2 * Copyright IBM Corp. 2007, 2011
3 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
6 #include <linux/sched.h>
7 #include <linux/kernel.h>
8 #include <linux/errno.h>
11 #include <linux/swap.h>
12 #include <linux/smp.h>
13 #include <linux/highmem.h>
14 #include <linux/pagemap.h>
15 #include <linux/spinlock.h>
16 #include <linux/module.h>
17 #include <linux/quicklist.h>
18 #include <linux/rcupdate.h>
19 #include <linux/slab.h>
20 #include <linux/swapops.h>
22 #include <asm/pgtable.h>
23 #include <asm/pgalloc.h>
25 #include <asm/tlbflush.h>
26 #include <asm/mmu_context.h>
30 #define FRAG_MASK 0x0f
33 #define FRAG_MASK 0x03
37 unsigned long *crst_table_alloc(struct mm_struct *mm)
39 struct page *page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
43 return (unsigned long *) page_to_phys(page);
46 void crst_table_free(struct mm_struct *mm, unsigned long *table)
48 free_pages((unsigned long) table, ALLOC_ORDER);
52 static void __crst_table_upgrade(void *arg)
54 struct mm_struct *mm = arg;
56 if (current->active_mm == mm) {
63 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
65 unsigned long *table, *pgd;
69 BUG_ON(limit > (1UL << 53));
72 table = crst_table_alloc(mm);
75 spin_lock_bh(&mm->page_table_lock);
76 if (mm->context.asce_limit < limit) {
77 pgd = (unsigned long *) mm->pgd;
78 if (mm->context.asce_limit <= (1UL << 31)) {
79 entry = _REGION3_ENTRY_EMPTY;
80 mm->context.asce_limit = 1UL << 42;
81 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
85 entry = _REGION2_ENTRY_EMPTY;
86 mm->context.asce_limit = 1UL << 53;
87 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
91 crst_table_init(table, entry);
92 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
93 mm->pgd = (pgd_t *) table;
94 mm->task_size = mm->context.asce_limit;
98 spin_unlock_bh(&mm->page_table_lock);
100 crst_table_free(mm, table);
101 if (mm->context.asce_limit < limit)
104 on_each_cpu(__crst_table_upgrade, mm, 0);
108 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
112 if (current->active_mm == mm) {
116 while (mm->context.asce_limit > limit) {
118 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
119 case _REGION_ENTRY_TYPE_R2:
120 mm->context.asce_limit = 1UL << 42;
121 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
125 case _REGION_ENTRY_TYPE_R3:
126 mm->context.asce_limit = 1UL << 31;
127 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
134 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
135 mm->task_size = mm->context.asce_limit;
136 crst_table_free(mm, (unsigned long *) pgd);
138 if (current->active_mm == mm)
146 * gmap_alloc - allocate a guest address space
147 * @mm: pointer to the parent mm_struct
148 * @limit: maximum size of the gmap address space
150 * Returns a guest address space structure.
152 struct gmap *gmap_alloc(struct mm_struct *mm, unsigned long limit)
156 unsigned long *table;
157 unsigned long etype, atype;
159 if (limit < (1UL << 31)) {
160 limit = (1UL << 31) - 1;
161 atype = _ASCE_TYPE_SEGMENT;
162 etype = _SEGMENT_ENTRY_EMPTY;
163 } else if (limit < (1UL << 42)) {
164 limit = (1UL << 42) - 1;
165 atype = _ASCE_TYPE_REGION3;
166 etype = _REGION3_ENTRY_EMPTY;
167 } else if (limit < (1UL << 53)) {
168 limit = (1UL << 53) - 1;
169 atype = _ASCE_TYPE_REGION2;
170 etype = _REGION2_ENTRY_EMPTY;
173 atype = _ASCE_TYPE_REGION1;
174 etype = _REGION1_ENTRY_EMPTY;
176 gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL);
179 INIT_LIST_HEAD(&gmap->crst_list);
180 INIT_RADIX_TREE(&gmap->guest_to_host, GFP_KERNEL);
181 INIT_RADIX_TREE(&gmap->host_to_guest, GFP_ATOMIC);
182 spin_lock_init(&gmap->guest_table_lock);
184 page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
188 list_add(&page->lru, &gmap->crst_list);
189 table = (unsigned long *) page_to_phys(page);
190 crst_table_init(table, etype);
192 gmap->asce = atype | _ASCE_TABLE_LENGTH |
193 _ASCE_USER_BITS | __pa(table);
194 gmap->asce_end = limit;
195 down_write(&mm->mmap_sem);
196 list_add(&gmap->list, &mm->context.gmap_list);
197 up_write(&mm->mmap_sem);
205 EXPORT_SYMBOL_GPL(gmap_alloc);
207 static void gmap_flush_tlb(struct gmap *gmap)
209 if (MACHINE_HAS_IDTE)
210 __tlb_flush_asce(gmap->mm, gmap->asce);
212 __tlb_flush_global();
215 static void gmap_radix_tree_free(struct radix_tree_root *root)
217 struct radix_tree_iter iter;
218 unsigned long indices[16];
223 /* A radix tree is freed by deleting all of its entries */
227 radix_tree_for_each_slot(slot, root, &iter, index) {
