1 // SPDX-License-Identifier: GPL-2.0
3 * Page table allocation functions
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
9 #include <linux/sysctl.h>
10 #include <linux/slab.h>
12 #include <asm/mmu_context.h>
13 #include <asm/pgalloc.h>
16 #include <asm/tlbflush.h>
20 static int page_table_allocate_pgste_min
= 0;
21 static int page_table_allocate_pgste_max
= 1;
22 int page_table_allocate_pgste
= 0;
23 EXPORT_SYMBOL(page_table_allocate_pgste
);
25 static struct ctl_table page_table_sysctl
[] = {
27 .procname
= "allocate_pgste",
28 .data
= &page_table_allocate_pgste
,
29 .maxlen
= sizeof(int),
30 .mode
= S_IRUGO
| S_IWUSR
,
31 .proc_handler
= proc_dointvec_minmax
,
32 .extra1
= &page_table_allocate_pgste_min
,
33 .extra2
= &page_table_allocate_pgste_max
,
38 static struct ctl_table page_table_sysctl_dir
[] = {
43 .child
= page_table_sysctl
,
48 static int __init
page_table_register_sysctl(void)
50 return register_sysctl_table(page_table_sysctl_dir
) ? 0 : -ENOMEM
;
52 __initcall(page_table_register_sysctl
);
54 #endif /* CONFIG_PGSTE */
56 unsigned long *crst_table_alloc(struct mm_struct
*mm
)
58 struct page
*page
= alloc_pages(GFP_KERNEL
, 2);
62 arch_set_page_dat(page
, 2);
63 return (unsigned long *) page_to_phys(page
);
66 void crst_table_free(struct mm_struct
*mm
, unsigned long *table
)
68 free_pages((unsigned long) table
, 2);
71 static void __crst_table_upgrade(void *arg
)
73 struct mm_struct
*mm
= arg
;
75 if (current
->active_mm
== mm
)
80 int crst_table_upgrade(struct mm_struct
*mm
, unsigned long end
)
82 unsigned long *table
, *pgd
;
85 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
86 VM_BUG_ON(mm
->context
.asce_limit
< _REGION2_SIZE
);
89 while (mm
->context
.asce_limit
< end
) {
90 table
= crst_table_alloc(mm
);
95 spin_lock_bh(&mm
->page_table_lock
);
96 pgd
= (unsigned long *) mm
->pgd
;
97 if (mm
->context
.asce_limit
== _REGION2_SIZE
) {
98 crst_table_init(table
, _REGION2_ENTRY_EMPTY
);
99 p4d_populate(mm
, (p4d_t
*) table
, (pud_t
*) pgd
);
100 mm
->pgd
= (pgd_t
*) table
;
101 mm
->context
.asce_limit
= _REGION1_SIZE
;
102 mm
->context
.asce
= __pa(mm
->pgd
) | _ASCE_TABLE_LENGTH
|
103 _ASCE_USER_BITS
| _ASCE_TYPE_REGION2
;
106 crst_table_init(table
, _REGION1_ENTRY_EMPTY
);
107 pgd_populate(mm
, (pgd_t
*) table
, (p4d_t
*) pgd
);
108 mm
->pgd
= (pgd_t
*) table
;
109 mm
->context
.asce_limit
= -PAGE_SIZE
;
110 mm
->context
.asce
= __pa(mm
->pgd
) | _ASCE_TABLE_LENGTH
|
111 _ASCE_USER_BITS
| _ASCE_TYPE_REGION1
;
114 spin_unlock_bh(&mm
->page_table_lock
);
117 on_each_cpu(__crst_table_upgrade
, mm
, 0);
121 void crst_table_downgrade(struct mm_struct
*mm
)
125 /* downgrade should only happen from 3 to 2 levels (compat only) */
126 VM_BUG_ON(mm
->context
.asce_limit
!= _REGION2_SIZE
);
128 if (current
->active_mm
== mm
) {
134 mm
->pgd
= (pgd_t
*) (pgd_val(*pgd
) & _REGION_ENTRY_ORIGIN
);
135 mm
->context
.asce_limit
= _REGION3_SIZE
;
136 mm
->context
.asce
= __pa(mm
->pgd
) | _ASCE_TABLE_LENGTH
|
137 _ASCE_USER_BITS
| _ASCE_TYPE_SEGMENT
;
138 crst_table_free(mm
, (unsigned long *) pgd
);
140 if (current
->active_mm
== mm
)
144 static inline unsigned int atomic_xor_bits(atomic_t
*v
, unsigned int bits
)
146 unsigned int old
, new;
149 old
= atomic_read(v
);
151 } while (atomic_cmpxchg(v
, old
, new) != old
);
157 struct page
*page_table_alloc_pgste(struct mm_struct
*mm
)
162 page
= alloc_page(GFP_KERNEL
);
164 table
= (u64
*)page_to_phys(page
);
165 memset64(table
, _PAGE_INVALID
, PTRS_PER_PTE
);
166 memset64(table
+ PTRS_PER_PTE
, 0, PTRS_PER_PTE
);
171 void page_table_free_pgste(struct page
*page
)
176 #endif /* CONFIG_PGSTE */
179 * page table entry allocation/free routines.
