xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision df2634f43f5106947f3735a0b61a6527a4b278cd)
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  *
13  * ... and the days got worse and worse and now you see
14  * I've gone completly out of my mind.
15  *
16  * They're coming to take me a away haha
17  * they're coming to take me a away hoho hihi haha
18  * to the funny farm where code is beautiful all the time ...
19  *
20  * (Condolences to Napoleon XIV)
21  */
22 
23 #include <linux/bug.h>
24 #include <linux/kernel.h>
25 #include <linux/types.h>
26 #include <linux/smp.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
29 #include <linux/cache.h>
30 
31 #include <asm/cacheflush.h>
32 #include <asm/pgtable.h>
33 #include <asm/war.h>
34 #include <asm/uasm.h>
35 
36 /*
37  * TLB load/store/modify handlers.
38  *
39  * Only the fastpath gets synthesized at runtime, the slowpath for
40  * do_page_fault remains normal asm.
41  */
42 extern void tlb_do_page_fault_0(void);
43 extern void tlb_do_page_fault_1(void);
44 
45 
46 static inline int r45k_bvahwbug(void)
47 {
48 	/* XXX: We should probe for the presence of this bug, but we don't. */
49 	return 0;
50 }
51 
52 static inline int r4k_250MHZhwbug(void)
53 {
54 	/* XXX: We should probe for the presence of this bug, but we don't. */
55 	return 0;
56 }
57 
58 static inline int __maybe_unused bcm1250_m3_war(void)
59 {
60 	return BCM1250_M3_WAR;
61 }
62 
63 static inline int __maybe_unused r10000_llsc_war(void)
64 {
65 	return R10000_LLSC_WAR;
66 }
67 
68 static int use_bbit_insns(void)
69 {
70 	switch (current_cpu_type()) {
71 	case CPU_CAVIUM_OCTEON:
72 	case CPU_CAVIUM_OCTEON_PLUS:
73 	case CPU_CAVIUM_OCTEON2:
74 		return 1;
75 	default:
76 		return 0;
77 	}
78 }
79 
80 static int use_lwx_insns(void)
81 {
82 	switch (current_cpu_type()) {
83 	case CPU_CAVIUM_OCTEON2:
84 		return 1;
85 	default:
86 		return 0;
87 	}
88 }
89 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
90     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
91 static bool scratchpad_available(void)
92 {
93 	return true;
94 }
95 static int scratchpad_offset(int i)
96 {
97 	/*
98 	 * CVMSEG starts at address -32768 and extends for
99 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
100 	 */
101 	i += 1; /* Kernel use starts at the top and works down. */
102 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
103 }
104 #else
105 static bool scratchpad_available(void)
106 {
107 	return false;
108 }
109 static int scratchpad_offset(int i)
110 {
111 	BUG();
112 }
113 #endif
114 /*
115  * Found by experiment: At least some revisions of the 4kc throw under
116  * some circumstances a machine check exception, triggered by invalid
117  * values in the index register.  Delaying the tlbp instruction until
118  * after the next branch,  plus adding an additional nop in front of
119  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
120  * why; it's not an issue caused by the core RTL.
121  *
122  */
123 static int __cpuinit m4kc_tlbp_war(void)
124 {
125 	return (current_cpu_data.processor_id & 0xffff00) ==
126 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
127 }
128 
129 /* Handle labels (which must be positive integers). */
130 enum label_id {
131 	label_second_part = 1,
132 	label_leave,
133 	label_vmalloc,
134 	label_vmalloc_done,
135 	label_tlbw_hazard,
136 	label_split,
137 	label_tlbl_goaround1,
138 	label_tlbl_goaround2,
139 	label_nopage_tlbl,
140 	label_nopage_tlbs,
141 	label_nopage_tlbm,
142 	label_smp_pgtable_change,
143 	label_r3000_write_probe_fail,
144 	label_large_segbits_fault,
145 #ifdef CONFIG_HUGETLB_PAGE
146 	label_tlb_huge_update,
147 #endif
148 };
149 
150 UASM_L_LA(_second_part)
151 UASM_L_LA(_leave)
152 UASM_L_LA(_vmalloc)
153 UASM_L_LA(_vmalloc_done)
154 UASM_L_LA(_tlbw_hazard)
155 UASM_L_LA(_split)
156 UASM_L_LA(_tlbl_goaround1)
157 UASM_L_LA(_tlbl_goaround2)
158 UASM_L_LA(_nopage_tlbl)
159 UASM_L_LA(_nopage_tlbs)
160 UASM_L_LA(_nopage_tlbm)
161 UASM_L_LA(_smp_pgtable_change)
162 UASM_L_LA(_r3000_write_probe_fail)
163 UASM_L_LA(_large_segbits_fault)
164 #ifdef CONFIG_HUGETLB_PAGE
165 UASM_L_LA(_tlb_huge_update)
166 #endif
167 
168 /*
169  * For debug purposes.
170  */
171 static inline void dump_handler(const u32 *handler, int count)
172 {
173 	int i;
174 
175 	pr_debug("\t.set push\n");
176 	pr_debug("\t.set noreorder\n");
177 
178 	for (i = 0; i < count; i++)
179 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
180 
181 	pr_debug("\t.set pop\n");
182 }
183 
184 /* The only general purpose registers allowed in TLB handlers. */
185 #define K0		26
186 #define K1		27
187 
188 /* Some CP0 registers */
189 #define C0_INDEX	0, 0
190 #define C0_ENTRYLO0	2, 0
191 #define C0_TCBIND	2, 2
192 #define C0_ENTRYLO1	3, 0
193 #define C0_CONTEXT	4, 0
194 #define C0_PAGEMASK	5, 0
195 #define C0_BADVADDR	8, 0
196 #define C0_ENTRYHI	10, 0
197 #define C0_EPC		14, 0
198 #define C0_XCONTEXT	20, 0
199 
200 #ifdef CONFIG_64BIT
201 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
202 #else
203 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
204 #endif
205 
206 /* The worst case length of the handler is around 18 instructions for
207  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
208  * Maximum space available is 32 instructions for R3000 and 64
209  * instructions for R4000.
210  *
211  * We deliberately chose a buffer size of 128, so we won't scribble
212  * over anything important on overflow before we panic.
213  */
214 static u32 tlb_handler[128] __cpuinitdata;
215 
216 /* simply assume worst case size for labels and relocs */
217 static struct uasm_label labels[128] __cpuinitdata;
218 static struct uasm_reloc relocs[128] __cpuinitdata;
219 
220 #ifdef CONFIG_64BIT
221 static int check_for_high_segbits __cpuinitdata;
222 #endif
223 
224 static int check_for_high_segbits __cpuinitdata;
225 
226 static unsigned int kscratch_used_mask __cpuinitdata;
227 
228 static int __cpuinit allocate_kscratch(void)
229 {
230 	int r;
231 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
232 
233 	r = ffs(a);
234 
235 	if (r == 0)
236 		return -1;
237 
238 	r--; /* make it zero based */
239 
240 	kscratch_used_mask |= (1 << r);
241 
242 	return r;
243 }
244 
245 static int scratch_reg __cpuinitdata;
246 static int pgd_reg __cpuinitdata;
247 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
248 
249 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
250 
251 /*
252  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
253  * we cannot do r3000 under these circumstances.
254  *
255  * Declare pgd_current here instead of including mmu_context.h to avoid type
256  * conflicts for tlbmiss_handler_setup_pgd
257  */
258 extern unsigned long pgd_current[];
259 
260 /*
261  * The R3000 TLB handler is simple.
