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/spinlock.h>
14 #include <linux/rcupdate.h>
15 #include <linux/slab.h>
16 #include <linux/swapops.h>
17 #include <linux/sysctl.h>
18 #include <linux/ksm.h>
19 #include <linux/mman.h>
21 #include <asm/pgtable.h>
22 #include <asm/pgalloc.h>
24 #include <asm/tlbflush.h>
25 #include <asm/mmu_context.h>
27 unsigned long *crst_table_alloc(struct mm_struct *mm)
29 struct page *page = alloc_pages(GFP_KERNEL, 2);
33 return (unsigned long *) page_to_phys(page);
36 void crst_table_free(struct mm_struct *mm, unsigned long *table)
38 free_pages((unsigned long) table, 2);
41 static void __crst_table_upgrade(void *arg)
43 struct mm_struct *mm = arg;
45 if (current->active_mm == mm) {
52 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
54 unsigned long *table, *pgd;
58 BUG_ON(limit > TASK_MAX_SIZE);
61 table = crst_table_alloc(mm);
64 spin_lock_bh(&mm->page_table_lock);
65 if (mm->context.asce_limit < limit) {
66 pgd = (unsigned long *) mm->pgd;
67 if (mm->context.asce_limit <= (1UL << 31)) {
68 entry = _REGION3_ENTRY_EMPTY;
69 mm->context.asce_limit = 1UL << 42;
70 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
74 entry = _REGION2_ENTRY_EMPTY;
75 mm->context.asce_limit = 1UL << 53;
76 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
80 crst_table_init(table, entry);
81 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
82 mm->pgd = (pgd_t *) table;
83 mm->task_size = mm->context.asce_limit;
87 spin_unlock_bh(&mm->page_table_lock);
89 crst_table_free(mm, table);
90 if (mm->context.asce_limit < limit)
93 on_each_cpu(__crst_table_upgrade, mm, 0);
97 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
101 if (current->active_mm == mm) {
105 while (mm->context.asce_limit > limit) {
107 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
108 case _REGION_ENTRY_TYPE_R2:
109 mm->context.asce_limit = 1UL << 42;
110 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
114 case _REGION_ENTRY_TYPE_R3:
115 mm->context.asce_limit = 1UL << 31;
116 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
123 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
124 mm->task_size = mm->context.asce_limit;
125 crst_table_free(mm, (unsigned long *) pgd);
127 if (current->active_mm == mm)
134 * gmap_alloc - allocate a guest address space
135 * @mm: pointer to the parent mm_struct
136 * @limit: maximum address of the gmap address space
138 * Returns a guest address space structure.
140 struct gmap *gmap_alloc(struct mm_struct *mm, unsigned long limit)
144 unsigned long *table;
145 unsigned long etype, atype;
147 if (limit < (1UL << 31)) {
148 limit = (1UL << 31) - 1;
149 atype = _ASCE_TYPE_SEGMENT;
150 etype = _SEGMENT_ENTRY_EMPTY;
151 } else if (limit < (1UL << 42)) {
152 limit = (1UL << 42) - 1;
153 atype = _ASCE_TYPE_REGION3;
154 etype = _REGION3_ENTRY_EMPTY;
155 } else if (limit < (1UL << 53)) {
156 limit = (1UL << 53) - 1;
157 atype = _ASCE_TYPE_REGION2;
158 etype = _REGION2_ENTRY_EMPTY;
161 atype = _ASCE_TYPE_REGION1;
162 etype = _REGION1_ENTRY_EMPTY;
164 gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL);
167 INIT_LIST_HEAD(&gmap->crst_list);
168 INIT_RADIX_TREE(&gmap->guest_to_host, GFP_KERNEL);
169 INIT_RADIX_TREE(&gmap->host_to_guest, GFP_ATOMIC);
170 spin_lock_init(&gmap->guest_table_lock);
172 page = alloc_pages(GFP_KERNEL, 2);
176 list_add(&page->lru, &gmap->crst_list);
177 table = (unsigned long *) page_to_phys(page);
178 crst_table_init(table, etype);
180 gmap->asce = atype | _ASCE_TABLE_LENGTH |
181 _ASCE_USER_BITS | __pa(table);
182 gmap->asce_end = limit;
183 down_write(&mm->mmap_sem);
184 list_add(&gmap->list, &mm->context.gmap_list);
185 up_write(&mm->mmap_sem);
193 EXPORT_SYMBOL_GPL(gmap_alloc);
195 static void gmap_flush_tlb(struct gmap *gmap)
197 if (MACHINE_HAS_IDTE)
198 __tlb_flush_asce(gmap->mm, gmap->asce);
200 __tlb_flush_global();
203 static void gmap_radix_tree_free(struct radix_tree_root *root)