181 unsigned long *page_table_alloc(struct mm_struct
*mm
)
183 unsigned long *table
;
185 unsigned int mask
, bit
;
187 /* Try to get a fragment of a 4K page as a 2K page table */
188 if (!mm_alloc_pgste(mm
)) {
190 spin_lock_bh(&mm
->context
.lock
);
191 if (!list_empty(&mm
->context
.pgtable_list
)) {
192 page
= list_first_entry(&mm
->context
.pgtable_list
,
194 mask
= atomic_read(&page
->_refcount
) >> 24;
195 mask
= (mask
| (mask
>> 4)) & 3;
197 table
= (unsigned long *) page_to_phys(page
);
198 bit
= mask
& 1; /* =1 -> second 2K */
200 table
+= PTRS_PER_PTE
;
201 atomic_xor_bits(&page
->_refcount
,
203 list_del(&page
->lru
);
206 spin_unlock_bh(&mm
->context
.lock
);
210 /* Allocate a fresh page */
211 page
= alloc_page(GFP_KERNEL
);
214 if (!pgtable_page_ctor(page
)) {
218 arch_set_page_dat(page
, 0);
219 /* Initialize page table */
220 table
= (unsigned long *) page_to_phys(page
);
221 if (mm_alloc_pgste(mm
)) {
222 /* Return 4K page table with PGSTEs */
223 atomic_xor_bits(&page
->_refcount
, 3 << 24);
224 memset64((u64
*)table
, _PAGE_INVALID
, PTRS_PER_PTE
);
225 memset64((u64
*)table
+ PTRS_PER_PTE
, 0, PTRS_PER_PTE
);
227 /* Return the first 2K fragment of the page */
228 atomic_xor_bits(&page
->_refcount
, 1 << 24);
229 memset64((u64
*)table
, _PAGE_INVALID
, 2 * PTRS_PER_PTE
);
230 spin_lock_bh(&mm
->context
.lock
);
231 list_add(&page
->lru
, &mm
->context
.pgtable_list
);
232 spin_unlock_bh(&mm
->context
.lock
);
237 void page_table_free(struct mm_struct
*mm
, unsigned long *table
)
240 unsigned int bit
, mask
;
242 page
= pfn_to_page(__pa(table
) >> PAGE_SHIFT
);
243 if (!mm_alloc_pgste(mm
)) {
244 /* Free 2K page table fragment of a 4K page */
245 bit
= (__pa(table
) & ~PAGE_MASK
)/(PTRS_PER_PTE
*sizeof(pte_t
));
246 spin_lock_bh(&mm
->context
.lock
);
247 mask
= atomic_xor_bits(&page
->_refcount
, 1U << (bit
+ 24));
250 list_add(&page
->lru
, &mm
->context
.pgtable_list
);
252 list_del(&page
->lru
);
253 spin_unlock_bh(&mm
->context
.lock
);
257 atomic_xor_bits(&page
->_refcount
, 3U << 24);
260 pgtable_page_dtor(page
);
264 void page_table_free_rcu(struct mmu_gather
*tlb
, unsigned long *table
,
265 unsigned long vmaddr
)
267 struct mm_struct
*mm
;
269 unsigned int bit
, mask
;
272 page
= pfn_to_page(__pa(table
) >> PAGE_SHIFT
);
273 if (mm_alloc_pgste(mm
)) {
274 gmap_unlink(mm
, table
, vmaddr
);
275 table
= (unsigned long *) (__pa(table
) | 3);
276 tlb_remove_table(tlb
, table
);
279 bit
= (__pa(table
) & ~PAGE_MASK
) / (PTRS_PER_PTE
*sizeof(pte_t
));
280 spin_lock_bh(&mm
->context
.lock
);
281 mask
= atomic_xor_bits(&page
->_refcount
, 0x11U
<< (bit
+ 24));
284 list_add_tail(&page
->lru
, &mm
->context
.pgtable_list
);
286 list_del(&page
->lru
);
287 spin_unlock_bh(&mm
->context
.lock
);
288 table
= (unsigned long *) (__pa(table
) | (1U << bit
));
289 tlb_remove_table(tlb
, table
);
292 static void __tlb_remove_table(void *_table
)
294 unsigned int mask
= (unsigned long) _table
& 3;
295 void *table
= (void *)((unsigned long) _table
^ mask
);
296 struct page
*page
= pfn_to_page(__pa(table
) >> PAGE_SHIFT
);
299 case 0: /* pmd, pud, or p4d */
300 free_pages((unsigned long) table
, 2);
302 case 1: /* lower 2K of a 4K page table */
303 case 2: /* higher 2K of a 4K page table */
304 mask
= atomic_xor_bits(&page
->_refcount
, mask
<< (4 + 24));
309 case 3: /* 4K page table with pgstes */
311 atomic_xor_bits(&page
->_refcount
, 3 << 24);
312 pgtable_page_dtor(page
);
318 static void tlb_remove_table_smp_sync(void *arg
)
320 /* Simply deliver the interrupt */
323 static void tlb_remove_table_one(void *table
)
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
329 * It is however sufficient for software page-table walkers that rely
330 * on IRQ disabling. See the comment near struct mmu_table_batch.
332 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
333 __tlb_remove_table(table
);
336 static void tlb_remove_table_rcu(struct rcu_head
*head
)
338 struct mmu_table_batch
*batch
;
341 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
343 for (i
= 0; i
< batch
->nr
; i
++)
344 __tlb_remove_table(batch
->tables
[i
]);
346 free_page((unsigned long)batch
);
349 void tlb_table_flush(struct mmu_gather
*tlb
)
351 struct mmu_table_batch
**batch
= &tlb
->batch
;
354 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
359 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
361 struct mmu_table_batch
**batch
= &tlb
->batch
;
363 tlb
->mm
->context
.flush_mm
= 1;
364 if (*batch
== NULL
) {
365 *batch
= (struct mmu_table_batch
*)
366 __get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
367 if (*batch
== NULL
) {
368 __tlb_flush_mm_lazy(tlb
->mm
);
369 tlb_remove_table_one(table
);
374 (*batch
)->tables
[(*batch
)->nr
++] = table
;
375 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
380 * Base infrastructure required to generate basic asces, region, segment,
381 * and page tables that do not make use of enhanced features like EDAT1.
384 static struct kmem_cache
*base_pgt_cache
;
386 static unsigned long base_pgt_alloc(void)
390 table
= kmem_cache_alloc(base_pgt_cache
, GFP_KERNEL
);
392 memset64(table
, _PAGE_INVALID
, PTRS_PER_PTE
);
393 return (unsigned long) table
;
396 static void base_pgt_free(unsigned long table
)
398 kmem_cache_free(base_pgt_cache
, (void *) table
);
401 static unsigned long base_crst_alloc(unsigned long val
)
405 table
= __get_free_pages(GFP_KERNEL
, CRST_ALLOC_ORDER
);
407 crst_table_init((unsigned long *)table
, val
);
411 static void base_crst_free(unsigned long table
)
413 free_pages(table
, CRST_ALLOC_ORDER
);
416 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
417 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
420 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
422 return (next - 1) < (end - 1) ? next : end; \
425 BASE_ADDR_END_FUNC(page
, _PAGE_SIZE
)
426 BASE_ADDR_END_FUNC(segment
, _SEGMENT_SIZE
)
427 BASE_ADDR_END_FUNC(region3
, _REGION3_SIZE
)
428 BASE_ADDR_END_FUNC(region2
, _REGION2_SIZE
)
429 BASE_ADDR_END_FUNC(region1
, _REGION1_SIZE
)
431 static inline unsigned long base_lra(unsigned long address
)
437 : "=d" (real
) : "a" (address
) : "cc");
441 static int base_page_walk(unsigned long origin
, unsigned long addr
,
442 unsigned long end
, int alloc
)
444 unsigned long *pte
, next
;
448 pte
= (unsigned long *) origin
;
449 pte
+= (addr
& _PAGE_INDEX
) >> _PAGE_SHIFT
;
451 next
= base_page_addr_end(addr
, end
);
452 *pte
= base_lra(addr
);
453 } while (pte
++, addr
= next
, addr
< end
);
457 static int base_segment_walk(unsigned long origin
, unsigned long addr
,
458 unsigned long end
, int alloc
)
460 unsigned long *ste
, next
, table
;
463 ste
= (unsigned long *) origin
;
464 ste
+= (addr
& _SEGMENT_INDEX
) >> _SEGMENT_SHIFT
;
466 next
= base_segment_addr_end(addr
, end
);
467 if (*ste
& _SEGMENT_ENTRY_INVALID
) {
470 table
= base_pgt_alloc();
473 *ste
= table
| _SEGMENT_ENTRY
;
475 table
= *ste
& _SEGMENT_ENTRY_ORIGIN
;
476 rc
= base_page_walk(table
, addr
, next
, alloc
);
480 base_pgt_free(table
);
482 } while (ste
++, addr
= next
, addr
< end
);
486 static int base_region3_walk(unsigned long origin
, unsigned long addr
,
487 unsigned long end
, int alloc
)
489 unsigned long *rtte
, next
, table
;
492 rtte
= (unsigned long *) origin
;
493 rtte
+= (addr
& _REGION3_INDEX
) >> _REGION3_SHIFT
;
495 next
= base_region3_addr_end(addr
, end
);
496 if (*rtte
& _REGION_ENTRY_INVALID
) {
499 table
= base_crst_alloc(_SEGMENT_ENTRY_EMPTY
);
502 *rtte
= table
| _REGION3_ENTRY
;
504 table
= *rtte
& _REGION_ENTRY_ORIGIN
;
505 rc
= base_segment_walk(table
, addr
, next
, alloc
);
509 base_crst_free(table
);
510 } while (rtte
++, addr
= next
, addr
< end
);
514 static int base_region2_walk(unsigned long origin
, unsigned long addr
,
515 unsigned long end
, int alloc
)
517 unsigned long *rste
, next
, table
;
520 rste
= (unsigned long *) origin
;
521 rste
+= (addr
& _REGION2_INDEX
) >> _REGION2_SHIFT
;
523 next
= base_region2_addr_end(addr
, end
);
524 if (*rste
& _REGION_ENTRY_INVALID
) {
527 table
= base_crst_alloc(_REGION3_ENTRY_EMPTY
);
530 *rste
= table
| _REGION2_ENTRY
;
532 table
= *rste
& _REGION_ENTRY_ORIGIN
;
533 rc
= base_region3_walk(table
, addr
, next
, alloc
);
537 base_crst_free(table
);
538 } while (rste
++, addr
= next
, addr
< end
);
542 static int base_region1_walk(unsigned long origin
, unsigned long addr
,
543 unsigned long end
, int alloc
)
545 unsigned long *rfte
, next
, table
;
548 rfte
= (unsigned long *) origin
;
549 rfte
+= (addr
& _REGION1_INDEX
) >> _REGION1_SHIFT
;
551 next
= base_region1_addr_end(addr
, end
);
552 if (*rfte
& _REGION_ENTRY_INVALID
) {
555 table
= base_crst_alloc(_REGION2_ENTRY_EMPTY
);
558 *rfte
= table
| _REGION1_ENTRY
;
560 table
= *rfte
& _REGION_ENTRY_ORIGIN
;
561 rc
= base_region2_walk(table
, addr
, next
, alloc
);
565 base_crst_free(table
);
566 } while (rfte
++, addr
= next
, addr
< end
);
571 * base_asce_free - free asce and tables returned from base_asce_alloc()
572 * @asce: asce to be freed
574 * Frees all region, segment, and page tables that were allocated with a
575 * corresponding base_asce_alloc() call.