262  */
263 static void __cpuinit build_r3000_tlb_refill_handler(void)
264 {
265 	long pgdc = (long)pgd_current;
266 	u32 *p;
267 
268 	memset(tlb_handler, 0, sizeof(tlb_handler));
269 	p = tlb_handler;
270 
271 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
272 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
273 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
274 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
275 	uasm_i_sll(&p, K0, K0, 2);
276 	uasm_i_addu(&p, K1, K1, K0);
277 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
278 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
279 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
280 	uasm_i_addu(&p, K1, K1, K0);
281 	uasm_i_lw(&p, K0, 0, K1);
282 	uasm_i_nop(&p); /* load delay */
283 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
284 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
285 	uasm_i_tlbwr(&p); /* cp0 delay */
286 	uasm_i_jr(&p, K1);
287 	uasm_i_rfe(&p); /* branch delay */
288 
289 	if (p > tlb_handler + 32)
290 		panic("TLB refill handler space exceeded");
291 
292 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
293 		 (unsigned int)(p - tlb_handler));
294 
295 	memcpy((void *)ebase, tlb_handler, 0x80);
296 
297 	dump_handler((u32 *)ebase, 32);
298 }
299 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
300 
301 /*
302  * The R4000 TLB handler is much more complicated. We have two
303  * consecutive handler areas with 32 instructions space each.
304  * Since they aren't used at the same time, we can overflow in the
305  * other one.To keep things simple, we first assume linear space,
306  * then we relocate it to the final handler layout as needed.
307  */
308 static u32 final_handler[64] __cpuinitdata;
309 
310 /*
311  * Hazards
312  *
313  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
314  * 2. A timing hazard exists for the TLBP instruction.
315  *
316  *      stalling_instruction
317  *      TLBP
318  *
319  * The JTLB is being read for the TLBP throughout the stall generated by the
320  * previous instruction. This is not really correct as the stalling instruction
321  * can modify the address used to access the JTLB.  The failure symptom is that
322  * the TLBP instruction will use an address created for the stalling instruction
323  * and not the address held in C0_ENHI and thus report the wrong results.
324  *
325  * The software work-around is to not allow the instruction preceding the TLBP
326  * to stall - make it an NOP or some other instruction guaranteed not to stall.
327  *
328  * Errata 2 will not be fixed.  This errata is also on the R5000.
329  *
330  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
331  */
332 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
333 {
334 	switch (current_cpu_type()) {
335 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
336 	case CPU_R4600:
337 	case CPU_R4700:
338 	case CPU_R5000:
339 	case CPU_R5000A:
340 	case CPU_NEVADA:
341 		uasm_i_nop(p);
342 		uasm_i_tlbp(p);
343 		break;
344 
345 	default:
346 		uasm_i_tlbp(p);
347 		break;
348 	}
349 }
350 
351 /*
352  * Write random or indexed TLB entry, and care about the hazards from
353  * the preceeding mtc0 and for the following eret.
354  */
355 enum tlb_write_entry { tlb_random, tlb_indexed };
356 
357 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
358 					 struct uasm_reloc **r,
359 					 enum tlb_write_entry wmode)
360 {
361 	void(*tlbw)(u32 **) = NULL;
362 
363 	switch (wmode) {
364 	case tlb_random: tlbw = uasm_i_tlbwr; break;
365 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
366 	}
367 
368 	if (cpu_has_mips_r2) {
369 		if (cpu_has_mips_r2_exec_hazard)
370 			uasm_i_ehb(p);
371 		tlbw(p);
372 		return;
373 	}
374 
375 	switch (current_cpu_type()) {
376 	case CPU_R4000PC:
377 	case CPU_R4000SC:
378 	case CPU_R4000MC:
379 	case CPU_R4400PC:
380 	case CPU_R4400SC:
381 	case CPU_R4400MC:
382 		/*
383 		 * This branch uses up a mtc0 hazard nop slot and saves
384 		 * two nops after the tlbw instruction.
385 		 */
386 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
387 		tlbw(p);
388 		uasm_l_tlbw_hazard(l, *p);
389 		uasm_i_nop(p);
390 		break;
391 
392 	case CPU_R4600:
393 	case CPU_R4700:
394 	case CPU_R5000:
395 	case CPU_R5000A:
396 		uasm_i_nop(p);
397 		tlbw(p);
398 		uasm_i_nop(p);
399 		break;
400 
401 	case CPU_R4300:
402 	case CPU_5KC:
403 	case CPU_TX49XX:
404 	case CPU_PR4450:
405 		uasm_i_nop(p);
406 		tlbw(p);
407 		break;
408 
409 	case CPU_R10000:
410 	case CPU_R12000:
411 	case CPU_R14000:
412 	case CPU_4KC:
413 	case CPU_4KEC:
414 	case CPU_SB1:
415 	case CPU_SB1A:
416 	case CPU_4KSC:
417 	case CPU_20KC:
418 	case CPU_25KF:
419 	case CPU_BMIPS32:
420 	case CPU_BMIPS3300:
421 	case CPU_BMIPS4350:
422 	case CPU_BMIPS4380:
423 	case CPU_BMIPS5000:
424 	case CPU_LOONGSON2:
425 	case CPU_R5500:
426 		if (m4kc_tlbp_war())
427 			uasm_i_nop(p);
428 	case CPU_ALCHEMY:
429 		tlbw(p);
430 		break;
431 
432 	case CPU_NEVADA:
433 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
434 		/*
435 		 * This branch uses up a mtc0 hazard nop slot and saves
436 		 * a nop after the tlbw instruction.
437 		 */
438 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
439 		tlbw(p);
440 		uasm_l_tlbw_hazard(l, *p);
441 		break;
442 
443 	case CPU_RM7000:
444 		uasm_i_nop(p);
445 		uasm_i_nop(p);
446 		uasm_i_nop(p);
447 		uasm_i_nop(p);
448 		tlbw(p);
449 		break;
450 
451 	case CPU_RM9000:
452 		/*
453 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
454 		 * use of the JTLB for instructions should not occur for 4
455 		 * cpu cycles and use for data translations should not occur
456 		 * for 3 cpu cycles.
457 		 */
458 		uasm_i_ssnop(p);
459 		uasm_i_ssnop(p);
460 		uasm_i_ssnop(p);
461 		uasm_i_ssnop(p);
462 		tlbw(p);
463 		uasm_i_ssnop(p);
464 		uasm_i_ssnop(p);
465 		uasm_i_ssnop(p);
466 		uasm_i_ssnop(p);
467 		break;
468 
469 	case CPU_VR4111:
470 	case CPU_VR4121:
471 	case CPU_VR4122:
472 	case CPU_VR4181:
473 	case CPU_VR4181A:
474 		uasm_i_nop(p);
475 		uasm_i_nop(p);
476 		tlbw(p);
477 		uasm_i_nop(p);
478 		uasm_i_nop(p);
479 		break;
480 
481 	case CPU_VR4131:
482 	case CPU_VR4133:
483 	case CPU_R5432:
484 		uasm_i_nop(p);
485 		uasm_i_nop(p);
486 		tlbw(p);
487 		break;
488 
489 	case CPU_JZRISC:
490 		tlbw(p);
491 		uasm_i_nop(p);
492 		break;
493 
494 	default:
495 		panic("No TLB refill handler yet (CPU type: %d)",
496 		      current_cpu_data.cputype);
497 		break;
498 	}
499 }
500 
501 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
502 								  unsigned int reg)
503 {
504 	if (kernel_uses_smartmips_rixi) {
505 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
506 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
507 	} else {
508 #ifdef CONFIG_64BIT_PHYS_ADDR
509 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
510 #else
511 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
512 #endif
513 	}
514 }
515 
516 #ifdef CONFIG_HUGETLB_PAGE
517 
518 static __cpuinit void build_restore_pagemask(u32 **p,
519 					     struct uasm_reloc **r,
520 					     unsigned int tmp,
521 					     enum label_id lid,
522 					     int restore_scratch)
523 {
524 	if (restore_scratch) {
525 		/* Reset default page size */
526 		if (PM_DEFAULT_MASK >> 16) {
527 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
528 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
529 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
530 			uasm_il_b(p, r, lid);
531 		} else if (PM_DEFAULT_MASK) {
532 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
533 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
534 			uasm_il_b(p, r, lid);
535 		} else {
536 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
537 			uasm_il_b(p, r, lid);
538 		}
539 		if (scratch_reg > 0)
540 			UASM_i_MFC0(p, 1, 31, scratch_reg);
541 		else
542 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
543 	} else {
544 		/* Reset default page size */
545 		if (PM_DEFAULT_MASK >> 16) {
546 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
547 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
548 			uasm_il_b(p, r, lid);
549 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
550 		} else if (PM_DEFAULT_MASK) {
551 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
552 			uasm_il_b(p, r, lid);
553 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
554 		} else {
555 			uasm_il_b(p, r, lid);
556 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
557 		}
558 	}
559 }
560 
561 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
562 						 struct uasm_label **l,
563 						 struct uasm_reloc **r,
564 						 unsigned int tmp,
565 						 enum tlb_write_entry wmode,
566 						 int restore_scratch)
567 {
568 	/* Set huge page tlb entry size */
569 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
570 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
571 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
572 
573 	build_tlb_write_entry(p, l, r, wmode);
574 
575 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
576 }
577 
578 /*
579  * Check if Huge PTE is present, if so then jump to LABEL.
580  */
581 static void __cpuinit
582 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
583 		unsigned int pmd, int lid)
584 {
585 	UASM_i_LW(p, tmp, 0, pmd);
586 	if (use_bbit_insns()) {
587 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
588 	} else {
589 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
590 		uasm_il_bnez(p, r, tmp, lid);
591 	}
592 }
593 
594 static __cpuinit void build_huge_update_entries(u32 **p,
595 						unsigned int pte,
596 						unsigned int tmp)
597 {
598 	int small_sequence;
599 
600 	/*
601 	 * A huge PTE describes an area the size of the
602 	 * configured huge page size. This is twice the
603 	 * of the large TLB entry size we intend to use.
604 	 * A TLB entry half the size of the configured
605 	 * huge page size is configured into entrylo0
606 	 * and entrylo1 to cover the contiguous huge PTE
607 	 * address space.
608 	 */
609 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
610 
611 	/* We can clobber tmp.  It isn't used after this.*/
612 	if (!small_sequence)
613 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
614 
615 	build_convert_pte_to_entrylo(p, pte);
616 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
617 	/* convert to entrylo1 */
618 	if (small_sequence)
619 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
620 	else
621 		UASM_i_ADDU(p, pte, pte, tmp);
622 
623 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
624 }
625 
626 static __cpuinit void build_huge_handler_tail(u32 **p,
627 					      struct uasm_reloc **r,
628 					      struct uasm_label **l,
629 					      unsigned int pte,
630 					      unsigned int ptr)
631 {
632 #ifdef CONFIG_SMP
633 	UASM_i_SC(p, pte, 0, ptr);
634 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
635 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
636 #else
637 	UASM_i_SW(p, pte, 0, ptr);
638 #endif
639 	build_huge_update_entries(p, pte, ptr);
640 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
641 }
642 #endif /* CONFIG_HUGETLB_PAGE */
643 
644 #ifdef CONFIG_64BIT
645 /*
646  * TMP and PTR are scratch.
647  * TMP will be clobbered, PTR will hold the pmd entry.
648  */
649 static void __cpuinit
650 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
651 		 unsigned int tmp, unsigned int ptr)
652 {
653 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
654 	long pgdc = (long)pgd_current;
655 #endif
656 	/*
657 	 * The vmalloc handling is not in the hotpath.
658 	 */
659 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
660 
661 	if (check_for_high_segbits) {
662 		/*
663 		 * The kernel currently implicitely assumes that the
664 		 * MIPS SEGBITS parameter for the processor is
665 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
666 		 * allocate virtual addresses outside the maximum
667 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
668 		 * that doesn't prevent user code from accessing the
669 		 * higher xuseg addresses.  Here, we make sure that
670 		 * everything but the lower xuseg addresses goes down
671 		 * the module_alloc/vmalloc path.
672 		 */
673 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
674 		uasm_il_bnez(p, r, ptr, label_vmalloc);
675 	} else {
676 		uasm_il_bltz(p, r, tmp, label_vmalloc);
677 	}
678 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
679 
680 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
681 	if (pgd_reg != -1) {
682 		/* pgd is in pgd_reg */
683 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
684 	} else {
685 		/*
686 		 * &pgd << 11 stored in CONTEXT [23..63].
687 		 */
688 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
689 
690 		/* Clear lower 23 bits of context. */
691 		uasm_i_dins(p, ptr, 0, 0, 23);
692 
693 		/* 1 0  1 0 1  << 6  xkphys cached */
694 		uasm_i_ori(p, ptr, ptr, 0x540);
695 		uasm_i_drotr(p, ptr, ptr, 11);
696 	}
697 #elif defined(CONFIG_SMP)
698 # ifdef  CONFIG_MIPS_MT_SMTC
699 	/*
700 	 * SMTC uses TCBind value as "CPU" index
701 	 */
702 	uasm_i_mfc0(p, ptr, C0_TCBIND);
703 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
704 # else
705 	/*
706 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
707 	 * stored in CONTEXT.
708 	 */
709 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
710 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
711 # endif
712 	UASM_i_LA_mostly(p, tmp, pgdc);
713 	uasm_i_daddu(p, ptr, ptr, tmp);
714 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
715 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
716 #else
717 	UASM_i_LA_mostly(p, ptr, pgdc);
718 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
719 #endif
720 
721 	uasm_l_vmalloc_done(l, *p);
722 
723 	/* get pgd offset in bytes */
724 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
725 
726 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
727 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
728 #ifndef __PAGETABLE_PMD_FOLDED
729 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
730 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
731 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
732 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
733 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
734 #endif
735 }
736 
737 /*
738  * BVADDR is the faulting address, PTR is scratch.
739  * PTR will hold the pgd for vmalloc.
740  */
741 static void __cpuinit
742 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
743 			unsigned int bvaddr, unsigned int ptr,
744 			enum vmalloc64_mode mode)
745 {
746 	long swpd = (long)swapper_pg_dir;
747 	int single_insn_swpd;
748 	int did_vmalloc_branch = 0;
749 
750 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
751 
752 	uasm_l_vmalloc(l, *p);
753 
754 	if (mode != not_refill && check_for_high_segbits) {
755 		if (single_insn_swpd) {
756 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
757 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
758 			did_vmalloc_branch = 1;
759 			/* fall through */
760 		} else {
761 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
762 		}
763 	}
764 	if (!did_vmalloc_branch) {
765 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
766 			uasm_il_b(p, r, label_vmalloc_done);
767 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
768 		} else {
769 			UASM_i_LA_mostly(p, ptr, swpd);
770 			uasm_il_b(p, r, label_vmalloc_done);
771 			if (uasm_in_compat_space_p(swpd))
772 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
773 			else
774 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
775 		}
776 	}
777 	if (mode != not_refill && check_for_high_segbits) {
778 		uasm_l_large_segbits_fault(l, *p);
779 		/*
780 		 * We get here if we are an xsseg address, or if we are
781 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
782 		 *
783 		 * Ignoring xsseg (assume disabled so would generate
784 		 * (address errors?), the only remaining possibility
785 		 * is the upper xuseg addresses.  On processors with
786 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
787 		 * addresses would have taken an address error. We try
788 		 * to mimic that here by taking a load/istream page
789 		 * fault.
790 		 */
791 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
792 		uasm_i_jr(p, ptr);
793 
794 		if (mode == refill_scratch) {
795 			if (scratch_reg > 0)
796 				UASM_i_MFC0(p, 1, 31, scratch_reg);
797 			else
798 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
799 		} else {
800 			uasm_i_nop(p);
801 		}
802 	}
803 }
804 
805 #else /* !CONFIG_64BIT */
806 
807 /*
808  * TMP and PTR are scratch.
