Merge branch 'hwmon-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jdelv...
[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/threads.h>
18 #include <linux/smp.h>
19
20 #include <asm/abs_addr.h>
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 static DEFINE_SPINLOCK(native_tlbie_lock);
41
42 static inline void __tlbie(unsigned long va, int psize, int ssize)
43 {
44 unsigned int penc;
45
46 /* clear top 16 bits, non SLS segment */
47 va &= ~(0xffffULL << 48);
48
49 switch (psize) {
50 case MMU_PAGE_4K:
51 va &= ~0xffful;
52 va |= ssize << 8;
53 asm volatile(ASM_MMU_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0),
54 %2)
55 : : "r" (va), "r"(0), "i" (MMU_FTR_TLBIE_206)
56 : "memory");
57 break;
58 default:
59 penc = mmu_psize_defs[psize].penc;
60 va &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
61 va |= penc << 12;
62 va |= ssize << 8;
63 va |= 1; /* L */
64 asm volatile(ASM_MMU_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0),
65 %2)
66 : : "r" (va), "r"(0), "i" (MMU_FTR_TLBIE_206)
67 : "memory");
68 break;
69 }
70 }
71
72 static inline void __tlbiel(unsigned long va, int psize, int ssize)
73 {
74 unsigned int penc;
75
76 /* clear top 16 bits, non SLS segment */
77 va &= ~(0xffffULL << 48);
78
79 switch (psize) {
80 case MMU_PAGE_4K:
81 va &= ~0xffful;
82 va |= ssize << 8;
83 asm volatile(".long 0x7c000224 | (%0 << 11) | (0 << 21)"
84 : : "r"(va) : "memory");
85 break;
86 default:
87 penc = mmu_psize_defs[psize].penc;
88 va &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
89 va |= penc << 12;
90 va |= ssize << 8;
91 va |= 1; /* L */
92 asm volatile(".long 0x7c000224 | (%0 << 11) | (1 << 21)"
93 : : "r"(va) : "memory");
94 break;
95 }
96
97 }
98
99 static inline void tlbie(unsigned long va, int psize, int ssize, int local)
100 {
101 unsigned int use_local = local && cpu_has_feature(CPU_FTR_TLBIEL);
102 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
103
104 if (use_local)
105 use_local = mmu_psize_defs[psize].tlbiel;
106 if (lock_tlbie && !use_local)
107 spin_lock(&native_tlbie_lock);
108 asm volatile("ptesync": : :"memory");
109 if (use_local) {
110 __tlbiel(va, psize, ssize);
111 asm volatile("ptesync": : :"memory");
112 } else {
113 __tlbie(va, psize, ssize);
114 asm volatile("eieio; tlbsync; ptesync": : :"memory");
115 }
116 if (lock_tlbie && !use_local)
117 spin_unlock(&native_tlbie_lock);
118 }
119
120 static inline void native_lock_hpte(struct hash_pte *hptep)
121 {
122 unsigned long *word = &hptep->v;
123
124 while (1) {
125 if (!test_and_set_bit(HPTE_LOCK_BIT, word))
126 break;
127 while(test_bit(HPTE_LOCK_BIT, word))
128 cpu_relax();
129 }
130 }
131
132 static inline void native_unlock_hpte(struct hash_pte *hptep)
133 {
134 unsigned long *word = &hptep->v;
135
136 asm volatile("lwsync":::"memory");
137 clear_bit(HPTE_LOCK_BIT, word);
138 }
139
140 static long native_hpte_insert(unsigned long hpte_group, unsigned long va,
141 unsigned long pa, unsigned long rflags,
142 unsigned long vflags, int psize, int ssize)
143 {
144 struct hash_pte *hptep = htab_address + hpte_group;
145 unsigned long hpte_v, hpte_r;
146 int i;
147
148 if (!(vflags & HPTE_V_BOLTED)) {
149 DBG_LOW(" insert(group=%lx, va=%016lx, pa=%016lx,"
150 " rflags=%lx, vflags=%lx, psize=%d)\n",
151 hpte_group, va, pa, rflags, vflags, psize);
152 }
153
154 for (i = 0; i < HPTES_PER_GROUP; i++) {
155 if (! (hptep->v & HPTE_V_VALID)) {
156 /* retry with lock held */
157 native_lock_hpte(hptep);
158 if (! (hptep->v & HPTE_V_VALID))
159 break;
160 native_unlock_hpte(hptep);
161 }
162
163 hptep++;
164 }
165
166 if (i == HPTES_PER_GROUP)
167 return -1;
168
169 hpte_v = hpte_encode_v(va, psize, ssize) | vflags | HPTE_V_VALID;
170 hpte_r = hpte_encode_r(pa, psize) | rflags;
171
172 if (!(vflags & HPTE_V_BOLTED)) {
173 DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
174 i, hpte_v, hpte_r);
175 }
176
177 hptep->r = hpte_r;
178 /* Guarantee the second dword is visible before the valid bit */
179 eieio();
180 /*
181 * Now set the first dword including the valid bit
182 * NOTE: this also unlocks the hpte
183 */
184 hptep->v = hpte_v;
185
186 __asm__ __volatile__ ("ptesync" : : : "memory");
187
188 return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
189 }
190
191 static long native_hpte_remove(unsigned long hpte_group)
192 {
193 struct hash_pte *hptep;
194 int i;
195 int slot_offset;
196 unsigned long hpte_v;
197
198 DBG_LOW(" remove(group=%lx)\n", hpte_group);
199
200 /* pick a random entry to start at */
201 slot_offset = mftb() & 0x7;
202
203 for (i = 0; i < HPTES_PER_GROUP; i++) {
204 hptep = htab_address + hpte_group + slot_offset;
205 hpte_v = hptep->v;
206
207 if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
208 /* retry with lock held */
209 native_lock_hpte(hptep);
210 hpte_v = hptep->v;
211 if ((hpte_v & HPTE_V_VALID)
212 && !(hpte_v & HPTE_V_BOLTED))
213 break;
214 native_unlock_hpte(hptep);
215 }
216
217 slot_offset++;
218 slot_offset &= 0x7;
219 }
220
221 if (i == HPTES_PER_GROUP)
222 return -1;
223
224 /* Invalidate the hpte. NOTE: this also unlocks it */
225 hptep->v = 0;
226
227 return i;
228 }
229
230 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
231 unsigned long va, int psize, int ssize,
232 int local)
233 {
234 struct hash_pte *hptep = htab_address + slot;
235 unsigned long hpte_v, want_v;
236 int ret = 0;
237
238 want_v = hpte_encode_v(va, psize, ssize);
239
240 DBG_LOW(" update(va=%016lx, avpnv=%016lx, hash=%016lx, newpp=%x)",
241 va, want_v & HPTE_V_AVPN, slot, newpp);
242
243 native_lock_hpte(hptep);
244
245 hpte_v = hptep->v;
246
247 /* Even if we miss, we need to invalidate the TLB */
248 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) {
249 DBG_LOW(" -> miss\n");
250 ret = -1;
251 } else {
252 DBG_LOW(" -> hit\n");
253 /* Update the HPTE */
254 hptep->r = (hptep->r & ~(HPTE_R_PP | HPTE_R_N)) |
255 (newpp & (HPTE_R_PP | HPTE_R_N | HPTE_R_C));
256 }
257 native_unlock_hpte(hptep);
258
259 /* Ensure it is out of the tlb too. */
260 tlbie(va, psize, ssize, local);
261
262 return ret;
263 }
264
265 static long native_hpte_find(unsigned long va, int psize, int ssize)
266 {
267 struct hash_pte *hptep;
268 unsigned long hash;
269 unsigned long i;
270 long slot;
271 unsigned long want_v, hpte_v;
272
273 hash = hpt_hash(va, mmu_psize_defs[psize].shift, ssize);
274 want_v = hpte_encode_v(va, psize, ssize);
275
276 /* Bolted mappings are only ever in the primary group */
277 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
278 for (i = 0; i < HPTES_PER_GROUP; i++) {
279 hptep = htab_address + slot;
280 hpte_v = hptep->v;
281
282 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
283 /* HPTE matches */
284 return slot;
285 ++slot;
286 }
287
288 return -1;
289 }
290
291 /*
292 * Update the page protection bits. Intended to be used to create
293 * guard pages for kernel data structures on pages which are bolted
294 * in the HPT. Assumes pages being operated on will not be stolen.
295 *
296 * No need to lock here because we should be the only user.
297 */
298 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea,
299 int psize, int ssize)
300 {
301 unsigned long vsid, va;
302 long slot;
303 struct hash_pte *hptep;
304
305 vsid = get_kernel_vsid(ea, ssize);
306 va = hpt_va(ea, vsid, ssize);
307
308 slot = native_hpte_find(va, psize, ssize);
309 if (slot == -1)
310 panic("could not find page to bolt\n");
311 hptep = htab_address + slot;
312
313 /* Update the HPTE */
314 hptep->r = (hptep->r & ~(HPTE_R_PP | HPTE_R_N)) |
315 (newpp & (HPTE_R_PP | HPTE_R_N));
316
317 /* Ensure it is out of the tlb too. */
318 tlbie(va, psize, ssize, 0);
319 }
320
321 static void native_hpte_invalidate(unsigned long slot, unsigned long va,
322 int psize, int ssize, int local)
323 {
324 struct hash_pte *hptep = htab_address + slot;
325 unsigned long hpte_v;
326 unsigned long want_v;
327 unsigned long flags;
328
329 local_irq_save(flags);
330
331 DBG_LOW(" invalidate(va=%016lx, hash: %x)\n", va, slot);
332
333 want_v = hpte_encode_v(va, psize, ssize);
334 native_lock_hpte(hptep);
335 hpte_v = hptep->v;
336
337 /* Even if we miss, we need to invalidate the TLB */
338 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
339 native_unlock_hpte(hptep);
340 else
341 /* Invalidate the hpte. NOTE: this also unlocks it */
342 hptep->v = 0;
343
344 /* Invalidate the TLB */
345 tlbie(va, psize, ssize, local);
346
347 local_irq_restore(flags);
348 }
349
350 #define LP_SHIFT 12
351 #define LP_BITS 8
352 #define LP_MASK(i) ((0xFF >> (i)) << LP_SHIFT)
353
354 static void hpte_decode(struct hash_pte *hpte, unsigned long slot,
355 int *psize, int *ssize, unsigned long *va)
356 {
357 unsigned long hpte_r = hpte->r;
358 unsigned long hpte_v = hpte->v;
359 unsigned long avpn;
360 int i, size, shift, penc;
361
362 if (!(hpte_v & HPTE_V_LARGE))
363 size = MMU_PAGE_4K;
364 else {
365 for (i = 0; i < LP_BITS; i++) {
366 if ((hpte_r & LP_MASK(i+1)) == LP_MASK(i+1))
367 break;
368 }
369 penc = LP_MASK(i+1) >> LP_SHIFT;
370 for (size = 0; size < MMU_PAGE_COUNT; size++) {
371
372 /* 4K pages are not represented by LP */
373 if (size == MMU_PAGE_4K)
374 continue;
375
376 /* valid entries have a shift value */
377 if (!mmu_psize_defs[size].shift)
378 continue;
379
380 if (penc == mmu_psize_defs[size].penc)
381 break;
382 }
383 }
384
385 /* This works for all page sizes, and for 256M and 1T segments */
386 shift = mmu_psize_defs[size].shift;
387 avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm) << 23;
388
389 if (shift < 23) {
390 unsigned long vpi, vsid, pteg;
391
392 pteg = slot / HPTES_PER_GROUP;
393 if (hpte_v & HPTE_V_SECONDARY)
394 pteg = ~pteg;
395 switch (hpte_v >> HPTE_V_SSIZE_SHIFT) {
396 case MMU_SEGSIZE_256M:
397 vpi = ((avpn >> 28) ^ pteg) & htab_hash_mask;
398 break;
399 case MMU_SEGSIZE_1T:
400 vsid = avpn >> 40;
401 vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask;
402 break;
403 default:
404 avpn = vpi = size = 0;
405 }
406 avpn |= (vpi << mmu_psize_defs[size].shift);
407 }
408
409 *va = avpn;
410 *psize = size;
411 *ssize = hpte_v >> HPTE_V_SSIZE_SHIFT;
412 }
413
414 /*
415 * clear all mappings on kexec. All cpus are in real mode (or they will
416 * be when they isi), and we are the only one left. We rely on our kernel
417 * mapping being 0xC0's and the hardware ignoring those two real bits.
418 *
419 * TODO: add batching support when enabled. remember, no dynamic memory here,
420 * athough there is the control page available...
421 */
422 static void native_hpte_clear(void)
423 {
424 unsigned long slot, slots, flags;
425 struct hash_pte *hptep = htab_address;
426 unsigned long hpte_v, va;
427 unsigned long pteg_count;
428 int psize, ssize;
429
430 pteg_count = htab_hash_mask + 1;
431
432 local_irq_save(flags);
433
434 /* we take the tlbie lock and hold it. Some hardware will
435 * deadlock if we try to tlbie from two processors at once.
436 */
437 spin_lock(&native_tlbie_lock);
438
439 slots = pteg_count * HPTES_PER_GROUP;
440
441 for (slot = 0; slot < slots; slot++, hptep++) {
442 /*
443 * we could lock the pte here, but we are the only cpu
444 * running, right? and for crash dump, we probably
445 * don't want to wait for a maybe bad cpu.
446 */
447 hpte_v = hptep->v;
448
449 /*
450 * Call __tlbie() here rather than tlbie() since we
451 * already hold the native_tlbie_lock.
452 */
453 if (hpte_v & HPTE_V_VALID) {
454 hpte_decode(hptep, slot, &psize, &ssize, &va);
455 hptep->v = 0;
456 __tlbie(va, psize, ssize);
457 }
458 }
459
460 asm volatile("eieio; tlbsync; ptesync":::"memory");
461 spin_unlock(&native_tlbie_lock);
462 local_irq_restore(flags);
463 }
464
465 /*
466 * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
467 * the lock all the time
468 */
469 static void native_flush_hash_range(unsigned long number, int local)
470 {
471 unsigned long va, hash, index, hidx, shift, slot;
472 struct hash_pte *hptep;
473 unsigned long hpte_v;
474 unsigned long want_v;
475 unsigned long flags;
476 real_pte_t pte;
477 struct ppc64_tlb_batch *batch = &__get_cpu_var(ppc64_tlb_batch);
478 unsigned long psize = batch->psize;
479 int ssize = batch->ssize;
480 int i;
481
482 local_irq_save(flags);
483
484 for (i = 0; i < number; i++) {
485 va = batch->vaddr[i];
486 pte = batch->pte[i];
487
488 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
489 hash = hpt_hash(va, shift, ssize);
490 hidx = __rpte_to_hidx(pte, index);
491 if (hidx & _PTEIDX_SECONDARY)
492 hash = ~hash;
493 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
494 slot += hidx & _PTEIDX_GROUP_IX;
495 hptep = htab_address + slot;
496 want_v = hpte_encode_v(va, psize, ssize);
497 native_lock_hpte(hptep);
498 hpte_v = hptep->v;
499 if (!HPTE_V_COMPARE(hpte_v, want_v) ||
500 !(hpte_v & HPTE_V_VALID))
501 native_unlock_hpte(hptep);
502 else
503 hptep->v = 0;
504 } pte_iterate_hashed_end();
505 }
506
507 if (cpu_has_feature(CPU_FTR_TLBIEL) &&
508 mmu_psize_defs[psize].tlbiel && local) {
509 asm volatile("ptesync":::"memory");
510 for (i = 0; i < number; i++) {
511 va = batch->vaddr[i];
512 pte = batch->pte[i];
513
514 pte_iterate_hashed_subpages(pte, psize, va, index,
515 shift) {
516 __tlbiel(va, psize, ssize);
517 } pte_iterate_hashed_end();
518 }
519 asm volatile("ptesync":::"memory");
520 } else {
521 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
522
523 if (lock_tlbie)
524 spin_lock(&native_tlbie_lock);
525
526 asm volatile("ptesync":::"memory");
527 for (i = 0; i < number; i++) {
528 va = batch->vaddr[i];
529 pte = batch->pte[i];
530
531 pte_iterate_hashed_subpages(pte, psize, va, index,
532 shift) {
533 __tlbie(va, psize, ssize);
534 } pte_iterate_hashed_end();
535 }
536 asm volatile("eieio; tlbsync; ptesync":::"memory");
537
538 if (lock_tlbie)
539 spin_unlock(&native_tlbie_lock);
540 }
541
542 local_irq_restore(flags);
543 }
544
545 #ifdef CONFIG_PPC_PSERIES
546 /* Disable TLB batching on nighthawk */
547 static inline int tlb_batching_enabled(void)
548 {
549 struct device_node *root = of_find_node_by_path("/");
550 int enabled = 1;
551
552 if (root) {
553 const char *model = of_get_property(root, "model", NULL);
554 if (model && !strcmp(model, "IBM,9076-N81"))
555 enabled = 0;
556 of_node_put(root);
557 }
558
559 return enabled;
560 }
561 #else
562 static inline int tlb_batching_enabled(void)
563 {
564 return 1;
565 }
566 #endif
567
568 void __init hpte_init_native(void)
569 {
570 ppc_md.hpte_invalidate = native_hpte_invalidate;
571 ppc_md.hpte_updatepp = native_hpte_updatepp;
572 ppc_md.hpte_updateboltedpp = native_hpte_updateboltedpp;
573 ppc_md.hpte_insert = native_hpte_insert;
574 ppc_md.hpte_remove = native_hpte_remove;
575 ppc_md.hpte_clear_all = native_hpte_clear;
576 if (tlb_batching_enabled())
577 ppc_md.flush_hash_range = native_flush_hash_range;
578 }
This page took 0.04369 seconds and 5 git commands to generate.