205 struct radix_tree_iter iter;
206 unsigned long indices[16];
211 /* A radix tree is freed by deleting all of its entries */
215 radix_tree_for_each_slot(slot, root, &iter, index) {
216 indices[nr] = iter.index;
220 for (i = 0; i < nr; i++) {
222 radix_tree_delete(root, index);
228 * gmap_free - free a guest address space
229 * @gmap: pointer to the guest address space structure
231 void gmap_free(struct gmap *gmap)
233 struct page *page, *next;
236 if (MACHINE_HAS_IDTE)
237 __tlb_flush_asce(gmap->mm, gmap->asce);
239 __tlb_flush_global();
241 /* Free all segment & region tables. */
242 list_for_each_entry_safe(page, next, &gmap->crst_list, lru)
243 __free_pages(page, 2);
244 gmap_radix_tree_free(&gmap->guest_to_host);
245 gmap_radix_tree_free(&gmap->host_to_guest);
246 down_write(&gmap->mm->mmap_sem);
247 list_del(&gmap->list);
248 up_write(&gmap->mm->mmap_sem);
251 EXPORT_SYMBOL_GPL(gmap_free);
254 * gmap_enable - switch primary space to the guest address space
255 * @gmap: pointer to the guest address space structure
257 void gmap_enable(struct gmap *gmap)
259 S390_lowcore.gmap = (unsigned long) gmap;
261 EXPORT_SYMBOL_GPL(gmap_enable);
264 * gmap_disable - switch back to the standard primary address space
265 * @gmap: pointer to the guest address space structure
267 void gmap_disable(struct gmap *gmap)
269 S390_lowcore.gmap = 0UL;
271 EXPORT_SYMBOL_GPL(gmap_disable);
274 * gmap_alloc_table is assumed to be called with mmap_sem held
276 static int gmap_alloc_table(struct gmap *gmap, unsigned long *table,
277 unsigned long init, unsigned long gaddr)
282 /* since we dont free the gmap table until gmap_free we can unlock */
283 page = alloc_pages(GFP_KERNEL, 2);
286 new = (unsigned long *) page_to_phys(page);
287 crst_table_init(new, init);
288 spin_lock(&gmap->mm->page_table_lock);
289 if (*table & _REGION_ENTRY_INVALID) {
290 list_add(&page->lru, &gmap->crst_list);
291 *table = (unsigned long) new | _REGION_ENTRY_LENGTH |
292 (*table & _REGION_ENTRY_TYPE_MASK);
296 spin_unlock(&gmap->mm->page_table_lock);
298 __free_pages(page, 2);
303 * __gmap_segment_gaddr - find virtual address from segment pointer
304 * @entry: pointer to a segment table entry in the guest address space
306 * Returns the virtual address in the guest address space for the segment
308 static unsigned long __gmap_segment_gaddr(unsigned long *entry)
311 unsigned long offset, mask;
313 offset = (unsigned long) entry / sizeof(unsigned long);
314 offset = (offset & (PTRS_PER_PMD - 1)) * PMD_SIZE;
315 mask = ~(PTRS_PER_PMD * sizeof(pmd_t) - 1);
316 page = virt_to_page((void *)((unsigned long) entry & mask));
317 return page->index + offset;
321 * __gmap_unlink_by_vmaddr - unlink a single segment via a host address
322 * @gmap: pointer to the guest address space structure
323 * @vmaddr: address in the host process address space
325 * Returns 1 if a TLB flush is required
327 static int __gmap_unlink_by_vmaddr(struct gmap *gmap, unsigned long vmaddr)
329 unsigned long *entry;
332 spin_lock(&gmap->guest_table_lock);
333 entry = radix_tree_delete(&gmap->host_to_guest, vmaddr >> PMD_SHIFT);
335 flush = (*entry != _SEGMENT_ENTRY_INVALID);
336 *entry = _SEGMENT_ENTRY_INVALID;
338 spin_unlock(&gmap->guest_table_lock);
343 * __gmap_unmap_by_gaddr - unmap a single segment via a guest address
344 * @gmap: pointer to the guest address space structure
345 * @gaddr: address in the guest address space
347 * Returns 1 if a TLB flush is required
349 static int __gmap_unmap_by_gaddr(struct gmap *gmap, unsigned long gaddr)
351 unsigned long vmaddr;
353 vmaddr = (unsigned long) radix_tree_delete(&gmap->guest_to_host,
355 return vmaddr ? __gmap_unlink_by_vmaddr(gmap, vmaddr) : 0;
359 * gmap_unmap_segment - unmap segment from the guest address space
360 * @gmap: pointer to the guest address space structure
361 * @to: address in the guest address space
362 * @len: length of the memory area to unmap
364 * Returns 0 if the unmap succeeded, -EINVAL if not.
