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>
21 #include <asm/pgtable.h>
22 #include <asm/pgalloc.h>
24 #include <asm/tlbflush.h>
25 #include <asm/mmu_context.h>
29 #define FRAG_MASK 0x0f
32 #define FRAG_MASK 0x03
36 unsigned long *crst_table_alloc(struct mm_struct *mm)
38 struct page *page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
42 return (unsigned long *) page_to_phys(page);
45 void crst_table_free(struct mm_struct *mm, unsigned long *table)
47 free_pages((unsigned long) table, ALLOC_ORDER);
51 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
53 unsigned long *table, *pgd;
56 BUG_ON(limit > (1UL << 53));
58 table = crst_table_alloc(mm);
61 spin_lock_bh(&mm->page_table_lock);
62 if (mm->context.asce_limit < limit) {
63 pgd = (unsigned long *) mm->pgd;
64 if (mm->context.asce_limit <= (1UL << 31)) {
65 entry = _REGION3_ENTRY_EMPTY;
66 mm->context.asce_limit = 1UL << 42;
67 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
71 entry = _REGION2_ENTRY_EMPTY;
72 mm->context.asce_limit = 1UL << 53;
73 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
77 crst_table_init(table, entry);
78 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
79 mm->pgd = (pgd_t *) table;
80 mm->task_size = mm->context.asce_limit;
83 spin_unlock_bh(&mm->page_table_lock);
85 crst_table_free(mm, table);
86 if (mm->context.asce_limit < limit)
91 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
95 while (mm->context.asce_limit > limit) {
97 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
98 case _REGION_ENTRY_TYPE_R2:
99 mm->context.asce_limit = 1UL << 42;
100 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
104 case _REGION_ENTRY_TYPE_R3:
105 mm->context.asce_limit = 1UL << 31;
106 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
113 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
114 mm->task_size = mm->context.asce_limit;
115 crst_table_free(mm, (unsigned long *) pgd);
123 * gmap_alloc - allocate a guest address space
124 * @mm: pointer to the parent mm_struct
126 * Returns a guest address space structure.
128 struct gmap *gmap_alloc(struct mm_struct *mm)
132 unsigned long *table;
134 gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL);
137 INIT_LIST_HEAD(&gmap->crst_list);
139 page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
142 list_add(&page->lru, &gmap->crst_list);
143 table = (unsigned long *) page_to_phys(page);
144 crst_table_init(table, _REGION1_ENTRY_EMPTY);
146 gmap->asce = _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH |
147 _ASCE_USER_BITS | __pa(table);
148 list_add(&gmap->list, &mm->context.gmap_list);
156 EXPORT_SYMBOL_GPL(gmap_alloc);
158 static int gmap_unlink_segment(struct gmap *gmap, unsigned long *table)
160 struct gmap_pgtable *mp;
161 struct gmap_rmap *rmap;
164 if (*table & _SEGMENT_ENTRY_INV)
166 page = pfn_to_page(*table >> PAGE_SHIFT);
167 mp = (struct gmap_pgtable *) page->index;
168 list_for_each_entry(rmap, &mp->mapper, list) {
169 if (rmap->entry != table)
171 list_del(&rmap->list);
175 *table = _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | mp->vmaddr;
179 static void gmap_flush_tlb(struct gmap *gmap)
181 if (MACHINE_HAS_IDTE)
182 __tlb_flush_idte((unsigned long) gmap->table |
185 __tlb_flush_global();
189 * gmap_free - free a guest address space
190 * @gmap: pointer to the guest address space structure
192 void gmap_free(struct gmap *gmap)
194 struct page *page, *next;
195 unsigned long *table;
200 if (MACHINE_HAS_IDTE)
201 __tlb_flush_idte((unsigned long) gmap->table |
