powerpc/mm: handle hugepage size correctly when invalidating hpte entries
[deliverable/linux.git] / arch / powerpc / mm / hash_native_64.c
1 /*
2 * native hashtable management.
3 *
4 * SMP scalability work:
5 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
11 */
12
13 #undef DEBUG_LOW
14
15 #include <linux/spinlock.h>
16 #include <linux/bitops.h>
17 #include <linux/of.h>
18 #include <linux/threads.h>
19 #include <linux/smp.h>
20
21 #include <asm/machdep.h>
22 #include <asm/mmu.h>
23 #include <asm/mmu_context.h>
24 #include <asm/pgtable.h>
25 #include <asm/tlbflush.h>
26 #include <asm/tlb.h>
27 #include <asm/cputable.h>
28 #include <asm/udbg.h>
29 #include <asm/kexec.h>
30 #include <asm/ppc-opcode.h>
31
32 #ifdef DEBUG_LOW
33 #define DBG_LOW(fmt...) udbg_printf(fmt)
34 #else
35 #define DBG_LOW(fmt...)
36 #endif
37
38 #define HPTE_LOCK_BIT 3
39
40 DEFINE_RAW_SPINLOCK(native_tlbie_lock);
41
42 static inline void __tlbie(unsigned long vpn, int psize, int apsize, int ssize)
43 {
44 unsigned long va;
45 unsigned int penc;
46
47 /*
48 * We need 14 to 65 bits of va for a tlibe of 4K page
49 * With vpn we ignore the lower VPN_SHIFT bits already.
50 * And top two bits are already ignored because we can
51 * only accomadate 76 bits in a 64 bit vpn with a VPN_SHIFT
52 * of 12.
53 */
54 va = vpn << VPN_SHIFT;
55 /*
56 * clear top 16 bits of 64bit va, non SLS segment
57 * Older versions of the architecture (2.02 and earler) require the
58 * masking of the top 16 bits.
59 */
60 va &= ~(0xffffULL << 48);
61
62 switch (psize) {
63 case MMU_PAGE_4K:
64 /* clear out bits after (52) [0....52.....63] */
65 va &= ~((1ul << (64 - 52)) - 1);
66 va |= ssize << 8;
67 va |= mmu_psize_defs[apsize].sllp << 6;
68 asm volatile(ASM_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0), %2)
69 : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
70 : "memory");
71 break;
72 default:
73 /* We need 14 to 14 + i bits of va */
74 penc = mmu_psize_defs[psize].penc[apsize];
75 va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
76 va |= penc << 12;
77 va |= ssize << 8;
78 /* Add AVAL part */
79 if (psize != apsize) {
80 /*
81 * MPSS, 64K base page size and 16MB parge page size
82 * We don't need all the bits, but rest of the bits
83 * must be ignored by the processor.
84 * vpn cover upto 65 bits of va. (0...65) and we need
85 * 58..64 bits of va.
86 */
87 va |= (vpn & 0xfe);
88 }
89 va |= 1; /* L */
90 asm volatile(ASM_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0), %2)
91 : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
92 : "memory");
93 break;
94 }
95 }
96
97 static inline void __tlbiel(unsigned long vpn, int psize, int apsize, int ssize)
98 {
99 unsigned long va;
100 unsigned int penc;
101
102 /* VPN_SHIFT can be atmost 12 */
103 va = vpn << VPN_SHIFT;
104 /*
105 * clear top 16 bits of 64 bit va, non SLS segment
106 * Older versions of the architecture (2.02 and earler) require the
107 * masking of the top 16 bits.
108 */
109 va &= ~(0xffffULL << 48);
110
111 switch (psize) {
112 case MMU_PAGE_4K:
113 /* clear out bits after(52) [0....52.....63] */
114 va &= ~((1ul << (64 - 52)) - 1);
115 va |= ssize << 8;
116 va |= mmu_psize_defs[apsize].sllp << 6;
117 asm volatile(".long 0x7c000224 | (%0 << 11) | (0 << 21)"
118 : : "r"(va) : "memory");
119 break;
120 default:
121 /* We need 14 to 14 + i bits of va */
122 penc = mmu_psize_defs[psize].penc[apsize];
123 va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
124 va |= penc << 12;
125 va |= ssize << 8;
126 /* Add AVAL part */
127 if (psize != apsize) {
128 /*
129 * MPSS, 64K base page size and 16MB parge page size
130 * We don't need all the bits, but rest of the bits
131 * must be ignored by the processor.
132 * vpn cover upto 65 bits of va. (0...65) and we need
133 * 58..64 bits of va.
134 */
135 va |= (vpn & 0xfe);
136 }
137 va |= 1; /* L */
138 asm volatile(".long 0x7c000224 | (%0 << 11) | (1 << 21)"
139 : : "r"(va) : "memory");
140 break;
141 }
142
143 }
144
145 static inline void tlbie(unsigned long vpn, int psize, int apsize,
146 int ssize, int local)
147 {
148 unsigned int use_local = local && mmu_has_feature(MMU_FTR_TLBIEL);
149 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
150
151 if (use_local)
152 use_local = mmu_psize_defs[psize].tlbiel;
153 if (lock_tlbie && !use_local)
154 raw_spin_lock(&native_tlbie_lock);
155 asm volatile("ptesync": : :"memory");
156 if (use_local) {
157 __tlbiel(vpn, psize, apsize, ssize);
158 asm volatile("ptesync": : :"memory");
159 } else {
160 __tlbie(vpn, psize, apsize, ssize);
161 asm volatile("eieio; tlbsync; ptesync": : :"memory");
162 }
163 if (lock_tlbie && !use_local)
164 raw_spin_unlock(&native_tlbie_lock);
165 }
166
167 static inline void native_lock_hpte(struct hash_pte *hptep)
168 {
169 unsigned long *word = &hptep->v;
170
171 while (1) {
172 if (!test_and_set_bit_lock(HPTE_LOCK_BIT, word))
173 break;
174 while(test_bit(HPTE_LOCK_BIT, word))
175 cpu_relax();
176 }
177 }
178
179 static inline void native_unlock_hpte(struct hash_pte *hptep)
180 {
181 unsigned long *word = &hptep->v;
182
183 clear_bit_unlock(HPTE_LOCK_BIT, word);
184 }
185
186 static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn,
187 unsigned long pa, unsigned long rflags,
188 unsigned long vflags, int psize, int apsize, int ssize)
189 {
190 struct hash_pte *hptep = htab_address + hpte_group;
191 unsigned long hpte_v, hpte_r;
192 int i;
193
194 if (!(vflags & HPTE_V_BOLTED)) {
195 DBG_LOW(" insert(group=%lx, vpn=%016lx, pa=%016lx,"
196 " rflags=%lx, vflags=%lx, psize=%d)\n",
197 hpte_group, vpn, pa, rflags, vflags, psize);
198 }
199
200 for (i = 0; i < HPTES_PER_GROUP; i++) {
201 if (! (hptep->v & HPTE_V_VALID)) {
202 /* retry with lock held */
203 native_lock_hpte(hptep);
204 if (! (hptep->v & HPTE_V_VALID))
205 break;
206 native_unlock_hpte(hptep);
207 }
208
209 hptep++;
210 }
211
212 if (i == HPTES_PER_GROUP)
213 return -1;
214
215 hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
216 hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
217
218 if (!(vflags & HPTE_V_BOLTED)) {
219 DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
220 i, hpte_v, hpte_r);
221 }
222
223 hptep->r = hpte_r;
224 /* Guarantee the second dword is visible before the valid bit */
225 eieio();
226 /*
227 * Now set the first dword including the valid bit
228 * NOTE: this also unlocks the hpte
229 */
230 hptep->v = hpte_v;
231
232 __asm__ __volatile__ ("ptesync" : : : "memory");
233
234 return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
235 }
236
237 static long native_hpte_remove(unsigned long hpte_group)
238 {
239 struct hash_pte *hptep;
240 int i;
241 int slot_offset;
242 unsigned long hpte_v;
243
244 DBG_LOW(" remove(group=%lx)\n", hpte_group);
245
246 /* pick a random entry to start at */
247 slot_offset = mftb() & 0x7;
248
249 for (i = 0; i < HPTES_PER_GROUP; i++) {
250 hptep = htab_address + hpte_group + slot_offset;
251 hpte_v = hptep->v;
252
253 if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
254 /* retry with lock held */
255 native_lock_hpte(hptep);
256 hpte_v = hptep->v;
257 if ((hpte_v & HPTE_V_VALID)
258 && !(hpte_v & HPTE_V_BOLTED))
259 break;
260 native_unlock_hpte(hptep);
261 }
262
263 slot_offset++;
264 slot_offset &= 0x7;
265 }
266
267 if (i == HPTES_PER_GROUP)
268 return -1;
269
270 /* Invalidate the hpte. NOTE: this also unlocks it */
271 hptep->v = 0;
272
273 return i;
274 }
275
276 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
277 unsigned long vpn, int bpsize,
278 int apsize, int ssize, int local)
279 {
280 struct hash_pte *hptep = htab_address + slot;
281 unsigned long hpte_v, want_v;
282 int ret = 0;
283
284 want_v = hpte_encode_avpn(vpn, bpsize, ssize);
285
286 DBG_LOW(" update(vpn=%016lx, avpnv=%016lx, group=%lx, newpp=%lx)",
287 vpn, want_v & HPTE_V_AVPN, slot, newpp);
288
289 native_lock_hpte(hptep);
290
291 hpte_v = hptep->v;
292 /*
293 * We need to invalidate the TLB always because hpte_remove doesn't do
294 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
295 * random entry from it. When we do that we don't invalidate the TLB
296 * (hpte_remove) because we assume the old translation is still
297 * technically "valid".
298 */
299 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) {
300 DBG_LOW(" -> miss\n");
301 ret = -1;
302 } else {
303 DBG_LOW(" -> hit\n");
304 /* Update the HPTE */
305 hptep->r = (hptep->r & ~(HPTE_R_PP | HPTE_R_N)) |
306 (newpp & (HPTE_R_PP | HPTE_R_N | HPTE_R_C));
307 }
308 native_unlock_hpte(hptep);
309
310 /* Ensure it is out of the tlb too. */
311 tlbie(vpn, bpsize, apsize, ssize, local);
312
313 return ret;
314 }
315
316 static long native_hpte_find(unsigned long vpn, int psize, int ssize)
317 {
318 struct hash_pte *hptep;
319 unsigned long hash;
320 unsigned long i;
321 long slot;
322 unsigned long want_v, hpte_v;
323
324 hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
325 want_v = hpte_encode_avpn(vpn, psize, ssize);
326
327 /* Bolted mappings are only ever in the primary group */
328 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
329 for (i = 0; i < HPTES_PER_GROUP; i++) {
330 hptep = htab_address + slot;
331 hpte_v = hptep->v;
332
333 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
334 /* HPTE matches */
335 return slot;
336 ++slot;
337 }
338
339 return -1;
340 }
341
342 /*
343 * Update the page protection bits. Intended to be used to create
344 * guard pages for kernel data structures on pages which are bolted
345 * in the HPT. Assumes pages being operated on will not be stolen.
346 *
347 * No need to lock here because we should be the only user.
348 */
349 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea,
350 int psize, int ssize)
351 {
352 unsigned long vpn;
353 unsigned long vsid;
354 long slot;
355 struct hash_pte *hptep;
356
357 vsid = get_kernel_vsid(ea, ssize);
358 vpn = hpt_vpn(ea, vsid, ssize);
359
360 slot = native_hpte_find(vpn, psize, ssize);
361 if (slot == -1)
362 panic("could not find page to bolt\n");
363 hptep = htab_address + slot;
364
365 /* Update the HPTE */
366 hptep->r = (hptep->r & ~(HPTE_R_PP | HPTE_R_N)) |
367 (newpp & (HPTE_R_PP | HPTE_R_N));
368 /*
369 * Ensure it is out of the tlb too. Bolted entries base and
370 * actual page size will be same.
371 */
372 tlbie(vpn, psize, psize, ssize, 0);
373 }
374
375 static void native_hpte_invalidate(unsigned long slot, unsigned long vpn,
376 int bpsize, int apsize, int ssize, int local)
377 {
378 struct hash_pte *hptep = htab_address + slot;
379 unsigned long hpte_v;
380 unsigned long want_v;
381 unsigned long flags;
382
383 local_irq_save(flags);
384
385 DBG_LOW(" invalidate(vpn=%016lx, hash: %lx)\n", vpn, slot);
386
387 want_v = hpte_encode_avpn(vpn, bpsize, ssize);
388 native_lock_hpte(hptep);
389 hpte_v = hptep->v;
390
391 /*
392 * We need to invalidate the TLB always because hpte_remove doesn't do
393 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
394 * random entry from it. When we do that we don't invalidate the TLB
395 * (hpte_remove) because we assume the old translation is still
396 * technically "valid".
397 */
398 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
399 native_unlock_hpte(hptep);
400 else
401 /* Invalidate the hpte. NOTE: this also unlocks it */
402 hptep->v = 0;
403
404 /* Invalidate the TLB */
405 tlbie(vpn, bpsize, apsize, ssize, local);
406
407 local_irq_restore(flags);
408 }
409
410 static inline int __hpte_actual_psize(unsigned int lp, int psize)
411 {
412 int i, shift;
413 unsigned int mask;
414
415 /* start from 1 ignoring MMU_PAGE_4K */
416 for (i = 1; i < MMU_PAGE_COUNT; i++) {
417
418 /* invalid penc */
419 if (mmu_psize_defs[psize].penc[i] == -1)
420 continue;
421 /*
422 * encoding bits per actual page size
423 * PTE LP actual page size
424 * rrrr rrrz >=8KB
425 * rrrr rrzz >=16KB
426 * rrrr rzzz >=32KB
427 * rrrr zzzz >=64KB
428 * .......
429 */
430 shift = mmu_psize_defs[i].shift - LP_SHIFT;
431 if (shift > LP_BITS)
432 shift = LP_BITS;
433 mask = (1 << shift) - 1;
434 if ((lp & mask) == mmu_psize_defs[psize].penc[i])
435 return i;
436 }
437 return -1;
438 }
439
440 static void hpte_decode(struct hash_pte *hpte, unsigned long slot,
441 int *psize, int *apsize, int *ssize, unsigned long *vpn)
442 {
443 unsigned long avpn, pteg, vpi;
444 unsigned long hpte_v = hpte->v;
445 unsigned long vsid, seg_off;
446 int size, a_size, shift;
447 /* Look at the 8 bit LP value */
448 unsigned int lp = (hpte->r >> LP_SHIFT) & ((1 << LP_BITS) - 1);
449
450 if (!(hpte_v & HPTE_V_LARGE)) {
451 size = MMU_PAGE_4K;
452 a_size = MMU_PAGE_4K;
453 } else {
454 for (size = 0; size < MMU_PAGE_COUNT; size++) {
455
456 /* valid entries have a shift value */
457 if (!mmu_psize_defs[size].shift)
458 continue;
459
460 a_size = __hpte_actual_psize(lp, size);
461 if (a_size != -1)
462 break;
463 }
464 }
465 /* This works for all page sizes, and for 256M and 1T segments */
466 *ssize = hpte_v >> HPTE_V_SSIZE_SHIFT;
467 shift = mmu_psize_defs[size].shift;
468
469 avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm);
470 pteg = slot / HPTES_PER_GROUP;
471 if (hpte_v & HPTE_V_SECONDARY)
472 pteg = ~pteg;
473
474 switch (*ssize) {
475 case MMU_SEGSIZE_256M:
476 /* We only have 28 - 23 bits of seg_off in avpn */
477 seg_off = (avpn & 0x1f) << 23;
478 vsid = avpn >> 5;
479 /* We can find more bits from the pteg value */
480 if (shift < 23) {
481 vpi = (vsid ^ pteg) & htab_hash_mask;
482 seg_off |= vpi << shift;
483 }
484 *vpn = vsid << (SID_SHIFT - VPN_SHIFT) | seg_off >> VPN_SHIFT;
485 case MMU_SEGSIZE_1T:
486 /* We only have 40 - 23 bits of seg_off in avpn */
487 seg_off = (avpn & 0x1ffff) << 23;
488 vsid = avpn >> 17;
489 if (shift < 23) {
490 vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask;
491 seg_off |= vpi << shift;
492 }
493 *vpn = vsid << (SID_SHIFT_1T - VPN_SHIFT) | seg_off >> VPN_SHIFT;
494 default:
495 *vpn = size = 0;
496 }
497 *psize = size;
498 *apsize = a_size;
499 }
500
501 /*
502 * clear all mappings on kexec. All cpus are in real mode (or they will
503 * be when they isi), and we are the only one left. We rely on our kernel
504 * mapping being 0xC0's and the hardware ignoring those two real bits.
505 *
506 * TODO: add batching support when enabled. remember, no dynamic memory here,
507 * athough there is the control page available...
508 */
509 static void native_hpte_clear(void)
510 {
511 unsigned long vpn = 0;
512 unsigned long slot, slots, flags;
513 struct hash_pte *hptep = htab_address;
514 unsigned long hpte_v;
515 unsigned long pteg_count;
516 int psize, apsize, ssize;
517
518 pteg_count = htab_hash_mask + 1;
519
520 local_irq_save(flags);
521
522 /* we take the tlbie lock and hold it. Some hardware will
523 * deadlock if we try to tlbie from two processors at once.
524 */
525 raw_spin_lock(&native_tlbie_lock);
526
527 slots = pteg_count * HPTES_PER_GROUP;
528
529 for (slot = 0; slot < slots; slot++, hptep++) {
530 /*
531 * we could lock the pte here, but we are the only cpu
532 * running, right? and for crash dump, we probably
533 * don't want to wait for a maybe bad cpu.
534 */
535 hpte_v = hptep->v;
536
537 /*
538 * Call __tlbie() here rather than tlbie() since we
539 * already hold the native_tlbie_lock.
540 */
541 if (hpte_v & HPTE_V_VALID) {
542 hpte_decode(hptep, slot, &psize, &apsize, &ssize, &vpn);
543 hptep->v = 0;
544 __tlbie(vpn, psize, apsize, ssize);
545 }
546 }
547
548 asm volatile("eieio; tlbsync; ptesync":::"memory");
549 raw_spin_unlock(&native_tlbie_lock);
550 local_irq_restore(flags);
551 }
552
553 /*
554 * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
555 * the lock all the time
556 */
557 static void native_flush_hash_range(unsigned long number, int local)
558 {
559 unsigned long vpn;
560 unsigned long hash, index, hidx, shift, slot;
561 struct hash_pte *hptep;
562 unsigned long hpte_v;
563 unsigned long want_v;
564 unsigned long flags;
565 real_pte_t pte;
566 struct ppc64_tlb_batch *batch = &__get_cpu_var(ppc64_tlb_batch);
567 unsigned long psize = batch->psize;
568 int ssize = batch->ssize;
569 int i;
570
571 local_irq_save(flags);
572
573 for (i = 0; i < number; i++) {
574 vpn = batch->vpn[i];
575 pte = batch->pte[i];
576
577 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
578 hash = hpt_hash(vpn, shift, ssize);
579 hidx = __rpte_to_hidx(pte, index);
580 if (hidx & _PTEIDX_SECONDARY)
581 hash = ~hash;
582 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
583 slot += hidx & _PTEIDX_GROUP_IX;
584 hptep = htab_address + slot;
585 want_v = hpte_encode_avpn(vpn, psize, ssize);
586 native_lock_hpte(hptep);
587 hpte_v = hptep->v;
588 if (!HPTE_V_COMPARE(hpte_v, want_v) ||
589 !(hpte_v & HPTE_V_VALID))
590 native_unlock_hpte(hptep);
591 else
592 hptep->v = 0;
593 } pte_iterate_hashed_end();
594 }
595
596 if (mmu_has_feature(MMU_FTR_TLBIEL) &&
597 mmu_psize_defs[psize].tlbiel && local) {
598 asm volatile("ptesync":::"memory");
599 for (i = 0; i < number; i++) {
600 vpn = batch->vpn[i];
601 pte = batch->pte[i];
602
603 pte_iterate_hashed_subpages(pte, psize,
604 vpn, index, shift) {
605 __tlbiel(vpn, psize, psize, ssize);
606 } pte_iterate_hashed_end();
607 }
608 asm volatile("ptesync":::"memory");
609 } else {
610 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
611
612 if (lock_tlbie)
613 raw_spin_lock(&native_tlbie_lock);
614
615 asm volatile("ptesync":::"memory");
616 for (i = 0; i < number; i++) {
617 vpn = batch->vpn[i];
618 pte = batch->pte[i];
619
620 pte_iterate_hashed_subpages(pte, psize,
621 vpn, index, shift) {
622 __tlbie(vpn, psize, psize, ssize);
623 } pte_iterate_hashed_end();
624 }
625 asm volatile("eieio; tlbsync; ptesync":::"memory");
626
627 if (lock_tlbie)
628 raw_spin_unlock(&native_tlbie_lock);
629 }
630
631 local_irq_restore(flags);
632 }
633
634 void __init hpte_init_native(void)
635 {
636 ppc_md.hpte_invalidate = native_hpte_invalidate;
637 ppc_md.hpte_updatepp = native_hpte_updatepp;
638 ppc_md.hpte_updateboltedpp = native_hpte_updateboltedpp;
639 ppc_md.hpte_insert = native_hpte_insert;
640 ppc_md.hpte_remove = native_hpte_remove;
641 ppc_md.hpte_clear_all = native_hpte_clear;
642 ppc_md.flush_hash_range = native_flush_hash_range;
643 }
This page took 0.073775 seconds and 5 git commands to generate.