577 void base_asce_free(unsigned long asce
)
579 unsigned long table
= asce
& _ASCE_ORIGIN
;
583 switch (asce
& _ASCE_TYPE_MASK
) {
584 case _ASCE_TYPE_SEGMENT
:
585 base_segment_walk(table
, 0, _REGION3_SIZE
, 0);
587 case _ASCE_TYPE_REGION3
:
588 base_region3_walk(table
, 0, _REGION2_SIZE
, 0);
590 case _ASCE_TYPE_REGION2
:
591 base_region2_walk(table
, 0, _REGION1_SIZE
, 0);
593 case _ASCE_TYPE_REGION1
:
594 base_region1_walk(table
, 0, -_PAGE_SIZE
, 0);
597 base_crst_free(table
);
600 static int base_pgt_cache_init(void)
602 static DEFINE_MUTEX(base_pgt_cache_mutex
);
603 unsigned long sz
= _PAGE_TABLE_SIZE
;
607 mutex_lock(&base_pgt_cache_mutex
);
609 base_pgt_cache
= kmem_cache_create("base_pgt", sz
, sz
, 0, NULL
);
610 mutex_unlock(&base_pgt_cache_mutex
);
611 return base_pgt_cache
? 0 : -ENOMEM
;
615 * base_asce_alloc - create kernel mapping without enhanced DAT features
616 * @addr: virtual start address of kernel mapping
617 * @num_pages: number of consecutive pages
619 * Generate an asce, including all required region, segment and page tables,
620 * that can be used to access the virtual kernel mapping. The difference is
621 * that the returned asce does not make use of any enhanced DAT features like
622 * e.g. large pages. This is required for some I/O functions that pass an
623 * asce, like e.g. some service call requests.
625 * Note: the returned asce may NEVER be attached to any cpu. It may only be
626 * used for I/O requests. tlb entries that might result because the
627 * asce was attached to a cpu won't be cleared.
629 unsigned long base_asce_alloc(unsigned long addr
, unsigned long num_pages
)
631 unsigned long asce
, table
, end
;
634 if (base_pgt_cache_init())
636 end
= addr
+ num_pages
* PAGE_SIZE
;
637 if (end
<= _REGION3_SIZE
) {
638 table
= base_crst_alloc(_SEGMENT_ENTRY_EMPTY
);
641 rc
= base_segment_walk(table
, addr
, end
, 1);
642 asce
= table
| _ASCE_TYPE_SEGMENT
| _ASCE_TABLE_LENGTH
;
643 } else if (end
<= _REGION2_SIZE
) {
644 table
= base_crst_alloc(_REGION3_ENTRY_EMPTY
);
647 rc
= base_region3_walk(table
, addr
, end
, 1);
648 asce
= table
| _ASCE_TYPE_REGION3
| _ASCE_TABLE_LENGTH
;
649 } else if (end
<= _REGION1_SIZE
) {
650 table
= base_crst_alloc(_REGION2_ENTRY_EMPTY
);
653 rc
= base_region2_walk(table
, addr
, end
, 1);
654 asce
= table
| _ASCE_TYPE_REGION2
| _ASCE_TABLE_LENGTH
;
656 table
= base_crst_alloc(_REGION1_ENTRY_EMPTY
);
659 rc
= base_region1_walk(table
, addr
, end
, 1);
660 asce
= table
| _ASCE_TYPE_REGION1
| _ASCE_TABLE_LENGTH
;
663 base_asce_free(asce
);