204 __tlb_flush_global();
206 /* Free all segment & region tables. */
207 down_read(&gmap->mm->mmap_sem);
208 spin_lock(&gmap->mm->page_table_lock);
209 list_for_each_entry_safe(page, next, &gmap->crst_list, lru) {
210 table = (unsigned long *) page_to_phys(page);
211 if ((*table & _REGION_ENTRY_TYPE_MASK) == 0)
212 /* Remove gmap rmap structures for segment table. */
213 for (i = 0; i < PTRS_PER_PMD; i++, table++)
214 gmap_unlink_segment(gmap, table);
215 __free_pages(page, ALLOC_ORDER);
217 spin_unlock(&gmap->mm->page_table_lock);
218 up_read(&gmap->mm->mmap_sem);
219 list_del(&gmap->list);
222 EXPORT_SYMBOL_GPL(gmap_free);
225 * gmap_enable - switch primary space to the guest address space
226 * @gmap: pointer to the guest address space structure
228 void gmap_enable(struct gmap *gmap)
230 S390_lowcore.gmap = (unsigned long) gmap;
232 EXPORT_SYMBOL_GPL(gmap_enable);
235 * gmap_disable - switch back to the standard primary address space
236 * @gmap: pointer to the guest address space structure
238 void gmap_disable(struct gmap *gmap)
240 S390_lowcore.gmap = 0UL;
242 EXPORT_SYMBOL_GPL(gmap_disable);
245 * gmap_alloc_table is assumed to be called with mmap_sem held
247 static int gmap_alloc_table(struct gmap *gmap,
248 unsigned long *table, unsigned long init)
253 /* since we dont free the gmap table until gmap_free we can unlock */
254 spin_unlock(&gmap->mm->page_table_lock);
255 page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
256 spin_lock(&gmap->mm->page_table_lock);
259 new = (unsigned long *) page_to_phys(page);
260 crst_table_init(new, init);
261 if (*table & _REGION_ENTRY_INV) {
262 list_add(&page->lru, &gmap->crst_list);
263 *table = (unsigned long) new | _REGION_ENTRY_LENGTH |
264 (*table & _REGION_ENTRY_TYPE_MASK);
266 __free_pages(page, ALLOC_ORDER);
271 * gmap_unmap_segment - unmap segment from the guest address space
272 * @gmap: pointer to the guest address space structure
273 * @addr: address in the guest address space
274 * @len: length of the memory area to unmap
276 * Returns 0 if the unmap succeded, -EINVAL if not.
278 int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
280 unsigned long *table;
284 if ((to | len) & (PMD_SIZE - 1))
286 if (len == 0 || to + len < to)
290 down_read(&gmap->mm->mmap_sem);
291 spin_lock(&gmap->mm->page_table_lock);
292 for (off = 0; off < len; off += PMD_SIZE) {
293 /* Walk the guest addr space page table */
294 table = gmap->table + (((to + off) >> 53) & 0x7ff);
295 if (*table & _REGION_ENTRY_INV)
297 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
298 table = table + (((to + off) >> 42) & 0x7ff);
299 if (*table & _REGION_ENTRY_INV)
301 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
302 table = table + (((to + off) >> 31) & 0x7ff);
303 if (*table & _REGION_ENTRY_INV)
305 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
306 table = table + (((to + off) >> 20) & 0x7ff);
308 /* Clear segment table entry in guest address space. */
309 flush |= gmap_unlink_segment(gmap, table);
310 *table = _SEGMENT_ENTRY_INV;
313 spin_unlock(&gmap->mm->page_table_lock);
314 up_read(&gmap->mm->mmap_sem);
316 gmap_flush_tlb(gmap);
319 EXPORT_SYMBOL_GPL(gmap_unmap_segment);
322 * gmap_mmap_segment - map a segment to the guest address space
323 * @gmap: pointer to the guest address space structure
324 * @from: source address in the parent address space
325 * @to: target address in the guest address space
327 * Returns 0 if the mmap succeded, -EINVAL or -ENOMEM if not.