228 indices[nr] = iter.index;
232 for (i = 0; i < nr; i++) {
234 radix_tree_delete(root, index);
240 * gmap_free - free a guest address space
241 * @gmap: pointer to the guest address space structure
243 void gmap_free(struct gmap *gmap)
245 struct page *page, *next;
248 if (MACHINE_HAS_IDTE)
249 __tlb_flush_asce(gmap->mm, gmap->asce);
251 __tlb_flush_global();
253 /* Free all segment & region tables. */
254 list_for_each_entry_safe(page, next, &gmap->crst_list, lru)
255 __free_pages(page, ALLOC_ORDER);
256 gmap_radix_tree_free(&gmap->guest_to_host);
257 gmap_radix_tree_free(&gmap->host_to_guest);
258 down_write(&gmap->mm->mmap_sem);
259 list_del(&gmap->list);
260 up_write(&gmap->mm->mmap_sem);
263 EXPORT_SYMBOL_GPL(gmap_free);
266 * gmap_enable - switch primary space to the guest address space
267 * @gmap: pointer to the guest address space structure
269 void gmap_enable(struct gmap *gmap)
271 S390_lowcore.gmap = (unsigned long) gmap;
273 EXPORT_SYMBOL_GPL(gmap_enable);
276 * gmap_disable - switch back to the standard primary address space
277 * @gmap: pointer to the guest address space structure
279 void gmap_disable(struct gmap *gmap)
281 S390_lowcore.gmap = 0UL;
283 EXPORT_SYMBOL_GPL(gmap_disable);
286 * gmap_alloc_table is assumed to be called with mmap_sem held
288 static int gmap_alloc_table(struct gmap *gmap, unsigned long *table,
289 unsigned long init, unsigned long gaddr)
294 /* since we dont free the gmap table until gmap_free we can unlock */
295 page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
298 new = (unsigned long *) page_to_phys(page);
299 crst_table_init(new, init);
300 spin_lock(&gmap->mm->page_table_lock);
301 if (*table & _REGION_ENTRY_INVALID) {
302 list_add(&page->lru, &gmap->crst_list);
303 *table = (unsigned long) new | _REGION_ENTRY_LENGTH |
304 (*table & _REGION_ENTRY_TYPE_MASK);
308 spin_unlock(&gmap->mm->page_table_lock);
310 __free_pages(page, ALLOC_ORDER);
315 * __gmap_segment_gaddr - find virtual address from segment pointer
316 * @entry: pointer to a segment table entry in the guest address space
318 * Returns the virtual address in the guest address space for the segment
320 static unsigned long __gmap_segment_gaddr(unsigned long *entry)
323 unsigned long offset;
325 offset = (unsigned long) entry / sizeof(unsigned long);
326 offset = (offset & (PTRS_PER_PMD - 1)) * PMD_SIZE;
327 page = pmd_to_page((pmd_t *) entry);
328 return page->index + offset;
332 * __gmap_unlink_by_vmaddr - unlink a single segment via a host address
333 * @gmap: pointer to the guest address space structure
334 * @vmaddr: address in the host process address space
336 * Returns 1 if a TLB flush is required
338 static int __gmap_unlink_by_vmaddr(struct gmap *gmap, unsigned long vmaddr)
340 unsigned long *entry;
343 spin_lock(&gmap->guest_table_lock);
344 entry = radix_tree_delete(&gmap->host_to_guest, vmaddr >> PMD_SHIFT);
346 flush = (*entry != _SEGMENT_ENTRY_INVALID);
347 *entry = _SEGMENT_ENTRY_INVALID;
349 spin_unlock(&gmap->guest_table_lock);
354 * __gmap_unmap_by_gaddr - unmap a single segment via a guest address
355 * @gmap: pointer to the guest address space structure
356 * @gaddr: address in the guest address space
358 * Returns 1 if a TLB flush is required
360 static int __gmap_unmap_by_gaddr(struct gmap *gmap, unsigned long gaddr)
362 unsigned long vmaddr;
364 vmaddr = (unsigned long) radix_tree_delete(&gmap->guest_to_host,
366 return vmaddr ? __gmap_unlink_by_vmaddr(gmap, vmaddr) : 0;
370 * gmap_unmap_segment - unmap segment from the guest address space
371 * @gmap: pointer to the guest address space structure
372 * @to: address in the guest address space
373 * @len: length of the memory area to unmap
375 * Returns 0 if the unmap succeeded, -EINVAL if not.
377 int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
382 if ((to | len) & (PMD_SIZE - 1))
384 if (len == 0 || to + len < to)
388 down_write(&gmap->mm->mmap_sem);
389 for (off = 0; off < len; off += PMD_SIZE)
390 flush |= __gmap_unmap_by_gaddr(gmap, to + off);
391 up_write(&gmap->mm->mmap_sem);
393 gmap_flush_tlb(gmap);
396 EXPORT_SYMBOL_GPL(gmap_unmap_segment);
399 * gmap_mmap_segment - map a segment to the guest address space
400 * @gmap: pointer to the guest address space structure
401 * @from: source address in the parent address space
402 * @to: target address in the guest address space
403 * @len: length of the memory area to map
405 * Returns 0 if the mmap succeeded, -EINVAL or -ENOMEM if not.