809  * TMP will be clobbered, PTR will hold the pgd entry.
810  */
811 static void __cpuinit __maybe_unused
812 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
813 {
814 	long pgdc = (long)pgd_current;
815 
816 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
817 #ifdef CONFIG_SMP
818 #ifdef  CONFIG_MIPS_MT_SMTC
819 	/*
820 	 * SMTC uses TCBind value as "CPU" index
821 	 */
822 	uasm_i_mfc0(p, ptr, C0_TCBIND);
823 	UASM_i_LA_mostly(p, tmp, pgdc);
824 	uasm_i_srl(p, ptr, ptr, 19);
825 #else
826 	/*
827 	 * smp_processor_id() << 3 is stored in CONTEXT.
828          */
829 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
830 	UASM_i_LA_mostly(p, tmp, pgdc);
831 	uasm_i_srl(p, ptr, ptr, 23);
832 #endif
833 	uasm_i_addu(p, ptr, tmp, ptr);
834 #else
835 	UASM_i_LA_mostly(p, ptr, pgdc);
836 #endif
837 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
838 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
839 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
840 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
841 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
842 }
843 
844 #endif /* !CONFIG_64BIT */
845 
846 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
847 {
848 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
849 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
850 
851 	switch (current_cpu_type()) {
852 	case CPU_VR41XX:
853 	case CPU_VR4111:
854 	case CPU_VR4121:
855 	case CPU_VR4122:
856 	case CPU_VR4131:
857 	case CPU_VR4181:
858 	case CPU_VR4181A:
859 	case CPU_VR4133:
860 		shift += 2;
861 		break;
862 
863 	default:
864 		break;
865 	}
866 
867 	if (shift)
868 		UASM_i_SRL(p, ctx, ctx, shift);
869 	uasm_i_andi(p, ctx, ctx, mask);
870 }
871 
872 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
873 {
874 	/*
875 	 * Bug workaround for the Nevada. It seems as if under certain
876 	 * circumstances the move from cp0_context might produce a
877 	 * bogus result when the mfc0 instruction and its consumer are
878 	 * in a different cacheline or a load instruction, probably any
879 	 * memory reference, is between them.
880 	 */
881 	switch (current_cpu_type()) {
882 	case CPU_NEVADA:
883 		UASM_i_LW(p, ptr, 0, ptr);
884 		GET_CONTEXT(p, tmp); /* get context reg */
885 		break;
886 
887 	default:
888 		GET_CONTEXT(p, tmp); /* get context reg */
889 		UASM_i_LW(p, ptr, 0, ptr);
890 		break;
891 	}
892 
893 	build_adjust_context(p, tmp);
894 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
895 }
896 
897 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
898 					unsigned int ptep)
899 {
900 	/*
901 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
902 	 * Kernel is a special case. Only a few CPUs use it.
903 	 */
904 #ifdef CONFIG_64BIT_PHYS_ADDR
905 	if (cpu_has_64bits) {
906 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
907 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
908 		if (kernel_uses_smartmips_rixi) {
909 			UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
910 			UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
911 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
912 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
913 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
914 		} else {
915 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
916 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
917 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
918 		}
919 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
920 	} else {
921 		int pte_off_even = sizeof(pte_t) / 2;
922 		int pte_off_odd = pte_off_even + sizeof(pte_t);
923 
924 		/* The pte entries are pre-shifted */
925 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
926 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
927 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
928 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
929 	}
930 #else
931 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
932 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
933 	if (r45k_bvahwbug())
934 		build_tlb_probe_entry(p);
935 	if (kernel_uses_smartmips_rixi) {
936 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
937 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
938 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
939 		if (r4k_250MHZhwbug())
940 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
941 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
942 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
943 	} else {
944 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
945 		if (r4k_250MHZhwbug())
946 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
947 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
948 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
949 		if (r45k_bvahwbug())
950 			uasm_i_mfc0(p, tmp, C0_INDEX);
951 	}
952 	if (r4k_250MHZhwbug())
953 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
954 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
955 #endif
956 }
957 
958 struct mips_huge_tlb_info {
959 	int huge_pte;
960 	int restore_scratch;
961 };
962 
963 static struct mips_huge_tlb_info __cpuinit
964 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
965 			       struct uasm_reloc **r, unsigned int tmp,
966 			       unsigned int ptr, int c0_scratch)
967 {
968 	struct mips_huge_tlb_info rv;
969 	unsigned int even, odd;
970 	int vmalloc_branch_delay_filled = 0;
971 	const int scratch = 1; /* Our extra working register */
972 
973 	rv.huge_pte = scratch;
974 	rv.restore_scratch = 0;
975 
976 	if (check_for_high_segbits) {
977 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
978 
979 		if (pgd_reg != -1)
980 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
981 		else
982 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
983 
984 		if (c0_scratch >= 0)
985 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
986 		else
987 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
988 
989 		uasm_i_dsrl_safe(p, scratch, tmp,
990 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
991 		uasm_il_bnez(p, r, scratch, label_vmalloc);
992 
993 		if (pgd_reg == -1) {
994 			vmalloc_branch_delay_filled = 1;
995 			/* Clear lower 23 bits of context. */
996 			uasm_i_dins(p, ptr, 0, 0, 23);
997 		}
998 	} else {
999 		if (pgd_reg != -1)
1000 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1001 		else
1002 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1003 
1004 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1005 
1006 		if (c0_scratch >= 0)
1007 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1008 		else
1009 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1010 
1011 		if (pgd_reg == -1)
1012 			/* Clear lower 23 bits of context. */
1013 			uasm_i_dins(p, ptr, 0, 0, 23);
1014 
1015 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1016 	}
1017 
1018 	if (pgd_reg == -1) {
1019 		vmalloc_branch_delay_filled = 1;
1020 		/* 1 0  1 0 1  << 6  xkphys cached */
1021 		uasm_i_ori(p, ptr, ptr, 0x540);
1022 		uasm_i_drotr(p, ptr, ptr, 11);
1023 	}
1024 
1025 #ifdef __PAGETABLE_PMD_FOLDED
1026 #define LOC_PTEP scratch
1027 #else
1028 #define LOC_PTEP ptr
1029 #endif
1030 
1031 	if (!vmalloc_branch_delay_filled)
1032 		/* get pgd offset in bytes */
1033 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1034 
1035 	uasm_l_vmalloc_done(l, *p);
1036 
1037 	/*
1038 	 *                         tmp          ptr
1039 	 * fall-through case =   badvaddr  *pgd_current
1040 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1041 	 */
1042 
1043 	if (vmalloc_branch_delay_filled)
1044 		/* get pgd offset in bytes */
1045 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1046 
1047 #ifdef __PAGETABLE_PMD_FOLDED
1048 	GET_CONTEXT(p, tmp); /* get context reg */
1049 #endif
1050 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1051 
1052 	if (use_lwx_insns()) {
1053 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1054 	} else {
1055 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1056 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1057 	}
1058 
1059 #ifndef __PAGETABLE_PMD_FOLDED
1060 	/* get pmd offset in bytes */
1061 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1062 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1063 	GET_CONTEXT(p, tmp); /* get context reg */
1064 
1065 	if (use_lwx_insns()) {
1066 		UASM_i_LWX(p, scratch, scratch, ptr);
1067 	} else {
1068 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1069 		UASM_i_LW(p, scratch, 0, ptr);
1070 	}
1071 #endif
1072 	/* Adjust the context during the load latency. */
1073 	build_adjust_context(p, tmp);
1074 
1075 #ifdef CONFIG_HUGETLB_PAGE
1076 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1077 	/*
1078 	 * The in the LWX case we don't want to do the load in the
1079 	 * delay slot.  It cannot issue in the same cycle and may be
1080 	 * speculative and unneeded.
1081 	 */
1082 	if (use_lwx_insns())
1083 		uasm_i_nop(p);
1084 #endif /* CONFIG_HUGETLB_PAGE */
1085 
1086 
1087 	/* build_update_entries */
1088 	if (use_lwx_insns()) {
1089 		even = ptr;
1090 		odd = tmp;
1091 		UASM_i_LWX(p, even, scratch, tmp);
1092 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1093 		UASM_i_LWX(p, odd, scratch, tmp);
1094 	} else {
1095 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1096 		even = tmp;
1097 		odd = ptr;
1098 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1099 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1100 	}
1101 	if (kernel_uses_smartmips_rixi) {
1102 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_NO_EXEC));
1103 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_NO_EXEC));
1104 		uasm_i_drotr(p, even, even,
1105 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1106 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1107 		uasm_i_drotr(p, odd, odd,
1108 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1109 	} else {
1110 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1111 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1112 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1113 	}
1114 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1115 
1116 	if (c0_scratch >= 0) {
1117 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1118 		build_tlb_write_entry(p, l, r, tlb_random);
1119 		uasm_l_leave(l, *p);
1120 		rv.restore_scratch = 1;
1121 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1122 		build_tlb_write_entry(p, l, r, tlb_random);
1123 		uasm_l_leave(l, *p);
1124 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1125 	} else {
1126 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1127 		build_tlb_write_entry(p, l, r, tlb_random);
1128 		uasm_l_leave(l, *p);
1129 		rv.restore_scratch = 1;
1130 	}
1131 
1132 	uasm_i_eret(p); /* return from trap */
1133 
1134 	return rv;
1135 }
1136 
1137 /*
1138  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1139  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1140  * slots before the XTLB refill exception handler which belong to the
1141  * unused TLB refill exception.
1142  */
1143 #define MIPS64_REFILL_INSNS 32
1144 
1145 static void __cpuinit build_r4000_tlb_refill_handler(void)
1146 {
1147 	u32 *p = tlb_handler;
1148 	struct uasm_label *l = labels;
1149 	struct uasm_reloc *r = relocs;
1150 	u32 *f;
1151 	unsigned int final_len;
1152 	struct mips_huge_tlb_info htlb_info;
1153 	enum vmalloc64_mode vmalloc_mode;
1154 
1155 	memset(tlb_handler, 0, sizeof(tlb_handler));
1156 	memset(labels, 0, sizeof(labels));
1157 	memset(relocs, 0, sizeof(relocs));
1158 	memset(final_handler, 0, sizeof(final_handler));
1159 
1160 	if (scratch_reg == 0)
1161 		scratch_reg = allocate_kscratch();
1162 
1163 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1164 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1165 							  scratch_reg);
1166 		vmalloc_mode = refill_scratch;
1167 	} else {
1168 		htlb_info.huge_pte = K0;
1169 		htlb_info.restore_scratch = 0;
1170 		vmalloc_mode = refill_noscratch;
1171 		/*
1172 		 * create the plain linear handler
1173 		 */
1174 		if (bcm1250_m3_war()) {
1175 			unsigned int segbits = 44;
1176 
1177 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1178 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1179 			uasm_i_xor(&p, K0, K0, K1);
1180 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1181 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1182 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1183 			uasm_i_or(&p, K0, K0, K1);
1184 			uasm_il_bnez(&p, &r, K0, label_leave);
1185 			/* No need for uasm_i_nop */
1186 		}
1187 
1188 #ifdef CONFIG_64BIT
1189 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1190 #else
1191 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1192 #endif
1193 
1194 #ifdef CONFIG_HUGETLB_PAGE
1195 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1196 #endif
1197 
1198 		build_get_ptep(&p, K0, K1);
1199 		build_update_entries(&p, K0, K1);
1200 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1201 		uasm_l_leave(&l, p);
1202 		uasm_i_eret(&p); /* return from trap */
1203 	}
1204 #ifdef CONFIG_HUGETLB_PAGE
1205 	uasm_l_tlb_huge_update(&l, p);
1206 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1207 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1208 				   htlb_info.restore_scratch);
1209 #endif
1210 
1211 #ifdef CONFIG_64BIT
1212 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1213 #endif
1214 
1215 	/*
1216 	 * Overflow check: For the 64bit handler, we need at least one
1217 	 * free instruction slot for the wrap-around branch. In worst
1218 	 * case, if the intended insertion point is a delay slot, we
1219 	 * need three, with the second nop'ed and the third being
1220 	 * unused.
1221 	 */
1222 	/* Loongson2 ebase is different than r4k, we have more space */
1223 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1224 	if ((p - tlb_handler) > 64)
1225 		panic("TLB refill handler space exceeded");
1226 #else
1227 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1228 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1229 		&& uasm_insn_has_bdelay(relocs,
1230 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1231 		panic("TLB refill handler space exceeded");
1232 #endif
1233 
1234 	/*
1235 	 * Now fold the handler in the TLB refill handler space.
1236 	 */
1237 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1238 	f = final_handler;
1239 	/* Simplest case, just copy the handler. */
1240 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1241 	final_len = p - tlb_handler;
1242 #else /* CONFIG_64BIT */
1243 	f = final_handler + MIPS64_REFILL_INSNS;
1244 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1245 		/* Just copy the handler. */
1246 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1247 		final_len = p - tlb_handler;
1248 	} else {
1249 #if defined(CONFIG_HUGETLB_PAGE)
1250 		const enum label_id ls = label_tlb_huge_update;
1251 #else
1252 		const enum label_id ls = label_vmalloc;
1253 #endif
1254 		u32 *split;
1255 		int ov = 0;
1256 		int i;
1257 
1258 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1259 			;
1260 		BUG_ON(i == ARRAY_SIZE(labels));
1261 		split = labels[i].addr;
1262 
1263 		/*
1264 		 * See if we have overflown one way or the other.
1265 		 */
1266 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1267 		    split < p - MIPS64_REFILL_INSNS)
1268 			ov = 1;
1269 
1270 		if (ov) {
1271 			/*
1272 			 * Split two instructions before the end.  One
1273 			 * for the branch and one for the instruction
1274 			 * in the delay slot.
1275 			 */
1276 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1277 
1278 			/*
1279 			 * If the branch would fall in a delay slot,
1280 			 * we must back up an additional instruction
1281 			 * so that it is no longer in a delay slot.
1282 			 */
1283 			if (uasm_insn_has_bdelay(relocs, split - 1))
1284 				split--;
1285 		}
1286 		/* Copy first part of the handler. */
1287 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1288 		f += split - tlb_handler;
1289 
1290 		if (ov) {
1291 			/* Insert branch. */
1292 			uasm_l_split(&l, final_handler);
1293 			uasm_il_b(&f, &r, label_split);
1294 			if (uasm_insn_has_bdelay(relocs, split))
1295 				uasm_i_nop(&f);
1296 			else {
1297 				uasm_copy_handler(relocs, labels,
1298 						  split, split + 1, f);
1299 				uasm_move_labels(labels, f, f + 1, -1);
1300 				f++;
1301 				split++;
1302 			}
1303 		}
1304 
1305 		/* Copy the rest of the handler. */
1306 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1307 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1308 			    (p - split);
1309 	}
1310 #endif /* CONFIG_64BIT */
1311 
1312 	uasm_resolve_relocs(relocs, labels);
1313 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1314 		 final_len);
1315 
1316 	memcpy((void *)ebase, final_handler, 0x100);
1317 
1318 	dump_handler((u32 *)ebase, 64);
1319 }
1320 
1321 /*
1322  * 128 instructions for the fastpath handler is generous and should
1323  * never be exceeded.
1324  */
1325 #define FASTPATH_SIZE 128
1326 
1327 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1328 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1329 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1330 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1331 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1332 
1333 static void __cpuinit build_r4000_setup_pgd(void)
1334 {
1335 	const int a0 = 4;
1336 	const int a1 = 5;
1337 	u32 *p = tlbmiss_handler_setup_pgd;
1338 	struct uasm_label *l = labels;
1339 	struct uasm_reloc *r = relocs;
1340 
1341 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1342 	memset(labels, 0, sizeof(labels));
1343 	memset(relocs, 0, sizeof(relocs));
1344 
1345 	pgd_reg = allocate_kscratch();
1346 
1347 	if (pgd_reg == -1) {
1348 		/* PGD << 11 in c0_Context */
1349 		/*
1350 		 * If it is a ckseg0 address, convert to a physical
1351 		 * address.  Shifting right by 29 and adding 4 will
1352 		 * result in zero for these addresses.
1353 		 *
1354 		 */
1355 		UASM_i_SRA(&p, a1, a0, 29);
1356 		UASM_i_ADDIU(&p, a1, a1, 4);
1357 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1358 		uasm_i_nop(&p);
1359 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1360 		uasm_l_tlbl_goaround1(&l, p);
1361 		UASM_i_SLL(&p, a0, a0, 11);
1362 		uasm_i_jr(&p, 31);
1363 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1364 	} else {
1365 		/* PGD in c0_KScratch */
1366 		uasm_i_jr(&p, 31);
1367 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1368 	}
1369 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1370 		panic("tlbmiss_handler_setup_pgd space exceeded");
1371 	uasm_resolve_relocs(relocs, labels);
1372 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1373 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1374 
1375 	dump_handler(tlbmiss_handler_setup_pgd,
1376 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1377 }
1378 #endif
1379 
1380 static void __cpuinit
1381 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1382 {
1383 #ifdef CONFIG_SMP
1384 # ifdef CONFIG_64BIT_PHYS_ADDR
1385 	if (cpu_has_64bits)
1386 		uasm_i_lld(p, pte, 0, ptr);
1387 	else
1388 # endif
1389 		UASM_i_LL(p, pte, 0, ptr);
1390 #else
1391 # ifdef CONFIG_64BIT_PHYS_ADDR
1392 	if (cpu_has_64bits)
1393 		uasm_i_ld(p, pte, 0, ptr);
1394 	else
1395 # endif
1396 		UASM_i_LW(p, pte, 0, ptr);
1397 #endif
1398 }
1399 
1400 static void __cpuinit
1401 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1402 	unsigned int mode)
1403 {
1404 #ifdef CONFIG_64BIT_PHYS_ADDR
1405 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1406 #endif
1407 
1408 	uasm_i_ori(p, pte, pte, mode);
1409 #ifdef CONFIG_SMP
1410 # ifdef CONFIG_64BIT_PHYS_ADDR
1411 	if (cpu_has_64bits)
1412 		uasm_i_scd(p, pte, 0, ptr);
1413 	else
1414 # endif
1415 		UASM_i_SC(p, pte, 0, ptr);
1416 
1417 	if (r10000_llsc_war())
1418 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1419 	else
1420 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1421 
1422 # ifdef CONFIG_64BIT_PHYS_ADDR
1423 	if (!cpu_has_64bits) {
1424 		/* no uasm_i_nop needed */
1425 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1426 		uasm_i_ori(p, pte, pte, hwmode);
1427 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1428 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1429 		/* no uasm_i_nop needed */
1430 		uasm_i_lw(p, pte, 0, ptr);
1431 	} else
1432 		uasm_i_nop(p);
1433 # else
1434 	uasm_i_nop(p);
1435 # endif
1436 #else
1437 # ifdef CONFIG_64BIT_PHYS_ADDR
1438 	if (cpu_has_64bits)
1439 		uasm_i_sd(p, pte, 0, ptr);
1440 	else
1441 # endif
1442 		UASM_i_SW(p, pte, 0, ptr);
1443 
1444 # ifdef CONFIG_64BIT_PHYS_ADDR
1445 	if (!cpu_has_64bits) {
1446 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1447 		uasm_i_ori(p, pte, pte, hwmode);
1448 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1449 		uasm_i_lw(p, pte, 0, ptr);
1450 	}
1451 # endif
1452 #endif
1453 }
1454 
1455 /*
1456  * Check if PTE is present, if not then jump to LABEL. PTR points to
1457  * the page table where this PTE is located, PTE will be re-loaded
1458  * with it's original value.
1459  */
1460 static void __cpuinit
1461 build_pte_present(u32 **p, struct uasm_reloc **r,
1462 		  unsigned int pte, unsigned int ptr, enum label_id lid)
1463 {
1464 	if (kernel_uses_smartmips_rixi) {
1465 		if (use_bbit_insns()) {
1466 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1467 			uasm_i_nop(p);
1468 		} else {
1469 			uasm_i_andi(p, pte, pte, _PAGE_PRESENT);
1470 			uasm_il_beqz(p, r, pte, lid);
1471 			iPTE_LW(p, pte, ptr);
1472 		}
1473 	} else {
1474 		uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1475 		uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1476 		uasm_il_bnez(p, r, pte, lid);
1477 		iPTE_LW(p, pte, ptr);
1478 	}
1479 }
1480 
1481 /* Make PTE valid, store result in PTR. */
1482 static void __cpuinit
1483 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1484 		 unsigned int ptr)
1485 {
1486 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1487 
1488 	iPTE_SW(p, r, pte, ptr, mode);
1489 }
1490 
1491 /*
1492  * Check if PTE can be written to, if not branch to LABEL. Regardless
1493  * restore PTE with value from PTR when done.
1494  */
1495 static void __cpuinit
1496 build_pte_writable(u32 **p, struct uasm_reloc **r,
1497 		   unsigned int pte, unsigned int ptr, enum label_id lid)
1498 {
1499 	if (use_bbit_insns()) {
1500 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1501 		uasm_i_nop(p);
1502 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1503 		uasm_i_nop(p);
1504 	} else {
1505 		uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1506 		uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1507 		uasm_il_bnez(p, r, pte, lid);
1508 		iPTE_LW(p, pte, ptr);
1509 	}
1510 }
1511 
1512 /* Make PTE writable, update software status bits as well, then store
1513  * at PTR.
1514  */
1515 static void __cpuinit
1516 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1517 		 unsigned int ptr)
1518 {
1519 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1520 			     | _PAGE_DIRTY);
1521 
1522 	iPTE_SW(p, r, pte, ptr, mode);
1523 }
1524 
1525 /*
1526  * Check if PTE can be modified, if not branch to LABEL. Regardless
1527  * restore PTE with value from PTR when done.
1528  */
1529 static void __cpuinit
1530 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1531 		     unsigned int pte, unsigned int ptr, enum label_id lid)
1532 {
1533 	if (use_bbit_insns()) {
1534 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1535 		uasm_i_nop(p);
1536 	} else {
1537 		uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1538 		uasm_il_beqz(p, r, pte, lid);
1539 		iPTE_LW(p, pte, ptr);
1540 	}
1541 }
1542 
1543 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1544 
1545 
1546 /*
1547  * R3000 style TLB load/store/modify handlers.
1548  */
1549 
1550 /*
1551  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1552  * Then it returns.
1553  */
1554 static void __cpuinit
1555 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1556 {
1557 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1558 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1559 	uasm_i_tlbwi(p);
1560 	uasm_i_jr(p, tmp);
1561 	uasm_i_rfe(p); /* branch delay */
1562 }
1563 
1564 /*
1565  * This places the pte into ENTRYLO0 and writes it with tlbwi
1566  * or tlbwr as appropriate.  This is because the index register
1567  * may have the probe fail bit set as a result of a trap on a
1568  * kseg2 access, i.e. without refill.  Then it returns.
1569  */
1570 static void __cpuinit
1571 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1572 			     struct uasm_reloc **r, unsigned int pte,
1573 			     unsigned int tmp)
1574 {
1575 	uasm_i_mfc0(p, tmp, C0_INDEX);
1576 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1577 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1578 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1579 	uasm_i_tlbwi(p); /* cp0 delay */
1580 	uasm_i_jr(p, tmp);
1581 	uasm_i_rfe(p); /* branch delay */
1582 	uasm_l_r3000_write_probe_fail(l, *p);
1583 	uasm_i_tlbwr(p); /* cp0 delay */
1584 	uasm_i_jr(p, tmp);
1585 	uasm_i_rfe(p); /* branch delay */
1586 }
1587 
1588 static void __cpuinit
1589 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1590 				   unsigned int ptr)
1591 {
1592 	long pgdc = (long)pgd_current;
1593 
1594 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1595 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1596 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1597 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1598 	uasm_i_sll(p, pte, pte, 2);
1599 	uasm_i_addu(p, ptr, ptr, pte);
1600 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1601 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1602 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1603 	uasm_i_addu(p, ptr, ptr, pte);
1604 	uasm_i_lw(p, pte, 0, ptr);
1605 	uasm_i_tlbp(p); /* load delay */
1606 }
1607 
1608 static void __cpuinit build_r3000_tlb_load_handler(void)
1609 {
1610 	u32 *p = handle_tlbl;
1611 	struct uasm_label *l = labels;
1612 	struct uasm_reloc *r = relocs;
1613 
1614 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1615 	memset(labels, 0, sizeof(labels));
1616 	memset(relocs, 0, sizeof(relocs));
1617 
1618 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1619 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1620 	uasm_i_nop(&p); /* load delay */
1621 	build_make_valid(&p, &r, K0, K1);
1622 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1623 
1624 	uasm_l_nopage_tlbl(&l, p);
1625 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1626 	uasm_i_nop(&p);
1627 
1628 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1629 		panic("TLB load handler fastpath space exceeded");
1630 
1631 	uasm_resolve_relocs(relocs, labels);
1632 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1633 		 (unsigned int)(p - handle_tlbl));
1634 
1635 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1636 }
1637 
1638 static void __cpuinit build_r3000_tlb_store_handler(void)
1639 {
1640 	u32 *p = handle_tlbs;
1641 	struct uasm_label *l = labels;
1642 	struct uasm_reloc *r = relocs;
1643 
1644 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1645 	memset(labels, 0, sizeof(labels));
1646 	memset(relocs, 0, sizeof(relocs));
1647 
1648 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1649 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1650 	uasm_i_nop(&p); /* load delay */
1651 	build_make_write(&p, &r, K0, K1);
1652 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1653 
1654 	uasm_l_nopage_tlbs(&l, p);
1655 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1656 	uasm_i_nop(&p);
1657 
1658 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1659 		panic("TLB store handler fastpath space exceeded");
1660 
1661 	uasm_resolve_relocs(relocs, labels);
1662 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1663 		 (unsigned int)(p - handle_tlbs));
1664 
1665 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1666 }
1667 
1668 static void __cpuinit build_r3000_tlb_modify_handler(void)
1669 {
1670 	u32 *p = handle_tlbm;
1671 	struct uasm_label *l = labels;
1672 	struct uasm_reloc *r = relocs;
1673 
1674 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1675 	memset(labels, 0, sizeof(labels));
1676 	memset(relocs, 0, sizeof(relocs));
1677 
1678 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1679 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1680 	uasm_i_nop(&p); /* load delay */
1681 	build_make_write(&p, &r, K0, K1);
1682 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1683 
1684 	uasm_l_nopage_tlbm(&l, p);
1685 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1686 	uasm_i_nop(&p);
1687 
1688 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1689 		panic("TLB modify handler fastpath space exceeded");
1690 
1691 	uasm_resolve_relocs(relocs, labels);
1692 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1693 		 (unsigned int)(p - handle_tlbm));
1694 
1695 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1696 }
1697 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1698 
1699 /*
1700  * R4000 style TLB load/store/modify handlers.
1701  */
1702 static void __cpuinit
1703 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1704 				   struct uasm_reloc **r, unsigned int pte,
1705 				   unsigned int ptr)
1706 {
1707 #ifdef CONFIG_64BIT
1708 	build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1709 #else
1710 	build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1711 #endif
1712 
1713 #ifdef CONFIG_HUGETLB_PAGE
1714 	/*
1715 	 * For huge tlb entries, pmd doesn't contain an address but
1716 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1717 	 * see if we need to jump to huge tlb processing.
1718 	 */
1719 	build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1720 #endif
1721 
1722 	UASM_i_MFC0(p, pte, C0_BADVADDR);
1723 	UASM_i_LW(p, ptr, 0, ptr);
1724 	UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1725 	uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1726 	UASM_i_ADDU(p, ptr, ptr, pte);
1727 
1728 #ifdef CONFIG_SMP
1729 	uasm_l_smp_pgtable_change(l, *p);
1730 #endif
1731 	iPTE_LW(p, pte, ptr); /* get even pte */
1732 	if (!m4kc_tlbp_war())
1733 		build_tlb_probe_entry(p);
1734 }
1735 
1736 static void __cpuinit
1737 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1738 				   struct uasm_reloc **r, unsigned int tmp,
1739 				   unsigned int ptr)
1740 {
1741 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1742 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1743 	build_update_entries(p, tmp, ptr);
1744 	build_tlb_write_entry(p, l, r, tlb_indexed);
1745 	uasm_l_leave(l, *p);
1746 	uasm_i_eret(p); /* return from trap */
1747 
1748 #ifdef CONFIG_64BIT
1749 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1750 #endif
1751 }
1752 
1753 static void __cpuinit build_r4000_tlb_load_handler(void)
1754 {
1755 	u32 *p = handle_tlbl;
1756 	struct uasm_label *l = labels;
1757 	struct uasm_reloc *r = relocs;
1758 
1759 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1760 	memset(labels, 0, sizeof(labels));
1761 	memset(relocs, 0, sizeof(relocs));
1762 
1763 	if (bcm1250_m3_war()) {
1764 		unsigned int segbits = 44;
1765 
1766 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1767 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1768 		uasm_i_xor(&p, K0, K0, K1);
1769 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1770 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1771 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1772 		uasm_i_or(&p, K0, K0, K1);
1773 		uasm_il_bnez(&p, &r, K0, label_leave);
1774 		/* No need for uasm_i_nop */
1775 	}
1776 
1777 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1778 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1779 	if (m4kc_tlbp_war())
1780 		build_tlb_probe_entry(&p);
1781 
1782 	if (kernel_uses_smartmips_rixi) {
1783 		/*
1784 		 * If the page is not _PAGE_VALID, RI or XI could not
1785 		 * have triggered it.  Skip the expensive test..
1786 		 */
1787 		if (use_bbit_insns()) {
1788 			uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
1789 				      label_tlbl_goaround1);
1790 		} else {
1791 			uasm_i_andi(&p, K0, K0, _PAGE_VALID);
1792 			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround1);
1793 		}
1794 		uasm_i_nop(&p);
1795 
1796 		uasm_i_tlbr(&p);
1797 		/* Examine  entrylo 0 or 1 based on ptr. */
1798 		if (use_bbit_insns()) {
1799 			uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
1800 		} else {
1801 			uasm_i_andi(&p, K0, K1, sizeof(pte_t));
1802 			uasm_i_beqz(&p, K0, 8);
1803 		}
1804 
1805 		UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
1806 		UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
1807 		/*
1808 		 * If the entryLo (now in K0) is valid (bit 1), RI or
1809 		 * XI must have triggered it.
1810 		 */
1811 		if (use_bbit_insns()) {
1812 			uasm_il_bbit1(&p, &r, K0, 1, label_nopage_tlbl);
1813 			/* Reload the PTE value */
1814 			iPTE_LW(&p, K0, K1);
1815 			uasm_l_tlbl_goaround1(&l, p);
1816 		} else {
1817 			uasm_i_andi(&p, K0, K0, 2);
1818 			uasm_il_bnez(&p, &r, K0, label_nopage_tlbl);
1819 			uasm_l_tlbl_goaround1(&l, p);
1820 			/* Reload the PTE value */
1821 			iPTE_LW(&p, K0, K1);
1822 		}
1823 	}
1824 	build_make_valid(&p, &r, K0, K1);
1825 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1826 
1827 #ifdef CONFIG_HUGETLB_PAGE
1828 	/*
1829 	 * This is the entry point when build_r4000_tlbchange_handler_head
1830 	 * spots a huge page.
1831 	 */
1832 	uasm_l_tlb_huge_update(&l, p);
1833 	iPTE_LW(&p, K0, K1);
1834 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1835 	build_tlb_probe_entry(&p);
1836 
1837 	if (kernel_uses_smartmips_rixi) {
1838 		/*
1839 		 * If the page is not _PAGE_VALID, RI or XI could not
1840 		 * have triggered it.  Skip the expensive test..
1841 		 */
1842 		if (use_bbit_insns()) {
1843 			uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
1844 				      label_tlbl_goaround2);
1845 		} else {
1846 			uasm_i_andi(&p, K0, K0, _PAGE_VALID);
1847 			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
1848 		}
1849 		uasm_i_nop(&p);
1850 
1851 		uasm_i_tlbr(&p);
1852 		/* Examine  entrylo 0 or 1 based on ptr. */
1853 		if (use_bbit_insns()) {
1854 			uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
1855 		} else {
1856 			uasm_i_andi(&p, K0, K1, sizeof(pte_t));
1857 			uasm_i_beqz(&p, K0, 8);
1858 		}
1859 		UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
1860 		UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
1861 		/*
1862 		 * If the entryLo (now in K0) is valid (bit 1), RI or
1863 		 * XI must have triggered it.
1864 		 */
1865 		if (use_bbit_insns()) {
1866 			uasm_il_bbit0(&p, &r, K0, 1, label_tlbl_goaround2);
1867 		} else {
1868 			uasm_i_andi(&p, K0, K0, 2);
1869 			uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
1870 		}
1871 		/* Reload the PTE value */
1872 		iPTE_LW(&p, K0, K1);
1873 
1874 		/*
1875 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
1876 		 * it is restored in build_huge_tlb_write_entry.
1877 		 */
1878 		build_restore_pagemask(&p, &r, K0, label_nopage_tlbl, 0);
1879 
1880 		uasm_l_tlbl_goaround2(&l, p);
1881 	}
1882 	uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1883 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1884 #endif
1885 
1886 	uasm_l_nopage_tlbl(&l, p);
1887 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1888 	uasm_i_nop(&p);
1889 
1890 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1891 		panic("TLB load handler fastpath space exceeded");
1892 
1893 	uasm_resolve_relocs(relocs, labels);
1894 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1895 		 (unsigned int)(p - handle_tlbl));
1896 
1897 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1898 }
1899 
1900 static void __cpuinit build_r4000_tlb_store_handler(void)
1901 {
1902 	u32 *p = handle_tlbs;
1903 	struct uasm_label *l = labels;
1904 	struct uasm_reloc *r = relocs;
1905 
1906 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1907 	memset(labels, 0, sizeof(labels));
1908 	memset(relocs, 0, sizeof(relocs));
1909 
1910 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1911 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1912 	if (m4kc_tlbp_war())
1913 		build_tlb_probe_entry(&p);
1914 	build_make_write(&p, &r, K0, K1);
1915 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1916 
1917 #ifdef CONFIG_HUGETLB_PAGE
1918 	/*
1919 	 * This is the entry point when
1920 	 * build_r4000_tlbchange_handler_head spots a huge page.
1921 	 */
1922 	uasm_l_tlb_huge_update(&l, p);
1923 	iPTE_LW(&p, K0, K1);
1924 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1925 	build_tlb_probe_entry(&p);
1926 	uasm_i_ori(&p, K0, K0,
1927 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1928 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1929 #endif
1930 
1931 	uasm_l_nopage_tlbs(&l, p);
1932 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1933 	uasm_i_nop(&p);
1934 
1935 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1936 		panic("TLB store handler fastpath space exceeded");
1937 
1938 	uasm_resolve_relocs(relocs, labels);
1939 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1940 		 (unsigned int)(p - handle_tlbs));
1941 
1942 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1943 }
1944 
1945 static void __cpuinit build_r4000_tlb_modify_handler(void)
1946 {
1947 	u32 *p = handle_tlbm;
1948 	struct uasm_label *l = labels;
1949 	struct uasm_reloc *r = relocs;
1950 
1951 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1952 	memset(labels, 0, sizeof(labels));
1953 	memset(relocs, 0, sizeof(relocs));
1954 
1955 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1956 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1957 	if (m4kc_tlbp_war())
1958 		build_tlb_probe_entry(&p);
1959 	/* Present and writable bits set, set accessed and dirty bits. */
1960 	build_make_write(&p, &r, K0, K1);
1961 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1962 
1963 #ifdef CONFIG_HUGETLB_PAGE
1964 	/*
1965 	 * This is the entry point when
1966 	 * build_r4000_tlbchange_handler_head spots a huge page.
1967 	 */
1968 	uasm_l_tlb_huge_update(&l, p);
1969 	iPTE_LW(&p, K0, K1);
1970 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1971 	build_tlb_probe_entry(&p);
1972 	uasm_i_ori(&p, K0, K0,
1973 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1974 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1975 #endif
1976 
1977 	uasm_l_nopage_tlbm(&l, p);
1978 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1979 	uasm_i_nop(&p);
1980 
1981 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1982 		panic("TLB modify handler fastpath space exceeded");
1983 
1984 	uasm_resolve_relocs(relocs, labels);
1985 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1986 		 (unsigned int)(p - handle_tlbm));
1987 
1988 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1989 }
1990 
1991 void __cpuinit build_tlb_refill_handler(void)
1992 {
1993 	/*
1994 	 * The refill handler is generated per-CPU, multi-node systems
1995 	 * may have local storage for it. The other handlers are only
1996 	 * needed once.
1997 	 */
1998 	static int run_once = 0;
1999 
2000 #ifdef CONFIG_64BIT
2001 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2002 #endif
2003 
2004 	switch (current_cpu_type()) {
2005 	case CPU_R2000:
2006 	case CPU_R3000:
2007 	case CPU_R3000A:
2008 	case CPU_R3081E:
2009 	case CPU_TX3912:
2010 	case CPU_TX3922:
2011 	case CPU_TX3927:
2012 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2013 		build_r3000_tlb_refill_handler();
2014 		if (!run_once) {
2015 			build_r3000_tlb_load_handler();
2016 			build_r3000_tlb_store_handler();
2017 			build_r3000_tlb_modify_handler();
2018 			run_once++;
2019 		}
2020 #else
2021 		panic("No R3000 TLB refill handler");
2022 #endif
2023 		break;
2024 
2025 	case CPU_R6000:
2026 	case CPU_R6000A:
2027 		panic("No R6000 TLB refill handler yet");
2028 		break;
2029 
2030 	case CPU_R8000:
2031 		panic("No R8000 TLB refill handler yet");
2032 		break;
2033 
2034 	default:
2035 		if (!run_once) {
2036 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2037 			build_r4000_setup_pgd();
2038 #endif
2039 			build_r4000_tlb_load_handler();
2040 			build_r4000_tlb_store_handler();
2041 			build_r4000_tlb_modify_handler();
2042 			run_once++;
2043 		}
2044 		build_r4000_tlb_refill_handler();
2045 	}
2046 }
2047 
2048 void __cpuinit flush_tlb_handlers(void)
2049 {
2050 	local_flush_icache_range((unsigned long)handle_tlbl,
2051 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2052 	local_flush_icache_range((unsigned long)handle_tlbs,
2053 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2054 	local_flush_icache_range((unsigned long)handle_tlbm,
2055 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2056 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2057 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2058 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2059 #endif
2060 }
2061