1 // SPDX-License-Identifier: GPL-2.0
2 /* arch/sparc64/mm/tsb.c
4 * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
7 #include <linux/kernel.h>
8 #include <linux/preempt.h>
9 #include <linux/slab.h>
10 #include <linux/mm_types.h>
13 #include <asm/pgtable.h>
14 #include <asm/mmu_context.h>
15 #include <asm/setup.h>
18 #include <asm/oplib.h>
20 extern struct tsb swapper_tsb
[KERNEL_TSB_NENTRIES
];
22 static inline unsigned long tsb_hash(unsigned long vaddr
, unsigned long hash_shift
, unsigned long nentries
)
25 return vaddr
& (nentries
- 1);
28 static inline int tag_compare(unsigned long tag
, unsigned long vaddr
)
30 return (tag
== (vaddr
>> 22));
33 static void flush_tsb_kernel_range_scan(unsigned long start
, unsigned long end
)
37 for (idx
= 0; idx
< KERNEL_TSB_NENTRIES
; idx
++) {
38 struct tsb
*ent
= &swapper_tsb
[idx
];
39 unsigned long match
= idx
<< 13;
41 match
|= (ent
->tag
<< 22);
42 if (match
>= start
&& match
< end
)
43 ent
->tag
= (1UL << TSB_TAG_INVALID_BIT
);
47 /* TSB flushes need only occur on the processor initiating the address
48 * space modification, not on each cpu the address space has run on.
49 * Only the TLB flush needs that treatment.
52 void flush_tsb_kernel_range(unsigned long start
, unsigned long end
)
56 if ((end
- start
) >> PAGE_SHIFT
>= 2 * KERNEL_TSB_NENTRIES
)
57 return flush_tsb_kernel_range_scan(start
, end
);
59 for (v
= start
; v
< end
; v
+= PAGE_SIZE
) {
60 unsigned long hash
= tsb_hash(v
, PAGE_SHIFT
,
62 struct tsb
*ent
= &swapper_tsb
[hash
];
64 if (tag_compare(ent
->tag
, v
))
65 ent
->tag
= (1UL << TSB_TAG_INVALID_BIT
);
69 static void __flush_tsb_one_entry(unsigned long tsb
, unsigned long v
,
70 unsigned long hash_shift
,
71 unsigned long nentries
)
73 unsigned long tag
, ent
, hash
;
76 hash
= tsb_hash(v
, hash_shift
, nentries
);
77 ent
= tsb
+ (hash
* sizeof(struct tsb
));
83 static void __flush_tsb_one(struct tlb_batch
*tb
, unsigned long hash_shift
,
84 unsigned long tsb
, unsigned long nentries
)
88 for (i
= 0; i
< tb
->tlb_nr
; i
++)
89 __flush_tsb_one_entry(tsb
, tb
->vaddrs
[i
], hash_shift
, nentries
);
92 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
93 static void __flush_huge_tsb_one_entry(unsigned long tsb
, unsigned long v
,
94 unsigned long hash_shift
,
95 unsigned long nentries
,
96 unsigned int hugepage_shift
)
98 unsigned int hpage_entries
;
101 hpage_entries
= 1 << (hugepage_shift
- hash_shift
);
102 for (i
= 0; i
< hpage_entries
; i
++)
103 __flush_tsb_one_entry(tsb
, v
+ (i
<< hash_shift
), hash_shift
,
107 static void __flush_huge_tsb_one(struct tlb_batch
*tb
, unsigned long hash_shift
,
108 unsigned long tsb
, unsigned long nentries
,
109 unsigned int hugepage_shift
)
113 for (i
= 0; i
< tb
->tlb_nr
; i
++)
114 __flush_huge_tsb_one_entry(tsb
, tb
->vaddrs
[i
], hash_shift
,
115 nentries
, hugepage_shift
);
119 void flush_tsb_user(struct tlb_batch
*tb
)
121 struct mm_struct
*mm
= tb
->mm
;
122 unsigned long nentries
, base
, flags
;
124 spin_lock_irqsave(&mm
->context
.lock
, flags
);
126 if (tb
->hugepage_shift
< REAL_HPAGE_SHIFT
) {
127 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
;
128 nentries
= mm
->context
.tsb_block
[MM_TSB_BASE
].tsb_nentries
;
129 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
131 if (tb
->hugepage_shift
== PAGE_SHIFT
)
132 __flush_tsb_one(tb
, PAGE_SHIFT
, base
, nentries
);
133 #if defined(CONFIG_HUGETLB_PAGE)
135 __flush_huge_tsb_one(tb
, PAGE_SHIFT
, base
, nentries
,
139 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
140 else if (mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
) {
141 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
;
142 nentries
= mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb_nentries
;
143 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
145 __flush_huge_tsb_one(tb
, REAL_HPAGE_SHIFT
, base
, nentries
,
149 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
152 void flush_tsb_user_page(struct mm_struct
*mm
, unsigned long vaddr
,
153 unsigned int hugepage_shift
)
155 unsigned long nentries
, base
, flags
;
157 spin_lock_irqsave(&mm
->context
.lock
, flags
);
159 if (hugepage_shift
< REAL_HPAGE_SHIFT
) {
160 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
;
161 nentries
= mm
->context
.tsb_block
[MM_TSB_BASE
].tsb_nentries
;
162 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
164 if (hugepage_shift
== PAGE_SHIFT
)
165 __flush_tsb_one_entry(base
, vaddr
, PAGE_SHIFT
,
167 #if defined(CONFIG_HUGETLB_PAGE)
169 __flush_huge_tsb_one_entry(base
, vaddr
, PAGE_SHIFT
,
170 nentries
, hugepage_shift
);
173 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
174 else if (mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
) {
175 base
= (unsigned long) mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb
;
176 nentries
= mm
->context
.tsb_block
[MM_TSB_HUGE
].tsb_nentries
;
177 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
)
179 __flush_huge_tsb_one_entry(base
, vaddr
, REAL_HPAGE_SHIFT
,
180 nentries
, hugepage_shift
);
183 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
186 #define HV_PGSZ_IDX_BASE HV_PGSZ_IDX_8K
187 #define HV_PGSZ_MASK_BASE HV_PGSZ_MASK_8K
189 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
190 #define HV_PGSZ_IDX_HUGE HV_PGSZ_IDX_4MB
191 #define HV_PGSZ_MASK_HUGE HV_PGSZ_MASK_4MB
194 static void setup_tsb_params(struct mm_struct
*mm
, unsigned long tsb_idx
, unsigned long tsb_bytes
)
196 unsigned long tsb_reg
, base
, tsb_paddr
;
197 unsigned long page_sz
, tte
;
199 mm
->context
.tsb_block
[tsb_idx
].tsb_nentries
=
200 tsb_bytes
/ sizeof(struct tsb
);
204 base
= TSBMAP_8K_BASE
;
206 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
208 base
= TSBMAP_4M_BASE
;
215 tte
= pgprot_val(PAGE_KERNEL_LOCKED
);
216 tsb_paddr
= __pa(mm
->context
.tsb_block
[tsb_idx
].tsb
);
217 BUG_ON(tsb_paddr
& (tsb_bytes
- 1UL));
219 /* Use the smallest page size that can map the whole TSB
225 #ifdef DCACHE_ALIASING_POSSIBLE
226 base
+= (tsb_paddr
& 8192);
248 page_sz
= 512 * 1024;
253 page_sz
= 512 * 1024;
258 page_sz
= 512 * 1024;
263 page_sz
= 4 * 1024 * 1024;
267 printk(KERN_ERR
"TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
268 current
->comm
, current
->pid
, tsb_bytes
);
271 tte
|= pte_sz_bits(page_sz
);
273 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
) {
274 /* Physical mapping, no locked TLB entry for TSB. */
275 tsb_reg
|= tsb_paddr
;
277 mm
->context
.tsb_block
[tsb_idx
].tsb_reg_val
= tsb_reg
;
278 mm
->context
.tsb_block
[tsb_idx
].tsb_map_vaddr
= 0;
279 mm
->context
.tsb_block
[tsb_idx
].tsb_map_pte
= 0;
282 tsb_reg
|= (tsb_paddr
& (page_sz
- 1UL));
283 tte
|= (tsb_paddr
& ~(page_sz
- 1UL));
285 mm
->context
.tsb_block
[tsb_idx
].tsb_reg_val
= tsb_reg
;
286 mm
->context
.tsb_block
[tsb_idx
].tsb_map_vaddr
= base
;
287 mm
->context
.tsb_block
[tsb_idx
].tsb_map_pte
= tte
;
290 /* Setup the Hypervisor TSB descriptor. */
291 if (tlb_type
== hypervisor
) {
292 struct hv_tsb_descr
*hp
= &mm
->context
.tsb_descr
[tsb_idx
];
296 hp
->pgsz_idx
= HV_PGSZ_IDX_BASE
;
298 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
300 hp
->pgsz_idx
= HV_PGSZ_IDX_HUGE
;
307 hp
->num_ttes
= tsb_bytes
/ 16;
311 hp
->pgsz_mask
= HV_PGSZ_MASK_BASE
;
313 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
315 hp
->pgsz_mask
= HV_PGSZ_MASK_HUGE
;
321 hp
->tsb_base
= tsb_paddr
;
326 struct kmem_cache
*pgtable_cache __read_mostly
;
328 static struct kmem_cache
*tsb_caches
[8] __read_mostly
;
330 static const char *tsb_cache_names
[8] = {
341 void __init
pgtable_cache_init(void)
345 pgtable_cache
= kmem_cache_create("pgtable_cache",
346 PAGE_SIZE
, PAGE_SIZE
,
349 if (!pgtable_cache
) {
350 prom_printf("pgtable_cache_init(): Could not create!\n");
354 for (i
= 0; i
< ARRAY_SIZE(tsb_cache_names
); i
++) {
355 unsigned long size
= 8192 << i
;
356 const char *name
= tsb_cache_names
[i
];
358 tsb_caches
[i
] = kmem_cache_create(name
,
361 if (!tsb_caches
[i
]) {
362 prom_printf("Could not create %s cache\n", name
);
368 int sysctl_tsb_ratio
= -2;
370 static unsigned long tsb_size_to_rss_limit(unsigned long new_size
)
372 unsigned long num_ents
= (new_size
/ sizeof(struct tsb
));
374 if (sysctl_tsb_ratio
< 0)
375 return num_ents
- (num_ents
>> -sysctl_tsb_ratio
);
377 return num_ents
+ (num_ents
>> sysctl_tsb_ratio
);
380 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
381 * do_sparc64_fault() invokes this routine to try and grow it.
383 * When we reach the maximum TSB size supported, we stick ~0UL into
384 * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
385 * will not trigger any longer.
387 * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
388 * of two. The TSB must be aligned to it's size, so f.e. a 512K TSB
389 * must be 512K aligned. It also must be physically contiguous, so we
390 * cannot use vmalloc().
392 * The idea here is to grow the TSB when the RSS of the process approaches
393 * the number of entries that the current TSB can hold at once. Currently,
394 * we trigger when the RSS hits 3/4 of the TSB capacity.
396 void tsb_grow(struct mm_struct
*mm
, unsigned long tsb_index
, unsigned long rss
)
398 unsigned long max_tsb_size
= 1 * 1024 * 1024;
399 unsigned long new_size
, old_size
, flags
;
400 struct tsb
*old_tsb
, *new_tsb
;
401 unsigned long new_cache_index
, old_cache_index
;
402 unsigned long new_rss_limit
;
405 if (max_tsb_size
> (PAGE_SIZE
<< MAX_ORDER
))
406 max_tsb_size
= (PAGE_SIZE
<< MAX_ORDER
);
409 for (new_size
= 8192; new_size
< max_tsb_size
; new_size
<<= 1UL) {
410 new_rss_limit
= tsb_size_to_rss_limit(new_size
);
411 if (new_rss_limit
> rss
)
416 if (new_size
== max_tsb_size
)
417 new_rss_limit
= ~0UL;
420 gfp_flags
= GFP_KERNEL
;
421 if (new_size
> (PAGE_SIZE
* 2))
422 gfp_flags
|= __GFP_NOWARN
| __GFP_NORETRY
;
424 new_tsb
= kmem_cache_alloc_node(tsb_caches
[new_cache_index
],
425 gfp_flags
, numa_node_id());
426 if (unlikely(!new_tsb
)) {
427 /* Not being able to fork due to a high-order TSB
428 * allocation failure is very bad behavior. Just back
429 * down to a 0-order allocation and force no TSB
430 * growing for this address space.
432 if (mm
->context
.tsb_block
[tsb_index
].tsb
== NULL
&&
433 new_cache_index
> 0) {
436 new_rss_limit
= ~0UL;
437 goto retry_tsb_alloc
;
440 /* If we failed on a TSB grow, we are under serious
441 * memory pressure so don't try to grow any more.
443 if (mm
->context
.tsb_block
[tsb_index
].tsb
!= NULL
)
444 mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
= ~0UL;
448 /* Mark all tags as invalid. */
449 tsb_init(new_tsb
, new_size
);
451 /* Ok, we are about to commit the changes. If we are
452 * growing an existing TSB the locking is very tricky,
455 * We have to hold mm->context.lock while committing to the
456 * new TSB, this synchronizes us with processors in
457 * flush_tsb_user() and switch_mm() for this address space.
459 * But even with that lock held, processors run asynchronously
460 * accessing the old TSB via TLB miss handling. This is OK
461 * because those actions are just propagating state from the
462 * Linux page tables into the TSB, page table mappings are not
463 * being changed. If a real fault occurs, the processor will
464 * synchronize with us when it hits flush_tsb_user(), this is
465 * also true for the case where vmscan is modifying the page
466 * tables. The only thing we need to be careful with is to
467 * skip any locked TSB entries during copy_tsb().
469 * When we finish committing to the new TSB, we have to drop
470 * the lock and ask all other cpus running this address space
471 * to run tsb_context_switch() to see the new TSB table.
473 spin_lock_irqsave(&mm
->context
.lock
, flags
);
475 old_tsb
= mm
->context
.tsb_block
[tsb_index
].tsb
;
477 (mm
->context
.tsb_block
[tsb_index
].tsb_reg_val
& 0x7UL
);
478 old_size
= (mm
->context
.tsb_block
[tsb_index
].tsb_nentries
*
482 /* Handle multiple threads trying to grow the TSB at the same time.
483 * One will get in here first, and bump the size and the RSS limit.
484 * The others will get in here next and hit this check.
486 if (unlikely(old_tsb
&&
487 (rss
< mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
))) {
488 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
490 kmem_cache_free(tsb_caches
[new_cache_index
], new_tsb
);
494 mm
->context
.tsb_block
[tsb_index
].tsb_rss_limit
= new_rss_limit
;
497 extern void copy_tsb(unsigned long old_tsb_base
,
498 unsigned long old_tsb_size
,
499 unsigned long new_tsb_base
,
500 unsigned long new_tsb_size
,
501 unsigned long page_size_shift
);
502 unsigned long old_tsb_base
= (unsigned long) old_tsb
;
503 unsigned long new_tsb_base
= (unsigned long) new_tsb
;
505 if (tlb_type
== cheetah_plus
|| tlb_type
== hypervisor
) {
506 old_tsb_base
= __pa(old_tsb_base
);
507 new_tsb_base
= __pa(new_tsb_base
);
509 copy_tsb(old_tsb_base
, old_size
, new_tsb_base
, new_size
,
510 tsb_index
== MM_TSB_BASE
?
511 PAGE_SHIFT
: REAL_HPAGE_SHIFT
);
514 mm
->context
.tsb_block
[tsb_index
].tsb
= new_tsb
;
515 setup_tsb_params(mm
, tsb_index
, new_size
);
517 spin_unlock_irqrestore(&mm
->context
.lock
, flags
);
519 /* If old_tsb is NULL, we're being invoked for the first time
520 * from init_new_context().
523 /* Reload it on the local cpu. */
524 tsb_context_switch(mm
);
526 /* Now force other processors to do the same. */
531 /* Now it is safe to free the old tsb. */
532 kmem_cache_free(tsb_caches
[old_cache_index
], old_tsb
);
536 int init_new_context(struct task_struct
*tsk
, struct mm_struct
*mm
)
538 unsigned long mm_rss
= get_mm_rss(mm
);
539 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
540 unsigned long saved_hugetlb_pte_count
;
541 unsigned long saved_thp_pte_count
;
545 spin_lock_init(&mm
->context
.lock
);
547 mm
->context
.sparc64_ctx_val
= 0UL;
549 mm
->context
.tag_store
= NULL
;
550 spin_lock_init(&mm
->context
.tag_lock
);
552 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
553 /* We reset them to zero because the fork() page copying
554 * will re-increment the counters as the parent PTEs are
555 * copied into the child address space.
557 saved_hugetlb_pte_count
= mm
->context
.hugetlb_pte_count
;
558 saved_thp_pte_count
= mm
->context
.thp_pte_count
;
559 mm
->context
.hugetlb_pte_count
= 0;
560 mm
->context
.thp_pte_count
= 0;
562 mm_rss
-= saved_thp_pte_count
* (HPAGE_SIZE
/ PAGE_SIZE
);
565 /* copy_mm() copies over the parent's mm_struct before calling
566 * us, so we need to zero out the TSB pointer or else tsb_grow()
567 * will be confused and think there is an older TSB to free up.
569 for (i
= 0; i
< MM_NUM_TSBS
; i
++)
570 mm
->context
.tsb_block
[i
].tsb
= NULL
;
572 /* If this is fork, inherit the parent's TSB size. We would
573 * grow it to that size on the first page fault anyways.
575 tsb_grow(mm
, MM_TSB_BASE
, mm_rss
);
577 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
578 if (unlikely(saved_hugetlb_pte_count
+ saved_thp_pte_count
))
579 tsb_grow(mm
, MM_TSB_HUGE
,
580 (saved_hugetlb_pte_count
+ saved_thp_pte_count
) *
581 REAL_HPAGE_PER_HPAGE
);
584 if (unlikely(!mm
->context
.tsb_block
[MM_TSB_BASE
].tsb
))
590 static void tsb_destroy_one(struct tsb_config
*tp
)
592 unsigned long cache_index
;
596 cache_index
= tp
->tsb_reg_val
& 0x7UL
;
597 kmem_cache_free(tsb_caches
[cache_index
], tp
->tsb
);
599 tp
->tsb_reg_val
= 0UL;
602 void destroy_context(struct mm_struct
*mm
)
604 unsigned long flags
, i
;
606 for (i
= 0; i
< MM_NUM_TSBS
; i
++)
607 tsb_destroy_one(&mm
->context
.tsb_block
[i
]);
609 spin_lock_irqsave(&ctx_alloc_lock
, flags
);
611 if (CTX_VALID(mm
->context
)) {
612 unsigned long nr
= CTX_NRBITS(mm
->context
);
613 mmu_context_bmap
[nr
>>6] &= ~(1UL << (nr
& 63));
616 spin_unlock_irqrestore(&ctx_alloc_lock
, flags
);
618 /* If ADI tag storage was allocated for this task, free it */
619 if (mm
->context
.tag_store
) {
620 tag_storage_desc_t
*tag_desc
;
621 unsigned long max_desc
;
624 tag_desc
= mm
->context
.tag_store
;
625 max_desc
= PAGE_SIZE
/sizeof(tag_storage_desc_t
);
626 for (i
= 0; i
< max_desc
; i
++) {
627 tags
= tag_desc
->tags
;
628 tag_desc
->tags
= NULL
;
632 kfree(mm
->context
.tag_store
);
633 mm
->context
.tag_store
= NULL
;