2 * native hashtable management.
4 * SMP scalability work:
5 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
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.
15 #include <linux/spinlock.h>
16 #include <linux/bitops.h>
17 #include <linux/threads.h>
18 #include <linux/smp.h>
20 #include <asm/abs_addr.h>
21 #include <asm/machdep.h>
23 #include <asm/mmu_context.h>
24 #include <asm/pgtable.h>
25 #include <asm/tlbflush.h>
27 #include <asm/cputable.h>
29 #include <asm/kexec.h>
32 #define DBG_LOW(fmt...) udbg_printf(fmt)
34 #define DBG_LOW(fmt...)
37 #define HPTE_LOCK_BIT 3
39 static DEFINE_SPINLOCK(native_tlbie_lock
);
41 static inline void __tlbie(unsigned long va
, unsigned int psize
)
45 /* clear top 16 bits, non SLS segment */
46 va
&= ~(0xffffULL
<< 48);
51 asm volatile("tlbie %0,0" : : "r" (va
) : "memory");
54 penc
= mmu_psize_defs
[psize
].penc
;
55 va
&= ~((1ul << mmu_psize_defs
[psize
].shift
) - 1);
57 asm volatile("tlbie %0,1" : : "r" (va
) : "memory");
62 static inline void __tlbiel(unsigned long va
, unsigned int psize
)
66 /* clear top 16 bits, non SLS segment */
67 va
&= ~(0xffffULL
<< 48);
72 asm volatile(".long 0x7c000224 | (%0 << 11) | (0 << 21)"
73 : : "r"(va
) : "memory");
76 penc
= mmu_psize_defs
[psize
].penc
;
77 va
&= ~((1ul << mmu_psize_defs
[psize
].shift
) - 1);
79 asm volatile(".long 0x7c000224 | (%0 << 11) | (1 << 21)"
80 : : "r"(va
) : "memory");
86 static inline void tlbie(unsigned long va
, int psize
, int local
)
88 unsigned int use_local
= local
&& cpu_has_feature(CPU_FTR_TLBIEL
);
89 int lock_tlbie
= !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE
);
92 use_local
= mmu_psize_defs
[psize
].tlbiel
;
93 if (lock_tlbie
&& !use_local
)
94 spin_lock(&native_tlbie_lock
);
95 asm volatile("ptesync": : :"memory");
98 asm volatile("ptesync": : :"memory");
101 asm volatile("eieio; tlbsync; ptesync": : :"memory");
103 if (lock_tlbie
&& !use_local
)
104 spin_unlock(&native_tlbie_lock
);
107 static inline void native_lock_hpte(hpte_t
*hptep
)
109 unsigned long *word
= &hptep
->v
;
112 if (!test_and_set_bit(HPTE_LOCK_BIT
, word
))
114 while(test_bit(HPTE_LOCK_BIT
, word
))
119 static inline void native_unlock_hpte(hpte_t
*hptep
)
121 unsigned long *word
= &hptep
->v
;
123 asm volatile("lwsync":::"memory");
124 clear_bit(HPTE_LOCK_BIT
, word
);
127 static long native_hpte_insert(unsigned long hpte_group
, unsigned long va
,
128 unsigned long pa
, unsigned long rflags
,
129 unsigned long vflags
, int psize
)
131 hpte_t
*hptep
= htab_address
+ hpte_group
;
132 unsigned long hpte_v
, hpte_r
;
135 if (!(vflags
& HPTE_V_BOLTED
)) {
136 DBG_LOW(" insert(group=%lx, va=%016lx, pa=%016lx,"
137 " rflags=%lx, vflags=%lx, psize=%d)\n",
138 hpte_group
, va
, pa
, rflags
, vflags
, psize
);
141 for (i
= 0; i
< HPTES_PER_GROUP
; i
++) {
142 if (! (hptep
->v
& HPTE_V_VALID
)) {
143 /* retry with lock held */
144 native_lock_hpte(hptep
);
145 if (! (hptep
->v
& HPTE_V_VALID
))
147 native_unlock_hpte(hptep
);
153 if (i
== HPTES_PER_GROUP
)
156 hpte_v
= hpte_encode_v(va
, psize
) | vflags
| HPTE_V_VALID
;
157 hpte_r
= hpte_encode_r(pa
, psize
) | rflags
;
159 if (!(vflags
& HPTE_V_BOLTED
)) {
160 DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
165 /* Guarantee the second dword is visible before the valid bit */
166 __asm__
__volatile__ ("eieio" : : : "memory");
168 * Now set the first dword including the valid bit
169 * NOTE: this also unlocks the hpte
173 __asm__
__volatile__ ("ptesync" : : : "memory");
175 return i
| (!!(vflags
& HPTE_V_SECONDARY
) << 3);
178 static long native_hpte_remove(unsigned long hpte_group
)
183 unsigned long hpte_v
;
185 DBG_LOW(" remove(group=%lx)\n", hpte_group
);
187 /* pick a random entry to start at */
188 slot_offset
= mftb() & 0x7;
190 for (i
= 0; i
< HPTES_PER_GROUP
; i
++) {
191 hptep
= htab_address
+ hpte_group
+ slot_offset
;
194 if ((hpte_v
& HPTE_V_VALID
) && !(hpte_v
& HPTE_V_BOLTED
)) {
195 /* retry with lock held */
196 native_lock_hpte(hptep
);
198 if ((hpte_v
& HPTE_V_VALID
)
199 && !(hpte_v
& HPTE_V_BOLTED
))
201 native_unlock_hpte(hptep
);
208 if (i
== HPTES_PER_GROUP
)
211 /* Invalidate the hpte. NOTE: this also unlocks it */
217 static long native_hpte_updatepp(unsigned long slot
, unsigned long newpp
,
218 unsigned long va
, int psize
, int local
)
220 hpte_t
*hptep
= htab_address
+ slot
;
221 unsigned long hpte_v
, want_v
;
224 want_v
= hpte_encode_v(va
, psize
);
226 DBG_LOW(" update(va=%016lx, avpnv=%016lx, hash=%016lx, newpp=%x)",
227 va
, want_v
& HPTE_V_AVPN
, slot
, newpp
);
229 native_lock_hpte(hptep
);
233 /* Even if we miss, we need to invalidate the TLB */
234 if (!HPTE_V_COMPARE(hpte_v
, want_v
) || !(hpte_v
& HPTE_V_VALID
)) {
235 DBG_LOW(" -> miss\n");
236 native_unlock_hpte(hptep
);
239 DBG_LOW(" -> hit\n");
240 /* Update the HPTE */
241 hptep
->r
= (hptep
->r
& ~(HPTE_R_PP
| HPTE_R_N
)) |
242 (newpp
& (HPTE_R_PP
| HPTE_R_N
| HPTE_R_C
));
243 native_unlock_hpte(hptep
);
246 /* Ensure it is out of the tlb too. */
247 tlbie(va
, psize
, local
);
252 static long native_hpte_find(unsigned long va
, int psize
)
258 unsigned long want_v
, hpte_v
;
260 hash
= hpt_hash(va
, mmu_psize_defs
[psize
].shift
);
261 want_v
= hpte_encode_v(va
, psize
);
263 for (j
= 0; j
< 2; j
++) {
264 slot
= (hash
& htab_hash_mask
) * HPTES_PER_GROUP
;
265 for (i
= 0; i
< HPTES_PER_GROUP
; i
++) {
266 hptep
= htab_address
+ slot
;
269 if (HPTE_V_COMPARE(hpte_v
, want_v
)
270 && (hpte_v
& HPTE_V_VALID
)
271 && ( !!(hpte_v
& HPTE_V_SECONDARY
) == j
)) {
286 * Update the page protection bits. Intended to be used to create
287 * guard pages for kernel data structures on pages which are bolted
288 * in the HPT. Assumes pages being operated on will not be stolen.
290 * No need to lock here because we should be the only user.
292 static void native_hpte_updateboltedpp(unsigned long newpp
, unsigned long ea
,
295 unsigned long vsid
, va
;
299 vsid
= get_kernel_vsid(ea
);
300 va
= (vsid
<< 28) | (ea
& 0x0fffffff);
302 slot
= native_hpte_find(va
, psize
);
304 panic("could not find page to bolt\n");
305 hptep
= htab_address
+ slot
;
307 /* Update the HPTE */
308 hptep
->r
= (hptep
->r
& ~(HPTE_R_PP
| HPTE_R_N
)) |
309 (newpp
& (HPTE_R_PP
| HPTE_R_N
));
311 /* Ensure it is out of the tlb too. */
315 static void native_hpte_invalidate(unsigned long slot
, unsigned long va
,
316 int psize
, int local
)
318 hpte_t
*hptep
= htab_address
+ slot
;
319 unsigned long hpte_v
;
320 unsigned long want_v
;
323 local_irq_save(flags
);
325 DBG_LOW(" invalidate(va=%016lx, hash: %x)\n", va
, slot
);
327 want_v
= hpte_encode_v(va
, psize
);
328 native_lock_hpte(hptep
);
331 /* Even if we miss, we need to invalidate the TLB */
332 if (!HPTE_V_COMPARE(hpte_v
, want_v
) || !(hpte_v
& HPTE_V_VALID
))
333 native_unlock_hpte(hptep
);
335 /* Invalidate the hpte. NOTE: this also unlocks it */
338 /* Invalidate the TLB */
339 tlbie(va
, psize
, local
);
341 local_irq_restore(flags
);
346 #define LP_MASK(i) ((0xFF >> (i)) << LP_SHIFT)
348 static void hpte_decode(hpte_t
*hpte
, unsigned long slot
,
349 int *psize
, unsigned long *va
)
351 unsigned long hpte_r
= hpte
->r
;
352 unsigned long hpte_v
= hpte
->v
;
354 int i
, size
, shift
, penc
, avpnm_bits
;
356 if (!(hpte_v
& HPTE_V_LARGE
))
359 for (i
= 0; i
< LP_BITS
; i
++) {
360 if ((hpte_r
& LP_MASK(i
+1)) == LP_MASK(i
+1))
363 penc
= LP_MASK(i
+1) >> LP_SHIFT
;
364 for (size
= 0; size
< MMU_PAGE_COUNT
; size
++) {
366 /* 4K pages are not represented by LP */
367 if (size
== MMU_PAGE_4K
)
370 /* valid entries have a shift value */
371 if (!mmu_psize_defs
[size
].shift
)
374 if (penc
== mmu_psize_defs
[size
].penc
)
380 * FIXME, the code below works for 16M, 64K, and 4K pages as these
381 * fall under the p<=23 rules for calculating the virtual address.
382 * In the case of 16M pages, an extra bit is stolen from the AVPN
383 * field to achieve the requisite 24 bits.
385 * Does not work for 16G pages or 1 TB segments.
387 shift
= mmu_psize_defs
[size
].shift
;
388 if (mmu_psize_defs
[size
].avpnm
)
389 avpnm_bits
= __ilog2_u64(mmu_psize_defs
[size
].avpnm
) + 1;
392 if (shift
- avpnm_bits
<= 23) {
393 avpn
= HPTE_V_AVPN_VAL(hpte_v
) << 23;
396 unsigned long vpi
, pteg
;
398 pteg
= slot
/ HPTES_PER_GROUP
;
399 if (hpte_v
& HPTE_V_SECONDARY
)
401 vpi
= ((avpn
>> 28) ^ pteg
) & htab_hash_mask
;
402 avpn
|= (vpi
<< mmu_psize_defs
[size
].shift
);
411 * clear all mappings on kexec. All cpus are in real mode (or they will
412 * be when they isi), and we are the only one left. We rely on our kernel
413 * mapping being 0xC0's and the hardware ignoring those two real bits.
415 * TODO: add batching support when enabled. remember, no dynamic memory here,
416 * athough there is the control page available...
418 static void native_hpte_clear(void)
420 unsigned long slot
, slots
, flags
;
421 hpte_t
*hptep
= htab_address
;
422 unsigned long hpte_v
, va
;
423 unsigned long pteg_count
;
426 pteg_count
= htab_hash_mask
+ 1;
428 local_irq_save(flags
);
430 /* we take the tlbie lock and hold it. Some hardware will
431 * deadlock if we try to tlbie from two processors at once.
433 spin_lock(&native_tlbie_lock
);
435 slots
= pteg_count
* HPTES_PER_GROUP
;
437 for (slot
= 0; slot
< slots
; slot
++, hptep
++) {
439 * we could lock the pte here, but we are the only cpu
440 * running, right? and for crash dump, we probably
441 * don't want to wait for a maybe bad cpu.
446 * Call __tlbie() here rather than tlbie() since we
447 * already hold the native_tlbie_lock.
449 if (hpte_v
& HPTE_V_VALID
) {
450 hpte_decode(hptep
, slot
, &psize
, &va
);
456 asm volatile("eieio; tlbsync; ptesync":::"memory");
457 spin_unlock(&native_tlbie_lock
);
458 local_irq_restore(flags
);
462 * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
463 * the lock all the time
465 static void native_flush_hash_range(unsigned long number
, int local
)
467 unsigned long va
, hash
, index
, hidx
, shift
, slot
;
469 unsigned long hpte_v
;
470 unsigned long want_v
;
473 struct ppc64_tlb_batch
*batch
= &__get_cpu_var(ppc64_tlb_batch
);
474 unsigned long psize
= batch
->psize
;
477 local_irq_save(flags
);
479 for (i
= 0; i
< number
; i
++) {
480 va
= batch
->vaddr
[i
];
483 pte_iterate_hashed_subpages(pte
, psize
, va
, index
, shift
) {
484 hash
= hpt_hash(va
, shift
);
485 hidx
= __rpte_to_hidx(pte
, index
);
486 if (hidx
& _PTEIDX_SECONDARY
)
488 slot
= (hash
& htab_hash_mask
) * HPTES_PER_GROUP
;
489 slot
+= hidx
& _PTEIDX_GROUP_IX
;
490 hptep
= htab_address
+ slot
;
491 want_v
= hpte_encode_v(va
, psize
);
492 native_lock_hpte(hptep
);
494 if (!HPTE_V_COMPARE(hpte_v
, want_v
) ||
495 !(hpte_v
& HPTE_V_VALID
))
496 native_unlock_hpte(hptep
);
499 } pte_iterate_hashed_end();
502 if (cpu_has_feature(CPU_FTR_TLBIEL
) &&
503 mmu_psize_defs
[psize
].tlbiel
&& local
) {
504 asm volatile("ptesync":::"memory");
505 for (i
= 0; i
< number
; i
++) {
506 va
= batch
->vaddr
[i
];
509 pte_iterate_hashed_subpages(pte
, psize
, va
, index
,
512 } pte_iterate_hashed_end();
514 asm volatile("ptesync":::"memory");
516 int lock_tlbie
= !cpu_has_feature(CPU_FTR_LOCKLESS_TLBIE
);
519 spin_lock(&native_tlbie_lock
);
521 asm volatile("ptesync":::"memory");
522 for (i
= 0; i
< number
; i
++) {
523 va
= batch
->vaddr
[i
];
526 pte_iterate_hashed_subpages(pte
, psize
, va
, index
,
529 } pte_iterate_hashed_end();
531 asm volatile("eieio; tlbsync; ptesync":::"memory");
534 spin_unlock(&native_tlbie_lock
);
537 local_irq_restore(flags
);
540 #ifdef CONFIG_PPC_PSERIES
541 /* Disable TLB batching on nighthawk */
542 static inline int tlb_batching_enabled(void)
544 struct device_node
*root
= of_find_node_by_path("/");
548 const char *model
= of_get_property(root
, "model", NULL
);
549 if (model
&& !strcmp(model
, "IBM,9076-N81"))
557 static inline int tlb_batching_enabled(void)
563 void __init
hpte_init_native(void)
565 ppc_md
.hpte_invalidate
= native_hpte_invalidate
;
566 ppc_md
.hpte_updatepp
= native_hpte_updatepp
;
567 ppc_md
.hpte_updateboltedpp
= native_hpte_updateboltedpp
;
568 ppc_md
.hpte_insert
= native_hpte_insert
;
569 ppc_md
.hpte_remove
= native_hpte_remove
;
570 ppc_md
.hpte_clear_all
= native_hpte_clear
;
571 if (tlb_batching_enabled())
572 ppc_md
.flush_hash_range
= native_flush_hash_range
;