407 int gmap_map_segment(struct gmap *gmap, unsigned long from,
408 unsigned long to, unsigned long len)
413 if ((from | to | len) & (PMD_SIZE - 1))
415 if (len == 0 || from + len < from || to + len < to ||
416 from + len > TASK_MAX_SIZE || to + len > gmap->asce_end)
420 down_write(&gmap->mm->mmap_sem);
421 for (off = 0; off < len; off += PMD_SIZE) {
422 /* Remove old translation */
423 flush |= __gmap_unmap_by_gaddr(gmap, to + off);
424 /* Store new translation */
425 if (radix_tree_insert(&gmap->guest_to_host,
426 (to + off) >> PMD_SHIFT,
427 (void *) from + off))
430 up_write(&gmap->mm->mmap_sem);
432 gmap_flush_tlb(gmap);
435 gmap_unmap_segment(gmap, to, len);
438 EXPORT_SYMBOL_GPL(gmap_map_segment);
441 * __gmap_translate - translate a guest address to a user space address
442 * @gmap: pointer to guest mapping meta data structure
443 * @gaddr: guest address
445 * Returns user space address which corresponds to the guest address or
446 * -EFAULT if no such mapping exists.
447 * This function does not establish potentially missing page table entries.
448 * The mmap_sem of the mm that belongs to the address space must be held
449 * when this function gets called.
451 unsigned long __gmap_translate(struct gmap *gmap, unsigned long gaddr)
453 unsigned long vmaddr;
455 vmaddr = (unsigned long)
456 radix_tree_lookup(&gmap->guest_to_host, gaddr >> PMD_SHIFT);
457 return vmaddr ? (vmaddr | (gaddr & ~PMD_MASK)) : -EFAULT;
459 EXPORT_SYMBOL_GPL(__gmap_translate);
462 * gmap_translate - translate a guest address to a user space address
463 * @gmap: pointer to guest mapping meta data structure
464 * @gaddr: guest address
466 * Returns user space address which corresponds to the guest address or
467 * -EFAULT if no such mapping exists.
468 * This function does not establish potentially missing page table entries.
470 unsigned long gmap_translate(struct gmap *gmap, unsigned long gaddr)
474 down_read(&gmap->mm->mmap_sem);
475 rc = __gmap_translate(gmap, gaddr);
476 up_read(&gmap->mm->mmap_sem);
479 EXPORT_SYMBOL_GPL(gmap_translate);
482 * gmap_unlink - disconnect a page table from the gmap shadow tables
483 * @gmap: pointer to guest mapping meta data structure
484 * @table: pointer to the host page table
485 * @vmaddr: vm address associated with the host page table
487 static void gmap_unlink(struct mm_struct *mm, unsigned long *table,
488 unsigned long vmaddr)
493 list_for_each_entry(gmap, &mm->context.gmap_list, list) {
494 flush = __gmap_unlink_by_vmaddr(gmap, vmaddr);
496 gmap_flush_tlb(gmap);
501 * gmap_link - set up shadow page tables to connect a host to a guest address
502 * @gmap: pointer to guest mapping meta data structure
503 * @gaddr: guest address
504 * @vmaddr: vm address
506 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
507 * if the vm address is already mapped to a different guest segment.
508 * The mmap_sem of the mm that belongs to the address space must be held
509 * when this function gets called.
511 int __gmap_link(struct gmap *gmap, unsigned long gaddr, unsigned long vmaddr)
513 struct mm_struct *mm;
514 unsigned long *table;
521 /* Create higher level tables in the gmap page table */
523 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION1) {
524 table += (gaddr >> 53) & 0x7ff;
525 if ((*table & _REGION_ENTRY_INVALID) &&
526 gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY,
527 gaddr & 0xffe0000000000000))
529 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
531 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION2) {
532 table += (gaddr >> 42) & 0x7ff;
533 if ((*table & _REGION_ENTRY_INVALID) &&
534 gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY,
535 gaddr & 0xfffffc0000000000))
537 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
539 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION3) {
540 table += (gaddr >> 31) & 0x7ff;
541 if ((*table & _REGION_ENTRY_INVALID) &&
542 gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY,
543 gaddr & 0xffffffff80000000))
545 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
547 table += (gaddr >> 20) & 0x7ff;
548 /* Walk the parent mm page table */
550 pgd = pgd_offset(mm, vmaddr);
551 VM_BUG_ON(pgd_none(*pgd));
552 pud = pud_offset(pgd, vmaddr);
553 VM_BUG_ON(pud_none(*pud));
554 pmd = pmd_offset(pud, vmaddr);
555 VM_BUG_ON(pmd_none(*pmd));
556 /* large pmds cannot yet be handled */
559 /* Link gmap segment table entry location to page table. */
560 rc = radix_tree_preload(GFP_KERNEL);
563 ptl = pmd_lock(mm, pmd);
564 spin_lock(&gmap->guest_table_lock);
565 if (*table == _SEGMENT_ENTRY_INVALID) {
566 rc = radix_tree_insert(&gmap->host_to_guest,
567 vmaddr >> PMD_SHIFT, table);
569 *table = pmd_val(*pmd);
572 spin_unlock(&gmap->guest_table_lock);
574 radix_tree_preload_end();
579 * gmap_fault - resolve a fault on a guest address
580 * @gmap: pointer to guest mapping meta data structure
581 * @gaddr: guest address
582 * @fault_flags: flags to pass down to handle_mm_fault()
584 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
585 * if the vm address is already mapped to a different guest segment.
587 int gmap_fault(struct gmap *gmap, unsigned long gaddr,
588 unsigned int fault_flags)
590 unsigned long vmaddr;
593 down_read(&gmap->mm->mmap_sem);
594 vmaddr = __gmap_translate(gmap, gaddr);
595 if (IS_ERR_VALUE(vmaddr)) {
599 if (fixup_user_fault(current, gmap->mm, vmaddr, fault_flags)) {
603 rc = __gmap_link(gmap, gaddr, vmaddr);
605 up_read(&gmap->mm->mmap_sem);
608 EXPORT_SYMBOL_GPL(gmap_fault);
610 static void gmap_zap_swap_entry(swp_entry_t entry, struct mm_struct *mm)
612 if (!non_swap_entry(entry))
613 dec_mm_counter(mm, MM_SWAPENTS);
614 else if (is_migration_entry(entry)) {
615 struct page *page = migration_entry_to_page(entry);
618 dec_mm_counter(mm, MM_ANONPAGES);
620 dec_mm_counter(mm, MM_FILEPAGES);
622 free_swap_and_cache(entry);
626 * this function is assumed to be called with mmap_sem held
628 void __gmap_zap(struct gmap *gmap, unsigned long gaddr)
630 unsigned long vmaddr, ptev, pgstev;
635 /* Find the vm address for the guest address */
636 vmaddr = (unsigned long) radix_tree_lookup(&gmap->guest_to_host,
640 vmaddr |= gaddr & ~PMD_MASK;
641 /* Get pointer to the page table entry */
642 ptep = get_locked_pte(gmap->mm, vmaddr, &ptl);
648 /* Zap unused and logically-zero pages */
649 pgste = pgste_get_lock(ptep);
650 pgstev = pgste_val(pgste);
652 if (((pgstev & _PGSTE_GPS_USAGE_MASK) == _PGSTE_GPS_USAGE_UNUSED) ||
653 ((pgstev & _PGSTE_GPS_ZERO) && (ptev & _PAGE_INVALID))) {
654 gmap_zap_swap_entry(pte_to_swp_entry(pte), gmap->mm);
655 pte_clear(gmap->mm, vmaddr, ptep);
657 pgste_set_unlock(ptep, pgste);
659 pte_unmap_unlock(*ptep, ptl);
661 EXPORT_SYMBOL_GPL(__gmap_zap);
663 void gmap_discard(struct gmap *gmap, unsigned long from, unsigned long to)
665 unsigned long gaddr, vmaddr, size;
666 struct vm_area_struct *vma;
668 down_read(&gmap->mm->mmap_sem);
669 for (gaddr = from; gaddr < to;
670 gaddr = (gaddr + PMD_SIZE) & PMD_MASK) {
671 /* Find the vm address for the guest address */
672 vmaddr = (unsigned long)
673 radix_tree_lookup(&gmap->guest_to_host,
677 vmaddr |= gaddr & ~PMD_MASK;
678 /* Find vma in the parent mm */
679 vma = find_vma(gmap->mm, vmaddr);
680 size = min(to - gaddr, PMD_SIZE - (gaddr & ~PMD_MASK));
681 zap_page_range(vma, vmaddr, size, NULL);
683 up_read(&gmap->mm->mmap_sem);
685 EXPORT_SYMBOL_GPL(gmap_discard);
687 static LIST_HEAD(gmap_notifier_list);
688 static DEFINE_SPINLOCK(gmap_notifier_lock);
691 * gmap_register_ipte_notifier - register a pte invalidation callback
692 * @nb: pointer to the gmap notifier block
694 void gmap_register_ipte_notifier(struct gmap_notifier *nb)
696 spin_lock(&gmap_notifier_lock);
697 list_add(&nb->list, &gmap_notifier_list);
698 spin_unlock(&gmap_notifier_lock);
700 EXPORT_SYMBOL_GPL(gmap_register_ipte_notifier);
703 * gmap_unregister_ipte_notifier - remove a pte invalidation callback
704 * @nb: pointer to the gmap notifier block
706 void gmap_unregister_ipte_notifier(struct gmap_notifier *nb)
708 spin_lock(&gmap_notifier_lock);
709 list_del_init(&nb->list);
710 spin_unlock(&gmap_notifier_lock);
712 EXPORT_SYMBOL_GPL(gmap_unregister_ipte_notifier);
715 * gmap_ipte_notify - mark a range of ptes for invalidation notification
716 * @gmap: pointer to guest mapping meta data structure
717 * @gaddr: virtual address in the guest address space
720 * Returns 0 if for each page in the given range a gmap mapping exists and
721 * the invalidation notification could be set. If the gmap mapping is missing
722 * for one or more pages -EFAULT is returned. If no memory could be allocated
723 * -ENOMEM is returned. This function establishes missing page table entries.
725 int gmap_ipte_notify(struct gmap *gmap, unsigned long gaddr, unsigned long len)
733 if ((gaddr & ~PAGE_MASK) || (len & ~PAGE_MASK))
735 down_read(&gmap->mm->mmap_sem);
737 /* Convert gmap address and connect the page tables */
738 addr = __gmap_translate(gmap, gaddr);
739 if (IS_ERR_VALUE(addr)) {
743 /* Get the page mapped */
744 if (fixup_user_fault(current, gmap->mm, addr, FAULT_FLAG_WRITE)) {
748 rc = __gmap_link(gmap, gaddr, addr);
751 /* Walk the process page table, lock and get pte pointer */
752 ptep = get_locked_pte(gmap->mm, addr, &ptl);
755 /* Set notification bit in the pgste of the pte */
757 if ((pte_val(entry) & (_PAGE_INVALID | _PAGE_PROTECT)) == 0) {
758 pgste = pgste_get_lock(ptep);
759 pgste_val(pgste) |= PGSTE_IN_BIT;
760 pgste_set_unlock(ptep, pgste);
766 up_read(&gmap->mm->mmap_sem);
769 EXPORT_SYMBOL_GPL(gmap_ipte_notify);
772 * gmap_do_ipte_notify - call all invalidation callbacks for a specific pte.
773 * @mm: pointer to the process mm_struct
774 * @addr: virtual address in the process address space
775 * @pte: pointer to the page table entry
777 * This function is assumed to be called with the page table lock held
778 * for the pte to notify.
780 void gmap_do_ipte_notify(struct mm_struct *mm, unsigned long vmaddr, pte_t *pte)
782 unsigned long offset, gaddr;
783 unsigned long *table;
784 struct gmap_notifier *nb;
787 offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
788 offset = offset * (4096 / sizeof(pte_t));
789 spin_lock(&gmap_notifier_lock);
790 list_for_each_entry(gmap, &mm->context.gmap_list, list) {
791 table = radix_tree_lookup(&gmap->host_to_guest,
792 vmaddr >> PMD_SHIFT);
795 gaddr = __gmap_segment_gaddr(table) + offset;
796 list_for_each_entry(nb, &gmap_notifier_list, list)
797 nb->notifier_call(gmap, gaddr);
799 spin_unlock(&gmap_notifier_lock);
801 EXPORT_SYMBOL_GPL(gmap_do_ipte_notify);
803 static inline int page_table_with_pgste(struct page *page)
805 return atomic_read(&page->_mapcount) == 0;
808 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm)
811 unsigned long *table;
813 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
816 if (!pgtable_page_ctor(page)) {
820 atomic_set(&page->_mapcount, 0);
821 table = (unsigned long *) page_to_phys(page);
822 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
823 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
827 static inline void page_table_free_pgste(unsigned long *table)
831 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
832 pgtable_page_dtor(page);
833 atomic_set(&page->_mapcount, -1);
837 static inline unsigned long page_table_reset_pte(struct mm_struct *mm, pmd_t *pmd,
838 unsigned long addr, unsigned long end, bool init_skey)
840 pte_t *start_pte, *pte;
844 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
847 pgste = pgste_get_lock(pte);
848 pgste_val(pgste) &= ~_PGSTE_GPS_USAGE_MASK;
850 unsigned long address;
852 pgste_val(pgste) &= ~(PGSTE_ACC_BITS | PGSTE_FP_BIT |
853 PGSTE_GR_BIT | PGSTE_GC_BIT);
855 /* skip invalid and not writable pages */
856 if (pte_val(*pte) & _PAGE_INVALID ||
857 !(pte_val(*pte) & _PAGE_WRITE)) {
858 pgste_set_unlock(pte, pgste);
862 address = pte_val(*pte) & PAGE_MASK;
863 page_set_storage_key(address, PAGE_DEFAULT_KEY, 1);
865 pgste_set_unlock(pte, pgste);
866 } while (pte++, addr += PAGE_SIZE, addr != end);
867 pte_unmap_unlock(start_pte, ptl);
872 static inline unsigned long page_table_reset_pmd(struct mm_struct *mm, pud_t *pud,
873 unsigned long addr, unsigned long end, bool init_skey)
878 pmd = pmd_offset(pud, addr);
880 next = pmd_addr_end(addr, end);
881 if (pmd_none_or_clear_bad(pmd))
883 next = page_table_reset_pte(mm, pmd, addr, next, init_skey);
884 } while (pmd++, addr = next, addr != end);
889 static inline unsigned long page_table_reset_pud(struct mm_struct *mm, pgd_t *pgd,
890 unsigned long addr, unsigned long end, bool init_skey)
895 pud = pud_offset(pgd, addr);
897 next = pud_addr_end(addr, end);
898 if (pud_none_or_clear_bad(pud))
900 next = page_table_reset_pmd(mm, pud, addr, next, init_skey);
901 } while (pud++, addr = next, addr != end);
906 void page_table_reset_pgste(struct mm_struct *mm, unsigned long start,
907 unsigned long end, bool init_skey)
909 unsigned long addr, next;
912 down_write(&mm->mmap_sem);
913 if (init_skey && mm_use_skey(mm))
916 pgd = pgd_offset(mm, addr);
918 next = pgd_addr_end(addr, end);
919 if (pgd_none_or_clear_bad(pgd))
921 next = page_table_reset_pud(mm, pgd, addr, next, init_skey);
922 } while (pgd++, addr = next, addr != end);
924 current->mm->context.use_skey = 1;
926 up_write(&mm->mmap_sem);
928 EXPORT_SYMBOL(page_table_reset_pgste);
930 int set_guest_storage_key(struct mm_struct *mm, unsigned long addr,
931 unsigned long key, bool nq)
937 down_read(&mm->mmap_sem);
939 ptep = get_locked_pte(current->mm, addr, &ptl);
940 if (unlikely(!ptep)) {
941 up_read(&mm->mmap_sem);
944 if (!(pte_val(*ptep) & _PAGE_INVALID) &&
945 (pte_val(*ptep) & _PAGE_PROTECT)) {
946 pte_unmap_unlock(*ptep, ptl);
947 if (fixup_user_fault(current, mm, addr, FAULT_FLAG_WRITE)) {
948 up_read(&mm->mmap_sem);
954 new = old = pgste_get_lock(ptep);
955 pgste_val(new) &= ~(PGSTE_GR_BIT | PGSTE_GC_BIT |
956 PGSTE_ACC_BITS | PGSTE_FP_BIT);
957 pgste_val(new) |= (key & (_PAGE_CHANGED | _PAGE_REFERENCED)) << 48;
958 pgste_val(new) |= (key & (_PAGE_ACC_BITS | _PAGE_FP_BIT)) << 56;
959 if (!(pte_val(*ptep) & _PAGE_INVALID)) {
960 unsigned long address, bits, skey;
962 address = pte_val(*ptep) & PAGE_MASK;
963 skey = (unsigned long) page_get_storage_key(address);
964 bits = skey & (_PAGE_CHANGED | _PAGE_REFERENCED);
965 skey = key & (_PAGE_ACC_BITS | _PAGE_FP_BIT);
966 /* Set storage key ACC and FP */
967 page_set_storage_key(address, skey, !nq);
968 /* Merge host changed & referenced into pgste */
969 pgste_val(new) |= bits << 52;
971 /* changing the guest storage key is considered a change of the page */
972 if ((pgste_val(new) ^ pgste_val(old)) &
973 (PGSTE_ACC_BITS | PGSTE_FP_BIT | PGSTE_GR_BIT | PGSTE_GC_BIT))
974 pgste_val(new) |= PGSTE_UC_BIT;
976 pgste_set_unlock(ptep, new);
977 pte_unmap_unlock(*ptep, ptl);
978 up_read(&mm->mmap_sem);
981 EXPORT_SYMBOL(set_guest_storage_key);
983 #else /* CONFIG_PGSTE */
985 static inline int page_table_with_pgste(struct page *page)
990 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm)
995 void page_table_reset_pgste(struct mm_struct *mm, unsigned long start,
996 unsigned long end, bool init_skey)
1000 static inline void page_table_free_pgste(unsigned long *table)
1004 static inline void gmap_unlink(struct mm_struct *mm, unsigned long *table,
1005 unsigned long vmaddr)
1009 #endif /* CONFIG_PGSTE */
1011 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
1013 unsigned int old, new;
1016 old = atomic_read(v);
1018 } while (atomic_cmpxchg(v, old, new) != old);
1023 * page table entry allocation/free routines.
1025 unsigned long *page_table_alloc(struct mm_struct *mm)
1027 unsigned long *uninitialized_var(table);
1028 struct page *uninitialized_var(page);
1029 unsigned int mask, bit;
1031 if (mm_has_pgste(mm))
1032 return page_table_alloc_pgste(mm);
1033 /* Allocate fragments of a 4K page as 1K/2K page table */
1034 spin_lock_bh(&mm->context.list_lock);
1036 if (!list_empty(&mm->context.pgtable_list)) {
1037 page = list_first_entry(&mm->context.pgtable_list,
1039 table = (unsigned long *) page_to_phys(page);
1040 mask = atomic_read(&page->_mapcount);
1041 mask = mask | (mask >> 4);
1043 if ((mask & FRAG_MASK) == FRAG_MASK) {
1044 spin_unlock_bh(&mm->context.list_lock);
1045 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
1048 if (!pgtable_page_ctor(page)) {
1052 atomic_set(&page->_mapcount, 1);
1053 table = (unsigned long *) page_to_phys(page);
1054 clear_table(table, _PAGE_INVALID, PAGE_SIZE);
1055 spin_lock_bh(&mm->context.list_lock);
1056 list_add(&page->lru, &mm->context.pgtable_list);
1058 for (bit = 1; mask & bit; bit <<= 1)
1059 table += PTRS_PER_PTE;
1060 mask = atomic_xor_bits(&page->_mapcount, bit);
1061 if ((mask & FRAG_MASK) == FRAG_MASK)
1062 list_del(&page->lru);
1064 spin_unlock_bh(&mm->context.list_lock);
1068 void page_table_free(struct mm_struct *mm, unsigned long *table)
1071 unsigned int bit, mask;
1073 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1074 if (page_table_with_pgste(page))
1075 return page_table_free_pgste(table);
1076 /* Free 1K/2K page table fragment of a 4K page */
1077 bit = 1 << ((__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)));
1078 spin_lock_bh(&mm->context.list_lock);
1079 if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
1080 list_del(&page->lru);
1081 mask = atomic_xor_bits(&page->_mapcount, bit);
1082 if (mask & FRAG_MASK)
1083 list_add(&page->lru, &mm->context.pgtable_list);
1084 spin_unlock_bh(&mm->context.list_lock);
1086 pgtable_page_dtor(page);
1087 atomic_set(&page->_mapcount, -1);
1092 static void __page_table_free_rcu(void *table, unsigned bit)
1096 if (bit == FRAG_MASK)
1097 return page_table_free_pgste(table);
1098 /* Free 1K/2K page table fragment of a 4K page */
1099 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1100 if (atomic_xor_bits(&page->_mapcount, bit) == 0) {
1101 pgtable_page_dtor(page);
1102 atomic_set(&page->_mapcount, -1);
1107 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
1108 unsigned long vmaddr)
1110 struct mm_struct *mm;
1112 unsigned int bit, mask;
1115 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1116 if (page_table_with_pgste(page)) {
1117 gmap_unlink(mm, table, vmaddr);
1118 table = (unsigned long *) (__pa(table) | FRAG_MASK);
1119 tlb_remove_table(tlb, table);
1122 bit = 1 << ((__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)));
1123 spin_lock_bh(&mm->context.list_lock);
1124 if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
1125 list_del(&page->lru);
1126 mask = atomic_xor_bits(&page->_mapcount, bit | (bit << 4));
1127 if (mask & FRAG_MASK)
1128 list_add_tail(&page->lru, &mm->context.pgtable_list);
1129 spin_unlock_bh(&mm->context.list_lock);
1130 table = (unsigned long *) (__pa(table) | (bit << 4));
1131 tlb_remove_table(tlb, table);
1134 static void __tlb_remove_table(void *_table)
1136 const unsigned long mask = (FRAG_MASK << 4) | FRAG_MASK;
1137 void *table = (void *)((unsigned long) _table & ~mask);
1138 unsigned type = (unsigned long) _table & mask;
1141 __page_table_free_rcu(table, type);
1143 free_pages((unsigned long) table, ALLOC_ORDER);
1146 static void tlb_remove_table_smp_sync(void *arg)
1148 /* Simply deliver the interrupt */
1151 static void tlb_remove_table_one(void *table)
1154 * This isn't an RCU grace period and hence the page-tables cannot be
1155 * assumed to be actually RCU-freed.
1157 * It is however sufficient for software page-table walkers that rely
1158 * on IRQ disabling. See the comment near struct mmu_table_batch.
1160 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
1161 __tlb_remove_table(table);
1164 static void tlb_remove_table_rcu(struct rcu_head *head)
1166 struct mmu_table_batch *batch;
1169 batch = container_of(head, struct mmu_table_batch, rcu);
1171 for (i = 0; i < batch->nr; i++)
1172 __tlb_remove_table(batch->tables[i]);
1174 free_page((unsigned long)batch);
1177 void tlb_table_flush(struct mmu_gather *tlb)
1179 struct mmu_table_batch **batch = &tlb->batch;
1182 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
1187 void tlb_remove_table(struct mmu_gather *tlb, void *table)
1189 struct mmu_table_batch **batch = &tlb->batch;
1191 tlb->mm->context.flush_mm = 1;
1192 if (*batch == NULL) {
1193 *batch = (struct mmu_table_batch *)
1194 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
1195 if (*batch == NULL) {
1196 __tlb_flush_mm_lazy(tlb->mm);
1197 tlb_remove_table_one(table);
1202 (*batch)->tables[(*batch)->nr++] = table;
1203 if ((*batch)->nr == MAX_TABLE_BATCH)
1207 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1208 static inline void thp_split_vma(struct vm_area_struct *vma)
1212 for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE)
1213 follow_page(vma, addr, FOLL_SPLIT);
1216 static inline void thp_split_mm(struct mm_struct *mm)
1218 struct vm_area_struct *vma;
1220 for (vma = mm->mmap; vma != NULL; vma = vma->vm_next) {
1222 vma->vm_flags &= ~VM_HUGEPAGE;
1223 vma->vm_flags |= VM_NOHUGEPAGE;
1225 mm->def_flags |= VM_NOHUGEPAGE;
1228 static inline void thp_split_mm(struct mm_struct *mm)
1231 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1233 static unsigned long page_table_realloc_pmd(struct mmu_gather *tlb,
1234 struct mm_struct *mm, pud_t *pud,
1235 unsigned long addr, unsigned long end)
1237 unsigned long next, *table, *new;
1242 pmd = pmd_offset(pud, addr);
1244 next = pmd_addr_end(addr, end);
1246 if (pmd_none_or_clear_bad(pmd))
1248 table = (unsigned long *) pmd_deref(*pmd);
1249 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1250 if (page_table_with_pgste(page))
1252 /* Allocate new page table with pgstes */
1253 new = page_table_alloc_pgste(mm);
1257 ptl = pmd_lock(mm, pmd);
1258 if (likely((unsigned long *) pmd_deref(*pmd) == table)) {
1259 /* Nuke pmd entry pointing to the "short" page table */
1260 pmdp_flush_lazy(mm, addr, pmd);
1262 /* Copy ptes from old table to new table */
1263 memcpy(new, table, PAGE_SIZE/2);
1264 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
1265 /* Establish new table */
1266 pmd_populate(mm, pmd, (pte_t *) new);
1267 /* Free old table with rcu, there might be a walker! */
1268 page_table_free_rcu(tlb, table, addr);
1273 page_table_free_pgste(new);
1276 } while (pmd++, addr = next, addr != end);
1281 static unsigned long page_table_realloc_pud(struct mmu_gather *tlb,
1282 struct mm_struct *mm, pgd_t *pgd,
1283 unsigned long addr, unsigned long end)
1288 pud = pud_offset(pgd, addr);
1290 next = pud_addr_end(addr, end);
1291 if (pud_none_or_clear_bad(pud))
1293 next = page_table_realloc_pmd(tlb, mm, pud, addr, next);
1294 if (unlikely(IS_ERR_VALUE(next)))
1296 } while (pud++, addr = next, addr != end);
1301 static unsigned long page_table_realloc(struct mmu_gather *tlb, struct mm_struct *mm,
1302 unsigned long addr, unsigned long end)
1307 pgd = pgd_offset(mm, addr);
1309 next = pgd_addr_end(addr, end);
1310 if (pgd_none_or_clear_bad(pgd))
1312 next = page_table_realloc_pud(tlb, mm, pgd, addr, next);
1313 if (unlikely(IS_ERR_VALUE(next)))
1315 } while (pgd++, addr = next, addr != end);
1321 * switch on pgstes for its userspace process (for kvm)
1323 int s390_enable_sie(void)
1325 struct task_struct *tsk = current;
1326 struct mm_struct *mm = tsk->mm;
1327 struct mmu_gather tlb;
1329 /* Do we have pgstes? if yes, we are done */
1330 if (mm_has_pgste(tsk->mm))
1333 down_write(&mm->mmap_sem);
1334 /* split thp mappings and disable thp for future mappings */
1336 /* Reallocate the page tables with pgstes */
1337 tlb_gather_mmu(&tlb, mm, 0, TASK_SIZE);
1338 if (!page_table_realloc(&tlb, mm, 0, TASK_SIZE))
1339 mm->context.has_pgste = 1;
1340 tlb_finish_mmu(&tlb, 0, TASK_SIZE);
1341 up_write(&mm->mmap_sem);
1342 return mm->context.has_pgste ? 0 : -ENOMEM;
1344 EXPORT_SYMBOL_GPL(s390_enable_sie);
1347 * Enable storage key handling from now on and initialize the storage
1348 * keys with the default key.
1350 void s390_enable_skey(void)
1352 page_table_reset_pgste(current->mm, 0, TASK_SIZE, true);
1354 EXPORT_SYMBOL_GPL(s390_enable_skey);
1357 * Test and reset if a guest page is dirty
1359 bool gmap_test_and_clear_dirty(unsigned long address, struct gmap *gmap)
1365 pte = get_locked_pte(gmap->mm, address, &ptl);
1369 if (ptep_test_and_clear_user_dirty(gmap->mm, address, pte))
1375 EXPORT_SYMBOL_GPL(gmap_test_and_clear_dirty);
1377 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1378 int pmdp_clear_flush_young(struct vm_area_struct *vma, unsigned long address,
1381 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1382 /* No need to flush TLB
1383 * On s390 reference bits are in storage key and never in TLB */
1384 return pmdp_test_and_clear_young(vma, address, pmdp);
1387 int pmdp_set_access_flags(struct vm_area_struct *vma,
1388 unsigned long address, pmd_t *pmdp,
1389 pmd_t entry, int dirty)
1391 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1393 entry = pmd_mkyoung(entry);
1395 entry = pmd_mkdirty(entry);
1396 if (pmd_same(*pmdp, entry))
1398 pmdp_invalidate(vma, address, pmdp);
1399 set_pmd_at(vma->vm_mm, address, pmdp, entry);
1403 static void pmdp_splitting_flush_sync(void *arg)
1405 /* Simply deliver the interrupt */
1408 void pmdp_splitting_flush(struct vm_area_struct *vma, unsigned long address,
1411 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1412 if (!test_and_set_bit(_SEGMENT_ENTRY_SPLIT_BIT,
1413 (unsigned long *) pmdp)) {
1414 /* need to serialize against gup-fast (IRQ disabled) */
1415 smp_call_function(pmdp_splitting_flush_sync, NULL, 1);
1419 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1422 struct list_head *lh = (struct list_head *) pgtable;
1424 assert_spin_locked(pmd_lockptr(mm, pmdp));
1427 if (!pmd_huge_pte(mm, pmdp))
1430 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1431 pmd_huge_pte(mm, pmdp) = pgtable;
1434 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1436 struct list_head *lh;
1440 assert_spin_locked(pmd_lockptr(mm, pmdp));
1443 pgtable = pmd_huge_pte(mm, pmdp);
1444 lh = (struct list_head *) pgtable;
1446 pmd_huge_pte(mm, pmdp) = NULL;
1448 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1451 ptep = (pte_t *) pgtable;
1452 pte_val(*ptep) = _PAGE_INVALID;
1454 pte_val(*ptep) = _PAGE_INVALID;
1457 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */