xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 7e035230)
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,
152 	label_split,
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 UASM_L_LA(_tlbw_hazard)
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 /*
185  * For debug purposes.
186  */
187 static inline void dump_handler(const u32 *handler, int count)
188 {
189 	int i;
190 
191 	pr_debug("\t.set push\n");
192 	pr_debug("\t.set noreorder\n");
193 
194 	for (i = 0; i < count; i++)
195 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
196 
197 	pr_debug("\t.set pop\n");
198 }
199 
200 /* The only general purpose registers allowed in TLB handlers. */
201 #define K0		26
202 #define K1		27
203 
204 /* Some CP0 registers */
205 #define C0_INDEX	0, 0
206 #define C0_ENTRYLO0	2, 0
207 #define C0_TCBIND	2, 2
208 #define C0_ENTRYLO1	3, 0
209 #define C0_CONTEXT	4, 0
210 #define C0_PAGEMASK	5, 0
211 #define C0_BADVADDR	8, 0
212 #define C0_ENTRYHI	10, 0
213 #define C0_EPC		14, 0
214 #define C0_XCONTEXT	20, 0
215 
216 #ifdef CONFIG_64BIT
217 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
218 #else
219 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
220 #endif
221 
222 /* The worst case length of the handler is around 18 instructions for
223  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
224  * Maximum space available is 32 instructions for R3000 and 64
225  * instructions for R4000.
226  *
227  * We deliberately chose a buffer size of 128, so we won't scribble
228  * over anything important on overflow before we panic.
229  */
230 static u32 tlb_handler[128] __cpuinitdata;
231 
232 /* simply assume worst case size for labels and relocs */
233 static struct uasm_label labels[128] __cpuinitdata;
234 static struct uasm_reloc relocs[128] __cpuinitdata;
235 
236 #ifdef CONFIG_64BIT
237 static int check_for_high_segbits __cpuinitdata;
238 #endif
239 
240 static int check_for_high_segbits __cpuinitdata;
241 
242 static unsigned int kscratch_used_mask __cpuinitdata;
243 
244 static int __cpuinit allocate_kscratch(void)
245 {
246 	int r;
247 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
248 
249 	r = ffs(a);
250 
251 	if (r == 0)
252 		return -1;
253 
254 	r--; /* make it zero based */
255 
256 	kscratch_used_mask |= (1 << r);
257 
258 	return r;
259 }
260 
261 static int scratch_reg __cpuinitdata;
262 static int pgd_reg __cpuinitdata;
263 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
264 
265 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
266 {
267 	struct work_registers r;
268 
269 	int smp_processor_id_reg;
270 	int smp_processor_id_sel;
271 	int smp_processor_id_shift;
272 
273 	if (scratch_reg > 0) {
274 		/* Save in CPU local C0_KScratch? */
275 		UASM_i_MTC0(p, 1, 31, scratch_reg);
276 		r.r1 = K0;
277 		r.r2 = K1;
278 		r.r3 = 1;
279 		return r;
280 	}
281 
282 	if (num_possible_cpus() > 1) {
283 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
284 		smp_processor_id_shift = 51;
285 		smp_processor_id_reg = 20; /* XContext */
286 		smp_processor_id_sel = 0;
287 #else
288 # ifdef CONFIG_32BIT
289 		smp_processor_id_shift = 25;
290 		smp_processor_id_reg = 4; /* Context */
291 		smp_processor_id_sel = 0;
292 # endif
293 # ifdef CONFIG_64BIT
294 		smp_processor_id_shift = 26;
295 		smp_processor_id_reg = 4; /* Context */
296 		smp_processor_id_sel = 0;
297 # endif
298 #endif
299 		/* Get smp_processor_id */
300 		UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
301 		UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
302 
303 		/* handler_reg_save index in K0 */
304 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
305 
306 		UASM_i_LA(p, K1, (long)&handler_reg_save);
307 		UASM_i_ADDU(p, K0, K0, K1);
308 	} else {
309 		UASM_i_LA(p, K0, (long)&handler_reg_save);
310 	}
311 	/* K0 now points to save area, save $1 and $2  */
312 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
313 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
314 
315 	r.r1 = K1;
316 	r.r2 = 1;
317 	r.r3 = 2;
318 	return r;
319 }
320 
321 static void __cpuinit build_restore_work_registers(u32 **p)
322 {
323 	if (scratch_reg > 0) {
324 		UASM_i_MFC0(p, 1, 31, scratch_reg);
325 		return;
326 	}
327 	/* K0 already points to save area, restore $1 and $2  */
328 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
329 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
330 }
331 
332 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
333 
334 /*
335  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
336  * we cannot do r3000 under these circumstances.
337  *
338  * Declare pgd_current here instead of including mmu_context.h to avoid type
339  * conflicts for tlbmiss_handler_setup_pgd
340  */
341 extern unsigned long pgd_current[];
342 
343 /*
344  * The R3000 TLB handler is simple.
345  */
346 static void __cpuinit build_r3000_tlb_refill_handler(void)
347 {
348 	long pgdc = (long)pgd_current;
349 	u32 *p;
350 
351 	memset(tlb_handler, 0, sizeof(tlb_handler));
352 	p = tlb_handler;
353 
354 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
355 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
356 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
357 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
358 	uasm_i_sll(&p, K0, K0, 2);
359 	uasm_i_addu(&p, K1, K1, K0);
360 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
361 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
362 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
363 	uasm_i_addu(&p, K1, K1, K0);
364 	uasm_i_lw(&p, K0, 0, K1);
365 	uasm_i_nop(&p); /* load delay */
366 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
367 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
368 	uasm_i_tlbwr(&p); /* cp0 delay */
369 	uasm_i_jr(&p, K1);
370 	uasm_i_rfe(&p); /* branch delay */
371 
372 	if (p > tlb_handler + 32)
373 		panic("TLB refill handler space exceeded");
374 
375 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
376 		 (unsigned int)(p - tlb_handler));
377 
378 	memcpy((void *)ebase, tlb_handler, 0x80);
379 
380 	dump_handler((u32 *)ebase, 32);
381 }
382 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
383 
384 /*
385  * The R4000 TLB handler is much more complicated. We have two
386  * consecutive handler areas with 32 instructions space each.
387  * Since they aren't used at the same time, we can overflow in the
388  * other one.To keep things simple, we first assume linear space,
389  * then we relocate it to the final handler layout as needed.
390  */
391 static u32 final_handler[64] __cpuinitdata;
392 
393 /*
394  * Hazards
395  *
396  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
397  * 2. A timing hazard exists for the TLBP instruction.
398  *
399  *      stalling_instruction
400  *      TLBP
401  *
402  * The JTLB is being read for the TLBP throughout the stall generated by the
403  * previous instruction. This is not really correct as the stalling instruction
404  * can modify the address used to access the JTLB.  The failure symptom is that
405  * the TLBP instruction will use an address created for the stalling instruction
406  * and not the address held in C0_ENHI and thus report the wrong results.
407  *
408  * The software work-around is to not allow the instruction preceding the TLBP
409  * to stall - make it an NOP or some other instruction guaranteed not to stall.
410  *
411  * Errata 2 will not be fixed.  This errata is also on the R5000.
412  *
413  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
414  */
415 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
416 {
417 	switch (current_cpu_type()) {
418 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
419 	case CPU_R4600:
420 	case CPU_R4700:
421 	case CPU_R5000:
422 	case CPU_R5000A:
423 	case CPU_NEVADA:
424 		uasm_i_nop(p);
425 		uasm_i_tlbp(p);
426 		break;
427 
428 	default:
429 		uasm_i_tlbp(p);
430 		break;
431 	}
432 }
433 
434 /*
435  * Write random or indexed TLB entry, and care about the hazards from
436  * the preceding mtc0 and for the following eret.
437  */
438 enum tlb_write_entry { tlb_random, tlb_indexed };
439 
440 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
441 					 struct uasm_reloc **r,
442 					 enum tlb_write_entry wmode)
443 {
444 	void(*tlbw)(u32 **) = NULL;
445 
446 	switch (wmode) {
447 	case tlb_random: tlbw = uasm_i_tlbwr; break;
448 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
449 	}
450 
451 	if (cpu_has_mips_r2) {
452 		if (cpu_has_mips_r2_exec_hazard)
453 			uasm_i_ehb(p);
454 		tlbw(p);
455 		return;
456 	}
457 
458 	switch (current_cpu_type()) {
459 	case CPU_R4000PC:
460 	case CPU_R4000SC:
461 	case CPU_R4000MC:
462 	case CPU_R4400PC:
463 	case CPU_R4400SC:
464 	case CPU_R4400MC:
465 		/*
466 		 * This branch uses up a mtc0 hazard nop slot and saves
467 		 * two nops after the tlbw instruction.
468 		 */
469 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
470 		tlbw(p);
471 		uasm_l_tlbw_hazard(l, *p);
472 		uasm_i_nop(p);
473 		break;
474 
475 	case CPU_R4600:
476 	case CPU_R4700:
477 	case CPU_R5000:
478 	case CPU_R5000A:
479 		uasm_i_nop(p);
480 		tlbw(p);
481 		uasm_i_nop(p);
482 		break;
483 
484 	case CPU_R4300:
485 	case CPU_5KC:
486 	case CPU_TX49XX:
487 	case CPU_PR4450:
488 	case CPU_XLR:
489 		uasm_i_nop(p);
490 		tlbw(p);
491 		break;
492 
493 	case CPU_R10000:
494 	case CPU_R12000:
495 	case CPU_R14000:
496 	case CPU_4KC:
497 	case CPU_4KEC:
498 	case CPU_M14KC:
499 	case CPU_SB1:
500 	case CPU_SB1A:
501 	case CPU_4KSC:
502 	case CPU_20KC:
503 	case CPU_25KF:
504 	case CPU_BMIPS32:
505 	case CPU_BMIPS3300:
506 	case CPU_BMIPS4350:
507 	case CPU_BMIPS4380:
508 	case CPU_BMIPS5000:
509 	case CPU_LOONGSON2:
510 	case CPU_R5500:
511 		if (m4kc_tlbp_war())
512 			uasm_i_nop(p);
513 	case CPU_ALCHEMY:
514 		tlbw(p);
515 		break;
516 
517 	case CPU_NEVADA:
518 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
519 		/*
520 		 * This branch uses up a mtc0 hazard nop slot and saves
521 		 * a nop after the tlbw instruction.
522 		 */
523 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
524 		tlbw(p);
525 		uasm_l_tlbw_hazard(l, *p);
526 		break;
527 
528 	case CPU_RM7000:
529 		uasm_i_nop(p);
530 		uasm_i_nop(p);
531 		uasm_i_nop(p);
532 		uasm_i_nop(p);
533 		tlbw(p);
534 		break;
535 
536 	case CPU_RM9000:
537 		/*
538 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
539 		 * use of the JTLB for instructions should not occur for 4
540 		 * cpu cycles and use for data translations should not occur
541 		 * for 3 cpu cycles.
542 		 */
543 		uasm_i_ssnop(p);
544 		uasm_i_ssnop(p);
545 		uasm_i_ssnop(p);
546 		uasm_i_ssnop(p);
547 		tlbw(p);
548 		uasm_i_ssnop(p);
549 		uasm_i_ssnop(p);
550 		uasm_i_ssnop(p);
551 		uasm_i_ssnop(p);
552 		break;
553 
554 	case CPU_VR4111:
555 	case CPU_VR4121:
556 	case CPU_VR4122:
557 	case CPU_VR4181:
558 	case CPU_VR4181A:
559 		uasm_i_nop(p);
560 		uasm_i_nop(p);
561 		tlbw(p);
562 		uasm_i_nop(p);
563 		uasm_i_nop(p);
564 		break;
565 
566 	case CPU_VR4131:
567 	case CPU_VR4133:
568 	case CPU_R5432:
569 		uasm_i_nop(p);
570 		uasm_i_nop(p);
571 		tlbw(p);
572 		break;
573 
574 	case CPU_JZRISC:
575 		tlbw(p);
576 		uasm_i_nop(p);
577 		break;
578 
579 	default:
580 		panic("No TLB refill handler yet (CPU type: %d)",
581 		      current_cpu_data.cputype);
582 		break;
583 	}
584 }
585 
586 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
587 								  unsigned int reg)
588 {
589 	if (kernel_uses_smartmips_rixi) {
590 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
591 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
592 	} else {
593 #ifdef CONFIG_64BIT_PHYS_ADDR
594 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
595 #else
596 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
597 #endif
598 	}
599 }
600 
601 #ifdef CONFIG_HUGETLB_PAGE
602 
603 static __cpuinit void build_restore_pagemask(u32 **p,
604 					     struct uasm_reloc **r,
605 					     unsigned int tmp,
606 					     enum label_id lid,
607 					     int restore_scratch)
608 {
609 	if (restore_scratch) {
610 		/* Reset default page size */
611 		if (PM_DEFAULT_MASK >> 16) {
612 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
613 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
614 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
615 			uasm_il_b(p, r, lid);
616 		} else if (PM_DEFAULT_MASK) {
617 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
618 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
619 			uasm_il_b(p, r, lid);
620 		} else {
621 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
622 			uasm_il_b(p, r, lid);
623 		}
624 		if (scratch_reg > 0)
625 			UASM_i_MFC0(p, 1, 31, scratch_reg);
626 		else
627 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
628 	} else {
629 		/* Reset default page size */
630 		if (PM_DEFAULT_MASK >> 16) {
631 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
632 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
633 			uasm_il_b(p, r, lid);
634 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
635 		} else if (PM_DEFAULT_MASK) {
636 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
637 			uasm_il_b(p, r, lid);
638 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
639 		} else {
640 			uasm_il_b(p, r, lid);
641 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
642 		}
643 	}
644 }
645 
646 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
647 						 struct uasm_label **l,
648 						 struct uasm_reloc **r,
649 						 unsigned int tmp,
650 						 enum tlb_write_entry wmode,
651 						 int restore_scratch)
652 {
653 	/* Set huge page tlb entry size */
654 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
655 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
656 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
657 
658 	build_tlb_write_entry(p, l, r, wmode);
659 
660 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
661 }
662 
663 /*
664  * Check if Huge PTE is present, if so then jump to LABEL.
665  */
666 static void __cpuinit
667 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
668 		unsigned int pmd, int lid)
669 {
670 	UASM_i_LW(p, tmp, 0, pmd);
671 	if (use_bbit_insns()) {
672 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
673 	} else {
674 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
675 		uasm_il_bnez(p, r, tmp, lid);
676 	}
677 }
678 
679 static __cpuinit void build_huge_update_entries(u32 **p,
680 						unsigned int pte,
681 						unsigned int tmp)
682 {
683 	int small_sequence;
684 
685 	/*
686 	 * A huge PTE describes an area the size of the
687 	 * configured huge page size. This is twice the
688 	 * of the large TLB entry size we intend to use.
689 	 * A TLB entry half the size of the configured
690 	 * huge page size is configured into entrylo0
691 	 * and entrylo1 to cover the contiguous huge PTE
692 	 * address space.
693 	 */
694 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
695 
696 	/* We can clobber tmp.  It isn't used after this.*/
697 	if (!small_sequence)
698 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
699 
700 	build_convert_pte_to_entrylo(p, pte);
701 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
702 	/* convert to entrylo1 */
703 	if (small_sequence)
704 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
705 	else
706 		UASM_i_ADDU(p, pte, pte, tmp);
707 
708 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
709 }
710 
711 static __cpuinit void build_huge_handler_tail(u32 **p,
712 					      struct uasm_reloc **r,
713 					      struct uasm_label **l,
714 					      unsigned int pte,
715 					      unsigned int ptr)
716 {
717 #ifdef CONFIG_SMP
718 	UASM_i_SC(p, pte, 0, ptr);
719 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
720 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
721 #else
722 	UASM_i_SW(p, pte, 0, ptr);
723 #endif
724 	build_huge_update_entries(p, pte, ptr);
725 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
726 }
727 #endif /* CONFIG_HUGETLB_PAGE */
728 
729 #ifdef CONFIG_64BIT
730 /*
731  * TMP and PTR are scratch.
732  * TMP will be clobbered, PTR will hold the pmd entry.
733  */
734 static void __cpuinit
735 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
736 		 unsigned int tmp, unsigned int ptr)
737 {
738 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
739 	long pgdc = (long)pgd_current;
740 #endif
741 	/*
742 	 * The vmalloc handling is not in the hotpath.
743 	 */
744 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
745 
746 	if (check_for_high_segbits) {
747 		/*
748 		 * The kernel currently implicitely assumes that the
749 		 * MIPS SEGBITS parameter for the processor is
750 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
751 		 * allocate virtual addresses outside the maximum
752 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
753 		 * that doesn't prevent user code from accessing the
754 		 * higher xuseg addresses.  Here, we make sure that
755 		 * everything but the lower xuseg addresses goes down
756 		 * the module_alloc/vmalloc path.
757 		 */
758 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
759 		uasm_il_bnez(p, r, ptr, label_vmalloc);
760 	} else {
761 		uasm_il_bltz(p, r, tmp, label_vmalloc);
762 	}
763 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
764 
765 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
766 	if (pgd_reg != -1) {
767 		/* pgd is in pgd_reg */
768 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
769 	} else {
770 		/*
771 		 * &pgd << 11 stored in CONTEXT [23..63].
772 		 */
773 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
774 
775 		/* Clear lower 23 bits of context. */
776 		uasm_i_dins(p, ptr, 0, 0, 23);
777 
778 		/* 1 0  1 0 1  << 6  xkphys cached */
779 		uasm_i_ori(p, ptr, ptr, 0x540);
780 		uasm_i_drotr(p, ptr, ptr, 11);
781 	}
782 #elif defined(CONFIG_SMP)
783 # ifdef  CONFIG_MIPS_MT_SMTC
784 	/*
785 	 * SMTC uses TCBind value as "CPU" index
786 	 */
787 	uasm_i_mfc0(p, ptr, C0_TCBIND);
788 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
789 # else
790 	/*
791 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
792 	 * stored in CONTEXT.
793 	 */
794 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
795 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
796 # endif
797 	UASM_i_LA_mostly(p, tmp, pgdc);
798 	uasm_i_daddu(p, ptr, ptr, tmp);
799 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
800 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
801 #else
802 	UASM_i_LA_mostly(p, ptr, pgdc);
803 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
804 #endif
805 
806 	uasm_l_vmalloc_done(l, *p);
807 
808 	/* get pgd offset in bytes */
809 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
810 
811 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
812 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
813 #ifndef __PAGETABLE_PMD_FOLDED
814 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
815 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
816 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
817 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
818 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
819 #endif
820 }
821 
822 /*
823  * BVADDR is the faulting address, PTR is scratch.
824  * PTR will hold the pgd for vmalloc.
825  */
826 static void __cpuinit
827 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
828 			unsigned int bvaddr, unsigned int ptr,
829 			enum vmalloc64_mode mode)
830 {
831 	long swpd = (long)swapper_pg_dir;
832 	int single_insn_swpd;
833 	int did_vmalloc_branch = 0;
834 
835 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
836 
837 	uasm_l_vmalloc(l, *p);
838 
839 	if (mode != not_refill && check_for_high_segbits) {
840 		if (single_insn_swpd) {
841 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
842 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
843 			did_vmalloc_branch = 1;
844 			/* fall through */
845 		} else {
846 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
847 		}
848 	}
849 	if (!did_vmalloc_branch) {
850 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
851 			uasm_il_b(p, r, label_vmalloc_done);
852 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
853 		} else {
854 			UASM_i_LA_mostly(p, ptr, swpd);
855 			uasm_il_b(p, r, label_vmalloc_done);
856 			if (uasm_in_compat_space_p(swpd))
857 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
858 			else
859 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
860 		}
861 	}
862 	if (mode != not_refill && check_for_high_segbits) {
863 		uasm_l_large_segbits_fault(l, *p);
864 		/*
865 		 * We get here if we are an xsseg address, or if we are
866 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
867 		 *
868 		 * Ignoring xsseg (assume disabled so would generate
869 		 * (address errors?), the only remaining possibility
870 		 * is the upper xuseg addresses.  On processors with
871 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
872 		 * addresses would have taken an address error. We try
873 		 * to mimic that here by taking a load/istream page
874 		 * fault.
875 		 */
876 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
877 		uasm_i_jr(p, ptr);
878 
879 		if (mode == refill_scratch) {
880 			if (scratch_reg > 0)
881 				UASM_i_MFC0(p, 1, 31, scratch_reg);
882 			else
883 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
884 		} else {
885 			uasm_i_nop(p);
886 		}
887 	}
888 }
889 
890 #else /* !CONFIG_64BIT */
891 
892 /*
893  * TMP and PTR are scratch.
894  * TMP will be clobbered, PTR will hold the pgd entry.
895  */
896 static void __cpuinit __maybe_unused
897 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
898 {
899 	long pgdc = (long)pgd_current;
900 
901 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
902 #ifdef CONFIG_SMP
903 #ifdef  CONFIG_MIPS_MT_SMTC
904 	/*
905 	 * SMTC uses TCBind value as "CPU" index
906 	 */
907 	uasm_i_mfc0(p, ptr, C0_TCBIND);
908 	UASM_i_LA_mostly(p, tmp, pgdc);
909 	uasm_i_srl(p, ptr, ptr, 19);
910 #else
911 	/*
912 	 * smp_processor_id() << 3 is stored in CONTEXT.
913          */
914 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
915 	UASM_i_LA_mostly(p, tmp, pgdc);
916 	uasm_i_srl(p, ptr, ptr, 23);
917 #endif
918 	uasm_i_addu(p, ptr, tmp, ptr);
919 #else
920 	UASM_i_LA_mostly(p, ptr, pgdc);
921 #endif
922 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
923 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
924 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
925 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
926 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
927 }
928 
929 #endif /* !CONFIG_64BIT */
930 
931 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
932 {
933 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
934 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
935 
936 	switch (current_cpu_type()) {
937 	case CPU_VR41XX:
938 	case CPU_VR4111:
939 	case CPU_VR4121:
940 	case CPU_VR4122:
941 	case CPU_VR4131:
942 	case CPU_VR4181:
943 	case CPU_VR4181A:
944 	case CPU_VR4133:
945 		shift += 2;
946 		break;
947 
948 	default:
949 		break;
950 	}
951 
952 	if (shift)
953 		UASM_i_SRL(p, ctx, ctx, shift);
954 	uasm_i_andi(p, ctx, ctx, mask);
955 }
956 
957 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
958 {
959 	/*
960 	 * Bug workaround for the Nevada. It seems as if under certain
961 	 * circumstances the move from cp0_context might produce a
962 	 * bogus result when the mfc0 instruction and its consumer are
963 	 * in a different cacheline or a load instruction, probably any
964 	 * memory reference, is between them.
965 	 */
966 	switch (current_cpu_type()) {
967 	case CPU_NEVADA:
968 		UASM_i_LW(p, ptr, 0, ptr);
969 		GET_CONTEXT(p, tmp); /* get context reg */
970 		break;
971 
972 	default:
973 		GET_CONTEXT(p, tmp); /* get context reg */
974 		UASM_i_LW(p, ptr, 0, ptr);
975 		break;
976 	}
977 
978 	build_adjust_context(p, tmp);
979 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
980 }
981 
982 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
983 					unsigned int ptep)
984 {
985 	/*
986 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
987 	 * Kernel is a special case. Only a few CPUs use it.
988 	 */
989 #ifdef CONFIG_64BIT_PHYS_ADDR
990 	if (cpu_has_64bits) {
991 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
992 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
993 		if (kernel_uses_smartmips_rixi) {
994 			UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
995 			UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
996 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
997 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
998 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
999 		} else {
1000 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1001 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1002 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1003 		}
1004 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1005 	} else {
1006 		int pte_off_even = sizeof(pte_t) / 2;
1007 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1008 
1009 		/* The pte entries are pre-shifted */
1010 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1011 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1012 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1013 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1014 	}
1015 #else
1016 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1017 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1018 	if (r45k_bvahwbug())
1019 		build_tlb_probe_entry(p);
1020 	if (kernel_uses_smartmips_rixi) {
1021 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
1022 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
1023 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1024 		if (r4k_250MHZhwbug())
1025 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1026 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1027 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1028 	} else {
1029 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1030 		if (r4k_250MHZhwbug())
1031 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1032 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1033 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1034 		if (r45k_bvahwbug())
1035 			uasm_i_mfc0(p, tmp, C0_INDEX);
1036 	}
1037 	if (r4k_250MHZhwbug())
1038 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1039 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1040 #endif
1041 }
1042 
1043 struct mips_huge_tlb_info {
1044 	int huge_pte;
1045 	int restore_scratch;
1046 };
1047 
1048 static struct mips_huge_tlb_info __cpuinit
1049 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1050 			       struct uasm_reloc **r, unsigned int tmp,
1051 			       unsigned int ptr, int c0_scratch)
1052 {
1053 	struct mips_huge_tlb_info rv;
1054 	unsigned int even, odd;
1055 	int vmalloc_branch_delay_filled = 0;
1056 	const int scratch = 1; /* Our extra working register */
1057 
1058 	rv.huge_pte = scratch;
1059 	rv.restore_scratch = 0;
1060 
1061 	if (check_for_high_segbits) {
1062 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1063 
1064 		if (pgd_reg != -1)
1065 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1066 		else
1067 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1068 
1069 		if (c0_scratch >= 0)
1070 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1071 		else
1072 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1073 
1074 		uasm_i_dsrl_safe(p, scratch, tmp,
1075 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1076 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1077 
1078 		if (pgd_reg == -1) {
1079 			vmalloc_branch_delay_filled = 1;
1080 			/* Clear lower 23 bits of context. */
1081 			uasm_i_dins(p, ptr, 0, 0, 23);
1082 		}
1083 	} else {
1084 		if (pgd_reg != -1)
1085 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1086 		else
1087 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1088 
1089 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1090 
1091 		if (c0_scratch >= 0)
1092 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1093 		else
1094 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1095 
1096 		if (pgd_reg == -1)
1097 			/* Clear lower 23 bits of context. */
1098 			uasm_i_dins(p, ptr, 0, 0, 23);
1099 
1100 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1101 	}
1102 
1103 	if (pgd_reg == -1) {
1104 		vmalloc_branch_delay_filled = 1;
1105 		/* 1 0  1 0 1  << 6  xkphys cached */
1106 		uasm_i_ori(p, ptr, ptr, 0x540);
1107 		uasm_i_drotr(p, ptr, ptr, 11);
1108 	}
1109 
1110 #ifdef __PAGETABLE_PMD_FOLDED
1111 #define LOC_PTEP scratch
1112 #else
1113 #define LOC_PTEP ptr
1114 #endif
1115 
1116 	if (!vmalloc_branch_delay_filled)
1117 		/* get pgd offset in bytes */
1118 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1119 
1120 	uasm_l_vmalloc_done(l, *p);
1121 
1122 	/*
1123 	 *                         tmp          ptr
1124 	 * fall-through case =   badvaddr  *pgd_current
1125 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1126 	 */
1127 
1128 	if (vmalloc_branch_delay_filled)
1129 		/* get pgd offset in bytes */
1130 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1131 
1132 #ifdef __PAGETABLE_PMD_FOLDED
1133 	GET_CONTEXT(p, tmp); /* get context reg */
1134 #endif
1135 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1136 
1137 	if (use_lwx_insns()) {
1138 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1139 	} else {
1140 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1141 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1142 	}
1143 
1144 #ifndef __PAGETABLE_PMD_FOLDED
1145 	/* get pmd offset in bytes */
1146 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1147 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1148 	GET_CONTEXT(p, tmp); /* get context reg */
1149 
1150 	if (use_lwx_insns()) {
1151 		UASM_i_LWX(p, scratch, scratch, ptr);
1152 	} else {
1153 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1154 		UASM_i_LW(p, scratch, 0, ptr);
1155 	}
1156 #endif
1157 	/* Adjust the context during the load latency. */
1158 	build_adjust_context(p, tmp);
1159 
1160 #ifdef CONFIG_HUGETLB_PAGE
1161 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1162 	/*
1163 	 * The in the LWX case we don't want to do the load in the
1164 	 * delay slot.  It cannot issue in the same cycle and may be
1165 	 * speculative and unneeded.
1166 	 */
1167 	if (use_lwx_insns())
1168 		uasm_i_nop(p);
1169 #endif /* CONFIG_HUGETLB_PAGE */
1170 
1171 
1172 	/* build_update_entries */
1173 	if (use_lwx_insns()) {
1174 		even = ptr;
1175 		odd = tmp;
1176 		UASM_i_LWX(p, even, scratch, tmp);
1177 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1178 		UASM_i_LWX(p, odd, scratch, tmp);
1179 	} else {
1180 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1181 		even = tmp;
1182 		odd = ptr;
1183 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1184 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1185 	}
1186 	if (kernel_uses_smartmips_rixi) {
1187 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_NO_EXEC));
1188 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_NO_EXEC));
1189 		uasm_i_drotr(p, even, even,
1190 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1191 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1192 		uasm_i_drotr(p, odd, odd,
1193 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1194 	} else {
1195 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1196 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1197 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1198 	}
1199 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1200 
1201 	if (c0_scratch >= 0) {
1202 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1203 		build_tlb_write_entry(p, l, r, tlb_random);
1204 		uasm_l_leave(l, *p);
1205 		rv.restore_scratch = 1;
1206 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1207 		build_tlb_write_entry(p, l, r, tlb_random);
1208 		uasm_l_leave(l, *p);
1209 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1210 	} else {
1211 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1212 		build_tlb_write_entry(p, l, r, tlb_random);
1213 		uasm_l_leave(l, *p);
1214 		rv.restore_scratch = 1;
1215 	}
1216 
1217 	uasm_i_eret(p); /* return from trap */
1218 
1219 	return rv;
1220 }
1221 
1222 /*
1223  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1224  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1225  * slots before the XTLB refill exception handler which belong to the
1226  * unused TLB refill exception.
1227  */
1228 #define MIPS64_REFILL_INSNS 32
1229 
1230 static void __cpuinit build_r4000_tlb_refill_handler(void)
1231 {
1232 	u32 *p = tlb_handler;
1233 	struct uasm_label *l = labels;
1234 	struct uasm_reloc *r = relocs;
1235 	u32 *f;
1236 	unsigned int final_len;
1237 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1238 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1239 
1240 	memset(tlb_handler, 0, sizeof(tlb_handler));
1241 	memset(labels, 0, sizeof(labels));
1242 	memset(relocs, 0, sizeof(relocs));
1243 	memset(final_handler, 0, sizeof(final_handler));
1244 
1245 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1246 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1247 							  scratch_reg);
1248 		vmalloc_mode = refill_scratch;
1249 	} else {
1250 		htlb_info.huge_pte = K0;
1251 		htlb_info.restore_scratch = 0;
1252 		vmalloc_mode = refill_noscratch;
1253 		/*
1254 		 * create the plain linear handler
1255 		 */
1256 		if (bcm1250_m3_war()) {
1257 			unsigned int segbits = 44;
1258 
1259 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1260 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1261 			uasm_i_xor(&p, K0, K0, K1);
1262 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1263 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1264 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1265 			uasm_i_or(&p, K0, K0, K1);
1266 			uasm_il_bnez(&p, &r, K0, label_leave);
1267 			/* No need for uasm_i_nop */
1268 		}
1269 
1270 #ifdef CONFIG_64BIT
1271 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1272 #else
1273 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1274 #endif
1275 
1276 #ifdef CONFIG_HUGETLB_PAGE
1277 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1278 #endif
1279 
1280 		build_get_ptep(&p, K0, K1);
1281 		build_update_entries(&p, K0, K1);
1282 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1283 		uasm_l_leave(&l, p);
1284 		uasm_i_eret(&p); /* return from trap */
1285 	}
1286 #ifdef CONFIG_HUGETLB_PAGE
1287 	uasm_l_tlb_huge_update(&l, p);
1288 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1289 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1290 				   htlb_info.restore_scratch);
1291 #endif
1292 
1293 #ifdef CONFIG_64BIT
1294 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1295 #endif
1296 
1297 	/*
1298 	 * Overflow check: For the 64bit handler, we need at least one
1299 	 * free instruction slot for the wrap-around branch. In worst
1300 	 * case, if the intended insertion point is a delay slot, we
1301 	 * need three, with the second nop'ed and the third being
1302 	 * unused.
1303 	 */
1304 	/* Loongson2 ebase is different than r4k, we have more space */
1305 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1306 	if ((p - tlb_handler) > 64)
1307 		panic("TLB refill handler space exceeded");
1308 #else
1309 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1310 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1311 		&& uasm_insn_has_bdelay(relocs,
1312 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1313 		panic("TLB refill handler space exceeded");
1314 #endif
1315 
1316 	/*
1317 	 * Now fold the handler in the TLB refill handler space.
1318 	 */
1319 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1320 	f = final_handler;
1321 	/* Simplest case, just copy the handler. */
1322 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1323 	final_len = p - tlb_handler;
1324 #else /* CONFIG_64BIT */
1325 	f = final_handler + MIPS64_REFILL_INSNS;
1326 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1327 		/* Just copy the handler. */
1328 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1329 		final_len = p - tlb_handler;
1330 	} else {
1331 #if defined(CONFIG_HUGETLB_PAGE)
1332 		const enum label_id ls = label_tlb_huge_update;
1333 #else
1334 		const enum label_id ls = label_vmalloc;
1335 #endif
1336 		u32 *split;
1337 		int ov = 0;
1338 		int i;
1339 
1340 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1341 			;
1342 		BUG_ON(i == ARRAY_SIZE(labels));
1343 		split = labels[i].addr;
1344 
1345 		/*
1346 		 * See if we have overflown one way or the other.
1347 		 */
1348 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1349 		    split < p - MIPS64_REFILL_INSNS)
1350 			ov = 1;
1351 
1352 		if (ov) {
1353 			/*
1354 			 * Split two instructions before the end.  One
1355 			 * for the branch and one for the instruction
1356 			 * in the delay slot.
1357 			 */
1358 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1359 
1360 			/*
1361 			 * If the branch would fall in a delay slot,
1362 			 * we must back up an additional instruction
1363 			 * so that it is no longer in a delay slot.
1364 			 */
1365 			if (uasm_insn_has_bdelay(relocs, split - 1))
1366 				split--;
1367 		}
1368 		/* Copy first part of the handler. */
1369 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1370 		f += split - tlb_handler;
1371 
1372 		if (ov) {
1373 			/* Insert branch. */
1374 			uasm_l_split(&l, final_handler);
1375 			uasm_il_b(&f, &r, label_split);
1376 			if (uasm_insn_has_bdelay(relocs, split))
1377 				uasm_i_nop(&f);
1378 			else {
1379 				uasm_copy_handler(relocs, labels,
1380 						  split, split + 1, f);
1381 				uasm_move_labels(labels, f, f + 1, -1);
1382 				f++;
1383 				split++;
1384 			}
1385 		}
1386 
1387 		/* Copy the rest of the handler. */
1388 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1389 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1390 			    (p - split);
1391 	}
1392 #endif /* CONFIG_64BIT */
1393 
1394 	uasm_resolve_relocs(relocs, labels);
1395 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1396 		 final_len);
1397 
1398 	memcpy((void *)ebase, final_handler, 0x100);
1399 
1400 	dump_handler((u32 *)ebase, 64);
1401 }
1402 
1403 /*
1404  * 128 instructions for the fastpath handler is generous and should
1405  * never be exceeded.
1406  */
1407 #define FASTPATH_SIZE 128
1408 
1409 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1410 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1411 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1412 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1413 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1414 
1415 static void __cpuinit build_r4000_setup_pgd(void)
1416 {
1417 	const int a0 = 4;
1418 	const int a1 = 5;
1419 	u32 *p = tlbmiss_handler_setup_pgd;
1420 	struct uasm_label *l = labels;
1421 	struct uasm_reloc *r = relocs;
1422 
1423 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1424 	memset(labels, 0, sizeof(labels));
1425 	memset(relocs, 0, sizeof(relocs));
1426 
1427 	pgd_reg = allocate_kscratch();
1428 
1429 	if (pgd_reg == -1) {
1430 		/* PGD << 11 in c0_Context */
1431 		/*
1432 		 * If it is a ckseg0 address, convert to a physical
1433 		 * address.  Shifting right by 29 and adding 4 will
1434 		 * result in zero for these addresses.
1435 		 *
1436 		 */
1437 		UASM_i_SRA(&p, a1, a0, 29);
1438 		UASM_i_ADDIU(&p, a1, a1, 4);
1439 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1440 		uasm_i_nop(&p);
1441 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1442 		uasm_l_tlbl_goaround1(&l, p);
1443 		UASM_i_SLL(&p, a0, a0, 11);
1444 		uasm_i_jr(&p, 31);
1445 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1446 	} else {
1447 		/* PGD in c0_KScratch */
1448 		uasm_i_jr(&p, 31);
1449 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1450 	}
1451 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1452 		panic("tlbmiss_handler_setup_pgd space exceeded");
1453 	uasm_resolve_relocs(relocs, labels);
1454 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1455 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1456 
1457 	dump_handler(tlbmiss_handler_setup_pgd,
1458 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1459 }
1460 #endif
1461 
1462 static void __cpuinit
1463 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1464 {
1465 #ifdef CONFIG_SMP
1466 # ifdef CONFIG_64BIT_PHYS_ADDR
1467 	if (cpu_has_64bits)
1468 		uasm_i_lld(p, pte, 0, ptr);
1469 	else
1470 # endif
1471 		UASM_i_LL(p, pte, 0, ptr);
1472 #else
1473 # ifdef CONFIG_64BIT_PHYS_ADDR
1474 	if (cpu_has_64bits)
1475 		uasm_i_ld(p, pte, 0, ptr);
1476 	else
1477 # endif
1478 		UASM_i_LW(p, pte, 0, ptr);
1479 #endif
1480 }
1481 
1482 static void __cpuinit
1483 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1484 	unsigned int mode)
1485 {
1486 #ifdef CONFIG_64BIT_PHYS_ADDR
1487 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1488 #endif
1489 
1490 	uasm_i_ori(p, pte, pte, mode);
1491 #ifdef CONFIG_SMP
1492 # ifdef CONFIG_64BIT_PHYS_ADDR
1493 	if (cpu_has_64bits)
1494 		uasm_i_scd(p, pte, 0, ptr);
1495 	else
1496 # endif
1497 		UASM_i_SC(p, pte, 0, ptr);
1498 
1499 	if (r10000_llsc_war())
1500 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1501 	else
1502 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1503 
1504 # ifdef CONFIG_64BIT_PHYS_ADDR
1505 	if (!cpu_has_64bits) {
1506 		/* no uasm_i_nop needed */
1507 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1508 		uasm_i_ori(p, pte, pte, hwmode);
1509 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1510 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1511 		/* no uasm_i_nop needed */
1512 		uasm_i_lw(p, pte, 0, ptr);
1513 	} else
1514 		uasm_i_nop(p);
1515 # else
1516 	uasm_i_nop(p);
1517 # endif
1518 #else
1519 # ifdef CONFIG_64BIT_PHYS_ADDR
1520 	if (cpu_has_64bits)
1521 		uasm_i_sd(p, pte, 0, ptr);
1522 	else
1523 # endif
1524 		UASM_i_SW(p, pte, 0, ptr);
1525 
1526 # ifdef CONFIG_64BIT_PHYS_ADDR
1527 	if (!cpu_has_64bits) {
1528 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1529 		uasm_i_ori(p, pte, pte, hwmode);
1530 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1531 		uasm_i_lw(p, pte, 0, ptr);
1532 	}
1533 # endif
1534 #endif
1535 }
1536 
1537 /*
1538  * Check if PTE is present, if not then jump to LABEL. PTR points to
1539  * the page table where this PTE is located, PTE will be re-loaded
1540  * with it's original value.
1541  */
1542 static void __cpuinit
1543 build_pte_present(u32 **p, struct uasm_reloc **r,
1544 		  int pte, int ptr, int scratch, enum label_id lid)
1545 {
1546 	int t = scratch >= 0 ? scratch : pte;
1547 
1548 	if (kernel_uses_smartmips_rixi) {
1549 		if (use_bbit_insns()) {
1550 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1551 			uasm_i_nop(p);
1552 		} else {
1553 			uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1554 			uasm_il_beqz(p, r, t, lid);
1555 			if (pte == t)
1556 				/* You lose the SMP race :-(*/
1557 				iPTE_LW(p, pte, ptr);
1558 		}
1559 	} else {
1560 		uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1561 		uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1562 		uasm_il_bnez(p, r, t, lid);
1563 		if (pte == t)
1564 			/* You lose the SMP race :-(*/
1565 			iPTE_LW(p, pte, ptr);
1566 	}
1567 }
1568 
1569 /* Make PTE valid, store result in PTR. */
1570 static void __cpuinit
1571 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1572 		 unsigned int ptr)
1573 {
1574 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1575 
1576 	iPTE_SW(p, r, pte, ptr, mode);
1577 }
1578 
1579 /*
1580  * Check if PTE can be written to, if not branch to LABEL. Regardless
1581  * restore PTE with value from PTR when done.
1582  */
1583 static void __cpuinit
1584 build_pte_writable(u32 **p, struct uasm_reloc **r,
1585 		   unsigned int pte, unsigned int ptr, int scratch,
1586 		   enum label_id lid)
1587 {
1588 	int t = scratch >= 0 ? scratch : pte;
1589 
1590 	uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1591 	uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1592 	uasm_il_bnez(p, r, t, lid);
1593 	if (pte == t)
1594 		/* You lose the SMP race :-(*/
1595 		iPTE_LW(p, pte, ptr);
1596 	else
1597 		uasm_i_nop(p);
1598 }
1599 
1600 /* Make PTE writable, update software status bits as well, then store
1601  * at PTR.
1602  */
1603 static void __cpuinit
1604 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1605 		 unsigned int ptr)
1606 {
1607 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1608 			     | _PAGE_DIRTY);
1609 
1610 	iPTE_SW(p, r, pte, ptr, mode);
1611 }
1612 
1613 /*
1614  * Check if PTE can be modified, if not branch to LABEL. Regardless
1615  * restore PTE with value from PTR when done.
1616  */
1617 static void __cpuinit
1618 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1619 		     unsigned int pte, unsigned int ptr, int scratch,
1620 		     enum label_id lid)
1621 {
1622 	if (use_bbit_insns()) {
1623 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1624 		uasm_i_nop(p);
1625 	} else {
1626 		int t = scratch >= 0 ? scratch : pte;
1627 		uasm_i_andi(p, t, pte, _PAGE_WRITE);
1628 		uasm_il_beqz(p, r, t, lid);
1629 		if (pte == t)
1630 			/* You lose the SMP race :-(*/
1631 			iPTE_LW(p, pte, ptr);
1632 	}
1633 }
1634 
1635 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1636 
1637 
1638 /*
1639  * R3000 style TLB load/store/modify handlers.
1640  */
1641 
1642 /*
1643  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1644  * Then it returns.
1645  */
1646 static void __cpuinit
1647 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1648 {
1649 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1650 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1651 	uasm_i_tlbwi(p);
1652 	uasm_i_jr(p, tmp);
1653 	uasm_i_rfe(p); /* branch delay */
1654 }
1655 
1656 /*
1657  * This places the pte into ENTRYLO0 and writes it with tlbwi
1658  * or tlbwr as appropriate.  This is because the index register
1659  * may have the probe fail bit set as a result of a trap on a
1660  * kseg2 access, i.e. without refill.  Then it returns.
1661  */
1662 static void __cpuinit
1663 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1664 			     struct uasm_reloc **r, unsigned int pte,
1665 			     unsigned int tmp)
1666 {
1667 	uasm_i_mfc0(p, tmp, C0_INDEX);
1668 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1669 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1670 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1671 	uasm_i_tlbwi(p); /* cp0 delay */
1672 	uasm_i_jr(p, tmp);
1673 	uasm_i_rfe(p); /* branch delay */
1674 	uasm_l_r3000_write_probe_fail(l, *p);
1675 	uasm_i_tlbwr(p); /* cp0 delay */
1676 	uasm_i_jr(p, tmp);
1677 	uasm_i_rfe(p); /* branch delay */
1678 }
1679 
1680 static void __cpuinit
1681 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1682 				   unsigned int ptr)
1683 {
1684 	long pgdc = (long)pgd_current;
1685 
1686 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1687 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1688 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1689 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1690 	uasm_i_sll(p, pte, pte, 2);
1691 	uasm_i_addu(p, ptr, ptr, pte);
1692 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1693 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1694 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1695 	uasm_i_addu(p, ptr, ptr, pte);
1696 	uasm_i_lw(p, pte, 0, ptr);
1697 	uasm_i_tlbp(p); /* load delay */
1698 }
1699 
1700 static void __cpuinit build_r3000_tlb_load_handler(void)
1701 {
1702 	u32 *p = handle_tlbl;
1703 	struct uasm_label *l = labels;
1704 	struct uasm_reloc *r = relocs;
1705 
1706 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1707 	memset(labels, 0, sizeof(labels));
1708 	memset(relocs, 0, sizeof(relocs));
1709 
1710 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1711 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1712 	uasm_i_nop(&p); /* load delay */
1713 	build_make_valid(&p, &r, K0, K1);
1714 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1715 
1716 	uasm_l_nopage_tlbl(&l, p);
1717 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1718 	uasm_i_nop(&p);
1719 
1720 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1721 		panic("TLB load handler fastpath space exceeded");
1722 
1723 	uasm_resolve_relocs(relocs, labels);
1724 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1725 		 (unsigned int)(p - handle_tlbl));
1726 
1727 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1728 }
1729 
1730 static void __cpuinit build_r3000_tlb_store_handler(void)
1731 {
1732 	u32 *p = handle_tlbs;
1733 	struct uasm_label *l = labels;
1734 	struct uasm_reloc *r = relocs;
1735 
1736 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1737 	memset(labels, 0, sizeof(labels));
1738 	memset(relocs, 0, sizeof(relocs));
1739 
1740 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1741 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1742 	uasm_i_nop(&p); /* load delay */
1743 	build_make_write(&p, &r, K0, K1);
1744 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1745 
1746 	uasm_l_nopage_tlbs(&l, p);
1747 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1748 	uasm_i_nop(&p);
1749 
1750 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1751 		panic("TLB store handler fastpath space exceeded");
1752 
1753 	uasm_resolve_relocs(relocs, labels);
1754 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1755 		 (unsigned int)(p - handle_tlbs));
1756 
1757 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1758 }
1759 
1760 static void __cpuinit build_r3000_tlb_modify_handler(void)
1761 {
1762 	u32 *p = handle_tlbm;
1763 	struct uasm_label *l = labels;
1764 	struct uasm_reloc *r = relocs;
1765 
1766 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1767 	memset(labels, 0, sizeof(labels));
1768 	memset(relocs, 0, sizeof(relocs));
1769 
1770 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1771 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1772 	uasm_i_nop(&p); /* load delay */
1773 	build_make_write(&p, &r, K0, K1);
1774 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1775 
1776 	uasm_l_nopage_tlbm(&l, p);
1777 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1778 	uasm_i_nop(&p);
1779 
1780 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1781 		panic("TLB modify handler fastpath space exceeded");
1782 
1783 	uasm_resolve_relocs(relocs, labels);
1784 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1785 		 (unsigned int)(p - handle_tlbm));
1786 
1787 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1788 }
1789 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1790 
1791 /*
1792  * R4000 style TLB load/store/modify handlers.
1793  */
1794 static struct work_registers __cpuinit
1795 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1796 				   struct uasm_reloc **r)
1797 {
1798 	struct work_registers wr = build_get_work_registers(p);
1799 
1800 #ifdef CONFIG_64BIT
1801 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1802 #else
1803 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1804 #endif
1805 
1806 #ifdef CONFIG_HUGETLB_PAGE
1807 	/*
1808 	 * For huge tlb entries, pmd doesn't contain an address but
1809 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1810 	 * see if we need to jump to huge tlb processing.
1811 	 */
1812 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1813 #endif
1814 
1815 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1816 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1817 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1818 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1819 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1820 
1821 #ifdef CONFIG_SMP
1822 	uasm_l_smp_pgtable_change(l, *p);
1823 #endif
1824 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1825 	if (!m4kc_tlbp_war())
1826 		build_tlb_probe_entry(p);
1827 	return wr;
1828 }
1829 
1830 static void __cpuinit
1831 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1832 				   struct uasm_reloc **r, unsigned int tmp,
1833 				   unsigned int ptr)
1834 {
1835 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1836 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1837 	build_update_entries(p, tmp, ptr);
1838 	build_tlb_write_entry(p, l, r, tlb_indexed);
1839 	uasm_l_leave(l, *p);
1840 	build_restore_work_registers(p);
1841 	uasm_i_eret(p); /* return from trap */
1842 
1843 #ifdef CONFIG_64BIT
1844 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1845 #endif
1846 }
1847 
1848 static void __cpuinit build_r4000_tlb_load_handler(void)
1849 {
1850 	u32 *p = handle_tlbl;
1851 	struct uasm_label *l = labels;
1852 	struct uasm_reloc *r = relocs;
1853 	struct work_registers wr;
1854 
1855 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1856 	memset(labels, 0, sizeof(labels));
1857 	memset(relocs, 0, sizeof(relocs));
1858 
1859 	if (bcm1250_m3_war()) {
1860 		unsigned int segbits = 44;
1861 
1862 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1863 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1864 		uasm_i_xor(&p, K0, K0, K1);
1865 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1866 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1867 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1868 		uasm_i_or(&p, K0, K0, K1);
1869 		uasm_il_bnez(&p, &r, K0, label_leave);
1870 		/* No need for uasm_i_nop */
1871 	}
1872 
1873 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1874 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1875 	if (m4kc_tlbp_war())
1876 		build_tlb_probe_entry(&p);
1877 
1878 	if (kernel_uses_smartmips_rixi) {
1879 		/*
1880 		 * If the page is not _PAGE_VALID, RI or XI could not
1881 		 * have triggered it.  Skip the expensive test..
1882 		 */
1883 		if (use_bbit_insns()) {
1884 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1885 				      label_tlbl_goaround1);
1886 		} else {
1887 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1888 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1889 		}
1890 		uasm_i_nop(&p);
1891 
1892 		uasm_i_tlbr(&p);
1893 		/* Examine  entrylo 0 or 1 based on ptr. */
1894 		if (use_bbit_insns()) {
1895 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1896 		} else {
1897 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1898 			uasm_i_beqz(&p, wr.r3, 8);
1899 		}
1900 		/* load it in the delay slot*/
1901 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1902 		/* load it if ptr is odd */
1903 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1904 		/*
1905 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1906 		 * XI must have triggered it.
1907 		 */
1908 		if (use_bbit_insns()) {
1909 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1910 			uasm_i_nop(&p);
1911 			uasm_l_tlbl_goaround1(&l, p);
1912 		} else {
1913 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1914 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1915 			uasm_i_nop(&p);
1916 		}
1917 		uasm_l_tlbl_goaround1(&l, p);
1918 	}
1919 	build_make_valid(&p, &r, wr.r1, wr.r2);
1920 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1921 
1922 #ifdef CONFIG_HUGETLB_PAGE
1923 	/*
1924 	 * This is the entry point when build_r4000_tlbchange_handler_head
1925 	 * spots a huge page.
1926 	 */
1927 	uasm_l_tlb_huge_update(&l, p);
1928 	iPTE_LW(&p, wr.r1, wr.r2);
1929 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1930 	build_tlb_probe_entry(&p);
1931 
1932 	if (kernel_uses_smartmips_rixi) {
1933 		/*
1934 		 * If the page is not _PAGE_VALID, RI or XI could not
1935 		 * have triggered it.  Skip the expensive test..
1936 		 */
1937 		if (use_bbit_insns()) {
1938 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1939 				      label_tlbl_goaround2);
1940 		} else {
1941 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1942 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1943 		}
1944 		uasm_i_nop(&p);
1945 
1946 		uasm_i_tlbr(&p);
1947 		/* Examine  entrylo 0 or 1 based on ptr. */
1948 		if (use_bbit_insns()) {
1949 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1950 		} else {
1951 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1952 			uasm_i_beqz(&p, wr.r3, 8);
1953 		}
1954 		/* load it in the delay slot*/
1955 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1956 		/* load it if ptr is odd */
1957 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1958 		/*
1959 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1960 		 * XI must have triggered it.
1961 		 */
1962 		if (use_bbit_insns()) {
1963 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
1964 		} else {
1965 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1966 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1967 		}
1968 		if (PM_DEFAULT_MASK == 0)
1969 			uasm_i_nop(&p);
1970 		/*
1971 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
1972 		 * it is restored in build_huge_tlb_write_entry.
1973 		 */
1974 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
1975 
1976 		uasm_l_tlbl_goaround2(&l, p);
1977 	}
1978 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
1979 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
1980 #endif
1981 
1982 	uasm_l_nopage_tlbl(&l, p);
1983 	build_restore_work_registers(&p);
1984 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1985 	uasm_i_nop(&p);
1986 
1987 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1988 		panic("TLB load handler fastpath space exceeded");
1989 
1990 	uasm_resolve_relocs(relocs, labels);
1991 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1992 		 (unsigned int)(p - handle_tlbl));
1993 
1994 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1995 }
1996 
1997 static void __cpuinit build_r4000_tlb_store_handler(void)
1998 {
1999 	u32 *p = handle_tlbs;
2000 	struct uasm_label *l = labels;
2001 	struct uasm_reloc *r = relocs;
2002 	struct work_registers wr;
2003 
2004 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
2005 	memset(labels, 0, sizeof(labels));
2006 	memset(relocs, 0, sizeof(relocs));
2007 
2008 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2009 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2010 	if (m4kc_tlbp_war())
2011 		build_tlb_probe_entry(&p);
2012 	build_make_write(&p, &r, wr.r1, wr.r2);
2013 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2014 
2015 #ifdef CONFIG_HUGETLB_PAGE
2016 	/*
2017 	 * This is the entry point when
2018 	 * build_r4000_tlbchange_handler_head spots a huge page.
2019 	 */
2020 	uasm_l_tlb_huge_update(&l, p);
2021 	iPTE_LW(&p, wr.r1, wr.r2);
2022 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2023 	build_tlb_probe_entry(&p);
2024 	uasm_i_ori(&p, wr.r1, wr.r1,
2025 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2026 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2027 #endif
2028 
2029 	uasm_l_nopage_tlbs(&l, p);
2030 	build_restore_work_registers(&p);
2031 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2032 	uasm_i_nop(&p);
2033 
2034 	if ((p - handle_tlbs) > FASTPATH_SIZE)
2035 		panic("TLB store handler fastpath space exceeded");
2036 
2037 	uasm_resolve_relocs(relocs, labels);
2038 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2039 		 (unsigned int)(p - handle_tlbs));
2040 
2041 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
2042 }
2043 
2044 static void __cpuinit build_r4000_tlb_modify_handler(void)
2045 {
2046 	u32 *p = handle_tlbm;
2047 	struct uasm_label *l = labels;
2048 	struct uasm_reloc *r = relocs;
2049 	struct work_registers wr;
2050 
2051 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
2052 	memset(labels, 0, sizeof(labels));
2053 	memset(relocs, 0, sizeof(relocs));
2054 
2055 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2056 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2057 	if (m4kc_tlbp_war())
2058 		build_tlb_probe_entry(&p);
2059 	/* Present and writable bits set, set accessed and dirty bits. */
2060 	build_make_write(&p, &r, wr.r1, wr.r2);
2061 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2062 
2063 #ifdef CONFIG_HUGETLB_PAGE
2064 	/*
2065 	 * This is the entry point when
2066 	 * build_r4000_tlbchange_handler_head spots a huge page.
2067 	 */
2068 	uasm_l_tlb_huge_update(&l, p);
2069 	iPTE_LW(&p, wr.r1, wr.r2);
2070 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2071 	build_tlb_probe_entry(&p);
2072 	uasm_i_ori(&p, wr.r1, wr.r1,
2073 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2074 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2075 #endif
2076 
2077 	uasm_l_nopage_tlbm(&l, p);
2078 	build_restore_work_registers(&p);
2079 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2080 	uasm_i_nop(&p);
2081 
2082 	if ((p - handle_tlbm) > FASTPATH_SIZE)
2083 		panic("TLB modify handler fastpath space exceeded");
2084 
2085 	uasm_resolve_relocs(relocs, labels);
2086 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2087 		 (unsigned int)(p - handle_tlbm));
2088 
2089 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
2090 }
2091 
2092 void __cpuinit build_tlb_refill_handler(void)
2093 {
2094 	/*
2095 	 * The refill handler is generated per-CPU, multi-node systems
2096 	 * may have local storage for it. The other handlers are only
2097 	 * needed once.
2098 	 */
2099 	static int run_once = 0;
2100 
2101 #ifdef CONFIG_64BIT
2102 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2103 #endif
2104 
2105 	switch (current_cpu_type()) {
2106 	case CPU_R2000:
2107 	case CPU_R3000:
2108 	case CPU_R3000A:
2109 	case CPU_R3081E:
2110 	case CPU_TX3912:
2111 	case CPU_TX3922:
2112 	case CPU_TX3927:
2113 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2114 		build_r3000_tlb_refill_handler();
2115 		if (!run_once) {
2116 			build_r3000_tlb_load_handler();
2117 			build_r3000_tlb_store_handler();
2118 			build_r3000_tlb_modify_handler();
2119 			run_once++;
2120 		}
2121 #else
2122 		panic("No R3000 TLB refill handler");
2123 #endif
2124 		break;
2125 
2126 	case CPU_R6000:
2127 	case CPU_R6000A:
2128 		panic("No R6000 TLB refill handler yet");
2129 		break;
2130 
2131 	case CPU_R8000:
2132 		panic("No R8000 TLB refill handler yet");
2133 		break;
2134 
2135 	default:
2136 		if (!run_once) {
2137 			scratch_reg = allocate_kscratch();
2138 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2139 			build_r4000_setup_pgd();
2140 #endif
2141 			build_r4000_tlb_load_handler();
2142 			build_r4000_tlb_store_handler();
2143 			build_r4000_tlb_modify_handler();
2144 			run_once++;
2145 		}
2146 		build_r4000_tlb_refill_handler();
2147 	}
2148 }
2149 
2150 void __cpuinit flush_tlb_handlers(void)
2151 {
2152 	local_flush_icache_range((unsigned long)handle_tlbl,
2153 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2154 	local_flush_icache_range((unsigned long)handle_tlbs,
2155 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2156 	local_flush_icache_range((unsigned long)handle_tlbm,
2157 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2158 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2159 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2160 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2161 #endif
2162 }
2163