
770 	long pgdc = (long)pgd_current;
771 #endif
772 	/*
773 	 * The vmalloc handling is not in the hotpath.
774 	 */
775 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
776 
777 	if (check_for_high_segbits) {
778 		/*
779 		 * The kernel currently implicitely assumes that the
780 		 * MIPS SEGBITS parameter for the processor is
781 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
782 		 * allocate virtual addresses outside the maximum
783 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
784 		 * that doesn't prevent user code from accessing the
785 		 * higher xuseg addresses.  Here, we make sure that
786 		 * everything but the lower xuseg addresses goes down
787 		 * the module_alloc/vmalloc path.
788 		 */
789 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
790 		uasm_il_bnez(p, r, ptr, label_vmalloc);
791 	} else {
792 		uasm_il_bltz(p, r, tmp, label_vmalloc);
793 	}
794 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
795 
796 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
797 	if (pgd_reg != -1) {
798 		/* pgd is in pgd_reg */
799 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
800 	} else {
801 		/*
802 		 * &pgd << 11 stored in CONTEXT [23..63].
803 		 */
804 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
805 
806 		/* Clear lower 23 bits of context. */
807 		uasm_i_dins(p, ptr, 0, 0, 23);
808 
809 		/* 1 0  1 0 1  << 6  xkphys cached */
810 		uasm_i_ori(p, ptr, ptr, 0x540);
811 		uasm_i_drotr(p, ptr, ptr, 11);
812 	}
813 #elif defined(CONFIG_SMP)
814 # ifdef  CONFIG_MIPS_MT_SMTC
815 	/*
816 	 * SMTC uses TCBind value as "CPU" index
817 	 */
818 	uasm_i_mfc0(p, ptr, C0_TCBIND);
819 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
820 # else
821 	/*
822 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
823 	 * stored in CONTEXT.
824 	 */
825 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
826 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
827 # endif
828 	UASM_i_LA_mostly(p, tmp, pgdc);
829 	uasm_i_daddu(p, ptr, ptr, tmp);
830 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
831 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
832 #else
833 	UASM_i_LA_mostly(p, ptr, pgdc);
834 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
835 #endif
836 
837 	uasm_l_vmalloc_done(l, *p);
838 
839 	/* get pgd offset in bytes */
840 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
841 
842 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
843 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
844 #ifndef __PAGETABLE_PMD_FOLDED
845 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
846 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
847 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
848 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
849 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
850 #endif
851 }
852 
853 /*
854  * BVADDR is the faulting address, PTR is scratch.
855  * PTR will hold the pgd for vmalloc.
856  */
857 static void __cpuinit
858 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
859 			unsigned int bvaddr, unsigned int ptr,
860 			enum vmalloc64_mode mode)
861 {
862 	long swpd = (long)swapper_pg_dir;
863 	int single_insn_swpd;
864 	int did_vmalloc_branch = 0;
865 
866 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
867 
868 	uasm_l_vmalloc(l, *p);
869 
870 	if (mode != not_refill && check_for_high_segbits) {
871 		if (single_insn_swpd) {
872 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
873 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
874 			did_vmalloc_branch = 1;
875 			/* fall through */
876 		} else {
877 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
878 		}
879 	}
880 	if (!did_vmalloc_branch) {
881 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
882 			uasm_il_b(p, r, label_vmalloc_done);
883 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
884 		} else {
885 			UASM_i_LA_mostly(p, ptr, swpd);
886 			uasm_il_b(p, r, label_vmalloc_done);
887 			if (uasm_in_compat_space_p(swpd))
888 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
889 			else
890 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
891 		}
892 	}
893 	if (mode != not_refill && check_for_high_segbits) {
894 		uasm_l_large_segbits_fault(l, *p);
895 		/*
896 		 * We get here if we are an xsseg address, or if we are
897 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
898 		 *
899 		 * Ignoring xsseg (assume disabled so would generate
900 		 * (address errors?), the only remaining possibility
901 		 * is the upper xuseg addresses.  On processors with
902 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
903 		 * addresses would have taken an address error. We try
904 		 * to mimic that here by taking a load/istream page
905 		 * fault.
906 		 */
907 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
908 		uasm_i_jr(p, ptr);
909 
910 		if (mode == refill_scratch) {
911 			if (scratch_reg > 0)
912 				UASM_i_MFC0(p, 1, 31, scratch_reg);
913 			else
914 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
915 		} else {
916 			uasm_i_nop(p);
917 		}
918 	}
919 }
920 
921 #else /* !CONFIG_64BIT */
922 
923 /*
924  * TMP and PTR are scratch.
925  * TMP will be clobbered, PTR will hold the pgd entry.
926  */
927 static void __cpuinit __maybe_unused
928 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
929 {
930 	long pgdc = (long)pgd_current;
931 
932 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
933 #ifdef CONFIG_SMP
934 #ifdef  CONFIG_MIPS_MT_SMTC
935 	/*
936 	 * SMTC uses TCBind value as "CPU" index
937 	 */
938 	uasm_i_mfc0(p, ptr, C0_TCBIND);
939 	UASM_i_LA_mostly(p, tmp, pgdc);
940 	uasm_i_srl(p, ptr, ptr, 19);
941 #else
942 	/*
943 	 * smp_processor_id() << 3 is stored in CONTEXT.
944          */
945 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
946 	UASM_i_LA_mostly(p, tmp, pgdc);
947 	uasm_i_srl(p, ptr, ptr, 23);
948 #endif
949 	uasm_i_addu(p, ptr, tmp, ptr);
950 #else
951 	UASM_i_LA_mostly(p, ptr, pgdc);
952 #endif
953 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
954 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
955 
956 	if (cpu_has_mips_r2) {
957 		uasm_i_ext(p, tmp, tmp, PGDIR_SHIFT, (32 - PGDIR_SHIFT));
958 		uasm_i_ins(p, ptr, tmp, PGD_T_LOG2, (32 - PGDIR_SHIFT));
959 		return;
960 	}
961 
962 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
963 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
964 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
965 }
966 
967 #endif /* !CONFIG_64BIT */
968 
969 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
970 {
971 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
972 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
973 
974 	switch (current_cpu_type()) {
975 	case CPU_VR41XX:
976 	case CPU_VR4111:
977 	case CPU_VR4121:
978 	case CPU_VR4122:
979 	case CPU_VR4131:
980 	case CPU_VR4181:
981 	case CPU_VR4181A:
982 	case CPU_VR4133:
983 		shift += 2;
984 		break;
985 
986 	default:
987 		break;
988 	}
989 
990 	if (shift)
991 		UASM_i_SRL(p, ctx, ctx, shift);
992 	uasm_i_andi(p, ctx, ctx, mask);
993 }
994 
995 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
996 {
997 	if (cpu_has_mips_r2) {
998 		/* PTE ptr offset is obtained from BadVAddr */
999 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1000 		UASM_i_LW(p, ptr, 0, ptr);
1001 		uasm_i_ext(p, tmp, tmp, PAGE_SHIFT+1, PGDIR_SHIFT-PAGE_SHIFT-1);
1002 		uasm_i_ins(p, ptr, tmp, PTE_T_LOG2+1, PGDIR_SHIFT-PAGE_SHIFT-1);
1003 		return;
1004 	}
1005 
1006 	/*
1007 	 * Bug workaround for the Nevada. It seems as if under certain
1008 	 * circumstances the move from cp0_context might produce a
1009 	 * bogus result when the mfc0 instruction and its consumer are
1010 	 * in a different cacheline or a load instruction, probably any
1011 	 * memory reference, is between them.
1012 	 */
1013 	switch (current_cpu_type()) {
1014 	case CPU_NEVADA:
1015 		UASM_i_LW(p, ptr, 0, ptr);
1016 		GET_CONTEXT(p, tmp); /* get context reg */
1017 		break;
1018 
1019 	default:
1020 		GET_CONTEXT(p, tmp); /* get context reg */
1021 		UASM_i_LW(p, ptr, 0, ptr);
1022 		break;
1023 	}
1024 
1025 	build_adjust_context(p, tmp);
1026 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1027 }
1028 
1029 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
1030 					unsigned int ptep)
1031 {
1032 	/*
1033 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1034 	 * Kernel is a special case. Only a few CPUs use it.
1035 	 */
1036 #ifdef CONFIG_64BIT_PHYS_ADDR
1037 	if (cpu_has_64bits) {
1038 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1039 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1040 		if (cpu_has_rixi) {
1041 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1042 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1043 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1044 		} else {
1045 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1046 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1047 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1048 		}
1049 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1050 	} else {
1051 		int pte_off_even = sizeof(pte_t) / 2;
1052 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1053 
1054 		/* The pte entries are pre-shifted */
1055 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1056 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1057 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1058 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1059 	}
1060 #else
1061 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1062 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1063 	if (r45k_bvahwbug())
1064 		build_tlb_probe_entry(p);
1065 	if (cpu_has_rixi) {
1066 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1067 		if (r4k_250MHZhwbug())
1068 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1069 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1070 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1071 	} else {
1072 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1073 		if (r4k_250MHZhwbug())
1074 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1075 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1076 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1077 		if (r45k_bvahwbug())
1078 			uasm_i_mfc0(p, tmp, C0_INDEX);
1079 	}
1080 	if (r4k_250MHZhwbug())
1081 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1082 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1083 #endif
1084 }
1085 
1086 struct mips_huge_tlb_info {
1087 	int huge_pte;
1088 	int restore_scratch;
1089 };
1090 
1091 static struct mips_huge_tlb_info __cpuinit
1092 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1093 			       struct uasm_reloc **r, unsigned int tmp,
1094 			       unsigned int ptr, int c0_scratch)
1095 {
1096 	struct mips_huge_tlb_info rv;
1097 	unsigned int even, odd;
1098 	int vmalloc_branch_delay_filled = 0;
1099 	const int scratch = 1; /* Our extra working register */
1100 
1101 	rv.huge_pte = scratch;
1102 	rv.restore_scratch = 0;
1103 
1104 	if (check_for_high_segbits) {
1105 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1106 
1107 		if (pgd_reg != -1)
1108 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1109 		else
1110 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1111 
1112 		if (c0_scratch >= 0)
1113 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1114 		else
1115 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1116 
1117 		uasm_i_dsrl_safe(p, scratch, tmp,
1118 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1119 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1120 
1121 		if (pgd_reg == -1) {
1122 			vmalloc_branch_delay_filled = 1;
1123 			/* Clear lower 23 bits of context. */
1124 			uasm_i_dins(p, ptr, 0, 0, 23);
1125 		}
1126 	} else {
1127 		if (pgd_reg != -1)
1128 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1129 		else
1130 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1131 
1132 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1133 
1134 		if (c0_scratch >= 0)
1135 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1136 		else
1137 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1138 
1139 		if (pgd_reg == -1)
1140 			/* Clear lower 23 bits of context. */
1141 			uasm_i_dins(p, ptr, 0, 0, 23);
1142 
1143 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1144 	}
1145 
1146 	if (pgd_reg == -1) {
1147 		vmalloc_branch_delay_filled = 1;
1148 		/* 1 0  1 0 1  << 6  xkphys cached */
1149 		uasm_i_ori(p, ptr, ptr, 0x540);
1150 		uasm_i_drotr(p, ptr, ptr, 11);
1151 	}
1152 
1153 #ifdef __PAGETABLE_PMD_FOLDED
1154 #define LOC_PTEP scratch
1155 #else
1156 #define LOC_PTEP ptr
1157 #endif
1158 
1159 	if (!vmalloc_branch_delay_filled)
1160 		/* get pgd offset in bytes */
1161 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1162 
1163 	uasm_l_vmalloc_done(l, *p);
1164 
1165 	/*
1166 	 *                         tmp          ptr
1167 	 * fall-through case =   badvaddr  *pgd_current
1168 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1169 	 */
1170 
1171 	if (vmalloc_branch_delay_filled)
1172 		/* get pgd offset in bytes */
1173 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1174 
1175 #ifdef __PAGETABLE_PMD_FOLDED
1176 	GET_CONTEXT(p, tmp); /* get context reg */
1177 #endif
1178 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1179 
1180 	if (use_lwx_insns()) {
1181 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1182 	} else {
1183 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1184 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1185 	}
1186 
1187 #ifndef __PAGETABLE_PMD_FOLDED
1188 	/* get pmd offset in bytes */
1189 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1190 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1191 	GET_CONTEXT(p, tmp); /* get context reg */
1192 
1193 	if (use_lwx_insns()) {
1194 		UASM_i_LWX(p, scratch, scratch, ptr);
1195 	} else {
1196 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1197 		UASM_i_LW(p, scratch, 0, ptr);
1198 	}
1199 #endif
1200 	/* Adjust the context during the load latency. */
1201 	build_adjust_context(p, tmp);
1202 
1203 #ifdef CONFIG_HUGETLB_PAGE
1204 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1205 	/*
1206 	 * The in the LWX case we don't want to do the load in the
1207 	 * delay slot.  It cannot issue in the same cycle and may be
1208 	 * speculative and unneeded.
1209 	 */
1210 	if (use_lwx_insns())
1211 		uasm_i_nop(p);
1212 #endif /* CONFIG_HUGETLB_PAGE */
1213 
1214 
1215 	/* build_update_entries */
1216 	if (use_lwx_insns()) {
1217 		even = ptr;
1218 		odd = tmp;
1219 		UASM_i_LWX(p, even, scratch, tmp);
1220 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1221 		UASM_i_LWX(p, odd, scratch, tmp);
1222 	} else {
1223 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1224 		even = tmp;
1225 		odd = ptr;
1226 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1227 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1228 	}
1229 	if (cpu_has_rixi) {
1230 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1231 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1232 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1233 	} else {
1234 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1235 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1236 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1237 	}
1238 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1239 
1240 	if (c0_scratch >= 0) {
1241 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1242 		build_tlb_write_entry(p, l, r, tlb_random);
1243 		uasm_l_leave(l, *p);
1244 		rv.restore_scratch = 1;
1245 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1246 		build_tlb_write_entry(p, l, r, tlb_random);
1247 		uasm_l_leave(l, *p);
1248 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1249 	} else {
1250 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1251 		build_tlb_write_entry(p, l, r, tlb_random);
1252 		uasm_l_leave(l, *p);
1253 		rv.restore_scratch = 1;
1254 	}
1255 
1256 	uasm_i_eret(p); /* return from trap */
1257 
1258 	return rv;
1259 }
1260 
1261 /*
1262  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1263  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1264  * slots before the XTLB refill exception handler which belong to the
1265  * unused TLB refill exception.
1266  */
1267 #define MIPS64_REFILL_INSNS 32
1268 
1269 static void __cpuinit build_r4000_tlb_refill_handler(void)
1270 {
1271 	u32 *p = tlb_handler;
1272 	struct uasm_label *l = labels;
1273 	struct uasm_reloc *r = relocs;
1274 	u32 *f;
1275 	unsigned int final_len;
1276 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1277 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1278 
1279 	memset(tlb_handler, 0, sizeof(tlb_handler));
1280 	memset(labels, 0, sizeof(labels));
1281 	memset(relocs, 0, sizeof(relocs));
1282 	memset(final_handler, 0, sizeof(final_handler));
1283 
1284 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1285 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1286 							  scratch_reg);
1287 		vmalloc_mode = refill_scratch;
1288 	} else {
1289 		htlb_info.huge_pte = K0;
1290 		htlb_info.restore_scratch = 0;
1291 		vmalloc_mode = refill_noscratch;
1292 		/*
1293 		 * create the plain linear handler
1294 		 */
1295 		if (bcm1250_m3_war()) {
1296 			unsigned int segbits = 44;
1297 
1298 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1299 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1300 			uasm_i_xor(&p, K0, K0, K1);
1301 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1302 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1303 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1304 			uasm_i_or(&p, K0, K0, K1);
1305 			uasm_il_bnez(&p, &r, K0, label_leave);
1306 			/* No need for uasm_i_nop */
1307 		}
1308 
1309 #ifdef CONFIG_64BIT
1310 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1311 #else
1312 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1313 #endif
1314 
1315 #ifdef CONFIG_HUGETLB_PAGE
1316 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1317 #endif
1318 
1319 		build_get_ptep(&p, K0, K1);
1320 		build_update_entries(&p, K0, K1);
1321 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1322 		uasm_l_leave(&l, p);
1323 		uasm_i_eret(&p); /* return from trap */
1324 	}
1325 #ifdef CONFIG_HUGETLB_PAGE
1326 	uasm_l_tlb_huge_update(&l, p);
1327 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1328 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1329 				   htlb_info.restore_scratch);
1330 #endif
1331 
1332 #ifdef CONFIG_64BIT
1333 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1334 #endif
1335 
1336 	/*
1337 	 * Overflow check: For the 64bit handler, we need at least one
1338 	 * free instruction slot for the wrap-around branch. In worst
1339 	 * case, if the intended insertion point is a delay slot, we
1340 	 * need three, with the second nop'ed and the third being
1341 	 * unused.
1342 	 */
1343 	/* Loongson2 ebase is different than r4k, we have more space */
1344 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1345 	if ((p - tlb_handler) > 64)
1346 		panic("TLB refill handler space exceeded");
1347 #else
1348 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1349 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1350 		&& uasm_insn_has_bdelay(relocs,
1351 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1352 		panic("TLB refill handler space exceeded");
1353 #endif
1354 
1355 	/*
1356 	 * Now fold the handler in the TLB refill handler space.
1357 	 */
1358 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1359 	f = final_handler;
1360 	/* Simplest case, just copy the handler. */
1361 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1362 	final_len = p - tlb_handler;
1363 #else /* CONFIG_64BIT */
1364 	f = final_handler + MIPS64_REFILL_INSNS;
1365 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1366 		/* Just copy the handler. */
1367 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1368 		final_len = p - tlb_handler;
1369 	} else {
1370 #if defined(CONFIG_HUGETLB_PAGE)
1371 		const enum label_id ls = label_tlb_huge_update;
1372 #else
1373 		const enum label_id ls = label_vmalloc;
1374 #endif
1375 		u32 *split;
1376 		int ov = 0;
1377 		int i;
1378 
1379 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1380 			;
1381 		BUG_ON(i == ARRAY_SIZE(labels));
1382 		split = labels[i].addr;
1383 
1384 		/*
1385 		 * See if we have overflown one way or the other.
1386 		 */
1387 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1388 		    split < p - MIPS64_REFILL_INSNS)
1389 			ov = 1;
1390 
1391 		if (ov) {
1392 			/*
1393 			 * Split two instructions before the end.  One
1394 			 * for the branch and one for the instruction
1395 			 * in the delay slot.
1396 			 */
1397 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1398 
1399 			/*
1400 			 * If the branch would fall in a delay slot,
1401 			 * we must back up an additional instruction
1402 			 * so that it is no longer in a delay slot.
1403 			 */
1404 			if (uasm_insn_has_bdelay(relocs, split - 1))
1405 				split--;
1406 		}
1407 		/* Copy first part of the handler. */
1408 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1409 		f += split - tlb_handler;
1410 
1411 		if (ov) {
1412 			/* Insert branch. */
1413 			uasm_l_split(&l, final_handler);
1414 			uasm_il_b(&f, &r, label_split);
1415 			if (uasm_insn_has_bdelay(relocs, split))
1416 				uasm_i_nop(&f);
1417 			else {
1418 				uasm_copy_handler(relocs, labels,
1419 						  split, split + 1, f);
1420 				uasm_move_labels(labels, f, f + 1, -1);
1421 				f++;
1422 				split++;
1423 			}
1424 		}
1425 
1426 		/* Copy the rest of the handler. */
1427 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1428 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1429 			    (p - split);
1430 	}
1431 #endif /* CONFIG_64BIT */
1432 
1433 	uasm_resolve_relocs(relocs, labels);
1434 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1435 		 final_len);
1436 
1437 	memcpy((void *)ebase, final_handler, 0x100);
1438 
1439 	dump_handler((u32 *)ebase, 64);
1440 }
1441 
1442 /*
1443  * 128 instructions for the fastpath handler is generous and should
1444  * never be exceeded.
1445  */
1446 #define FASTPATH_SIZE 128
1447 
1448 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1449 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1450 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1451 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1452 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1453 
1454 static void __cpuinit build_r4000_setup_pgd(void)
1455 {
1456 	const int a0 = 4;
1457 	const int a1 = 5;
1458 	u32 *p = tlbmiss_handler_setup_pgd;
1459 	struct uasm_label *l = labels;
1460 	struct uasm_reloc *r = relocs;
1461 
1462 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1463 	memset(labels, 0, sizeof(labels));
1464 	memset(relocs, 0, sizeof(relocs));
1465 
1466 	pgd_reg = allocate_kscratch();
1467 
1468 	if (pgd_reg == -1) {
1469 		/* PGD << 11 in c0_Context */
1470 		/*
1471 		 * If it is a ckseg0 address, convert to a physical
1472 		 * address.  Shifting right by 29 and adding 4 will
1473 		 * result in zero for these addresses.
1474 		 *
1475 		 */
1476 		UASM_i_SRA(&p, a1, a0, 29);
1477 		UASM_i_ADDIU(&p, a1, a1, 4);
1478 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1479 		uasm_i_nop(&p);
1480 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1481 		uasm_l_tlbl_goaround1(&l, p);
1482 		UASM_i_SLL(&p, a0, a0, 11);
1483 		uasm_i_jr(&p, 31);
1484 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1485 	} else {
1486 		/* PGD in c0_KScratch */
1487 		uasm_i_jr(&p, 31);
1488 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1489 	}
1490 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1491 		panic("tlbmiss_handler_setup_pgd space exceeded");
1492 	uasm_resolve_relocs(relocs, labels);
1493 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1494 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1495 
1496 	dump_handler(tlbmiss_handler_setup_pgd,
1497 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1498 }
1499 #endif
1500 
1501 static void __cpuinit
1502 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1503 {
1504 #ifdef CONFIG_SMP
1505 # ifdef CONFIG_64BIT_PHYS_ADDR
1506 	if (cpu_has_64bits)
1507 		uasm_i_lld(p, pte, 0, ptr);
1508 	else
1509 # endif
1510 		UASM_i_LL(p, pte, 0, ptr);
1511 #else
1512 # ifdef CONFIG_64BIT_PHYS_ADDR
1513 	if (cpu_has_64bits)
1514 		uasm_i_ld(p, pte, 0, ptr);
1515 	else
1516 # endif
1517 		UASM_i_LW(p, pte, 0, ptr);
1518 #endif
1519 }
1520 
1521 static void __cpuinit
1522 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1523 	unsigned int mode)
1524 {
1525 #ifdef CONFIG_64BIT_PHYS_ADDR
1526 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1527 #endif
1528 
1529 	uasm_i_ori(p, pte, pte, mode);
1530 #ifdef CONFIG_SMP
1531 # ifdef CONFIG_64BIT_PHYS_ADDR
1532 	if (cpu_has_64bits)
1533 		uasm_i_scd(p, pte, 0, ptr);
1534 	else
1535 # endif
1536 		UASM_i_SC(p, pte, 0, ptr);
1537 
1538 	if (r10000_llsc_war())
1539 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1540 	else
1541 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1542 
1543 # ifdef CONFIG_64BIT_PHYS_ADDR
1544 	if (!cpu_has_64bits) {
1545 		/* no uasm_i_nop needed */
1546 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1547 		uasm_i_ori(p, pte, pte, hwmode);
1548 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1549 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1550 		/* no uasm_i_nop needed */
1551 		uasm_i_lw(p, pte, 0, ptr);
1552 	} else
1553 		uasm_i_nop(p);
1554 # else
1555 	uasm_i_nop(p);
1556 # endif
1557 #else
1558 # ifdef CONFIG_64BIT_PHYS_ADDR
1559 	if (cpu_has_64bits)
1560 		uasm_i_sd(p, pte, 0, ptr);
1561 	else
1562 # endif
1563 		UASM_i_SW(p, pte, 0, ptr);
1564 
1565 # ifdef CONFIG_64BIT_PHYS_ADDR
1566 	if (!cpu_has_64bits) {
1567 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1568 		uasm_i_ori(p, pte, pte, hwmode);
1569 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1570 		uasm_i_lw(p, pte, 0, ptr);
1571 	}
1572 # endif
1573 #endif
1574 }
1575 
1576 /*
1577  * Check if PTE is present, if not then jump to LABEL. PTR points to
1578  * the page table where this PTE is located, PTE will be re-loaded
1579  * with it's original value.
1580  */
1581 static void __cpuinit
1582 build_pte_present(u32 **p, struct uasm_reloc **r,
1583 		  int pte, int ptr, int scratch, enum label_id lid)
1584 {
1585 	int t = scratch >= 0 ? scratch : pte;
1586 
1587 	if (cpu_has_rixi) {
1588 		if (use_bbit_insns()) {
1589 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1590 			uasm_i_nop(p);
1591 		} else {
1592 			uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1593 			uasm_il_beqz(p, r, t, lid);
1594 			if (pte == t)
1595 				/* You lose the SMP race :-(*/
1596 				iPTE_LW(p, pte, ptr);
1597 		}
1598 	} else {
1599 		uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1600 		uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1601 		uasm_il_bnez(p, r, t, lid);
1602 		if (pte == t)
1603 			/* You lose the SMP race :-(*/
1604 			iPTE_LW(p, pte, ptr);
1605 	}
1606 }
1607 
1608 /* Make PTE valid, store result in PTR. */
1609 static void __cpuinit
1610 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1611 		 unsigned int ptr)
1612 {
1613 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1614 
1615 	iPTE_SW(p, r, pte, ptr, mode);
1616 }
1617 
1618 /*
1619  * Check if PTE can be written to, if not branch to LABEL. Regardless
1620  * restore PTE with value from PTR when done.
1621  */
1622 static void __cpuinit
1623 build_pte_writable(u32 **p, struct uasm_reloc **r,
1624 		   unsigned int pte, unsigned int ptr, int scratch,
1625 		   enum label_id lid)
1626 {
1627 	int t = scratch >= 0 ? scratch : pte;
1628 
1629 	uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1630 	uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1631 	uasm_il_bnez(p, r, t, lid);
1632 	if (pte == t)
1633 		/* You lose the SMP race :-(*/
1634 		iPTE_LW(p, pte, ptr);
1635 	else
1636 		uasm_i_nop(p);
1637 }
1638 
1639 /* Make PTE writable, update software status bits as well, then store
1640  * at PTR.
1641  */
1642 static void __cpuinit
1643 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1644 		 unsigned int ptr)
1645 {
1646 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1647 			     | _PAGE_DIRTY);
1648 
1649 	iPTE_SW(p, r, pte, ptr, mode);
1650 }
1651 
1652 /*
1653  * Check if PTE can be modified, if not branch to LABEL. Regardless
1654  * restore PTE with value from PTR when done.
1655  */
1656 static void __cpuinit
1657 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1658 		     unsigned int pte, unsigned int ptr, int scratch,
1659 		     enum label_id lid)
1660 {
1661 	if (use_bbit_insns()) {
1662 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1663 		uasm_i_nop(p);
1664 	} else {
1665 		int t = scratch >= 0 ? scratch : pte;
1666 		uasm_i_andi(p, t, pte, _PAGE_WRITE);
1667 		uasm_il_beqz(p, r, t, lid);
1668 		if (pte == t)
1669 			/* You lose the SMP race :-(*/
1670 			iPTE_LW(p, pte, ptr);
1671 	}
1672 }
1673 
1674 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1675 
1676 
1677 /*
1678  * R3000 style TLB load/store/modify handlers.
1679  */
1680 
1681 /*
1682  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1683  * Then it returns.
1684  */
1685 static void __cpuinit
1686 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1687 {
1688 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1689 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1690 	uasm_i_tlbwi(p);
1691 	uasm_i_jr(p, tmp);
1692 	uasm_i_rfe(p); /* branch delay */
1693 }
1694 
1695 /*
1696  * This places the pte into ENTRYLO0 and writes it with tlbwi
1697  * or tlbwr as appropriate.  This is because the index register
1698  * may have the probe fail bit set as a result of a trap on a
1699  * kseg2 access, i.e. without refill.  Then it returns.
1700  */
1701 static void __cpuinit
1702 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1703 			     struct uasm_reloc **r, unsigned int pte,
1704 			     unsigned int tmp)
1705 {
1706 	uasm_i_mfc0(p, tmp, C0_INDEX);
1707 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1708 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1709 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1710 	uasm_i_tlbwi(p); /* cp0 delay */
1711 	uasm_i_jr(p, tmp);
1712 	uasm_i_rfe(p); /* branch delay */
1713 	uasm_l_r3000_write_probe_fail(l, *p);
1714 	uasm_i_tlbwr(p); /* cp0 delay */
1715 	uasm_i_jr(p, tmp);
1716 	uasm_i_rfe(p); /* branch delay */
1717 }
1718 
1719 static void __cpuinit
1720 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1721 				   unsigned int ptr)
1722 {
1723 	long pgdc = (long)pgd_current;
1724 
1725 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1726 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1727 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1728 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1729 	uasm_i_sll(p, pte, pte, 2);
1730 	uasm_i_addu(p, ptr, ptr, pte);
1731 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1732 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1733 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1734 	uasm_i_addu(p, ptr, ptr, pte);
1735 	uasm_i_lw(p, pte, 0, ptr);
1736 	uasm_i_tlbp(p); /* load delay */
1737 }
1738 
1739 static void __cpuinit build_r3000_tlb_load_handler(void)
1740 {
1741 	u32 *p = handle_tlbl;
1742 	struct uasm_label *l = labels;
1743 	struct uasm_reloc *r = relocs;
1744 
1745 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1746 	memset(labels, 0, sizeof(labels));
1747 	memset(relocs, 0, sizeof(relocs));
1748 
1749 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1750 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1751 	uasm_i_nop(&p); /* load delay */
1752 	build_make_valid(&p, &r, K0, K1);
1753 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1754 
1755 	uasm_l_nopage_tlbl(&l, p);
1756 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1757 	uasm_i_nop(&p);
1758 
1759 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1760 		panic("TLB load handler fastpath space exceeded");
1761 
1762 	uasm_resolve_relocs(relocs, labels);
1763 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1764 		 (unsigned int)(p - handle_tlbl));
1765 
1766 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1767 }
1768 
1769 static void __cpuinit build_r3000_tlb_store_handler(void)
1770 {
1771 	u32 *p = handle_tlbs;
1772 	struct uasm_label *l = labels;
1773 	struct uasm_reloc *r = relocs;
1774 
1775 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1776 	memset(labels, 0, sizeof(labels));
1777 	memset(relocs, 0, sizeof(relocs));
1778 
1779 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1780 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1781 	uasm_i_nop(&p); /* load delay */
1782 	build_make_write(&p, &r, K0, K1);
1783 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1784 
1785 	uasm_l_nopage_tlbs(&l, p);
1786 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1787 	uasm_i_nop(&p);
1788 
1789 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1790 		panic("TLB store handler fastpath space exceeded");
1791 
1792 	uasm_resolve_relocs(relocs, labels);
1793 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1794 		 (unsigned int)(p - handle_tlbs));
1795 
1796 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1797 }
1798 
1799 static void __cpuinit build_r3000_tlb_modify_handler(void)
1800 {
1801 	u32 *p = handle_tlbm;
1802 	struct uasm_label *l = labels;
1803 	struct uasm_reloc *r = relocs;
1804 
1805 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1806 	memset(labels, 0, sizeof(labels));
1807 	memset(relocs, 0, sizeof(relocs));
1808 
1809 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1810 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1811 	uasm_i_nop(&p); /* load delay */
1812 	build_make_write(&p, &r, K0, K1);
1813 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1814 
1815 	uasm_l_nopage_tlbm(&l, p);
1816 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1817 	uasm_i_nop(&p);
1818 
1819 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1820 		panic("TLB modify handler fastpath space exceeded");
1821 
1822 	uasm_resolve_relocs(relocs, labels);
1823 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1824 		 (unsigned int)(p - handle_tlbm));
1825 
1826 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1827 }
1828 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1829 
1830 /*
1831  * R4000 style TLB load/store/modify handlers.
1832  */
1833 static struct work_registers __cpuinit
1834 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1835 				   struct uasm_reloc **r)
1836 {
1837 	struct work_registers wr = build_get_work_registers(p);
1838 
1839 #ifdef CONFIG_64BIT
1840 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1841 #else
1842 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1843 #endif
1844 
1845 #ifdef CONFIG_HUGETLB_PAGE
1846 	/*
1847 	 * For huge tlb entries, pmd doesn't contain an address but
1848 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1849 	 * see if we need to jump to huge tlb processing.
1850 	 */
1851 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1852 #endif
1853 
1854 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1855 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1856 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1857 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1858 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1859 
1860 #ifdef CONFIG_SMP
1861 	uasm_l_smp_pgtable_change(l, *p);
1862 #endif
1863 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1864 	if (!m4kc_tlbp_war())
1865 		build_tlb_probe_entry(p);
1866 	return wr;
1867 }
1868 
1869 static void __cpuinit
1870 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1871 				   struct uasm_reloc **r, unsigned int tmp,
1872 				   unsigned int ptr)
1873 {
1874 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1875 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1876 	build_update_entries(p, tmp, ptr);
1877 	build_tlb_write_entry(p, l, r, tlb_indexed);
1878 	uasm_l_leave(l, *p);
1879 	build_restore_work_registers(p);
1880 	uasm_i_eret(p); /* return from trap */
1881 
1882 #ifdef CONFIG_64BIT
1883 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1884 #endif
1885 }
1886 
1887 static void __cpuinit build_r4000_tlb_load_handler(void)
1888 {
1889 	u32 *p = handle_tlbl;
1890 	struct uasm_label *l = labels;
1891 	struct uasm_reloc *r = relocs;
1892 	struct work_registers wr;
1893 
1894 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1895 	memset(labels, 0, sizeof(labels));
1896 	memset(relocs, 0, sizeof(relocs));
1897 
1898 	if (bcm1250_m3_war()) {
1899 		unsigned int segbits = 44;
1900 
1901 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1902 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1903 		uasm_i_xor(&p, K0, K0, K1);
1904 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1905 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1906 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1907 		uasm_i_or(&p, K0, K0, K1);
1908 		uasm_il_bnez(&p, &r, K0, label_leave);
1909 		/* No need for uasm_i_nop */
1910 	}
1911 
1912 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1913 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1914 	if (m4kc_tlbp_war())
1915 		build_tlb_probe_entry(&p);
1916 
1917 	if (cpu_has_rixi) {
1918 		/*
1919 		 * If the page is not _PAGE_VALID, RI or XI could not
1920 		 * have triggered it.  Skip the expensive test..
1921 		 */
1922 		if (use_bbit_insns()) {
1923 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1924 				      label_tlbl_goaround1);
1925 		} else {
1926 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1927 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1928 		}
1929 		uasm_i_nop(&p);
1930 
1931 		uasm_i_tlbr(&p);
1932 		/* Examine  entrylo 0 or 1 based on ptr. */
1933 		if (use_bbit_insns()) {
1934 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1935 		} else {
1936 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1937 			uasm_i_beqz(&p, wr.r3, 8);
1938 		}
1939 		/* load it in the delay slot*/
1940 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1941 		/* load it if ptr is odd */
1942 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1943 		/*
1944 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1945 		 * XI must have triggered it.
1946 		 */
1947 		if (use_bbit_insns()) {
1948 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1949 			uasm_i_nop(&p);
1950 			uasm_l_tlbl_goaround1(&l, p);
1951 		} else {
1952 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1953 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1954 			uasm_i_nop(&p);
1955 		}
1956 		uasm_l_tlbl_goaround1(&l, p);
1957 	}
1958 	build_make_valid(&p, &r, wr.r1, wr.r2);
1959 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1960 
1961 #ifdef CONFIG_HUGETLB_PAGE
1962 	/*
1963 	 * This is the entry point when build_r4000_tlbchange_handler_head
1964 	 * spots a huge page.
1965 	 */
1966 	uasm_l_tlb_huge_update(&l, p);
1967 	iPTE_LW(&p, wr.r1, wr.r2);
1968 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1969 	build_tlb_probe_entry(&p);
1970 
1971 	if (cpu_has_rixi) {
1972 		/*
1973 		 * If the page is not _PAGE_VALID, RI or XI could not
1974 		 * have triggered it.  Skip the expensive test..
1975 		 */
1976 		if (use_bbit_insns()) {
1977 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1978 				      label_tlbl_goaround2);
1979 		} else {
1980 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1981 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1982 		}
1983 		uasm_i_nop(&p);
1984 
1985 		uasm_i_tlbr(&p);
1986 		/* Examine  entrylo 0 or 1 based on ptr. */
1987 		if (use_bbit_insns()) {
1988 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1989 		} else {
1990 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1991 			uasm_i_beqz(&p, wr.r3, 8);
1992 		}
1993 		/* load it in the delay slot*/
1994 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1995 		/* load it if ptr is odd */
1996 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1997 		/*
1998 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1999 		 * XI must have triggered it.
2000 		 */
2001 		if (use_bbit_insns()) {
2002 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2003 		} else {
2004 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2005 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2006 		}
2007 		if (PM_DEFAULT_MASK == 0)
2008 			uasm_i_nop(&p);
2009 		/*
2010 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2011 		 * it is restored in build_huge_tlb_write_entry.
2012 		 */
2013 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2014 
2015 		uasm_l_tlbl_goaround2(&l, p);
2016 	}
2017 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2018 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2019 #endif
2020 
2021 	uasm_l_nopage_tlbl(&l, p);
2022 	build_restore_work_registers(&p);
2023 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2024 	uasm_i_nop(&p);
2025 
2026 	if ((p - handle_tlbl) > FASTPATH_SIZE)
2027 		panic("TLB load handler fastpath space exceeded");
2028 
2029 	uasm_resolve_relocs(relocs, labels);
2030 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2031 		 (unsigned int)(p - handle_tlbl));
2032 
2033 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
2034 }
2035 
2036 static void __cpuinit build_r4000_tlb_store_handler(void)
2037 {
2038 	u32 *p = handle_tlbs;
2039 	struct uasm_label *l = labels;
2040 	struct uasm_reloc *r = relocs;
2041 	struct work_registers wr;
2042 
2043 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
2044 	memset(labels, 0, sizeof(labels));
2045 	memset(relocs, 0, sizeof(relocs));
2046 
2047 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2048 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2049 	if (m4kc_tlbp_war())
2050 		build_tlb_probe_entry(&p);
2051 	build_make_write(&p, &r, wr.r1, wr.r2);
2052 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2053 
2054 #ifdef CONFIG_HUGETLB_PAGE
2055 	/*
2056 	 * This is the entry point when
2057 	 * build_r4000_tlbchange_handler_head spots a huge page.
2058 	 */
2059 	uasm_l_tlb_huge_update(&l, p);
2060 	iPTE_LW(&p, wr.r1, wr.r2);
2061 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2062 	build_tlb_probe_entry(&p);
2063 	uasm_i_ori(&p, wr.r1, wr.r1,
2064 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2065 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2066 #endif
2067 
2068 	uasm_l_nopage_tlbs(&l, p);
2069 	build_restore_work_registers(&p);
2070 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2071 	uasm_i_nop(&p);
2072 
2073 	if ((p - handle_tlbs) > FASTPATH_SIZE)
2074 		panic("TLB store handler fastpath space exceeded");
2075 
2076 	uasm_resolve_relocs(relocs, labels);
2077 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2078 		 (unsigned int)(p - handle_tlbs));
2079 
2080 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
2081 }
2082 
2083 static void __cpuinit build_r4000_tlb_modify_handler(void)
2084 {
2085 	u32 *p = handle_tlbm;
2086 	struct uasm_label *l = labels;
2087 	struct uasm_reloc *r = relocs;
2088 	struct work_registers wr;
2089 
2090 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
2091 	memset(labels, 0, sizeof(labels));
2092 	memset(relocs, 0, sizeof(relocs));
2093 
2094 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2095 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2096 	if (m4kc_tlbp_war())
2097 		build_tlb_probe_entry(&p);
2098 	/* Present and writable bits set, set accessed and dirty bits. */
2099 	build_make_write(&p, &r, wr.r1, wr.r2);
2100 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2101 
2102 #ifdef CONFIG_HUGETLB_PAGE
2103 	/*
2104 	 * This is the entry point when
2105 	 * build_r4000_tlbchange_handler_head spots a huge page.
2106 	 */
2107 	uasm_l_tlb_huge_update(&l, p);
2108 	iPTE_LW(&p, wr.r1, wr.r2);
2109 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2110 	build_tlb_probe_entry(&p);
2111 	uasm_i_ori(&p, wr.r1, wr.r1,
2112 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2113 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2114 #endif
2115 
2116 	uasm_l_nopage_tlbm(&l, p);
2117 	build_restore_work_registers(&p);
2118 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2119 	uasm_i_nop(&p);
2120 
2121 	if ((p - handle_tlbm) > FASTPATH_SIZE)
2122 		panic("TLB modify handler fastpath space exceeded");
2123 
2124 	uasm_resolve_relocs(relocs, labels);
2125 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2126 		 (unsigned int)(p - handle_tlbm));
2127 
2128 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
2129 }
2130 
2131 void __cpuinit build_tlb_refill_handler(void)
2132 {
2133 	/*
2134 	 * The refill handler is generated per-CPU, multi-node systems
2135 	 * may have local storage for it. The other handlers are only
2136 	 * needed once.
2137 	 */
2138 	static int run_once = 0;
2139 
2140 #ifdef CONFIG_64BIT
2141 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2142 #endif
2143 
2144 	switch (current_cpu_type()) {
2145 	case CPU_R2000:
2146 	case CPU_R3000:
2147 	case CPU_R3000A:
2148 	case CPU_R3081E:
2149 	case CPU_TX3912:
2150 	case CPU_TX3922:
2151 	case CPU_TX3927:
2152 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2153 		build_r3000_tlb_refill_handler();
2154 		if (!run_once) {
2155 			build_r3000_tlb_load_handler();
2156 			build_r3000_tlb_store_handler();
2157 			build_r3000_tlb_modify_handler();
2158 			run_once++;
2159 		}
2160 #else
2161 		panic("No R3000 TLB refill handler");
2162 #endif
2163 		break;
2164 
2165 	case CPU_R6000:
2166 	case CPU_R6000A:
2167 		panic("No R6000 TLB refill handler yet");
2168 		break;
2169 
2170 	case CPU_R8000:
2171 		panic("No R8000 TLB refill handler yet");
2172 		break;
2173 
2174 	default:
2175 		if (!run_once) {
2176 			scratch_reg = allocate_kscratch();
2177 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2178 			build_r4000_setup_pgd();
2179 #endif
2180 			build_r4000_tlb_load_handler();
2181 			build_r4000_tlb_store_handler();
2182 			build_r4000_tlb_modify_handler();
2183 			run_once++;
2184 		}
2185 		build_r4000_tlb_refill_handler();
2186 	}
2187 }
2188 
2189 void __cpuinit flush_tlb_handlers(void)
2190 {
2191 	local_flush_icache_range((unsigned long)handle_tlbl,
2192 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2193 	local_flush_icache_range((unsigned long)handle_tlbs,
2194 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2195 	local_flush_icache_range((unsigned long)handle_tlbm,
2196 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2197 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2198 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2199 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2200 #endif
2201 }
2202