[PATCH] ppc64: Store virtual address in TLB flush batches
[deliverable/linux.git] / arch / ppc64 / mm / hash_native.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 #include <linux/spinlock.h>
13 #include <linux/bitops.h>
14 #include <linux/threads.h>
15 #include <linux/smp.h>
16
17 #include <asm/abs_addr.h>
18 #include <asm/machdep.h>
19 #include <asm/mmu.h>
20 #include <asm/mmu_context.h>
21 #include <asm/pgtable.h>
22 #include <asm/tlbflush.h>
23 #include <asm/tlb.h>
24 #include <asm/cputable.h>
25
26 #define HPTE_LOCK_BIT 3
27
28 static DEFINE_SPINLOCK(native_tlbie_lock);
29
30 static inline void native_lock_hpte(hpte_t *hptep)
31 {
32 unsigned long *word = &hptep->v;
33
34 while (1) {
35 if (!test_and_set_bit(HPTE_LOCK_BIT, word))
36 break;
37 while(test_bit(HPTE_LOCK_BIT, word))
38 cpu_relax();
39 }
40 }
41
42 static inline void native_unlock_hpte(hpte_t *hptep)
43 {
44 unsigned long *word = &hptep->v;
45
46 asm volatile("lwsync":::"memory");
47 clear_bit(HPTE_LOCK_BIT, word);
48 }
49
50 long native_hpte_insert(unsigned long hpte_group, unsigned long va,
51 unsigned long prpn, unsigned long vflags,
52 unsigned long rflags)
53 {
54 hpte_t *hptep = htab_address + hpte_group;
55 unsigned long hpte_v, hpte_r;
56 int i;
57
58 for (i = 0; i < HPTES_PER_GROUP; i++) {
59 if (! (hptep->v & HPTE_V_VALID)) {
60 /* retry with lock held */
61 native_lock_hpte(hptep);
62 if (! (hptep->v & HPTE_V_VALID))
63 break;
64 native_unlock_hpte(hptep);
65 }
66
67 hptep++;
68 }
69
70 if (i == HPTES_PER_GROUP)
71 return -1;
72
73 hpte_v = (va >> 23) << HPTE_V_AVPN_SHIFT | vflags | HPTE_V_VALID;
74 if (vflags & HPTE_V_LARGE)
75 va &= ~(1UL << HPTE_V_AVPN_SHIFT);
76 hpte_r = (prpn << HPTE_R_RPN_SHIFT) | rflags;
77
78 hptep->r = hpte_r;
79 /* Guarantee the second dword is visible before the valid bit */
80 __asm__ __volatile__ ("eieio" : : : "memory");
81 /*
82 * Now set the first dword including the valid bit
83 * NOTE: this also unlocks the hpte
84 */
85 hptep->v = hpte_v;
86
87 __asm__ __volatile__ ("ptesync" : : : "memory");
88
89 return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
90 }
91
92 static long native_hpte_remove(unsigned long hpte_group)
93 {
94 hpte_t *hptep;
95 int i;
96 int slot_offset;
97 unsigned long hpte_v;
98
99 /* pick a random entry to start at */
100 slot_offset = mftb() & 0x7;
101
102 for (i = 0; i < HPTES_PER_GROUP; i++) {
103 hptep = htab_address + hpte_group + slot_offset;
104 hpte_v = hptep->v;
105
106 if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
107 /* retry with lock held */
108 native_lock_hpte(hptep);
109 hpte_v = hptep->v;
110 if ((hpte_v & HPTE_V_VALID)
111 && !(hpte_v & HPTE_V_BOLTED))
112 break;
113 native_unlock_hpte(hptep);
114 }
115
116 slot_offset++;
117 slot_offset &= 0x7;
118 }
119
120 if (i == HPTES_PER_GROUP)
121 return -1;
122
123 /* Invalidate the hpte. NOTE: this also unlocks it */
124 hptep->v = 0;
125
126 return i;
127 }
128
129 static inline void set_pp_bit(unsigned long pp, hpte_t *addr)
130 {
131 unsigned long old;
132 unsigned long *p = &addr->r;
133
134 __asm__ __volatile__(
135 "1: ldarx %0,0,%3\n\
136 rldimi %0,%2,0,61\n\
137 stdcx. %0,0,%3\n\
138 bne 1b"
139 : "=&r" (old), "=m" (*p)
140 : "r" (pp), "r" (p), "m" (*p)
141 : "cc");
142 }
143
144 /*
145 * Only works on small pages. Yes its ugly to have to check each slot in
146 * the group but we only use this during bootup.
147 */
148 static long native_hpte_find(unsigned long vpn)
149 {
150 hpte_t *hptep;
151 unsigned long hash;
152 unsigned long i, j;
153 long slot;
154 unsigned long hpte_v;
155
156 hash = hpt_hash(vpn, 0);
157
158 for (j = 0; j < 2; j++) {
159 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
160 for (i = 0; i < HPTES_PER_GROUP; i++) {
161 hptep = htab_address + slot;
162 hpte_v = hptep->v;
163
164 if ((HPTE_V_AVPN_VAL(hpte_v) == (vpn >> 11))
165 && (hpte_v & HPTE_V_VALID)
166 && ( !!(hpte_v & HPTE_V_SECONDARY) == j)) {
167 /* HPTE matches */
168 if (j)
169 slot = -slot;
170 return slot;
171 }
172 ++slot;
173 }
174 hash = ~hash;
175 }
176
177 return -1;
178 }
179
180 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
181 unsigned long va, int large, int local)
182 {
183 hpte_t *hptep = htab_address + slot;
184 unsigned long hpte_v;
185 unsigned long avpn = va >> 23;
186 int ret = 0;
187
188 if (large)
189 avpn &= ~1;
190
191 native_lock_hpte(hptep);
192
193 hpte_v = hptep->v;
194
195 /* Even if we miss, we need to invalidate the TLB */
196 if ((HPTE_V_AVPN_VAL(hpte_v) != avpn)
197 || !(hpte_v & HPTE_V_VALID)) {
198 native_unlock_hpte(hptep);
199 ret = -1;
200 } else {
201 set_pp_bit(newpp, hptep);
202 native_unlock_hpte(hptep);
203 }
204
205 /* Ensure it is out of the tlb too */
206 if (cpu_has_feature(CPU_FTR_TLBIEL) && !large && local) {
207 tlbiel(va);
208 } else {
209 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
210
211 if (lock_tlbie)
212 spin_lock(&native_tlbie_lock);
213 tlbie(va, large);
214 if (lock_tlbie)
215 spin_unlock(&native_tlbie_lock);
216 }
217
218 return ret;
219 }
220
221 /*
222 * Update the page protection bits. Intended to be used to create
223 * guard pages for kernel data structures on pages which are bolted
224 * in the HPT. Assumes pages being operated on will not be stolen.
225 * Does not work on large pages.
226 *
227 * No need to lock here because we should be the only user.
228 */
229 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea)
230 {
231 unsigned long vsid, va, vpn, flags = 0;
232 long slot;
233 hpte_t *hptep;
234 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
235
236 vsid = get_kernel_vsid(ea);
237 va = (vsid << 28) | (ea & 0x0fffffff);
238 vpn = va >> PAGE_SHIFT;
239
240 slot = native_hpte_find(vpn);
241 if (slot == -1)
242 panic("could not find page to bolt\n");
243 hptep = htab_address + slot;
244
245 set_pp_bit(newpp, hptep);
246
247 /* Ensure it is out of the tlb too */
248 if (lock_tlbie)
249 spin_lock_irqsave(&native_tlbie_lock, flags);
250 tlbie(va, 0);
251 if (lock_tlbie)
252 spin_unlock_irqrestore(&native_tlbie_lock, flags);
253 }
254
255 static void native_hpte_invalidate(unsigned long slot, unsigned long va,
256 int large, int local)
257 {
258 hpte_t *hptep = htab_address + slot;
259 unsigned long hpte_v;
260 unsigned long avpn = va >> 23;
261 unsigned long flags;
262 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
263
264 if (large)
265 avpn &= ~1;
266
267 local_irq_save(flags);
268 native_lock_hpte(hptep);
269
270 hpte_v = hptep->v;
271
272 /* Even if we miss, we need to invalidate the TLB */
273 if ((HPTE_V_AVPN_VAL(hpte_v) != avpn)
274 || !(hpte_v & HPTE_V_VALID)) {
275 native_unlock_hpte(hptep);
276 } else {
277 /* Invalidate the hpte. NOTE: this also unlocks it */
278 hptep->v = 0;
279 }
280
281 /* Invalidate the tlb */
282 if (cpu_has_feature(CPU_FTR_TLBIEL) && !large && local) {
283 tlbiel(va);
284 } else {
285 if (lock_tlbie)
286 spin_lock(&native_tlbie_lock);
287 tlbie(va, large);
288 if (lock_tlbie)
289 spin_unlock(&native_tlbie_lock);
290 }
291 local_irq_restore(flags);
292 }
293
294 /*
295 * clear all mappings on kexec. All cpus are in real mode (or they will
296 * be when they isi), and we are the only one left. We rely on our kernel
297 * mapping being 0xC0's and the hardware ignoring those two real bits.
298 *
299 * TODO: add batching support when enabled. remember, no dynamic memory here,
300 * athough there is the control page available...
301 */
302 static void native_hpte_clear(void)
303 {
304 unsigned long slot, slots, flags;
305 hpte_t *hptep = htab_address;
306 unsigned long hpte_v;
307 unsigned long pteg_count;
308
309 pteg_count = htab_hash_mask + 1;
310
311 local_irq_save(flags);
312
313 /* we take the tlbie lock and hold it. Some hardware will
314 * deadlock if we try to tlbie from two processors at once.
315 */
316 spin_lock(&native_tlbie_lock);
317
318 slots = pteg_count * HPTES_PER_GROUP;
319
320 for (slot = 0; slot < slots; slot++, hptep++) {
321 /*
322 * we could lock the pte here, but we are the only cpu
323 * running, right? and for crash dump, we probably
324 * don't want to wait for a maybe bad cpu.
325 */
326 hpte_v = hptep->v;
327
328 if (hpte_v & HPTE_V_VALID) {
329 hptep->v = 0;
330 tlbie(slot2va(hpte_v, slot), hpte_v & HPTE_V_LARGE);
331 }
332 }
333
334 spin_unlock(&native_tlbie_lock);
335 local_irq_restore(flags);
336 }
337
338 static void native_flush_hash_range(unsigned long number, int local)
339 {
340 unsigned long va, vpn, hash, secondary, slot, flags, avpn;
341 int i, j;
342 hpte_t *hptep;
343 unsigned long hpte_v;
344 struct ppc64_tlb_batch *batch = &__get_cpu_var(ppc64_tlb_batch);
345
346 /* XXX fix for large ptes */
347 unsigned long large = 0;
348
349 local_irq_save(flags);
350
351 j = 0;
352 for (i = 0; i < number; i++) {
353 va = batch->vaddr[j];
354 if (large)
355 vpn = va >> HPAGE_SHIFT;
356 else
357 vpn = va >> PAGE_SHIFT;
358 hash = hpt_hash(vpn, large);
359 secondary = (pte_val(batch->pte[i]) & _PAGE_SECONDARY) >> 15;
360 if (secondary)
361 hash = ~hash;
362 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
363 slot += (pte_val(batch->pte[i]) & _PAGE_GROUP_IX) >> 12;
364
365 hptep = htab_address + slot;
366
367 avpn = va >> 23;
368 if (large)
369 avpn &= ~0x1UL;
370
371 native_lock_hpte(hptep);
372
373 hpte_v = hptep->v;
374
375 /* Even if we miss, we need to invalidate the TLB */
376 if ((HPTE_V_AVPN_VAL(hpte_v) != avpn)
377 || !(hpte_v & HPTE_V_VALID)) {
378 native_unlock_hpte(hptep);
379 } else {
380 /* Invalidate the hpte. NOTE: this also unlocks it */
381 hptep->v = 0;
382 }
383
384 j++;
385 }
386
387 if (cpu_has_feature(CPU_FTR_TLBIEL) && !large && local) {
388 asm volatile("ptesync":::"memory");
389
390 for (i = 0; i < j; i++)
391 __tlbiel(batch->vaddr[i]);
392
393 asm volatile("ptesync":::"memory");
394 } else {
395 int lock_tlbie = !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE);
396
397 if (lock_tlbie)
398 spin_lock(&native_tlbie_lock);
399
400 asm volatile("ptesync":::"memory");
401
402 for (i = 0; i < j; i++)
403 __tlbie(batch->vaddr[i], 0);
404
405 asm volatile("eieio; tlbsync; ptesync":::"memory");
406
407 if (lock_tlbie)
408 spin_unlock(&native_tlbie_lock);
409 }
410
411 local_irq_restore(flags);
412 }
413
414 #ifdef CONFIG_PPC_PSERIES
415 /* Disable TLB batching on nighthawk */
416 static inline int tlb_batching_enabled(void)
417 {
418 struct device_node *root = of_find_node_by_path("/");
419 int enabled = 1;
420
421 if (root) {
422 const char *model = get_property(root, "model", NULL);
423 if (model && !strcmp(model, "IBM,9076-N81"))
424 enabled = 0;
425 of_node_put(root);
426 }
427
428 return enabled;
429 }
430 #else
431 static inline int tlb_batching_enabled(void)
432 {
433 return 1;
434 }
435 #endif
436
437 void hpte_init_native(void)
438 {
439 ppc_md.hpte_invalidate = native_hpte_invalidate;
440 ppc_md.hpte_updatepp = native_hpte_updatepp;
441 ppc_md.hpte_updateboltedpp = native_hpte_updateboltedpp;
442 ppc_md.hpte_insert = native_hpte_insert;
443 ppc_md.hpte_remove = native_hpte_remove;
444 ppc_md.hpte_clear_all = native_hpte_clear;
445 if (tlb_batching_enabled())
446 ppc_md.flush_hash_range = native_flush_hash_range;
447 htab_finish_init();
448 }
This page took 0.039983 seconds and 5 git commands to generate.