366 int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
371 if ((to | len) & (PMD_SIZE - 1))
373 if (len == 0 || to + len < to)
377 down_write(&gmap->mm->mmap_sem);
378 for (off = 0; off < len; off += PMD_SIZE)
379 flush |= __gmap_unmap_by_gaddr(gmap, to + off);
380 up_write(&gmap->mm->mmap_sem);
382 gmap_flush_tlb(gmap);
385 EXPORT_SYMBOL_GPL(gmap_unmap_segment);
388 * gmap_mmap_segment - map a segment to the guest address space
389 * @gmap: pointer to the guest address space structure
390 * @from: source address in the parent address space
391 * @to: target address in the guest address space
392 * @len: length of the memory area to map
394 * Returns 0 if the mmap succeeded, -EINVAL or -ENOMEM if not.
396 int gmap_map_segment(struct gmap *gmap, unsigned long from,
397 unsigned long to, unsigned long len)
402 if ((from | to | len) & (PMD_SIZE - 1))
404 if (len == 0 || from + len < from || to + len < to ||
405 from + len - 1 > TASK_MAX_SIZE || to + len - 1 > gmap->asce_end)
409 down_write(&gmap->mm->mmap_sem);
410 for (off = 0; off < len; off += PMD_SIZE) {
411 /* Remove old translation */
412 flush |= __gmap_unmap_by_gaddr(gmap, to + off);
413 /* Store new translation */
414 if (radix_tree_insert(&gmap->guest_to_host,
415 (to + off) >> PMD_SHIFT,
416 (void *) from + off))
419 up_write(&gmap->mm->mmap_sem);
421 gmap_flush_tlb(gmap);
424 gmap_unmap_segment(gmap, to, len);
427 EXPORT_SYMBOL_GPL(gmap_map_segment);
430 * __gmap_translate - translate a guest address to a user space address
431 * @gmap: pointer to guest mapping meta data structure
432 * @gaddr: guest address
434 * Returns user space address which corresponds to the guest address or
435 * -EFAULT if no such mapping exists.
436 * This function does not establish potentially missing page table entries.
437 * The mmap_sem of the mm that belongs to the address space must be held
438 * when this function gets called.
440 unsigned long __gmap_translate(struct gmap *gmap, unsigned long gaddr)
442 unsigned long vmaddr;
444 vmaddr = (unsigned long)
445 radix_tree_lookup(&gmap->guest_to_host, gaddr >> PMD_SHIFT);
446 return vmaddr ? (vmaddr | (gaddr & ~PMD_MASK)) : -EFAULT;
448 EXPORT_SYMBOL_GPL(__gmap_translate);
451 * gmap_translate - translate a guest address to a user space address
452 * @gmap: pointer to guest mapping meta data structure
453 * @gaddr: guest address
455 * Returns user space address which corresponds to the guest address or
456 * -EFAULT if no such mapping exists.
457 * This function does not establish potentially missing page table entries.
459 unsigned long gmap_translate(struct gmap *gmap, unsigned long gaddr)
463 down_read(&gmap->mm->mmap_sem);
464 rc = __gmap_translate(gmap, gaddr);
465 up_read(&gmap->mm->mmap_sem);
468 EXPORT_SYMBOL_GPL(gmap_translate);
471 * gmap_unlink - disconnect a page table from the gmap shadow tables
472 * @gmap: pointer to guest mapping meta data structure
473 * @table: pointer to the host page table
474 * @vmaddr: vm address associated with the host page table
476 static void gmap_unlink(struct mm_struct *mm, unsigned long *table,
477 unsigned long vmaddr)
482 list_for_each_entry(gmap, &mm->context.gmap_list, list) {
483 flush = __gmap_unlink_by_vmaddr(gmap, vmaddr);
485 gmap_flush_tlb(gmap);
490 * gmap_link - set up shadow page tables to connect a host to a guest address
491 * @gmap: pointer to guest mapping meta data structure
492 * @gaddr: guest address
493 * @vmaddr: vm address
495 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
496 * if the vm address is already mapped to a different guest segment.
497 * The mmap_sem of the mm that belongs to the address space must be held
498 * when this function gets called.
500 int __gmap_link(struct gmap *gmap, unsigned long gaddr, unsigned long vmaddr)
502 struct mm_struct *mm;
503 unsigned long *table;
510 /* Create higher level tables in the gmap page table */
512 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION1) {
513 table += (gaddr >> 53) & 0x7ff;
514 if ((*table & _REGION_ENTRY_INVALID) &&
515 gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY,
516 gaddr & 0xffe0000000000000UL))
518 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
520 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION2) {
521 table += (gaddr >> 42) & 0x7ff;
522 if ((*table & _REGION_ENTRY_INVALID) &&
523 gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY,
524 gaddr & 0xfffffc0000000000UL))
526 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
528 if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION3) {
529 table += (gaddr >> 31) & 0x7ff;
530 if ((*table & _REGION_ENTRY_INVALID) &&
531 gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY,
532 gaddr & 0xffffffff80000000UL))
534 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
536 table += (gaddr >> 20) & 0x7ff;
537 /* Walk the parent mm page table */
539 pgd = pgd_offset(mm, vmaddr);
540 VM_BUG_ON(pgd_none(*pgd));
541 pud = pud_offset(pgd, vmaddr);
542 VM_BUG_ON(pud_none(*pud));
543 pmd = pmd_offset(pud, vmaddr);
544 VM_BUG_ON(pmd_none(*pmd));
545 /* large pmds cannot yet be handled */
548 /* Link gmap segment table entry location to page table. */
549 rc = radix_tree_preload(GFP_KERNEL);
552 ptl = pmd_lock(mm, pmd);
553 spin_lock(&gmap->guest_table_lock);
554 if (*table == _SEGMENT_ENTRY_INVALID) {
555 rc = radix_tree_insert(&gmap->host_to_guest,
556 vmaddr >> PMD_SHIFT, table);
558 *table = pmd_val(*pmd);
561 spin_unlock(&gmap->guest_table_lock);
563 radix_tree_preload_end();
568 * gmap_fault - resolve a fault on a guest address
569 * @gmap: pointer to guest mapping meta data structure
570 * @gaddr: guest address
571 * @fault_flags: flags to pass down to handle_mm_fault()
573 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
574 * if the vm address is already mapped to a different guest segment.
576 int gmap_fault(struct gmap *gmap, unsigned long gaddr,
577 unsigned int fault_flags)
579 unsigned long vmaddr;
583 down_read(&gmap->mm->mmap_sem);
587 vmaddr = __gmap_translate(gmap, gaddr);
588 if (IS_ERR_VALUE(vmaddr)) {
592 if (fixup_user_fault(current, gmap->mm, vmaddr, fault_flags,
598 * In the case that fixup_user_fault unlocked the mmap_sem during
599 * faultin redo __gmap_translate to not race with a map/unmap_segment.
604 rc = __gmap_link(gmap, gaddr, vmaddr);
606 up_read(&gmap->mm->mmap_sem);
609 EXPORT_SYMBOL_GPL(gmap_fault);
611 static void gmap_zap_swap_entry(swp_entry_t entry, struct mm_struct *mm)
613 if (!non_swap_entry(entry))
614 dec_mm_counter(mm, MM_SWAPENTS);
615 else if (is_migration_entry(entry)) {
616 struct page *page = migration_entry_to_page(entry);
618 dec_mm_counter(mm, mm_counter(page));
620 free_swap_and_cache(entry);
624 * this function is assumed to be called with mmap_sem held
626 void __gmap_zap(struct gmap *gmap, unsigned long gaddr)
628 unsigned long vmaddr, ptev, pgstev;
633 /* Find the vm address for the guest address */
634 vmaddr = (unsigned long) radix_tree_lookup(&gmap->guest_to_host,
638 vmaddr |= gaddr & ~PMD_MASK;
639 /* Get pointer to the page table entry */
640 ptep = get_locked_pte(gmap->mm, vmaddr, &ptl);
646 /* Zap unused and logically-zero pages */
647 pgste = pgste_get_lock(ptep);
648 pgstev = pgste_val(pgste);
650 if (((pgstev & _PGSTE_GPS_USAGE_MASK) == _PGSTE_GPS_USAGE_UNUSED) ||
651 ((pgstev & _PGSTE_GPS_ZERO) && (ptev & _PAGE_INVALID))) {
652 gmap_zap_swap_entry(pte_to_swp_entry(pte), gmap->mm);
653 pte_clear(gmap->mm, vmaddr, ptep);
655 pgste_set_unlock(ptep, pgste);
657 pte_unmap_unlock(ptep, ptl);
659 EXPORT_SYMBOL_GPL(__gmap_zap);
661 void gmap_discard(struct gmap *gmap, unsigned long from, unsigned long to)
663 unsigned long gaddr, vmaddr, size;
664 struct vm_area_struct *vma;
666 down_read(&gmap->mm->mmap_sem);
667 for (gaddr = from; gaddr < to;
668 gaddr = (gaddr + PMD_SIZE) & PMD_MASK) {
669 /* Find the vm address for the guest address */
670 vmaddr = (unsigned long)
671 radix_tree_lookup(&gmap->guest_to_host,
675 vmaddr |= gaddr & ~PMD_MASK;
676 /* Find vma in the parent mm */
677 vma = find_vma(gmap->mm, vmaddr);
678 size = min(to - gaddr, PMD_SIZE - (gaddr & ~PMD_MASK));
679 zap_page_range(vma, vmaddr, size, NULL);
681 up_read(&gmap->mm->mmap_sem);
683 EXPORT_SYMBOL_GPL(gmap_discard);
685 static LIST_HEAD(gmap_notifier_list);
686 static DEFINE_SPINLOCK(gmap_notifier_lock);
689 * gmap_register_ipte_notifier - register a pte invalidation callback
690 * @nb: pointer to the gmap notifier block
692 void gmap_register_ipte_notifier(struct gmap_notifier *nb)
694 spin_lock(&gmap_notifier_lock);
695 list_add(&nb->list, &gmap_notifier_list);
696 spin_unlock(&gmap_notifier_lock);
698 EXPORT_SYMBOL_GPL(gmap_register_ipte_notifier);
701 * gmap_unregister_ipte_notifier - remove a pte invalidation callback
702 * @nb: pointer to the gmap notifier block
704 void gmap_unregister_ipte_notifier(struct gmap_notifier *nb)
706 spin_lock(&gmap_notifier_lock);
707 list_del_init(&nb->list);
708 spin_unlock(&gmap_notifier_lock);
710 EXPORT_SYMBOL_GPL(gmap_unregister_ipte_notifier);
713 * gmap_ipte_notify - mark a range of ptes for invalidation notification
714 * @gmap: pointer to guest mapping meta data structure
715 * @gaddr: virtual address in the guest address space
718 * Returns 0 if for each page in the given range a gmap mapping exists and
719 * the invalidation notification could be set. If the gmap mapping is missing
720 * for one or more pages -EFAULT is returned. If no memory could be allocated
721 * -ENOMEM is returned. This function establishes missing page table entries.
723 int gmap_ipte_notify(struct gmap *gmap, unsigned long gaddr, unsigned long len)
732 if ((gaddr & ~PAGE_MASK) || (len & ~PAGE_MASK))
734 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,
749 /* While trying to map mmap_sem got unlocked. Let us retry */
752 rc = __gmap_link(gmap, gaddr, addr);
755 /* Walk the process page table, lock and get pte pointer */
756 ptep = get_locked_pte(gmap->mm, addr, &ptl);
758 /* Set notification bit in the pgste of the pte */
760 if ((pte_val(entry) & (_PAGE_INVALID | _PAGE_PROTECT)) == 0) {
761 pgste = pgste_get_lock(ptep);
762 pgste_val(pgste) |= PGSTE_IN_BIT;
763 pgste_set_unlock(ptep, pgste);
767 pte_unmap_unlock(ptep, ptl);
769 up_read(&gmap->mm->mmap_sem);
772 EXPORT_SYMBOL_GPL(gmap_ipte_notify);
775 * gmap_do_ipte_notify - call all invalidation callbacks for a specific pte.
776 * @mm: pointer to the process mm_struct
777 * @addr: virtual address in the process address space
778 * @pte: pointer to the page table entry
780 * This function is assumed to be called with the page table lock held
781 * for the pte to notify.
783 void gmap_do_ipte_notify(struct mm_struct *mm, unsigned long vmaddr, pte_t *pte)
785 unsigned long offset, gaddr;
786 unsigned long *table;
787 struct gmap_notifier *nb;
790 offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
791 offset = offset * (4096 / sizeof(pte_t));
792 spin_lock(&gmap_notifier_lock);
793 list_for_each_entry(gmap, &mm->context.gmap_list, list) {
794 table = radix_tree_lookup(&gmap->host_to_guest,
795 vmaddr >> PMD_SHIFT);
798 gaddr = __gmap_segment_gaddr(table) + offset;
799 list_for_each_entry(nb, &gmap_notifier_list, list)
800 nb->notifier_call(gmap, gaddr);
802 spin_unlock(&gmap_notifier_lock);
804 EXPORT_SYMBOL_GPL(gmap_do_ipte_notify);
806 int set_guest_storage_key(struct mm_struct *mm, unsigned long addr,
807 unsigned long key, bool nq)
814 down_read(&mm->mmap_sem);
817 ptep = get_locked_pte(mm, addr, &ptl);
818 if (unlikely(!ptep)) {
819 up_read(&mm->mmap_sem);
822 if (!(pte_val(*ptep) & _PAGE_INVALID) &&
823 (pte_val(*ptep) & _PAGE_PROTECT)) {
824 pte_unmap_unlock(ptep, ptl);
826 * We do not really care about unlocked. We will retry either
827 * way. But this allows fixup_user_fault to enable userfaultfd.
829 if (fixup_user_fault(current, mm, addr, FAULT_FLAG_WRITE,
831 up_read(&mm->mmap_sem);
837 new = old = pgste_get_lock(ptep);
838 pgste_val(new) &= ~(PGSTE_GR_BIT | PGSTE_GC_BIT |
839 PGSTE_ACC_BITS | PGSTE_FP_BIT);
840 pgste_val(new) |= (key & (_PAGE_CHANGED | _PAGE_REFERENCED)) << 48;
841 pgste_val(new) |= (key & (_PAGE_ACC_BITS | _PAGE_FP_BIT)) << 56;
842 if (!(pte_val(*ptep) & _PAGE_INVALID)) {
843 unsigned long address, bits, skey;
845 address = pte_val(*ptep) & PAGE_MASK;
846 skey = (unsigned long) page_get_storage_key(address);
847 bits = skey & (_PAGE_CHANGED | _PAGE_REFERENCED);
848 skey = key & (_PAGE_ACC_BITS | _PAGE_FP_BIT);
849 /* Set storage key ACC and FP */
850 page_set_storage_key(address, skey, !nq);
851 /* Merge host changed & referenced into pgste */
852 pgste_val(new) |= bits << 52;
854 /* changing the guest storage key is considered a change of the page */
855 if ((pgste_val(new) ^ pgste_val(old)) &
856 (PGSTE_ACC_BITS | PGSTE_FP_BIT | PGSTE_GR_BIT | PGSTE_GC_BIT))
857 pgste_val(new) |= PGSTE_UC_BIT;
859 pgste_set_unlock(ptep, new);
860 pte_unmap_unlock(ptep, ptl);
861 up_read(&mm->mmap_sem);
864 EXPORT_SYMBOL(set_guest_storage_key);
866 unsigned long get_guest_storage_key(struct mm_struct *mm, unsigned long addr)
872 unsigned long key = 0;
874 down_read(&mm->mmap_sem);
875 ptep = get_locked_pte(mm, addr, &ptl);
876 if (unlikely(!ptep)) {
877 up_read(&mm->mmap_sem);
880 pgste = pgste_get_lock(ptep);
882 if (pte_val(*ptep) & _PAGE_INVALID) {
883 key |= (pgste_val(pgste) & PGSTE_ACC_BITS) >> 56;
884 key |= (pgste_val(pgste) & PGSTE_FP_BIT) >> 56;
885 key |= (pgste_val(pgste) & PGSTE_GR_BIT) >> 48;
886 key |= (pgste_val(pgste) & PGSTE_GC_BIT) >> 48;
888 physaddr = pte_val(*ptep) & PAGE_MASK;
889 key = page_get_storage_key(physaddr);
891 /* Reflect guest's logical view, not physical */
892 if (pgste_val(pgste) & PGSTE_GR_BIT)
893 key |= _PAGE_REFERENCED;
894 if (pgste_val(pgste) & PGSTE_GC_BIT)
895 key |= _PAGE_CHANGED;
898 pgste_set_unlock(ptep, pgste);
899 pte_unmap_unlock(ptep, ptl);
900 up_read(&mm->mmap_sem);
903 EXPORT_SYMBOL(get_guest_storage_key);
905 static int page_table_allocate_pgste_min = 0;
906 static int page_table_allocate_pgste_max = 1;
907 int page_table_allocate_pgste = 0;
908 EXPORT_SYMBOL(page_table_allocate_pgste);
910 static struct ctl_table page_table_sysctl[] = {
912 .procname = "allocate_pgste",
913 .data = &page_table_allocate_pgste,
914 .maxlen = sizeof(int),
915 .mode = S_IRUGO | S_IWUSR,
916 .proc_handler = proc_dointvec,
917 .extra1 = &page_table_allocate_pgste_min,
918 .extra2 = &page_table_allocate_pgste_max,
923 static struct ctl_table page_table_sysctl_dir[] = {
928 .child = page_table_sysctl,
933 static int __init page_table_register_sysctl(void)
935 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
937 __initcall(page_table_register_sysctl);
939 #else /* CONFIG_PGSTE */
941 static inline void gmap_unlink(struct mm_struct *mm, unsigned long *table,
942 unsigned long vmaddr)
946 #endif /* CONFIG_PGSTE */
948 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
950 unsigned int old, new;
953 old = atomic_read(v);
955 } while (atomic_cmpxchg(v, old, new) != old);
960 * page table entry allocation/free routines.
