1 /* arch/sparc64/mm/tsb.c
3 * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
6 #include <linux/kernel.h>
7 #include <linux/preempt.h>
8 #include <linux/slab.h>
9 #include <linux/mm_types.h>
12 #include <asm/pgtable.h>
13 #include <asm/mmu_context.h>
14 #include <asm/setup.h>
17 #include <asm/oplib.h>
19 extern struct tsb swapper_tsb
[KERNEL_TSB_NENTRIES
];
21 static inline unsigned long tsb_hash(unsigned long vaddr
, unsigned long hash_shift
, unsigned long nentries
)
24 return vaddr
& (nentries
- 1);
27 static inline int tag_compare(unsigned long tag
, unsigned long vaddr
)
29 return (tag
== (vaddr
>> 22));
32 static void flush_tsb_kernel_range_scan(unsigned long start
, unsigned long end
)
36 for (idx
= 0; idx
< KERNEL_TSB_NENTRIES
; idx
++) {
37 struct tsb
*ent
= &swapper_tsb
[idx
];
38 unsigned long match
= idx
<< 13;
40 match
|= (ent
->tag
<< 22);
41 if (match
>= start
&& match
< end
)
42 ent
->tag
= (1UL << TSB_TAG_INVALID_BIT
);
46 /* TSB flushes need only occur on the processor initiating the address
47 * space modification, not on each cpu the address space has run on.
48 * Only the TLB flush needs that treatment.
51 void flush_tsb_kernel_range(unsigned long start
, unsigned long end
)
55 if ((end
- start
) >> PAGE_SHIFT
>= 2 * KERNEL_TSB_NENTRIES
)
56 return flush_tsb_kernel_range_scan(start
, end
);
58 for (v
= start
; v
< end
; v
+= PAGE_SIZE
) {
59 unsigned long hash
= tsb_hash(v
, PAGE_SHIFT
,
61 struct tsb
*ent
= &swapper_tsb
[hash
];
63 if (tag_compare(ent
->tag
, v
))
64 ent
->tag
= (1UL << TSB_TAG_INVALID_BIT
);
68 static void __flush_tsb_one_entry(unsigned long tsb
, unsigned long v
,
69 unsigned long hash_shift
,
70 unsigned long nentries
)
72 unsigned long tag
, ent
, hash
;
75 hash
= tsb_hash(v
, hash_shift
, nentries
);
76 ent
= tsb
+ (hash
* sizeof(struct tsb
));
82 static void __flush_tsb_one(struct tlb_batch
*tb
, unsigned long hash_shift
,
83 unsigned long tsb
, unsigned long nentries
)
87 for (i
= 0; i
< tb
->tlb_nr
; i
++)
88 __flush_tsb_one_entry(tsb
, tb
->vaddrs
[i
], hash_shift
, nentries
);
91 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
92 static void __flush_huge_tsb_one_entry(unsigned long tsb
, unsigned long v
,
93 unsigned long hash_shift
,
94 unsigned long nentries
,
95 unsigned int hugepage_shift
)
97 unsigned int hpage_entries
;
100 hpage_entries
= 1 << (hugepage_shift
- hash_shift
);
101 for (i
= 0; i
< hpage_entries
; i
++)
102 __flush_tsb_one_entry(tsb
, v
+ (i
<< hash_shift
), hash_shift
,
106 static void __flush_huge_tsb_one(struct tlb_batch
*tb
, unsigned long hash_shift
,
107 unsigned long tsb
, unsigned long nentries
,
108 unsigned int hugepage_shift
)
112 for (i
= 0; i
< tb
->tlb_nr
; i
++)
113 __flush_huge_tsb_one_entry(tsb
, tb
->vaddrs
[i
], hash_shift
,
114 nentries
, hugepage_shift
);
118 void flush_tsb_user(struct tlb_batch
*tb
)
120 struct mm_struct
*mm
= tb
->mm
;
121 unsigned long nentries
, base
, flags
;
123 spin_lock_irqsave(&mm
->context
.lock
, flags
);
125 if (tb
->hugepage_shift
< REAL_HPAGE_SHIFT
) {
126 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
;
127 nentries
= mm
->context
.tsb_block
[MM_TSB_BASE
].tsb_nentries
;
128 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
130 if (tb
->hugepage_shift
== PAGE_SHIFT
)
131 __flush_tsb_one(tb
, PAGE_SHIFT
, base
, nentries
);
132 #if defined(CONFIG_HUGETLB_PAGE)
134 __flush_huge_tsb_one(tb
, PAGE_SHIFT
, base
, nentries
,
138 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
139 else if (mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
) {
140 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
;
141 nentries
= mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb_nentries
;
142 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
144 __flush_huge_tsb_one(tb
, REAL_HPAGE_SHIFT
, base
, nentries
,
148 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
151 void flush_tsb_user_page(struct mm_struct
*mm
, unsigned long vaddr
,
152 unsigned int hugepage_shift
)
154 unsigned long nentries
, base
, flags
;
156 spin_lock_irqsave(&mm
->context
.lock
, flags
);
158 if (hugepage_shift
< REAL_HPAGE_SHIFT
) {
159 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
;
160 nentries
= mm
->context
.tsb_block
[MM_TSB_BASE
].tsb_nentries
;
161 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
163 if (hugepage_shift
== PAGE_SHIFT
)
164 __flush_tsb_one_entry(base
, vaddr
, PAGE_SHIFT
,
166 #if defined(CONFIG_HUGETLB_PAGE)
168 __flush_huge_tsb_one_entry(base
, vaddr
, PAGE_SHIFT
,
169 nentries
, hugepage_shift
);
172 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
173 else if (mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
) {
174 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
;
175 nentries
= mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb_nentries
;
176 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
178 __flush_huge_tsb_one_entry(base
, vaddr
, REAL_HPAGE_SHIFT
,
179 nentries
, hugepage_shift
);
182 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
185 #define HV_PGSZ_IDX_BASE HV_PGSZ_IDX_8K
186 #define HV_PGSZ_MASK_BASE HV_PGSZ_MASK_8K
188 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
189 #define HV_PGSZ_IDX_HUGE HV_PGSZ_IDX_4MB
190 #define HV_PGSZ_MASK_HUGE HV_PGSZ_MASK_4MB
193 static void setup_tsb_params(struct mm_struct
*mm
, unsigned long tsb_idx
, unsigned long tsb_bytes
)
195 unsigned long tsb_reg
, base
, tsb_paddr
;
196 unsigned long page_sz
, tte
;
198 mm
->context
.tsb_block
[tsb_idx
].tsb_nentries
=
199 tsb_bytes
/ sizeof(struct tsb
);
203 base
= TSBMAP_8K_BASE
;
205 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
207 base
= TSBMAP_4M_BASE
;
214 tte
= pgprot_val(PAGE_KERNEL_LOCKED
);
215 tsb_paddr
= __pa(mm
->context
.tsb_block
[tsb_idx
].tsb
);
216 BUG_ON(tsb_paddr
& (tsb_bytes
- 1UL));
218 /* Use the smallest page size that can map the whole TSB
224 #ifdef DCACHE_ALIASING_POSSIBLE
225 base
+= (tsb_paddr
& 8192);
247 page_sz
= 512 * 1024;
252 page_sz
= 512 * 1024;
257 page_sz
= 512 * 1024;
262 page_sz
= 4 * 1024 * 1024;
266 printk(KERN_ERR
"TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
267 current
->comm
, current
->pid
, tsb_bytes
);
270 tte
|= pte_sz_bits(page_sz
);
272 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
) {
273 /* Physical mapping, no locked TLB entry for TSB. */
274 tsb_reg
|= tsb_paddr
;
276 mm
->context
.tsb_block
[tsb_idx
].tsb_reg_val
= tsb_reg
;
277 mm
->context
.tsb_block
[tsb_idx
].tsb_map_vaddr
= 0;
278 mm
->context
.tsb_block
[tsb_idx
].tsb_map_pte
= 0;
281 tsb_reg
|= (tsb_paddr
& (page_sz
- 1UL));
282 tte
|= (tsb_paddr
& ~(page_sz
- 1UL));
284 mm
->context
.tsb_block
[tsb_idx
].tsb_reg_val
= tsb_reg
;
285 mm
->context
.tsb_block
[tsb_idx
].tsb_map_vaddr
= base
;
286 mm
->context
.tsb_block
[tsb_idx
].tsb_map_pte
= tte
;
289 /* Setup the Hypervisor TSB descriptor. */
290 if (tlb_type
== hypervisor
) {
291 struct hv_tsb_descr
*hp
= &mm
->context
.tsb_descr
[tsb_idx
];
295 hp
->pgsz_idx
= HV_PGSZ_IDX_BASE
;
297 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
299 hp
->pgsz_idx
= HV_PGSZ_IDX_HUGE
;
306 hp
->num_ttes
= tsb_bytes
/ 16;
310 hp
->pgsz_mask
= HV_PGSZ_MASK_BASE
;
312 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
314 hp
->pgsz_mask
= HV_PGSZ_MASK_HUGE
;
320 hp
->tsb_base
= tsb_paddr
;
325 struct kmem_cache
*pgtable_cache __read_mostly
;
327 static struct kmem_cache
*tsb_caches
[8] __read_mostly
;
329 static const char *tsb_cache_names
[8] = {
340 void __init
pgtable_cache_init(void)
344 pgtable_cache
= kmem_cache_create("pgtable_cache",
345 PAGE_SIZE
, PAGE_SIZE
,
348 if (!pgtable_cache
) {
349 prom_printf("pgtable_cache_init(): Could not create!\n");
353 for (i
= 0; i
< ARRAY_SIZE(tsb_cache_names
); i
++) {
354 unsigned long size
= 8192 << i
;
355 const char *name
= tsb_cache_names
[i
];
357 tsb_caches
[i
] = kmem_cache_create(name
,
360 if (!tsb_caches
[i
]) {
361 prom_printf("Could not create %s cache\n", name
);
367 int sysctl_tsb_ratio
= -2;
369 static unsigned long tsb_size_to_rss_limit(unsigned long new_size
)
371 unsigned long num_ents
= (new_size
/ sizeof(struct tsb
));
373 if (sysctl_tsb_ratio
< 0)
374 return num_ents
- (num_ents
>> -sysctl_tsb_ratio
);
376 return num_ents
+ (num_ents
>> sysctl_tsb_ratio
);
379 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
380 * do_sparc64_fault() invokes this routine to try and grow it.
382 * When we reach the maximum TSB size supported, we stick ~0UL into
383 * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
384 * will not trigger any longer.
386 * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
387 * of two. The TSB must be aligned to it's size, so f.e. a 512K TSB
388 * must be 512K aligned. It also must be physically contiguous, so we
389 * cannot use vmalloc().
391 * The idea here is to grow the TSB when the RSS of the process approaches
392 * the number of entries that the current TSB can hold at once. Currently,
393 * we trigger when the RSS hits 3/4 of the TSB capacity.
395 void tsb_grow(struct mm_struct
*mm
, unsigned long tsb_index
, unsigned long rss
)
397 unsigned long max_tsb_size
= 1 * 1024 * 1024;
398 unsigned long new_size
, old_size
, flags
;
399 struct tsb
*old_tsb
, *new_tsb
;
400 unsigned long new_cache_index
, old_cache_index
;
401 unsigned long new_rss_limit
;
404 if (max_tsb_size
> (PAGE_SIZE
<< MAX_ORDER
))
405 max_tsb_size
= (PAGE_SIZE
<< MAX_ORDER
);
408 for (new_size
= 8192; new_size
< max_tsb_size
; new_size
<<= 1UL) {
409 new_rss_limit
= tsb_size_to_rss_limit(new_size
);
410 if (new_rss_limit
> rss
)
415 if (new_size
== max_tsb_size
)
416 new_rss_limit
= ~0UL;
419 gfp_flags
= GFP_KERNEL
;
420 if (new_size
> (PAGE_SIZE
* 2))
421 gfp_flags
|= __GFP_NOWARN
| __GFP_NORETRY
;
423 new_tsb
= kmem_cache_alloc_node(tsb_caches
[new_cache_index
],
424 gfp_flags
, numa_node_id());
425 if (unlikely(!new_tsb
)) {
426 /* Not being able to fork due to a high-order TSB
427 * allocation failure is very bad behavior. Just back
428 * down to a 0-order allocation and force no TSB
429 * growing for this address space.
431 if (mm
->context
.tsb_block
[tsb_index
].tsb
== NULL
&&
432 new_cache_index
> 0) {
435 new_rss_limit
= ~0UL;
436 goto retry_tsb_alloc
;
439 /* If we failed on a TSB grow, we are under serious
440 * memory pressure so don't try to grow any more.
442 if (mm
->context
.tsb_block
[tsb_index
].tsb
!= NULL
)
443 mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
= ~0UL;
447 /* Mark all tags as invalid. */
448 tsb_init(new_tsb
, new_size
);
450 /* Ok, we are about to commit the changes. If we are
451 * growing an existing TSB the locking is very tricky,
454 * We have to hold mm->context.lock while committing to the
455 * new TSB, this synchronizes us with processors in
456 * flush_tsb_user() and switch_mm() for this address space.
458 * But even with that lock held, processors run asynchronously
459 * accessing the old TSB via TLB miss handling. This is OK
460 * because those actions are just propagating state from the
461 * Linux page tables into the TSB, page table mappings are not
462 * being changed. If a real fault occurs, the processor will
463 * synchronize with us when it hits flush_tsb_user(), this is
464 * also true for the case where vmscan is modifying the page
465 * tables. The only thing we need to be careful with is to
466 * skip any locked TSB entries during copy_tsb().
468 * When we finish committing to the new TSB, we have to drop
469 * the lock and ask all other cpus running this address space
470 * to run tsb_context_switch() to see the new TSB table.
472 spin_lock_irqsave(&mm
->context
.lock
, flags
);
474 old_tsb
= mm
->context
.tsb_block
[tsb_index
].tsb
;
476 (mm
->context
.tsb_block
[tsb_index
].tsb_reg_val
& 0x7UL
);
477 old_size
= (mm
->context
.tsb_block
[tsb_index
].tsb_nentries
*
481 /* Handle multiple threads trying to grow the TSB at the same time.
482 * One will get in here first, and bump the size and the RSS limit.
483 * The others will get in here next and hit this check.
485 if (unlikely(old_tsb
&&
486 (rss
< mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
))) {
487 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
489 kmem_cache_free(tsb_caches
[new_cache_index
], new_tsb
);
493 mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
= new_rss_limit
;
496 extern void copy_tsb(unsigned long old_tsb_base
,
497 unsigned long old_tsb_size
,
498 unsigned long new_tsb_base
,
499 unsigned long new_tsb_size
,
500 unsigned long page_size_shift
);
501 unsigned long old_tsb_base
= (unsigned long) old_tsb
;
502 unsigned long new_tsb_base
= (unsigned long) new_tsb
;
504 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
) {
505 old_tsb_base
= __pa(old_tsb_base
);
506 new_tsb_base
= __pa(new_tsb_base
);
508 copy_tsb(old_tsb_base
, old_size
, new_tsb_base
, new_size
,
509 tsb_index
== MM_TSB_BASE
?
510 PAGE_SHIFT
: REAL_HPAGE_SHIFT
);
513 mm
->context
.tsb_block
[tsb_index
].tsb
= new_tsb
;
514 setup_tsb_params(mm
, tsb_index
, new_size
);
516 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
518 /* If old_tsb is NULL, we're being invoked for the first time
519 * from init_new_context().
522 /* Reload it on the local cpu. */
523 tsb_context_switch(mm
);
525 /* Now force other processors to do the same. */
530 /* Now it is safe to free the old tsb. */
531 kmem_cache_free(tsb_caches
[old_cache_index
], old_tsb
);
535 int init_new_context(struct task_struct
*tsk
, struct mm_struct
*mm
)
537 unsigned long mm_rss
= get_mm_rss(mm
);
538 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
539 unsigned long saved_hugetlb_pte_count
;
540 unsigned long saved_thp_pte_count
;
544 spin_lock_init(&mm
->context
.lock
);
546 mm
->context
.sparc64_ctx_val
= 0UL;
548 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
549 /* We reset them to zero because the fork() page copying
550 * will re-increment the counters as the parent PTEs are
551 * copied into the child address space.
553 saved_hugetlb_pte_count
= mm
->context
.hugetlb_pte_count
;
554 saved_thp_pte_count
= mm
->context
.thp_pte_count
;
555 mm
->context
.hugetlb_pte_count
= 0;
556 mm
->context
.thp_pte_count
= 0;
558 mm_rss
-= saved_thp_pte_count
* (HPAGE_SIZE
/ PAGE_SIZE
);
561 /* copy_mm() copies over the parent's mm_struct before calling
562 * us, so we need to zero out the TSB pointer or else tsb_grow()
563 * will be confused and think there is an older TSB to free up.
565 for (i
= 0; i
< MM_NUM_TSBS
; i
++)
566 mm
->context
.tsb_block
[i
].tsb
= NULL
;
568 /* If this is fork, inherit the parent's TSB size. We would
569 * grow it to that size on the first page fault anyways.
571 tsb_grow(mm
, MM_TSB_BASE
, mm_rss
);
573 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
574 if (unlikely(saved_hugetlb_pte_count
+ saved_thp_pte_count
))
575 tsb_grow(mm
, MM_TSB_HUGE
,
576 (saved_hugetlb_pte_count
+ saved_thp_pte_count
) *
577 REAL_HPAGE_PER_HPAGE
);
580 if (unlikely(!mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
))
586 static void tsb_destroy_one(struct tsb_config
*tp
)
588 unsigned long cache_index
;
592 cache_index
= tp
->tsb_reg_val
& 0x7UL
;
593 kmem_cache_free(tsb_caches
[cache_index
], tp
->tsb
);
595 tp
->tsb_reg_val
= 0UL;
598 void destroy_context(struct mm_struct
*mm
)
600 unsigned long flags
, i
;
602 for (i
= 0; i
< MM_NUM_TSBS
; i
++)
603 tsb_destroy_one(&mm
->context
.tsb_block
[i
]);
605 spin_lock_irqsave(&ctx_alloc_lock
, flags
);
607 if (CTX_VALID(mm
->context
)) {
608 unsigned long nr
= CTX_NRBITS(mm
->context
);
609 mmu_context_bmap
[nr
>>6] &= ~(1UL << (nr
& 63));
612 spin_unlock_irqrestore(&ctx_alloc_lock
, flags
);