329 int gmap_map_segment(struct gmap *gmap, unsigned long from,
330 unsigned long to, unsigned long len)
332 unsigned long *table;
336 if ((from | to | len) & (PMD_SIZE - 1))
338 if (len == 0 || from + len > PGDIR_SIZE ||
339 from + len < from || to + len < to)
343 down_read(&gmap->mm->mmap_sem);
344 spin_lock(&gmap->mm->page_table_lock);
345 for (off = 0; off < len; off += PMD_SIZE) {
346 /* Walk the gmap address space page table */
347 table = gmap->table + (((to + off) >> 53) & 0x7ff);
348 if ((*table & _REGION_ENTRY_INV) &&
349 gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY))
351 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
352 table = table + (((to + off) >> 42) & 0x7ff);
353 if ((*table & _REGION_ENTRY_INV) &&
354 gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY))
356 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
357 table = table + (((to + off) >> 31) & 0x7ff);
358 if ((*table & _REGION_ENTRY_INV) &&
359 gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY))
361 table = (unsigned long *) (*table & _REGION_ENTRY_ORIGIN);
362 table = table + (((to + off) >> 20) & 0x7ff);
364 /* Store 'from' address in an invalid segment table entry. */
365 flush |= gmap_unlink_segment(gmap, table);
366 *table = _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | (from + off);
368 spin_unlock(&gmap->mm->page_table_lock);
369 up_read(&gmap->mm->mmap_sem);
371 gmap_flush_tlb(gmap);
375 spin_unlock(&gmap->mm->page_table_lock);
376 up_read(&gmap->mm->mmap_sem);
377 gmap_unmap_segment(gmap, to, len);
380 EXPORT_SYMBOL_GPL(gmap_map_segment);
382 static unsigned long *gmap_table_walk(unsigned long address, struct gmap *gmap)
384 unsigned long *table;
386 table = gmap->table + ((address >> 53) & 0x7ff);
387 if (unlikely(*table & _REGION_ENTRY_INV))
388 return ERR_PTR(-EFAULT);
389 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
390 table = table + ((address >> 42) & 0x7ff);
391 if (unlikely(*table & _REGION_ENTRY_INV))
392 return ERR_PTR(-EFAULT);
393 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
394 table = table + ((address >> 31) & 0x7ff);
395 if (unlikely(*table & _REGION_ENTRY_INV))
396 return ERR_PTR(-EFAULT);
397 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
398 table = table + ((address >> 20) & 0x7ff);
403 * __gmap_translate - translate a guest address to a user space address
404 * @address: guest address
405 * @gmap: pointer to guest mapping meta data structure
407 * Returns user space address which corresponds to the guest address or
408 * -EFAULT if no such mapping exists.
409 * This function does not establish potentially missing page table entries.
410 * The mmap_sem of the mm that belongs to the address space must be held
411 * when this function gets called.
413 unsigned long __gmap_translate(unsigned long address, struct gmap *gmap)
415 unsigned long *segment_ptr, vmaddr, segment;
416 struct gmap_pgtable *mp;
419 current->thread.gmap_addr = address;
420 segment_ptr = gmap_table_walk(address, gmap);
421 if (IS_ERR(segment_ptr))
422 return PTR_ERR(segment_ptr);
423 /* Convert the gmap address to an mm address. */
424 segment = *segment_ptr;
425 if (!(segment & _SEGMENT_ENTRY_INV)) {
426 page = pfn_to_page(segment >> PAGE_SHIFT);
427 mp = (struct gmap_pgtable *) page->index;
428 return mp->vmaddr | (address & ~PMD_MASK);
429 } else if (segment & _SEGMENT_ENTRY_RO) {
430 vmaddr = segment & _SEGMENT_ENTRY_ORIGIN;
431 return vmaddr | (address & ~PMD_MASK);
435 EXPORT_SYMBOL_GPL(__gmap_translate);
438 * gmap_translate - translate a guest address to a user space address
439 * @address: guest address
440 * @gmap: pointer to guest mapping meta data structure
442 * Returns user space address which corresponds to the guest address or
443 * -EFAULT if no such mapping exists.
444 * This function does not establish potentially missing page table entries.
446 unsigned long gmap_translate(unsigned long address, struct gmap *gmap)
450 down_read(&gmap->mm->mmap_sem);
451 rc = __gmap_translate(address, gmap);
452 up_read(&gmap->mm->mmap_sem);
455 EXPORT_SYMBOL_GPL(gmap_translate);
457 static int gmap_connect_pgtable(unsigned long address, unsigned long segment,
458 unsigned long *segment_ptr, struct gmap *gmap)
460 unsigned long vmaddr;
461 struct vm_area_struct *vma;
462 struct gmap_pgtable *mp;
463 struct gmap_rmap *rmap;
464 struct mm_struct *mm;
471 vmaddr = segment & _SEGMENT_ENTRY_ORIGIN;
472 vma = find_vma(mm, vmaddr);
473 if (!vma || vma->vm_start > vmaddr)
475 /* Walk the parent mm page table */
476 pgd = pgd_offset(mm, vmaddr);
477 pud = pud_alloc(mm, pgd, vmaddr);
480 pmd = pmd_alloc(mm, pud, vmaddr);
483 if (!pmd_present(*pmd) &&
484 __pte_alloc(mm, vma, pmd, vmaddr))
486 /* pmd now points to a valid segment table entry. */
487 rmap = kmalloc(sizeof(*rmap), GFP_KERNEL|__GFP_REPEAT);
490 /* Link gmap segment table entry location to page table. */
491 page = pmd_page(*pmd);
492 mp = (struct gmap_pgtable *) page->index;
494 rmap->entry = segment_ptr;
495 rmap->vmaddr = address;
496 spin_lock(&mm->page_table_lock);
497 if (*segment_ptr == segment) {
498 list_add(&rmap->list, &mp->mapper);
499 /* Set gmap segment table entry to page table. */
500 *segment_ptr = pmd_val(*pmd) & PAGE_MASK;
503 spin_unlock(&mm->page_table_lock);
508 static void gmap_disconnect_pgtable(struct mm_struct *mm, unsigned long *table)
510 struct gmap_rmap *rmap, *next;
511 struct gmap_pgtable *mp;
516 spin_lock(&mm->page_table_lock);
517 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
518 mp = (struct gmap_pgtable *) page->index;
519 list_for_each_entry_safe(rmap, next, &mp->mapper, list) {
521 _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | mp->vmaddr;
522 list_del(&rmap->list);
526 spin_unlock(&mm->page_table_lock);
528 __tlb_flush_global();
532 * this function is assumed to be called with mmap_sem held
534 unsigned long __gmap_fault(unsigned long address, struct gmap *gmap)
536 unsigned long *segment_ptr, segment;
537 struct gmap_pgtable *mp;
541 current->thread.gmap_addr = address;
542 segment_ptr = gmap_table_walk(address, gmap);
543 if (IS_ERR(segment_ptr))
545 /* Convert the gmap address to an mm address. */
547 segment = *segment_ptr;
548 if (!(segment & _SEGMENT_ENTRY_INV)) {
549 /* Page table is present */
550 page = pfn_to_page(segment >> PAGE_SHIFT);
551 mp = (struct gmap_pgtable *) page->index;
552 return mp->vmaddr | (address & ~PMD_MASK);
554 if (!(segment & _SEGMENT_ENTRY_RO))
555 /* Nothing mapped in the gmap address space. */
557 rc = gmap_connect_pgtable(address, segment, segment_ptr, gmap);
564 unsigned long gmap_fault(unsigned long address, struct gmap *gmap)
568 down_read(&gmap->mm->mmap_sem);
569 rc = __gmap_fault(address, gmap);
570 up_read(&gmap->mm->mmap_sem);
574 EXPORT_SYMBOL_GPL(gmap_fault);
576 void gmap_discard(unsigned long from, unsigned long to, struct gmap *gmap)
579 unsigned long *table, address, size;
580 struct vm_area_struct *vma;
581 struct gmap_pgtable *mp;
584 down_read(&gmap->mm->mmap_sem);
586 while (address < to) {
587 /* Walk the gmap address space page table */
588 table = gmap->table + ((address >> 53) & 0x7ff);
589 if (unlikely(*table & _REGION_ENTRY_INV)) {
590 address = (address + PMD_SIZE) & PMD_MASK;
593 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