962 unsigned long *page_table_alloc(struct mm_struct *mm)
964 unsigned long *table;
966 unsigned int mask, bit;
968 /* Try to get a fragment of a 4K page as a 2K page table */
969 if (!mm_alloc_pgste(mm)) {
971 spin_lock_bh(&mm->context.list_lock);
972 if (!list_empty(&mm->context.pgtable_list)) {
973 page = list_first_entry(&mm->context.pgtable_list,
975 mask = atomic_read(&page->_mapcount);
976 mask = (mask | (mask >> 4)) & 3;
978 table = (unsigned long *) page_to_phys(page);
979 bit = mask & 1; /* =1 -> second 2K */
981 table += PTRS_PER_PTE;
982 atomic_xor_bits(&page->_mapcount, 1U << bit);
983 list_del(&page->lru);
986 spin_unlock_bh(&mm->context.list_lock);
990 /* Allocate a fresh page */
991 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
994 if (!pgtable_page_ctor(page)) {
998 /* Initialize page table */
999 table = (unsigned long *) page_to_phys(page);
1000 if (mm_alloc_pgste(mm)) {
1001 /* Return 4K page table with PGSTEs */
1002 atomic_set(&page->_mapcount, 3);
1003 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
1004 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
1006 /* Return the first 2K fragment of the page */
1007 atomic_set(&page->_mapcount, 1);
1008 clear_table(table, _PAGE_INVALID, PAGE_SIZE);
1009 spin_lock_bh(&mm->context.list_lock);
1010 list_add(&page->lru, &mm->context.pgtable_list);
1011 spin_unlock_bh(&mm->context.list_lock);
1016 void page_table_free(struct mm_struct *mm, unsigned long *table)
1019 unsigned int bit, mask;
1021 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1022 if (!mm_alloc_pgste(mm)) {
1023 /* Free 2K page table fragment of a 4K page */
1024 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
1025 spin_lock_bh(&mm->context.list_lock);
1026 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
1028 list_add(&page->lru, &mm->context.pgtable_list);
1030 list_del(&page->lru);
1031 spin_unlock_bh(&mm->context.list_lock);
1036 pgtable_page_dtor(page);
1037 atomic_set(&page->_mapcount, -1);
1041 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
1042 unsigned long vmaddr)
1044 struct mm_struct *mm;
1046 unsigned int bit, mask;
1049 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1050 if (mm_alloc_pgste(mm)) {
1051 gmap_unlink(mm, table, vmaddr);
1052 table = (unsigned long *) (__pa(table) | 3);
1053 tlb_remove_table(tlb, table);
1056 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
1057 spin_lock_bh(&mm->context.list_lock);
1058 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
1060 list_add_tail(&page->lru, &mm->context.pgtable_list);
1062 list_del(&page->lru);
1063 spin_unlock_bh(&mm->context.list_lock);
1064 table = (unsigned long *) (__pa(table) | (1U << bit));
1065 tlb_remove_table(tlb, table);
1068 static void __tlb_remove_table(void *_table)
1070 unsigned int mask = (unsigned long) _table & 3;
1071 void *table = (void *)((unsigned long) _table ^ mask);
1072 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
1075 case 0: /* pmd or pud */
1076 free_pages((unsigned long) table, 2);
1078 case 1: /* lower 2K of a 4K page table */
1079 case 2: /* higher 2K of a 4K page table */
1080 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
1083 case 3: /* 4K page table with pgstes */
1084 pgtable_page_dtor(page);
1085 atomic_set(&page->_mapcount, -1);
1091 static void tlb_remove_table_smp_sync(void *arg)
1093 /* Simply deliver the interrupt */
1096 static void tlb_remove_table_one(void *table)
1099 * This isn't an RCU grace period and hence the page-tables cannot be
1100 * assumed to be actually RCU-freed.
1102 * It is however sufficient for software page-table walkers that rely
1103 * on IRQ disabling. See the comment near struct mmu_table_batch.
1105 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
1106 __tlb_remove_table(table);
1109 static void tlb_remove_table_rcu(struct rcu_head *head)
1111 struct mmu_table_batch *batch;
1114 batch = container_of(head, struct mmu_table_batch, rcu);
1116 for (i = 0; i < batch->nr; i++)
1117 __tlb_remove_table(batch->tables[i]);
1119 free_page((unsigned long)batch);
1122 void tlb_table_flush(struct mmu_gather *tlb)
1124 struct mmu_table_batch **batch = &tlb->batch;
1127 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
1132 void tlb_remove_table(struct mmu_gather *tlb, void *table)
1134 struct mmu_table_batch **batch = &tlb->batch;
1136 tlb->mm->context.flush_mm = 1;
1137 if (*batch == NULL) {
1138 *batch = (struct mmu_table_batch *)
1139 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
1140 if (*batch == NULL) {
1141 __tlb_flush_mm_lazy(tlb->mm);
1142 tlb_remove_table_one(table);
1147 (*batch)->tables[(*batch)->nr++] = table;
1148 if ((*batch)->nr == MAX_TABLE_BATCH)
1152 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1153 static inline void thp_split_vma(struct vm_area_struct *vma)
1157 for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE)
1158 follow_page(vma, addr, FOLL_SPLIT);
1161 static inline void thp_split_mm(struct mm_struct *mm)
1163 struct vm_area_struct *vma;
1165 for (vma = mm->mmap; vma != NULL; vma = vma->vm_next) {
1167 vma->vm_flags &= ~VM_HUGEPAGE;
1168 vma->vm_flags |= VM_NOHUGEPAGE;
1170 mm->def_flags |= VM_NOHUGEPAGE;
1173 static inline void thp_split_mm(struct mm_struct *mm)
1176 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1179 * switch on pgstes for its userspace process (for kvm)
1181 int s390_enable_sie(void)
1183 struct mm_struct *mm = current->mm;
1185 /* Do we have pgstes? if yes, we are done */
1186 if (mm_has_pgste(mm))
1188 /* Fail if the page tables are 2K */
1189 if (!mm_alloc_pgste(mm))
1191 down_write(&mm->mmap_sem);
1192 mm->context.has_pgste = 1;
1193 /* split thp mappings and disable thp for future mappings */
1195 up_write(&mm->mmap_sem);
1198 EXPORT_SYMBOL_GPL(s390_enable_sie);
1201 * Enable storage key handling from now on and initialize the storage
1202 * keys with the default key.
1204 static int __s390_enable_skey(pte_t *pte, unsigned long addr,
1205 unsigned long next, struct mm_walk *walk)
1210 pgste = pgste_get_lock(pte);
1212 * Remove all zero page mappings,
1213 * after establishing a policy to forbid zero page mappings
1214 * following faults for that page will get fresh anonymous pages
1216 if (is_zero_pfn(pte_pfn(*pte))) {
1217 ptep_flush_direct(walk->mm, addr, pte);
1218 pte_val(*pte) = _PAGE_INVALID;
1220 /* Clear storage key */
1221 pgste_val(pgste) &= ~(PGSTE_ACC_BITS | PGSTE_FP_BIT |
1222 PGSTE_GR_BIT | PGSTE_GC_BIT);
1223 ptev = pte_val(*pte);
1224 if (!(ptev & _PAGE_INVALID) && (ptev & _PAGE_WRITE))
1225 page_set_storage_key(ptev & PAGE_MASK, PAGE_DEFAULT_KEY, 1);
1226 pgste_set_unlock(pte, pgste);
1230 int s390_enable_skey(void)
1232 struct mm_walk walk = { .pte_entry = __s390_enable_skey };
1233 struct mm_struct *mm = current->mm;
1234 struct vm_area_struct *vma;
1237 down_write(&mm->mmap_sem);
1238 if (mm_use_skey(mm))
1241 mm->context.use_skey = 1;
1242 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1243 if (ksm_madvise(vma, vma->vm_start, vma->vm_end,
1244 MADV_UNMERGEABLE, &vma->vm_flags)) {
1245 mm->context.use_skey = 0;
1250 mm->def_flags &= ~VM_MERGEABLE;
1253 walk_page_range(0, TASK_SIZE, &walk);
1256 up_write(&mm->mmap_sem);
1259 EXPORT_SYMBOL_GPL(s390_enable_skey);
1262 * Reset CMMA state, make all pages stable again.
1264 static int __s390_reset_cmma(pte_t *pte, unsigned long addr,
1265 unsigned long next, struct mm_walk *walk)
1269 pgste = pgste_get_lock(pte);
1270 pgste_val(pgste) &= ~_PGSTE_GPS_USAGE_MASK;
1271 pgste_set_unlock(pte, pgste);
1275 void s390_reset_cmma(struct mm_struct *mm)
1277 struct mm_walk walk = { .pte_entry = __s390_reset_cmma };
1279 down_write(&mm->mmap_sem);
1281 walk_page_range(0, TASK_SIZE, &walk);
1282 up_write(&mm->mmap_sem);
1284 EXPORT_SYMBOL_GPL(s390_reset_cmma);
1287 * Test and reset if a guest page is dirty
1289 bool gmap_test_and_clear_dirty(unsigned long address, struct gmap *gmap)
1295 pte = get_locked_pte(gmap->mm, address, &ptl);
1299 if (ptep_test_and_clear_user_dirty(gmap->mm, address, pte))
1305 EXPORT_SYMBOL_GPL(gmap_test_and_clear_dirty);
1307 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1308 int pmdp_clear_flush_young(struct vm_area_struct *vma, unsigned long address,
1311 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1312 /* No need to flush TLB
1313 * On s390 reference bits are in storage key and never in TLB */
1314 return pmdp_test_and_clear_young(vma, address, pmdp);
1317 int pmdp_set_access_flags(struct vm_area_struct *vma,
1318 unsigned long address, pmd_t *pmdp,
1319 pmd_t entry, int dirty)
1321 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1323 entry = pmd_mkyoung(entry);
1325 entry = pmd_mkdirty(entry);
1326 if (pmd_same(*pmdp, entry))
1328 pmdp_invalidate(vma, address, pmdp);
1329 set_pmd_at(vma->vm_mm, address, pmdp, entry);
1333 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1336 struct list_head *lh = (struct list_head *) pgtable;
1338 assert_spin_locked(pmd_lockptr(mm, pmdp));
1341 if (!pmd_huge_pte(mm, pmdp))
1344 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1345 pmd_huge_pte(mm, pmdp) = pgtable;
1348 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1350 struct list_head *lh;
1354 assert_spin_locked(pmd_lockptr(mm, pmdp));
1357 pgtable = pmd_huge_pte(mm, pmdp);
1358 lh = (struct list_head *) pgtable;
1360 pmd_huge_pte(mm, pmdp) = NULL;
1362 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1365 ptep = (pte_t *) pgtable;
1366 pte_val(*ptep) = _PAGE_INVALID;
1368 pte_val(*ptep) = _PAGE_INVALID;
1371 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */