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
r45k_bvahwbug(void)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
r4k_250MHZhwbug(void)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
bcm1250_m3_war(void)58 static inline int __maybe_unused bcm1250_m3_war(void)
59 {
60 return BCM1250_M3_WAR;
61 }
62
r10000_llsc_war(void)63 static inline int __maybe_unused r10000_llsc_war(void)
64 {
65 return R10000_LLSC_WAR;
66 }
67
use_bbit_insns(void)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
use_lwx_insns(void)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
scratchpad_available(void)91 static bool scratchpad_available(void)
92 {
93 return true;
94 }
scratchpad_offset(int i)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
scratchpad_available(void)105 static bool scratchpad_available(void)
106 {
107 return false;
108 }
scratchpad_offset(int i)109 static int scratchpad_offset(int i)
110 {
111 BUG();
112 /* Really unreachable, but evidently some GCC want this. */
113 return 0;
114 }
115 #endif
116 /*
117 * Found by experiment: At least some revisions of the 4kc throw under
118 * some circumstances a machine check exception, triggered by invalid
119 * values in the index register. Delaying the tlbp instruction until
120 * after the next branch, plus adding an additional nop in front of
121 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
122 * why; it's not an issue caused by the core RTL.
123 *
124 */
m4kc_tlbp_war(void)125 static int __cpuinit m4kc_tlbp_war(void)
126 {
127 return (current_cpu_data.processor_id & 0xffff00) ==
128 (PRID_COMP_MIPS | PRID_IMP_4KC);
129 }
130
131 /* Handle labels (which must be positive integers). */
132 enum label_id {
133 label_second_part = 1,
134 label_leave,
135 label_vmalloc,
136 label_vmalloc_done,
137 label_tlbw_hazard,
138 label_split,
139 label_tlbl_goaround1,
140 label_tlbl_goaround2,
141 label_nopage_tlbl,
142 label_nopage_tlbs,
143 label_nopage_tlbm,
144 label_smp_pgtable_change,
145 label_r3000_write_probe_fail,
146 label_large_segbits_fault,
147 #ifdef CONFIG_HUGETLB_PAGE
148 label_tlb_huge_update,
149 #endif
150 };
151
152 UASM_L_LA(_second_part)
UASM_L_LA(_leave)153 UASM_L_LA(_leave)
154 UASM_L_LA(_vmalloc)
155 UASM_L_LA(_vmalloc_done)
156 UASM_L_LA(_tlbw_hazard)
157 UASM_L_LA(_split)
158 UASM_L_LA(_tlbl_goaround1)
159 UASM_L_LA(_tlbl_goaround2)
160 UASM_L_LA(_nopage_tlbl)
161 UASM_L_LA(_nopage_tlbs)
162 UASM_L_LA(_nopage_tlbm)
163 UASM_L_LA(_smp_pgtable_change)
164 UASM_L_LA(_r3000_write_probe_fail)
165 UASM_L_LA(_large_segbits_fault)
166 #ifdef CONFIG_HUGETLB_PAGE
167 UASM_L_LA(_tlb_huge_update)
168 #endif
169
170 /*
171 * For debug purposes.
172 */
173 static inline void dump_handler(const u32 *handler, int count)
174 {
175 int i;
176
177 pr_debug("\t.set push\n");
178 pr_debug("\t.set noreorder\n");
179
180 for (i = 0; i < count; i++)
181 pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
182
183 pr_debug("\t.set pop\n");
184 }
185
186 /* The only general purpose registers allowed in TLB handlers. */
187 #define K0 26
188 #define K1 27
189
190 /* Some CP0 registers */
191 #define C0_INDEX 0, 0
192 #define C0_ENTRYLO0 2, 0
193 #define C0_TCBIND 2, 2
194 #define C0_ENTRYLO1 3, 0
195 #define C0_CONTEXT 4, 0
196 #define C0_PAGEMASK 5, 0
197 #define C0_BADVADDR 8, 0
198 #define C0_ENTRYHI 10, 0
199 #define C0_EPC 14, 0
200 #define C0_XCONTEXT 20, 0
201
202 #ifdef CONFIG_64BIT
203 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
204 #else
205 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
206 #endif
207
208 /* The worst case length of the handler is around 18 instructions for
209 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
210 * Maximum space available is 32 instructions for R3000 and 64
211 * instructions for R4000.
212 *
213 * We deliberately chose a buffer size of 128, so we won't scribble
214 * over anything important on overflow before we panic.
215 */
216 static u32 tlb_handler[128] __cpuinitdata;
217
218 /* simply assume worst case size for labels and relocs */
219 static struct uasm_label labels[128] __cpuinitdata;
220 static struct uasm_reloc relocs[128] __cpuinitdata;
221
222 #ifdef CONFIG_64BIT
223 static int check_for_high_segbits __cpuinitdata;
224 #endif
225
226 static int check_for_high_segbits __cpuinitdata;
227
228 static unsigned int kscratch_used_mask __cpuinitdata;
229
allocate_kscratch(void)230 static int __cpuinit allocate_kscratch(void)
231 {
232 int r;
233 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
234
235 r = ffs(a);
236
237 if (r == 0)
238 return -1;
239
240 r--; /* make it zero based */
241
242 kscratch_used_mask |= (1 << r);
243
244 return r;
245 }
246
247 static int scratch_reg __cpuinitdata;
248 static int pgd_reg __cpuinitdata;
249 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
250
251 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
252
253 /*
254 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
255 * we cannot do r3000 under these circumstances.
256 *
257 * Declare pgd_current here instead of including mmu_context.h to avoid type
258 * conflicts for tlbmiss_handler_setup_pgd
259 */
260 extern unsigned long pgd_current[];
261
262 /*
263 * The R3000 TLB handler is simple.
264 */
build_r3000_tlb_refill_handler(void)265 static void __cpuinit build_r3000_tlb_refill_handler(void)
266 {
267 long pgdc = (long)pgd_current;
268 u32 *p;
269
270 memset(tlb_handler, 0, sizeof(tlb_handler));
271 p = tlb_handler;
272
273 uasm_i_mfc0(&p, K0, C0_BADVADDR);
274 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
275 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
276 uasm_i_srl(&p, K0, K0, 22); /* load delay */
277 uasm_i_sll(&p, K0, K0, 2);
278 uasm_i_addu(&p, K1, K1, K0);
279 uasm_i_mfc0(&p, K0, C0_CONTEXT);
280 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
281 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
282 uasm_i_addu(&p, K1, K1, K0);
283 uasm_i_lw(&p, K0, 0, K1);
284 uasm_i_nop(&p); /* load delay */
285 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
286 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
287 uasm_i_tlbwr(&p); /* cp0 delay */
288 uasm_i_jr(&p, K1);
289 uasm_i_rfe(&p); /* branch delay */
290
291 if (p > tlb_handler + 32)
292 panic("TLB refill handler space exceeded");
293
294 pr_debug("Wrote TLB refill handler (%u instructions).\n",
295 (unsigned int)(p - tlb_handler));
296
297 memcpy((void *)ebase, tlb_handler, 0x80);
298
299 dump_handler((u32 *)ebase, 32);
300 }
301 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
302
303 /*
304 * The R4000 TLB handler is much more complicated. We have two
305 * consecutive handler areas with 32 instructions space each.
306 * Since they aren't used at the same time, we can overflow in the
307 * other one.To keep things simple, we first assume linear space,
308 * then we relocate it to the final handler layout as needed.
309 */
310 static u32 final_handler[64] __cpuinitdata;
311
312 /*
313 * Hazards
314 *
315 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
316 * 2. A timing hazard exists for the TLBP instruction.
317 *
318 * stalling_instruction
319 * TLBP
320 *
321 * The JTLB is being read for the TLBP throughout the stall generated by the
322 * previous instruction. This is not really correct as the stalling instruction
323 * can modify the address used to access the JTLB. The failure symptom is that
324 * the TLBP instruction will use an address created for the stalling instruction
325 * and not the address held in C0_ENHI and thus report the wrong results.
326 *
327 * The software work-around is to not allow the instruction preceding the TLBP
328 * to stall - make it an NOP or some other instruction guaranteed not to stall.
329 *
330 * Errata 2 will not be fixed. This errata is also on the R5000.
331 *
332 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
333 */
build_tlb_probe_entry(u32 ** p)334 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
335 {
336 switch (current_cpu_type()) {
337 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
338 case CPU_R4600:
339 case CPU_R4700:
340 case CPU_R5000:
341 case CPU_R5000A:
342 case CPU_NEVADA:
343 uasm_i_nop(p);
344 uasm_i_tlbp(p);
345 break;
346
347 default:
348 uasm_i_tlbp(p);
349 break;
350 }
351 }
352
353 /*
354 * Write random or indexed TLB entry, and care about the hazards from
355 * the preceding mtc0 and for the following eret.
356 */
357 enum tlb_write_entry { tlb_random, tlb_indexed };
358
build_tlb_write_entry(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,enum tlb_write_entry wmode)359 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
360 struct uasm_reloc **r,
361 enum tlb_write_entry wmode)
362 {
363 void(*tlbw)(u32 **) = NULL;
364
365 switch (wmode) {
366 case tlb_random: tlbw = uasm_i_tlbwr; break;
367 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
368 }
369
370 if (cpu_has_mips_r2) {
371 if (cpu_has_mips_r2_exec_hazard)
372 uasm_i_ehb(p);
373 tlbw(p);
374 return;
375 }
376
377 switch (current_cpu_type()) {
378 case CPU_R4000PC:
379 case CPU_R4000SC:
380 case CPU_R4000MC:
381 case CPU_R4400PC:
382 case CPU_R4400SC:
383 case CPU_R4400MC:
384 /*
385 * This branch uses up a mtc0 hazard nop slot and saves
386 * two nops after the tlbw instruction.
387 */
388 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
389 tlbw(p);
390 uasm_l_tlbw_hazard(l, *p);
391 uasm_i_nop(p);
392 break;
393
394 case CPU_R4600:
395 case CPU_R4700:
396 case CPU_R5000:
397 case CPU_R5000A:
398 uasm_i_nop(p);
399 tlbw(p);
400 uasm_i_nop(p);
401 break;
402
403 case CPU_R4300:
404 case CPU_5KC:
405 case CPU_TX49XX:
406 case CPU_PR4450:
407 uasm_i_nop(p);
408 tlbw(p);
409 break;
410
411 case CPU_R10000:
412 case CPU_R12000:
413 case CPU_R14000:
414 case CPU_4KC:
415 case CPU_4KEC:
416 case CPU_SB1:
417 case CPU_SB1A:
418 case CPU_4KSC:
419 case CPU_20KC:
420 case CPU_25KF:
421 case CPU_BMIPS32:
422 case CPU_BMIPS3300:
423 case CPU_BMIPS4350:
424 case CPU_BMIPS4380:
425 case CPU_BMIPS5000:
426 case CPU_LOONGSON2:
427 case CPU_R5500:
428 if (m4kc_tlbp_war())
429 uasm_i_nop(p);
430 case CPU_ALCHEMY:
431 tlbw(p);
432 break;
433
434 case CPU_NEVADA:
435 uasm_i_nop(p); /* QED specifies 2 nops hazard */
436 /*
437 * This branch uses up a mtc0 hazard nop slot and saves
438 * a nop after the tlbw instruction.
439 */
440 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
441 tlbw(p);
442 uasm_l_tlbw_hazard(l, *p);
443 break;
444
445 case CPU_RM7000:
446 uasm_i_nop(p);
447 uasm_i_nop(p);
448 uasm_i_nop(p);
449 uasm_i_nop(p);
450 tlbw(p);
451 break;
452
453 case CPU_RM9000:
454 /*
455 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
456 * use of the JTLB for instructions should not occur for 4
457 * cpu cycles and use for data translations should not occur
458 * for 3 cpu cycles.
459 */
460 uasm_i_ssnop(p);
461 uasm_i_ssnop(p);
462 uasm_i_ssnop(p);
463 uasm_i_ssnop(p);
464 tlbw(p);
465 uasm_i_ssnop(p);
466 uasm_i_ssnop(p);
467 uasm_i_ssnop(p);
468 uasm_i_ssnop(p);
469 break;
470
471 case CPU_VR4111:
472 case CPU_VR4121:
473 case CPU_VR4122:
474 case CPU_VR4181:
475 case CPU_VR4181A:
476 uasm_i_nop(p);
477 uasm_i_nop(p);
478 tlbw(p);
479 uasm_i_nop(p);
480 uasm_i_nop(p);
481 break;
482
483 case CPU_VR4131:
484 case CPU_VR4133:
485 case CPU_R5432:
486 uasm_i_nop(p);
487 uasm_i_nop(p);
488 tlbw(p);
489 break;
490
491 case CPU_JZRISC:
492 tlbw(p);
493 uasm_i_nop(p);
494 break;
495
496 default:
497 panic("No TLB refill handler yet (CPU type: %d)",
498 current_cpu_data.cputype);
499 break;
500 }
501 }
502
build_convert_pte_to_entrylo(u32 ** p,unsigned int reg)503 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
504 unsigned int reg)
505 {
506 if (kernel_uses_smartmips_rixi) {
507 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
508 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
509 } else {
510 #ifdef CONFIG_64BIT_PHYS_ADDR
511 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
512 #else
513 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
514 #endif
515 }
516 }
517
518 #ifdef CONFIG_HUGETLB_PAGE
519
build_restore_pagemask(u32 ** p,struct uasm_reloc ** r,unsigned int tmp,enum label_id lid,int restore_scratch)520 static __cpuinit void build_restore_pagemask(u32 **p,
521 struct uasm_reloc **r,
522 unsigned int tmp,
523 enum label_id lid,
524 int restore_scratch)
525 {
526 if (restore_scratch) {
527 /* Reset default page size */
528 if (PM_DEFAULT_MASK >> 16) {
529 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
530 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
531 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
532 uasm_il_b(p, r, lid);
533 } else if (PM_DEFAULT_MASK) {
534 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
535 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
536 uasm_il_b(p, r, lid);
537 } else {
538 uasm_i_mtc0(p, 0, C0_PAGEMASK);
539 uasm_il_b(p, r, lid);
540 }
541 if (scratch_reg > 0)
542 UASM_i_MFC0(p, 1, 31, scratch_reg);
543 else
544 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
545 } else {
546 /* Reset default page size */
547 if (PM_DEFAULT_MASK >> 16) {
548 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
549 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
550 uasm_il_b(p, r, lid);
551 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
552 } else if (PM_DEFAULT_MASK) {
553 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
554 uasm_il_b(p, r, lid);
555 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
556 } else {
557 uasm_il_b(p, r, lid);
558 uasm_i_mtc0(p, 0, C0_PAGEMASK);
559 }
560 }
561 }
562
build_huge_tlb_write_entry(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int tmp,enum tlb_write_entry wmode,int restore_scratch)563 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
564 struct uasm_label **l,
565 struct uasm_reloc **r,
566 unsigned int tmp,
567 enum tlb_write_entry wmode,
568 int restore_scratch)
569 {
570 /* Set huge page tlb entry size */
571 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
572 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
573 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
574
575 build_tlb_write_entry(p, l, r, wmode);
576
577 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
578 }
579
580 /*
581 * Check if Huge PTE is present, if so then jump to LABEL.
582 */
583 static void __cpuinit
build_is_huge_pte(u32 ** p,struct uasm_reloc ** r,unsigned int tmp,unsigned int pmd,int lid)584 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
585 unsigned int pmd, int lid)
586 {
587 UASM_i_LW(p, tmp, 0, pmd);
588 if (use_bbit_insns()) {
589 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
590 } else {
591 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
592 uasm_il_bnez(p, r, tmp, lid);
593 }
594 }
595
build_huge_update_entries(u32 ** p,unsigned int pte,unsigned int tmp)596 static __cpuinit void build_huge_update_entries(u32 **p,
597 unsigned int pte,
598 unsigned int tmp)
599 {
600 int small_sequence;
601
602 /*
603 * A huge PTE describes an area the size of the
604 * configured huge page size. This is twice the
605 * of the large TLB entry size we intend to use.
606 * A TLB entry half the size of the configured
607 * huge page size is configured into entrylo0
608 * and entrylo1 to cover the contiguous huge PTE
609 * address space.
610 */
611 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
612
613 /* We can clobber tmp. It isn't used after this.*/
614 if (!small_sequence)
615 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
616
617 build_convert_pte_to_entrylo(p, pte);
618 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
619 /* convert to entrylo1 */
620 if (small_sequence)
621 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
622 else
623 UASM_i_ADDU(p, pte, pte, tmp);
624
625 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
626 }
627
build_huge_handler_tail(u32 ** p,struct uasm_reloc ** r,struct uasm_label ** l,unsigned int pte,unsigned int ptr)628 static __cpuinit void build_huge_handler_tail(u32 **p,
629 struct uasm_reloc **r,
630 struct uasm_label **l,
631 unsigned int pte,
632 unsigned int ptr)
633 {
634 #ifdef CONFIG_SMP
635 UASM_i_SC(p, pte, 0, ptr);
636 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
637 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
638 #else
639 UASM_i_SW(p, pte, 0, ptr);
640 #endif
641 build_huge_update_entries(p, pte, ptr);
642 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
643 }
644 #endif /* CONFIG_HUGETLB_PAGE */
645
646 #ifdef CONFIG_64BIT
647 /*
648 * TMP and PTR are scratch.
649 * TMP will be clobbered, PTR will hold the pmd entry.
650 */
651 static void __cpuinit
build_get_pmde64(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int tmp,unsigned int ptr)652 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
653 unsigned int tmp, unsigned int ptr)
654 {
655 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
656 long pgdc = (long)pgd_current;
657 #endif
658 /*
659 * The vmalloc handling is not in the hotpath.
660 */
661 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
662
663 if (check_for_high_segbits) {
664 /*
665 * The kernel currently implicitely assumes that the
666 * MIPS SEGBITS parameter for the processor is
667 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
668 * allocate virtual addresses outside the maximum
669 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
670 * that doesn't prevent user code from accessing the
671 * higher xuseg addresses. Here, we make sure that
672 * everything but the lower xuseg addresses goes down
673 * the module_alloc/vmalloc path.
674 */
675 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
676 uasm_il_bnez(p, r, ptr, label_vmalloc);
677 } else {
678 uasm_il_bltz(p, r, tmp, label_vmalloc);
679 }
680 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
681
682 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
683 if (pgd_reg != -1) {
684 /* pgd is in pgd_reg */
685 UASM_i_MFC0(p, ptr, 31, pgd_reg);
686 } else {
687 /*
688 * &pgd << 11 stored in CONTEXT [23..63].
689 */
690 UASM_i_MFC0(p, ptr, C0_CONTEXT);
691
692 /* Clear lower 23 bits of context. */
693 uasm_i_dins(p, ptr, 0, 0, 23);
694
695 /* 1 0 1 0 1 << 6 xkphys cached */
696 uasm_i_ori(p, ptr, ptr, 0x540);
697 uasm_i_drotr(p, ptr, ptr, 11);
698 }
699 #elif defined(CONFIG_SMP)
700 # ifdef CONFIG_MIPS_MT_SMTC
701 /*
702 * SMTC uses TCBind value as "CPU" index
703 */
704 uasm_i_mfc0(p, ptr, C0_TCBIND);
705 uasm_i_dsrl_safe(p, ptr, ptr, 19);
706 # else
707 /*
708 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
709 * stored in CONTEXT.
710 */
711 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
712 uasm_i_dsrl_safe(p, ptr, ptr, 23);
713 # endif
714 UASM_i_LA_mostly(p, tmp, pgdc);
715 uasm_i_daddu(p, ptr, ptr, tmp);
716 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
717 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
718 #else
719 UASM_i_LA_mostly(p, ptr, pgdc);
720 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
721 #endif
722
723 uasm_l_vmalloc_done(l, *p);
724
725 /* get pgd offset in bytes */
726 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
727
728 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
729 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
730 #ifndef __PAGETABLE_PMD_FOLDED
731 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
732 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
733 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
734 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
735 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
736 #endif
737 }
738
739 /*
740 * BVADDR is the faulting address, PTR is scratch.
741 * PTR will hold the pgd for vmalloc.
742 */
743 static void __cpuinit
build_get_pgd_vmalloc64(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int bvaddr,unsigned int ptr,enum vmalloc64_mode mode)744 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
745 unsigned int bvaddr, unsigned int ptr,
746 enum vmalloc64_mode mode)
747 {
748 long swpd = (long)swapper_pg_dir;
749 int single_insn_swpd;
750 int did_vmalloc_branch = 0;
751
752 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
753
754 uasm_l_vmalloc(l, *p);
755
756 if (mode != not_refill && check_for_high_segbits) {
757 if (single_insn_swpd) {
758 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
759 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
760 did_vmalloc_branch = 1;
761 /* fall through */
762 } else {
763 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
764 }
765 }
766 if (!did_vmalloc_branch) {
767 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
768 uasm_il_b(p, r, label_vmalloc_done);
769 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
770 } else {
771 UASM_i_LA_mostly(p, ptr, swpd);
772 uasm_il_b(p, r, label_vmalloc_done);
773 if (uasm_in_compat_space_p(swpd))
774 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
775 else
776 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
777 }
778 }
779 if (mode != not_refill && check_for_high_segbits) {
780 uasm_l_large_segbits_fault(l, *p);
781 /*
782 * We get here if we are an xsseg address, or if we are
783 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
784 *
785 * Ignoring xsseg (assume disabled so would generate
786 * (address errors?), the only remaining possibility
787 * is the upper xuseg addresses. On processors with
788 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
789 * addresses would have taken an address error. We try
790 * to mimic that here by taking a load/istream page
791 * fault.
792 */
793 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
794 uasm_i_jr(p, ptr);
795
796 if (mode == refill_scratch) {
797 if (scratch_reg > 0)
798 UASM_i_MFC0(p, 1, 31, scratch_reg);
799 else
800 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
801 } else {
802 uasm_i_nop(p);
803 }
804 }
805 }
806
807 #else /* !CONFIG_64BIT */
808
809 /*
810 * TMP and PTR are scratch.
811 * TMP will be clobbered, PTR will hold the pgd entry.
812 */
813 static void __cpuinit __maybe_unused
build_get_pgde32(u32 ** p,unsigned int tmp,unsigned int ptr)814 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
815 {
816 long pgdc = (long)pgd_current;
817
818 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
819 #ifdef CONFIG_SMP
820 #ifdef CONFIG_MIPS_MT_SMTC
821 /*
822 * SMTC uses TCBind value as "CPU" index
823 */
824 uasm_i_mfc0(p, ptr, C0_TCBIND);
825 UASM_i_LA_mostly(p, tmp, pgdc);
826 uasm_i_srl(p, ptr, ptr, 19);
827 #else
828 /*
829 * smp_processor_id() << 3 is stored in CONTEXT.
830 */
831 uasm_i_mfc0(p, ptr, C0_CONTEXT);
832 UASM_i_LA_mostly(p, tmp, pgdc);
833 uasm_i_srl(p, ptr, ptr, 23);
834 #endif
835 uasm_i_addu(p, ptr, tmp, ptr);
836 #else
837 UASM_i_LA_mostly(p, ptr, pgdc);
838 #endif
839 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
840 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
841 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
842 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
843 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
844 }
845
846 #endif /* !CONFIG_64BIT */
847
build_adjust_context(u32 ** p,unsigned int ctx)848 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
849 {
850 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
851 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
852
853 switch (current_cpu_type()) {
854 case CPU_VR41XX:
855 case CPU_VR4111:
856 case CPU_VR4121:
857 case CPU_VR4122:
858 case CPU_VR4131:
859 case CPU_VR4181:
860 case CPU_VR4181A:
861 case CPU_VR4133:
862 shift += 2;
863 break;
864
865 default:
866 break;
867 }
868
869 if (shift)
870 UASM_i_SRL(p, ctx, ctx, shift);
871 uasm_i_andi(p, ctx, ctx, mask);
872 }
873
build_get_ptep(u32 ** p,unsigned int tmp,unsigned int ptr)874 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
875 {
876 /*
877 * Bug workaround for the Nevada. It seems as if under certain
878 * circumstances the move from cp0_context might produce a
879 * bogus result when the mfc0 instruction and its consumer are
880 * in a different cacheline or a load instruction, probably any
881 * memory reference, is between them.
882 */
883 switch (current_cpu_type()) {
884 case CPU_NEVADA:
885 UASM_i_LW(p, ptr, 0, ptr);
886 GET_CONTEXT(p, tmp); /* get context reg */
887 break;
888
889 default:
890 GET_CONTEXT(p, tmp); /* get context reg */
891 UASM_i_LW(p, ptr, 0, ptr);
892 break;
893 }
894
895 build_adjust_context(p, tmp);
896 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
897 }
898
build_update_entries(u32 ** p,unsigned int tmp,unsigned int ptep)899 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
900 unsigned int ptep)
901 {
902 /*
903 * 64bit address support (36bit on a 32bit CPU) in a 32bit
904 * Kernel is a special case. Only a few CPUs use it.
905 */
906 #ifdef CONFIG_64BIT_PHYS_ADDR
907 if (cpu_has_64bits) {
908 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
909 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
910 if (kernel_uses_smartmips_rixi) {
911 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
912 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
913 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
914 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
915 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
916 } else {
917 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
918 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
919 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
920 }
921 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
922 } else {
923 int pte_off_even = sizeof(pte_t) / 2;
924 int pte_off_odd = pte_off_even + sizeof(pte_t);
925
926 /* The pte entries are pre-shifted */
927 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
928 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
929 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
930 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
931 }
932 #else
933 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
934 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
935 if (r45k_bvahwbug())
936 build_tlb_probe_entry(p);
937 if (kernel_uses_smartmips_rixi) {
938 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
939 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
940 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
941 if (r4k_250MHZhwbug())
942 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
943 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
944 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
945 } else {
946 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
947 if (r4k_250MHZhwbug())
948 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
949 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
950 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
951 if (r45k_bvahwbug())
952 uasm_i_mfc0(p, tmp, C0_INDEX);
953 }
954 if (r4k_250MHZhwbug())
955 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
956 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
957 #endif
958 }
959
960 struct mips_huge_tlb_info {
961 int huge_pte;
962 int restore_scratch;
963 };
964
965 static struct mips_huge_tlb_info __cpuinit
build_fast_tlb_refill_handler(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int tmp,unsigned int ptr,int c0_scratch)966 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
967 struct uasm_reloc **r, unsigned int tmp,
968 unsigned int ptr, int c0_scratch)
969 {
970 struct mips_huge_tlb_info rv;
971 unsigned int even, odd;
972 int vmalloc_branch_delay_filled = 0;
973 const int scratch = 1; /* Our extra working register */
974
975 rv.huge_pte = scratch;
976 rv.restore_scratch = 0;
977
978 if (check_for_high_segbits) {
979 UASM_i_MFC0(p, tmp, C0_BADVADDR);
980
981 if (pgd_reg != -1)
982 UASM_i_MFC0(p, ptr, 31, pgd_reg);
983 else
984 UASM_i_MFC0(p, ptr, C0_CONTEXT);
985
986 if (c0_scratch >= 0)
987 UASM_i_MTC0(p, scratch, 31, c0_scratch);
988 else
989 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
990
991 uasm_i_dsrl_safe(p, scratch, tmp,
992 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
993 uasm_il_bnez(p, r, scratch, label_vmalloc);
994
995 if (pgd_reg == -1) {
996 vmalloc_branch_delay_filled = 1;
997 /* Clear lower 23 bits of context. */
998 uasm_i_dins(p, ptr, 0, 0, 23);
999 }
1000 } else {
1001 if (pgd_reg != -1)
1002 UASM_i_MFC0(p, ptr, 31, pgd_reg);
1003 else
1004 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1005
1006 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1007
1008 if (c0_scratch >= 0)
1009 UASM_i_MTC0(p, scratch, 31, c0_scratch);
1010 else
1011 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1012
1013 if (pgd_reg == -1)
1014 /* Clear lower 23 bits of context. */
1015 uasm_i_dins(p, ptr, 0, 0, 23);
1016
1017 uasm_il_bltz(p, r, tmp, label_vmalloc);
1018 }
1019
1020 if (pgd_reg == -1) {
1021 vmalloc_branch_delay_filled = 1;
1022 /* 1 0 1 0 1 << 6 xkphys cached */
1023 uasm_i_ori(p, ptr, ptr, 0x540);
1024 uasm_i_drotr(p, ptr, ptr, 11);
1025 }
1026
1027 #ifdef __PAGETABLE_PMD_FOLDED
1028 #define LOC_PTEP scratch
1029 #else
1030 #define LOC_PTEP ptr
1031 #endif
1032
1033 if (!vmalloc_branch_delay_filled)
1034 /* get pgd offset in bytes */
1035 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1036
1037 uasm_l_vmalloc_done(l, *p);
1038
1039 /*
1040 * tmp ptr
1041 * fall-through case = badvaddr *pgd_current
1042 * vmalloc case = badvaddr swapper_pg_dir
1043 */
1044
1045 if (vmalloc_branch_delay_filled)
1046 /* get pgd offset in bytes */
1047 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1048
1049 #ifdef __PAGETABLE_PMD_FOLDED
1050 GET_CONTEXT(p, tmp); /* get context reg */
1051 #endif
1052 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1053
1054 if (use_lwx_insns()) {
1055 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1056 } else {
1057 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1058 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1059 }
1060
1061 #ifndef __PAGETABLE_PMD_FOLDED
1062 /* get pmd offset in bytes */
1063 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1064 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1065 GET_CONTEXT(p, tmp); /* get context reg */
1066
1067 if (use_lwx_insns()) {
1068 UASM_i_LWX(p, scratch, scratch, ptr);
1069 } else {
1070 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1071 UASM_i_LW(p, scratch, 0, ptr);
1072 }
1073 #endif
1074 /* Adjust the context during the load latency. */
1075 build_adjust_context(p, tmp);
1076
1077 #ifdef CONFIG_HUGETLB_PAGE
1078 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1079 /*
1080 * The in the LWX case we don't want to do the load in the
1081 * delay slot. It cannot issue in the same cycle and may be
1082 * speculative and unneeded.
1083 */
1084 if (use_lwx_insns())
1085 uasm_i_nop(p);
1086 #endif /* CONFIG_HUGETLB_PAGE */
1087
1088
1089 /* build_update_entries */
1090 if (use_lwx_insns()) {
1091 even = ptr;
1092 odd = tmp;
1093 UASM_i_LWX(p, even, scratch, tmp);
1094 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1095 UASM_i_LWX(p, odd, scratch, tmp);
1096 } else {
1097 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1098 even = tmp;
1099 odd = ptr;
1100 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1101 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1102 }
1103 if (kernel_uses_smartmips_rixi) {
1104 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_NO_EXEC));
1105 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_NO_EXEC));
1106 uasm_i_drotr(p, even, even,
1107 ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1108 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1109 uasm_i_drotr(p, odd, odd,
1110 ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1111 } else {
1112 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1113 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1114 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1115 }
1116 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1117
1118 if (c0_scratch >= 0) {
1119 UASM_i_MFC0(p, scratch, 31, c0_scratch);
1120 build_tlb_write_entry(p, l, r, tlb_random);
1121 uasm_l_leave(l, *p);
1122 rv.restore_scratch = 1;
1123 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1124 build_tlb_write_entry(p, l, r, tlb_random);
1125 uasm_l_leave(l, *p);
1126 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1127 } else {
1128 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1129 build_tlb_write_entry(p, l, r, tlb_random);
1130 uasm_l_leave(l, *p);
1131 rv.restore_scratch = 1;
1132 }
1133
1134 uasm_i_eret(p); /* return from trap */
1135
1136 return rv;
1137 }
1138
1139 /*
1140 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1141 * because EXL == 0. If we wrap, we can also use the 32 instruction
1142 * slots before the XTLB refill exception handler which belong to the
1143 * unused TLB refill exception.
1144 */
1145 #define MIPS64_REFILL_INSNS 32
1146
build_r4000_tlb_refill_handler(void)1147 static void __cpuinit build_r4000_tlb_refill_handler(void)
1148 {
1149 u32 *p = tlb_handler;
1150 struct uasm_label *l = labels;
1151 struct uasm_reloc *r = relocs;
1152 u32 *f;
1153 unsigned int final_len;
1154 struct mips_huge_tlb_info htlb_info __maybe_unused;
1155 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1156
1157 memset(tlb_handler, 0, sizeof(tlb_handler));
1158 memset(labels, 0, sizeof(labels));
1159 memset(relocs, 0, sizeof(relocs));
1160 memset(final_handler, 0, sizeof(final_handler));
1161
1162 if (scratch_reg == 0)
1163 scratch_reg = allocate_kscratch();
1164
1165 if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1166 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1167 scratch_reg);
1168 vmalloc_mode = refill_scratch;
1169 } else {
1170 htlb_info.huge_pte = K0;
1171 htlb_info.restore_scratch = 0;
1172 vmalloc_mode = refill_noscratch;
1173 /*
1174 * create the plain linear handler
1175 */
1176 if (bcm1250_m3_war()) {
1177 unsigned int segbits = 44;
1178
1179 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1180 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1181 uasm_i_xor(&p, K0, K0, K1);
1182 uasm_i_dsrl_safe(&p, K1, K0, 62);
1183 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1184 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1185 uasm_i_or(&p, K0, K0, K1);
1186 uasm_il_bnez(&p, &r, K0, label_leave);
1187 /* No need for uasm_i_nop */
1188 }
1189
1190 #ifdef CONFIG_64BIT
1191 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1192 #else
1193 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1194 #endif
1195
1196 #ifdef CONFIG_HUGETLB_PAGE
1197 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1198 #endif
1199
1200 build_get_ptep(&p, K0, K1);
1201 build_update_entries(&p, K0, K1);
1202 build_tlb_write_entry(&p, &l, &r, tlb_random);
1203 uasm_l_leave(&l, p);
1204 uasm_i_eret(&p); /* return from trap */
1205 }
1206 #ifdef CONFIG_HUGETLB_PAGE
1207 uasm_l_tlb_huge_update(&l, p);
1208 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1209 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1210 htlb_info.restore_scratch);
1211 #endif
1212
1213 #ifdef CONFIG_64BIT
1214 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1215 #endif
1216
1217 /*
1218 * Overflow check: For the 64bit handler, we need at least one
1219 * free instruction slot for the wrap-around branch. In worst
1220 * case, if the intended insertion point is a delay slot, we
1221 * need three, with the second nop'ed and the third being
1222 * unused.
1223 */
1224 /* Loongson2 ebase is different than r4k, we have more space */
1225 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1226 if ((p - tlb_handler) > 64)
1227 panic("TLB refill handler space exceeded");
1228 #else
1229 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1230 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1231 && uasm_insn_has_bdelay(relocs,
1232 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1233 panic("TLB refill handler space exceeded");
1234 #endif
1235
1236 /*
1237 * Now fold the handler in the TLB refill handler space.
1238 */
1239 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1240 f = final_handler;
1241 /* Simplest case, just copy the handler. */
1242 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1243 final_len = p - tlb_handler;
1244 #else /* CONFIG_64BIT */
1245 f = final_handler + MIPS64_REFILL_INSNS;
1246 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1247 /* Just copy the handler. */
1248 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1249 final_len = p - tlb_handler;
1250 } else {
1251 #if defined(CONFIG_HUGETLB_PAGE)
1252 const enum label_id ls = label_tlb_huge_update;
1253 #else
1254 const enum label_id ls = label_vmalloc;
1255 #endif
1256 u32 *split;
1257 int ov = 0;
1258 int i;
1259
1260 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1261 ;
1262 BUG_ON(i == ARRAY_SIZE(labels));
1263 split = labels[i].addr;
1264
1265 /*
1266 * See if we have overflown one way or the other.
1267 */
1268 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1269 split < p - MIPS64_REFILL_INSNS)
1270 ov = 1;
1271
1272 if (ov) {
1273 /*
1274 * Split two instructions before the end. One
1275 * for the branch and one for the instruction
1276 * in the delay slot.
1277 */
1278 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1279
1280 /*
1281 * If the branch would fall in a delay slot,
1282 * we must back up an additional instruction
1283 * so that it is no longer in a delay slot.
1284 */
1285 if (uasm_insn_has_bdelay(relocs, split - 1))
1286 split--;
1287 }
1288 /* Copy first part of the handler. */
1289 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1290 f += split - tlb_handler;
1291
1292 if (ov) {
1293 /* Insert branch. */
1294 uasm_l_split(&l, final_handler);
1295 uasm_il_b(&f, &r, label_split);
1296 if (uasm_insn_has_bdelay(relocs, split))
1297 uasm_i_nop(&f);
1298 else {
1299 uasm_copy_handler(relocs, labels,
1300 split, split + 1, f);
1301 uasm_move_labels(labels, f, f + 1, -1);
1302 f++;
1303 split++;
1304 }
1305 }
1306
1307 /* Copy the rest of the handler. */
1308 uasm_copy_handler(relocs, labels, split, p, final_handler);
1309 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1310 (p - split);
1311 }
1312 #endif /* CONFIG_64BIT */
1313
1314 uasm_resolve_relocs(relocs, labels);
1315 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1316 final_len);
1317
1318 memcpy((void *)ebase, final_handler, 0x100);
1319
1320 dump_handler((u32 *)ebase, 64);
1321 }
1322
1323 /*
1324 * 128 instructions for the fastpath handler is generous and should
1325 * never be exceeded.
1326 */
1327 #define FASTPATH_SIZE 128
1328
1329 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1330 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1331 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1332 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1333 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1334
build_r4000_setup_pgd(void)1335 static void __cpuinit build_r4000_setup_pgd(void)
1336 {
1337 const int a0 = 4;
1338 const int a1 = 5;
1339 u32 *p = tlbmiss_handler_setup_pgd;
1340 struct uasm_label *l = labels;
1341 struct uasm_reloc *r = relocs;
1342
1343 memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1344 memset(labels, 0, sizeof(labels));
1345 memset(relocs, 0, sizeof(relocs));
1346
1347 pgd_reg = allocate_kscratch();
1348
1349 if (pgd_reg == -1) {
1350 /* PGD << 11 in c0_Context */
1351 /*
1352 * If it is a ckseg0 address, convert to a physical
1353 * address. Shifting right by 29 and adding 4 will
1354 * result in zero for these addresses.
1355 *
1356 */
1357 UASM_i_SRA(&p, a1, a0, 29);
1358 UASM_i_ADDIU(&p, a1, a1, 4);
1359 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1360 uasm_i_nop(&p);
1361 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1362 uasm_l_tlbl_goaround1(&l, p);
1363 UASM_i_SLL(&p, a0, a0, 11);
1364 uasm_i_jr(&p, 31);
1365 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1366 } else {
1367 /* PGD in c0_KScratch */
1368 uasm_i_jr(&p, 31);
1369 UASM_i_MTC0(&p, a0, 31, pgd_reg);
1370 }
1371 if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1372 panic("tlbmiss_handler_setup_pgd space exceeded");
1373 uasm_resolve_relocs(relocs, labels);
1374 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1375 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1376
1377 dump_handler(tlbmiss_handler_setup_pgd,
1378 ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1379 }
1380 #endif
1381
1382 static void __cpuinit
iPTE_LW(u32 ** p,unsigned int pte,unsigned int ptr)1383 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1384 {
1385 #ifdef CONFIG_SMP
1386 # ifdef CONFIG_64BIT_PHYS_ADDR
1387 if (cpu_has_64bits)
1388 uasm_i_lld(p, pte, 0, ptr);
1389 else
1390 # endif
1391 UASM_i_LL(p, pte, 0, ptr);
1392 #else
1393 # ifdef CONFIG_64BIT_PHYS_ADDR
1394 if (cpu_has_64bits)
1395 uasm_i_ld(p, pte, 0, ptr);
1396 else
1397 # endif
1398 UASM_i_LW(p, pte, 0, ptr);
1399 #endif
1400 }
1401
1402 static void __cpuinit
iPTE_SW(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr,unsigned int mode)1403 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1404 unsigned int mode)
1405 {
1406 #ifdef CONFIG_64BIT_PHYS_ADDR
1407 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1408 #endif
1409
1410 uasm_i_ori(p, pte, pte, mode);
1411 #ifdef CONFIG_SMP
1412 # ifdef CONFIG_64BIT_PHYS_ADDR
1413 if (cpu_has_64bits)
1414 uasm_i_scd(p, pte, 0, ptr);
1415 else
1416 # endif
1417 UASM_i_SC(p, pte, 0, ptr);
1418
1419 if (r10000_llsc_war())
1420 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1421 else
1422 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1423
1424 # ifdef CONFIG_64BIT_PHYS_ADDR
1425 if (!cpu_has_64bits) {
1426 /* no uasm_i_nop needed */
1427 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1428 uasm_i_ori(p, pte, pte, hwmode);
1429 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1430 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1431 /* no uasm_i_nop needed */
1432 uasm_i_lw(p, pte, 0, ptr);
1433 } else
1434 uasm_i_nop(p);
1435 # else
1436 uasm_i_nop(p);
1437 # endif
1438 #else
1439 # ifdef CONFIG_64BIT_PHYS_ADDR
1440 if (cpu_has_64bits)
1441 uasm_i_sd(p, pte, 0, ptr);
1442 else
1443 # endif
1444 UASM_i_SW(p, pte, 0, ptr);
1445
1446 # ifdef CONFIG_64BIT_PHYS_ADDR
1447 if (!cpu_has_64bits) {
1448 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1449 uasm_i_ori(p, pte, pte, hwmode);
1450 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1451 uasm_i_lw(p, pte, 0, ptr);
1452 }
1453 # endif
1454 #endif
1455 }
1456
1457 /*
1458 * Check if PTE is present, if not then jump to LABEL. PTR points to
1459 * the page table where this PTE is located, PTE will be re-loaded
1460 * with it's original value.
1461 */
1462 static void __cpuinit
build_pte_present(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr,enum label_id lid)1463 build_pte_present(u32 **p, struct uasm_reloc **r,
1464 unsigned int pte, unsigned int ptr, enum label_id lid)
1465 {
1466 if (kernel_uses_smartmips_rixi) {
1467 if (use_bbit_insns()) {
1468 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1469 uasm_i_nop(p);
1470 } else {
1471 uasm_i_andi(p, pte, pte, _PAGE_PRESENT);
1472 uasm_il_beqz(p, r, pte, lid);
1473 iPTE_LW(p, pte, ptr);
1474 }
1475 } else {
1476 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1477 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
1478 uasm_il_bnez(p, r, pte, lid);
1479 iPTE_LW(p, pte, ptr);
1480 }
1481 }
1482
1483 /* Make PTE valid, store result in PTR. */
1484 static void __cpuinit
build_make_valid(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr)1485 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1486 unsigned int ptr)
1487 {
1488 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1489
1490 iPTE_SW(p, r, pte, ptr, mode);
1491 }
1492
1493 /*
1494 * Check if PTE can be written to, if not branch to LABEL. Regardless
1495 * restore PTE with value from PTR when done.
1496 */
1497 static void __cpuinit
build_pte_writable(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr,enum label_id lid)1498 build_pte_writable(u32 **p, struct uasm_reloc **r,
1499 unsigned int pte, unsigned int ptr, enum label_id lid)
1500 {
1501 if (use_bbit_insns()) {
1502 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1503 uasm_i_nop(p);
1504 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1505 uasm_i_nop(p);
1506 } else {
1507 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1508 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1509 uasm_il_bnez(p, r, pte, lid);
1510 iPTE_LW(p, pte, ptr);
1511 }
1512 }
1513
1514 /* Make PTE writable, update software status bits as well, then store
1515 * at PTR.
1516 */
1517 static void __cpuinit
build_make_write(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr)1518 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1519 unsigned int ptr)
1520 {
1521 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1522 | _PAGE_DIRTY);
1523
1524 iPTE_SW(p, r, pte, ptr, mode);
1525 }
1526
1527 /*
1528 * Check if PTE can be modified, if not branch to LABEL. Regardless
1529 * restore PTE with value from PTR when done.
1530 */
1531 static void __cpuinit
build_pte_modifiable(u32 ** p,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr,enum label_id lid)1532 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1533 unsigned int pte, unsigned int ptr, enum label_id lid)
1534 {
1535 if (use_bbit_insns()) {
1536 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1537 uasm_i_nop(p);
1538 } else {
1539 uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1540 uasm_il_beqz(p, r, pte, lid);
1541 iPTE_LW(p, pte, ptr);
1542 }
1543 }
1544
1545 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1546
1547
1548 /*
1549 * R3000 style TLB load/store/modify handlers.
1550 */
1551
1552 /*
1553 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1554 * Then it returns.
1555 */
1556 static void __cpuinit
build_r3000_pte_reload_tlbwi(u32 ** p,unsigned int pte,unsigned int tmp)1557 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1558 {
1559 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1560 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1561 uasm_i_tlbwi(p);
1562 uasm_i_jr(p, tmp);
1563 uasm_i_rfe(p); /* branch delay */
1564 }
1565
1566 /*
1567 * This places the pte into ENTRYLO0 and writes it with tlbwi
1568 * or tlbwr as appropriate. This is because the index register
1569 * may have the probe fail bit set as a result of a trap on a
1570 * kseg2 access, i.e. without refill. Then it returns.
1571 */
1572 static void __cpuinit
build_r3000_tlb_reload_write(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int pte,unsigned int tmp)1573 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1574 struct uasm_reloc **r, unsigned int pte,
1575 unsigned int tmp)
1576 {
1577 uasm_i_mfc0(p, tmp, C0_INDEX);
1578 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1579 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1580 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1581 uasm_i_tlbwi(p); /* cp0 delay */
1582 uasm_i_jr(p, tmp);
1583 uasm_i_rfe(p); /* branch delay */
1584 uasm_l_r3000_write_probe_fail(l, *p);
1585 uasm_i_tlbwr(p); /* cp0 delay */
1586 uasm_i_jr(p, tmp);
1587 uasm_i_rfe(p); /* branch delay */
1588 }
1589
1590 static void __cpuinit
build_r3000_tlbchange_handler_head(u32 ** p,unsigned int pte,unsigned int ptr)1591 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1592 unsigned int ptr)
1593 {
1594 long pgdc = (long)pgd_current;
1595
1596 uasm_i_mfc0(p, pte, C0_BADVADDR);
1597 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1598 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1599 uasm_i_srl(p, pte, pte, 22); /* load delay */
1600 uasm_i_sll(p, pte, pte, 2);
1601 uasm_i_addu(p, ptr, ptr, pte);
1602 uasm_i_mfc0(p, pte, C0_CONTEXT);
1603 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1604 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1605 uasm_i_addu(p, ptr, ptr, pte);
1606 uasm_i_lw(p, pte, 0, ptr);
1607 uasm_i_tlbp(p); /* load delay */
1608 }
1609
build_r3000_tlb_load_handler(void)1610 static void __cpuinit build_r3000_tlb_load_handler(void)
1611 {
1612 u32 *p = handle_tlbl;
1613 struct uasm_label *l = labels;
1614 struct uasm_reloc *r = relocs;
1615
1616 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1617 memset(labels, 0, sizeof(labels));
1618 memset(relocs, 0, sizeof(relocs));
1619
1620 build_r3000_tlbchange_handler_head(&p, K0, K1);
1621 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1622 uasm_i_nop(&p); /* load delay */
1623 build_make_valid(&p, &r, K0, K1);
1624 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1625
1626 uasm_l_nopage_tlbl(&l, p);
1627 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1628 uasm_i_nop(&p);
1629
1630 if ((p - handle_tlbl) > FASTPATH_SIZE)
1631 panic("TLB load handler fastpath space exceeded");
1632
1633 uasm_resolve_relocs(relocs, labels);
1634 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1635 (unsigned int)(p - handle_tlbl));
1636
1637 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1638 }
1639
build_r3000_tlb_store_handler(void)1640 static void __cpuinit build_r3000_tlb_store_handler(void)
1641 {
1642 u32 *p = handle_tlbs;
1643 struct uasm_label *l = labels;
1644 struct uasm_reloc *r = relocs;
1645
1646 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1647 memset(labels, 0, sizeof(labels));
1648 memset(relocs, 0, sizeof(relocs));
1649
1650 build_r3000_tlbchange_handler_head(&p, K0, K1);
1651 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1652 uasm_i_nop(&p); /* load delay */
1653 build_make_write(&p, &r, K0, K1);
1654 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1655
1656 uasm_l_nopage_tlbs(&l, p);
1657 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1658 uasm_i_nop(&p);
1659
1660 if ((p - handle_tlbs) > FASTPATH_SIZE)
1661 panic("TLB store handler fastpath space exceeded");
1662
1663 uasm_resolve_relocs(relocs, labels);
1664 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1665 (unsigned int)(p - handle_tlbs));
1666
1667 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1668 }
1669
build_r3000_tlb_modify_handler(void)1670 static void __cpuinit build_r3000_tlb_modify_handler(void)
1671 {
1672 u32 *p = handle_tlbm;
1673 struct uasm_label *l = labels;
1674 struct uasm_reloc *r = relocs;
1675
1676 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1677 memset(labels, 0, sizeof(labels));
1678 memset(relocs, 0, sizeof(relocs));
1679
1680 build_r3000_tlbchange_handler_head(&p, K0, K1);
1681 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1682 uasm_i_nop(&p); /* load delay */
1683 build_make_write(&p, &r, K0, K1);
1684 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1685
1686 uasm_l_nopage_tlbm(&l, p);
1687 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1688 uasm_i_nop(&p);
1689
1690 if ((p - handle_tlbm) > FASTPATH_SIZE)
1691 panic("TLB modify handler fastpath space exceeded");
1692
1693 uasm_resolve_relocs(relocs, labels);
1694 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1695 (unsigned int)(p - handle_tlbm));
1696
1697 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1698 }
1699 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1700
1701 /*
1702 * R4000 style TLB load/store/modify handlers.
1703 */
1704 static void __cpuinit
build_r4000_tlbchange_handler_head(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int pte,unsigned int ptr)1705 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1706 struct uasm_reloc **r, unsigned int pte,
1707 unsigned int ptr)
1708 {
1709 #ifdef CONFIG_64BIT
1710 build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1711 #else
1712 build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1713 #endif
1714
1715 #ifdef CONFIG_HUGETLB_PAGE
1716 /*
1717 * For huge tlb entries, pmd doesn't contain an address but
1718 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1719 * see if we need to jump to huge tlb processing.
1720 */
1721 build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1722 #endif
1723
1724 UASM_i_MFC0(p, pte, C0_BADVADDR);
1725 UASM_i_LW(p, ptr, 0, ptr);
1726 UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1727 uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1728 UASM_i_ADDU(p, ptr, ptr, pte);
1729
1730 #ifdef CONFIG_SMP
1731 uasm_l_smp_pgtable_change(l, *p);
1732 #endif
1733 iPTE_LW(p, pte, ptr); /* get even pte */
1734 if (!m4kc_tlbp_war())
1735 build_tlb_probe_entry(p);
1736 }
1737
1738 static void __cpuinit
build_r4000_tlbchange_handler_tail(u32 ** p,struct uasm_label ** l,struct uasm_reloc ** r,unsigned int tmp,unsigned int ptr)1739 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1740 struct uasm_reloc **r, unsigned int tmp,
1741 unsigned int ptr)
1742 {
1743 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1744 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1745 build_update_entries(p, tmp, ptr);
1746 build_tlb_write_entry(p, l, r, tlb_indexed);
1747 uasm_l_leave(l, *p);
1748 uasm_i_eret(p); /* return from trap */
1749
1750 #ifdef CONFIG_64BIT
1751 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1752 #endif
1753 }
1754
build_r4000_tlb_load_handler(void)1755 static void __cpuinit build_r4000_tlb_load_handler(void)
1756 {
1757 u32 *p = handle_tlbl;
1758 struct uasm_label *l = labels;
1759 struct uasm_reloc *r = relocs;
1760
1761 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1762 memset(labels, 0, sizeof(labels));
1763 memset(relocs, 0, sizeof(relocs));
1764
1765 if (bcm1250_m3_war()) {
1766 unsigned int segbits = 44;
1767
1768 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1769 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1770 uasm_i_xor(&p, K0, K0, K1);
1771 uasm_i_dsrl_safe(&p, K1, K0, 62);
1772 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1773 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1774 uasm_i_or(&p, K0, K0, K1);
1775 uasm_il_bnez(&p, &r, K0, label_leave);
1776 /* No need for uasm_i_nop */
1777 }
1778
1779 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1780 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1781 if (m4kc_tlbp_war())
1782 build_tlb_probe_entry(&p);
1783
1784 if (kernel_uses_smartmips_rixi) {
1785 /*
1786 * If the page is not _PAGE_VALID, RI or XI could not
1787 * have triggered it. Skip the expensive test..
1788 */
1789 if (use_bbit_insns()) {
1790 uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
1791 label_tlbl_goaround1);
1792 } else {
1793 uasm_i_andi(&p, K0, K0, _PAGE_VALID);
1794 uasm_il_beqz(&p, &r, K0, label_tlbl_goaround1);
1795 }
1796 uasm_i_nop(&p);
1797
1798 uasm_i_tlbr(&p);
1799 /* Examine entrylo 0 or 1 based on ptr. */
1800 if (use_bbit_insns()) {
1801 uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
1802 } else {
1803 uasm_i_andi(&p, K0, K1, sizeof(pte_t));
1804 uasm_i_beqz(&p, K0, 8);
1805 }
1806
1807 UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
1808 UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
1809 /*
1810 * If the entryLo (now in K0) is valid (bit 1), RI or
1811 * XI must have triggered it.
1812 */
1813 if (use_bbit_insns()) {
1814 uasm_il_bbit1(&p, &r, K0, 1, label_nopage_tlbl);
1815 /* Reload the PTE value */
1816 iPTE_LW(&p, K0, K1);
1817 uasm_l_tlbl_goaround1(&l, p);
1818 } else {
1819 uasm_i_andi(&p, K0, K0, 2);
1820 uasm_il_bnez(&p, &r, K0, label_nopage_tlbl);
1821 uasm_l_tlbl_goaround1(&l, p);
1822 /* Reload the PTE value */
1823 iPTE_LW(&p, K0, K1);
1824 }
1825 }
1826 build_make_valid(&p, &r, K0, K1);
1827 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1828
1829 #ifdef CONFIG_HUGETLB_PAGE
1830 /*
1831 * This is the entry point when build_r4000_tlbchange_handler_head
1832 * spots a huge page.
1833 */
1834 uasm_l_tlb_huge_update(&l, p);
1835 iPTE_LW(&p, K0, K1);
1836 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1837 build_tlb_probe_entry(&p);
1838
1839 if (kernel_uses_smartmips_rixi) {
1840 /*
1841 * If the page is not _PAGE_VALID, RI or XI could not
1842 * have triggered it. Skip the expensive test..
1843 */
1844 if (use_bbit_insns()) {
1845 uasm_il_bbit0(&p, &r, K0, ilog2(_PAGE_VALID),
1846 label_tlbl_goaround2);
1847 } else {
1848 uasm_i_andi(&p, K0, K0, _PAGE_VALID);
1849 uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
1850 }
1851 uasm_i_nop(&p);
1852
1853 uasm_i_tlbr(&p);
1854 /* Examine entrylo 0 or 1 based on ptr. */
1855 if (use_bbit_insns()) {
1856 uasm_i_bbit0(&p, K1, ilog2(sizeof(pte_t)), 8);
1857 } else {
1858 uasm_i_andi(&p, K0, K1, sizeof(pte_t));
1859 uasm_i_beqz(&p, K0, 8);
1860 }
1861 UASM_i_MFC0(&p, K0, C0_ENTRYLO0); /* load it in the delay slot*/
1862 UASM_i_MFC0(&p, K0, C0_ENTRYLO1); /* load it if ptr is odd */
1863 /*
1864 * If the entryLo (now in K0) is valid (bit 1), RI or
1865 * XI must have triggered it.
1866 */
1867 if (use_bbit_insns()) {
1868 uasm_il_bbit0(&p, &r, K0, 1, label_tlbl_goaround2);
1869 } else {
1870 uasm_i_andi(&p, K0, K0, 2);
1871 uasm_il_beqz(&p, &r, K0, label_tlbl_goaround2);
1872 }
1873 /* Reload the PTE value */
1874 iPTE_LW(&p, K0, K1);
1875
1876 /*
1877 * We clobbered C0_PAGEMASK, restore it. On the other branch
1878 * it is restored in build_huge_tlb_write_entry.
1879 */
1880 build_restore_pagemask(&p, &r, K0, label_nopage_tlbl, 0);
1881
1882 uasm_l_tlbl_goaround2(&l, p);
1883 }
1884 uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1885 build_huge_handler_tail(&p, &r, &l, K0, K1);
1886 #endif
1887
1888 uasm_l_nopage_tlbl(&l, p);
1889 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1890 uasm_i_nop(&p);
1891
1892 if ((p - handle_tlbl) > FASTPATH_SIZE)
1893 panic("TLB load handler fastpath space exceeded");
1894
1895 uasm_resolve_relocs(relocs, labels);
1896 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1897 (unsigned int)(p - handle_tlbl));
1898
1899 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1900 }
1901
build_r4000_tlb_store_handler(void)1902 static void __cpuinit build_r4000_tlb_store_handler(void)
1903 {
1904 u32 *p = handle_tlbs;
1905 struct uasm_label *l = labels;
1906 struct uasm_reloc *r = relocs;
1907
1908 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1909 memset(labels, 0, sizeof(labels));
1910 memset(relocs, 0, sizeof(relocs));
1911
1912 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1913 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1914 if (m4kc_tlbp_war())
1915 build_tlb_probe_entry(&p);
1916 build_make_write(&p, &r, K0, K1);
1917 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1918
1919 #ifdef CONFIG_HUGETLB_PAGE
1920 /*
1921 * This is the entry point when
1922 * build_r4000_tlbchange_handler_head spots a huge page.
1923 */
1924 uasm_l_tlb_huge_update(&l, p);
1925 iPTE_LW(&p, K0, K1);
1926 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1927 build_tlb_probe_entry(&p);
1928 uasm_i_ori(&p, K0, K0,
1929 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1930 build_huge_handler_tail(&p, &r, &l, K0, K1);
1931 #endif
1932
1933 uasm_l_nopage_tlbs(&l, p);
1934 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1935 uasm_i_nop(&p);
1936
1937 if ((p - handle_tlbs) > FASTPATH_SIZE)
1938 panic("TLB store handler fastpath space exceeded");
1939
1940 uasm_resolve_relocs(relocs, labels);
1941 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1942 (unsigned int)(p - handle_tlbs));
1943
1944 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1945 }
1946
build_r4000_tlb_modify_handler(void)1947 static void __cpuinit build_r4000_tlb_modify_handler(void)
1948 {
1949 u32 *p = handle_tlbm;
1950 struct uasm_label *l = labels;
1951 struct uasm_reloc *r = relocs;
1952
1953 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1954 memset(labels, 0, sizeof(labels));
1955 memset(relocs, 0, sizeof(relocs));
1956
1957 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1958 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1959 if (m4kc_tlbp_war())
1960 build_tlb_probe_entry(&p);
1961 /* Present and writable bits set, set accessed and dirty bits. */
1962 build_make_write(&p, &r, K0, K1);
1963 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1964
1965 #ifdef CONFIG_HUGETLB_PAGE
1966 /*
1967 * This is the entry point when
1968 * build_r4000_tlbchange_handler_head spots a huge page.
1969 */
1970 uasm_l_tlb_huge_update(&l, p);
1971 iPTE_LW(&p, K0, K1);
1972 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1973 build_tlb_probe_entry(&p);
1974 uasm_i_ori(&p, K0, K0,
1975 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1976 build_huge_handler_tail(&p, &r, &l, K0, K1);
1977 #endif
1978
1979 uasm_l_nopage_tlbm(&l, p);
1980 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1981 uasm_i_nop(&p);
1982
1983 if ((p - handle_tlbm) > FASTPATH_SIZE)
1984 panic("TLB modify handler fastpath space exceeded");
1985
1986 uasm_resolve_relocs(relocs, labels);
1987 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1988 (unsigned int)(p - handle_tlbm));
1989
1990 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1991 }
1992
build_tlb_refill_handler(void)1993 void __cpuinit build_tlb_refill_handler(void)
1994 {
1995 /*
1996 * The refill handler is generated per-CPU, multi-node systems
1997 * may have local storage for it. The other handlers are only
1998 * needed once.
1999 */
2000 static int run_once = 0;
2001
2002 #ifdef CONFIG_64BIT
2003 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2004 #endif
2005
2006 switch (current_cpu_type()) {
2007 case CPU_R2000:
2008 case CPU_R3000:
2009 case CPU_R3000A:
2010 case CPU_R3081E:
2011 case CPU_TX3912:
2012 case CPU_TX3922:
2013 case CPU_TX3927:
2014 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2015 build_r3000_tlb_refill_handler();
2016 if (!run_once) {
2017 build_r3000_tlb_load_handler();
2018 build_r3000_tlb_store_handler();
2019 build_r3000_tlb_modify_handler();
2020 run_once++;
2021 }
2022 #else
2023 panic("No R3000 TLB refill handler");
2024 #endif
2025 break;
2026
2027 case CPU_R6000:
2028 case CPU_R6000A:
2029 panic("No R6000 TLB refill handler yet");
2030 break;
2031
2032 case CPU_R8000:
2033 panic("No R8000 TLB refill handler yet");
2034 break;
2035
2036 default:
2037 if (!run_once) {
2038 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2039 build_r4000_setup_pgd();
2040 #endif
2041 build_r4000_tlb_load_handler();
2042 build_r4000_tlb_store_handler();
2043 build_r4000_tlb_modify_handler();
2044 run_once++;
2045 }
2046 build_r4000_tlb_refill_handler();
2047 }
2048 }
2049
flush_tlb_handlers(void)2050 void __cpuinit flush_tlb_handlers(void)
2051 {
2052 local_flush_icache_range((unsigned long)handle_tlbl,
2053 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2054 local_flush_icache_range((unsigned long)handle_tlbs,
2055 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2056 local_flush_icache_range((unsigned long)handle_tlbm,
2057 (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2058 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2059 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2060 (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2061 #endif
2062 }
2063