594 table = table + ((address >> 42) & 0x7ff);
595 if (unlikely(*table & _REGION_ENTRY_INV)) {
596 address = (address + PMD_SIZE) & PMD_MASK;
599 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
600 table = table + ((address >> 31) & 0x7ff);
601 if (unlikely(*table & _REGION_ENTRY_INV)) {
602 address = (address + PMD_SIZE) & PMD_MASK;
605 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
606 table = table + ((address >> 20) & 0x7ff);
607 if (unlikely(*table & _SEGMENT_ENTRY_INV)) {
608 address = (address + PMD_SIZE) & PMD_MASK;
611 page = pfn_to_page(*table >> PAGE_SHIFT);
612 mp = (struct gmap_pgtable *) page->index;
613 vma = find_vma(gmap->mm, mp->vmaddr);
614 size = min(to - address, PMD_SIZE - (address & ~PMD_MASK));
615 zap_page_range(vma, mp->vmaddr | (address & ~PMD_MASK),
617 address = (address + PMD_SIZE) & PMD_MASK;
619 up_read(&gmap->mm->mmap_sem);
621 EXPORT_SYMBOL_GPL(gmap_discard);
623 static LIST_HEAD(gmap_notifier_list);
624 static DEFINE_SPINLOCK(gmap_notifier_lock);
627 * gmap_register_ipte_notifier - register a pte invalidation callback
628 * @nb: pointer to the gmap notifier block
630 void gmap_register_ipte_notifier(struct gmap_notifier *nb)
632 spin_lock(&gmap_notifier_lock);
633 list_add(&nb->list, &gmap_notifier_list);
634 spin_unlock(&gmap_notifier_lock);
636 EXPORT_SYMBOL_GPL(gmap_register_ipte_notifier);
639 * gmap_unregister_ipte_notifier - remove a pte invalidation callback
640 * @nb: pointer to the gmap notifier block
642 void gmap_unregister_ipte_notifier(struct gmap_notifier *nb)
644 spin_lock(&gmap_notifier_lock);
645 list_del_init(&nb->list);
646 spin_unlock(&gmap_notifier_lock);
648 EXPORT_SYMBOL_GPL(gmap_unregister_ipte_notifier);
651 * gmap_ipte_notify - mark a range of ptes for invalidation notification
652 * @gmap: pointer to guest mapping meta data structure
653 * @address: virtual address in the guest address space
656 * Returns 0 if for each page in the given range a gmap mapping exists and
657 * the invalidation notification could be set. If the gmap mapping is missing
658 * for one or more pages -EFAULT is returned. If no memory could be allocated
659 * -ENOMEM is returned. This function establishes missing page table entries.
661 int gmap_ipte_notify(struct gmap *gmap, unsigned long start, unsigned long len)
669 if ((start & ~PAGE_MASK) || (len & ~PAGE_MASK))
671 down_read(&gmap->mm->mmap_sem);
673 /* Convert gmap address and connect the page tables */
674 addr = __gmap_fault(start, gmap);
675 if (IS_ERR_VALUE(addr)) {
679 /* Get the page mapped */
680 if (fixup_user_fault(current, gmap->mm, addr, FAULT_FLAG_WRITE)) {
684 /* Walk the process page table, lock and get pte pointer */
685 ptep = get_locked_pte(gmap->mm, addr, &ptl);
688 /* Set notification bit in the pgste of the pte */
690 if ((pte_val(entry) & (_PAGE_INVALID | _PAGE_RO)) == 0) {
691 pgste = pgste_get_lock(ptep);
692 pgste_val(pgste) |= RCP_IN_BIT;
693 pgste_set_unlock(ptep, pgste);
699 up_read(&gmap->mm->mmap_sem);
702 EXPORT_SYMBOL_GPL(gmap_ipte_notify);
705 * gmap_do_ipte_notify - call all invalidation callbacks for a specific pte.
706 * @mm: pointer to the process mm_struct
707 * @addr: virtual address in the process address space
708 * @pte: pointer to the page table entry
710 * This function is assumed to be called with the page table lock held
711 * for the pte to notify.
713 void gmap_do_ipte_notify(struct mm_struct *mm, unsigned long addr, pte_t *pte)
715 unsigned long segment_offset;
716 struct gmap_notifier *nb;
717 struct gmap_pgtable *mp;
718 struct gmap_rmap *rmap;
721 segment_offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
722 segment_offset = segment_offset * (4096 / sizeof(pte_t));
723 page = pfn_to_page(__pa(pte) >> PAGE_SHIFT);
724 mp = (struct gmap_pgtable *) page->index;
725 spin_lock(&gmap_notifier_lock);
726 list_for_each_entry(rmap, &mp->mapper, list) {
727 list_for_each_entry(nb, &gmap_notifier_list, list)
728 nb->notifier_call(rmap->gmap,
729 rmap->vmaddr + segment_offset);
731 spin_unlock(&gmap_notifier_lock);
734 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm,
735 unsigned long vmaddr)
738 unsigned long *table;
739 struct gmap_pgtable *mp;
741 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
744 mp = kmalloc(sizeof(*mp), GFP_KERNEL|__GFP_REPEAT);
749 pgtable_page_ctor(page);
750 mp->vmaddr = vmaddr & PMD_MASK;
751 INIT_LIST_HEAD(&mp->mapper);
752 page->index = (unsigned long) mp;
753 atomic_set(&page->_mapcount, 3);
754 table = (unsigned long *) page_to_phys(page);
755 clear_table(table, _PAGE_TYPE_EMPTY, PAGE_SIZE/2);
756 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
760 static inline void page_table_free_pgste(unsigned long *table)
763 struct gmap_pgtable *mp;
765 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
766 mp = (struct gmap_pgtable *) page->index;
767 BUG_ON(!list_empty(&mp->mapper));
768 pgtable_page_dtor(page);
769 atomic_set(&page->_mapcount, -1);
774 #else /* CONFIG_PGSTE */
776 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm,
777 unsigned long vmaddr)
782 static inline void page_table_free_pgste(unsigned long *table)
786 static inline void gmap_disconnect_pgtable(struct mm_struct *mm,
787 unsigned long *table)
791 #endif /* CONFIG_PGSTE */
793 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
795 unsigned int old, new;
798 old = atomic_read(v);
800 } while (atomic_cmpxchg(v, old, new) != old);
805 * page table entry allocation/free routines.
807 unsigned long *page_table_alloc(struct mm_struct *mm, unsigned long vmaddr)
809 unsigned long *uninitialized_var(table);
810 struct page *uninitialized_var(page);
811 unsigned int mask, bit;
813 if (mm_has_pgste(mm))
814 return page_table_alloc_pgste(mm, vmaddr);
815 /* Allocate fragments of a 4K page as 1K/2K page table */
816 spin_lock_bh(&mm->context.list_lock);
818 if (!list_empty(&mm->context.pgtable_list)) {
819 page = list_first_entry(&mm->context.pgtable_list,
821 table = (unsigned long *) page_to_phys(page);
822 mask = atomic_read(&page->_mapcount);
823 mask = mask | (mask >> 4);
825 if ((mask & FRAG_MASK) == FRAG_MASK) {
826 spin_unlock_bh(&mm->context.list_lock);
827 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
830 pgtable_page_ctor(page);
831 atomic_set(&page->_mapcount, 1);
832 table = (unsigned long *) page_to_phys(page);
833 clear_table(table, _PAGE_TYPE_EMPTY, PAGE_SIZE);
834 spin_lock_bh(&mm->context.list_lock);
835 list_add(&page->lru, &mm->context.pgtable_list);
837 for (bit = 1; mask & bit; bit <<= 1)
838 table += PTRS_PER_PTE;
839 mask = atomic_xor_bits(&page->_mapcount, bit);
840 if ((mask & FRAG_MASK) == FRAG_MASK)
841 list_del(&page->lru);
843 spin_unlock_bh(&mm->context.list_lock);
847 void page_table_free(struct mm_struct *mm, unsigned long *table)
850 unsigned int bit, mask;
852 if (mm_has_pgste(mm)) {
853 gmap_disconnect_pgtable(mm, table);
854 return page_table_free_pgste(table);
856 /* Free 1K/2K page table fragment of a 4K page */
857 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
858 bit = 1 << ((__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)));
859 spin_lock_bh(&mm->context.list_lock);
860 if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
861 list_del(&page->lru);
862 mask = atomic_xor_bits(&page->_mapcount, bit);
863 if (mask & FRAG_MASK)
864 list_add(&page->lru, &mm->context.pgtable_list);
865 spin_unlock_bh(&mm->context.list_lock);
867 pgtable_page_dtor(page);
868 atomic_set(&page->_mapcount, -1);
873 static void __page_table_free_rcu(void *table, unsigned bit)
877 if (bit == FRAG_MASK)
878 return page_table_free_pgste(table);
879 /* Free 1K/2K page table fragment of a 4K page */
880 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
881 if (atomic_xor_bits(&page->_mapcount, bit) == 0) {
882 pgtable_page_dtor(page);
883 atomic_set(&page->_mapcount, -1);
888 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table)
890 struct mm_struct *mm;
892 unsigned int bit, mask;
895 if (mm_has_pgste(mm)) {
896 gmap_disconnect_pgtable(mm, table);
897 table = (unsigned long *) (__pa(table) | FRAG_MASK);
898 tlb_remove_table(tlb, table);
901 bit = 1 << ((__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)));
902 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
903 spin_lock_bh(&mm->context.list_lock);
904 if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
905 list_del(&page->lru);
906 mask = atomic_xor_bits(&page->_mapcount, bit | (bit << 4));
907 if (mask & FRAG_MASK)
908 list_add_tail(&page->lru, &mm->context.pgtable_list);
909 spin_unlock_bh(&mm->context.list_lock);
910 table = (unsigned long *) (__pa(table) | (bit << 4));
911 tlb_remove_table(tlb, table);
914 void __tlb_remove_table(void *_table)
916 const unsigned long mask = (FRAG_MASK << 4) | FRAG_MASK;
917 void *table = (void *)((unsigned long) _table & ~mask);
918 unsigned type = (unsigned long) _table & mask;
921 __page_table_free_rcu(table, type);
923 free_pages((unsigned long) table, ALLOC_ORDER);
926 static void tlb_remove_table_smp_sync(void *arg)
928 /* Simply deliver the interrupt */
931 static void tlb_remove_table_one(void *table)
934 * This isn't an RCU grace period and hence the page-tables cannot be
935 * assumed to be actually RCU-freed.
937 * It is however sufficient for software page-table walkers that rely
938 * on IRQ disabling. See the comment near struct mmu_table_batch.
940 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
941 __tlb_remove_table(table);
944 static void tlb_remove_table_rcu(struct rcu_head *head)
946 struct mmu_table_batch *batch;
949 batch = container_of(head, struct mmu_table_batch, rcu);
951 for (i = 0; i < batch->nr; i++)
952 __tlb_remove_table(batch->tables[i]);
954 free_page((unsigned long)batch);
957 void tlb_table_flush(struct mmu_gather *tlb)
959 struct mmu_table_batch **batch = &tlb->batch;
962 __tlb_flush_mm(tlb->mm);
963 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
968 void tlb_remove_table(struct mmu_gather *tlb, void *table)
970 struct mmu_table_batch **batch = &tlb->batch;
972 if (*batch == NULL) {
973 *batch = (struct mmu_table_batch *)
974 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
975 if (*batch == NULL) {
976 __tlb_flush_mm(tlb->mm);
977 tlb_remove_table_one(table);
982 (*batch)->tables[(*batch)->nr++] = table;
983 if ((*batch)->nr == MAX_TABLE_BATCH)
984 tlb_table_flush(tlb);
987 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
988 void thp_split_vma(struct vm_area_struct *vma)
993 for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
994 page = follow_page(vma, addr, FOLL_SPLIT);
998 void thp_split_mm(struct mm_struct *mm)
1000 struct vm_area_struct *vma = mm->mmap;
1002 while (vma != NULL) {
1004 vma->vm_flags &= ~VM_HUGEPAGE;
1005 vma->vm_flags |= VM_NOHUGEPAGE;
1009 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1012 * switch on pgstes for its userspace process (for kvm)
1014 int s390_enable_sie(void)
1016 struct task_struct *tsk = current;
1017 struct mm_struct *mm, *old_mm;
1019 /* Do we have switched amode? If no, we cannot do sie */
1020 if (s390_user_mode == HOME_SPACE_MODE)
1023 /* Do we have pgstes? if yes, we are done */
1024 if (mm_has_pgste(tsk->mm))
1027 /* lets check if we are allowed to replace the mm */
1029 if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
1031 !hlist_empty(&tsk->mm->ioctx_list) ||
1033 tsk->mm != tsk->active_mm) {
1039 /* we copy the mm and let dup_mm create the page tables with_pgstes */
1040 tsk->mm->context.alloc_pgste = 1;
1041 /* make sure that both mms have a correct rss state */
1042 sync_mm_rss(tsk->mm);
1044 tsk->mm->context.alloc_pgste = 0;
1048 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1049 /* split thp mappings and disable thp for future mappings */
1051 mm->def_flags |= VM_NOHUGEPAGE;
1054 /* Now lets check again if something happened */
1056 if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
1058 !hlist_empty(&tsk->mm->ioctx_list) ||
1060 tsk->mm != tsk->active_mm) {
1066 /* ok, we are alone. No ptrace, no threads, etc. */
1068 tsk->mm = tsk->active_mm = mm;
1071 atomic_inc(&mm->context.attach_count);
1072 atomic_dec(&old_mm->context.attach_count);
1073 cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
1079 EXPORT_SYMBOL_GPL(s390_enable_sie);
1081 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1082 int pmdp_clear_flush_young(struct vm_area_struct *vma, unsigned long address,
1085 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1086 /* No need to flush TLB
1087 * On s390 reference bits are in storage key and never in TLB */
1088 return pmdp_test_and_clear_young(vma, address, pmdp);
1091 int pmdp_set_access_flags(struct vm_area_struct *vma,
1092 unsigned long address, pmd_t *pmdp,
1093 pmd_t entry, int dirty)
1095 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1097 if (pmd_same(*pmdp, entry))
1099 pmdp_invalidate(vma, address, pmdp);
1100 set_pmd_at(vma->vm_mm, address, pmdp, entry);
1104 static void pmdp_splitting_flush_sync(void *arg)
1106 /* Simply deliver the interrupt */
1109 void pmdp_splitting_flush(struct vm_area_struct *vma, unsigned long address,
1112 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1113 if (!test_and_set_bit(_SEGMENT_ENTRY_SPLIT_BIT,
1114 (unsigned long *) pmdp)) {
1115 /* need to serialize against gup-fast (IRQ disabled) */
1116 smp_call_function(pmdp_splitting_flush_sync, NULL, 1);
1120 void pgtable_trans_huge_deposit(struct mm_struct *mm, pgtable_t pgtable)
1122 struct list_head *lh = (struct list_head *) pgtable;
1124 assert_spin_locked(&mm->page_table_lock);
1127 if (!mm->pmd_huge_pte)
1130 list_add(lh, (struct list_head *) mm->pmd_huge_pte);
1131 mm->pmd_huge_pte = pgtable;
1134 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm)
1136 struct list_head *lh;
1140 assert_spin_locked(&mm->page_table_lock);
1143 pgtable = mm->pmd_huge_pte;
1144 lh = (struct list_head *) pgtable;
1146 mm->pmd_huge_pte = NULL;
1148 mm->pmd_huge_pte = (pgtable_t) lh->next;
1151 ptep = (pte_t *) pgtable;
1152 pte_val(*ptep) = _PAGE_TYPE_EMPTY;
1154 pte_val(*ptep) = _PAGE_TYPE_EMPTY;
1157 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */