1/*
  2 * Lockless get_user_pages_fast for sparc, cribbed from powerpc
  3 *
  4 * Copyright (C) 2008 Nick Piggin
  5 * Copyright (C) 2008 Novell Inc.
  6 */
  7
  8#include <linux/sched.h>
  9#include <linux/mm.h>
 10#include <linux/vmstat.h>
 11#include <linux/pagemap.h>
 12#include <linux/rwsem.h>
 13#include <asm/pgtable.h>
 14
 15/*
 16 * The performance critical leaf functions are made noinline otherwise gcc
 17 * inlines everything into a single function which results in too much
 18 * register pressure.
 19 */
 20static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
 21		unsigned long end, int write, struct page **pages, int *nr)
 22{
 23	unsigned long mask, result;
 24	pte_t *ptep;
 25
 26	if (tlb_type == hypervisor) {
 27		result = _PAGE_PRESENT_4V|_PAGE_P_4V;
 28		if (write)
 29			result |= _PAGE_WRITE_4V;
 30	} else {
 31		result = _PAGE_PRESENT_4U|_PAGE_P_4U;
 32		if (write)
 33			result |= _PAGE_WRITE_4U;
 34	}
 35	mask = result | _PAGE_SPECIAL;
 36
 37	ptep = pte_offset_kernel(&pmd, addr);
 38	do {
 39		struct page *page, *head;
 40		pte_t pte = *ptep;
 41
 42		if ((pte_val(pte) & mask) != result)
 43			return 0;
 44		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 45
 46		/* The hugepage case is simplified on sparc64 because
 47		 * we encode the sub-page pfn offsets into the
 48		 * hugepage PTEs.  We could optimize this in the future
 49		 * use page_cache_add_speculative() for the hugepage case.
 50		 */
 51		page = pte_page(pte);
 52		head = compound_head(page);
 53		if (!page_cache_get_speculative(head))
 54			return 0;
 55		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 56			put_page(head);
 57			return 0;
 58		}
 59		if (head != page)
 60			get_huge_page_tail(page);
 61
 62		pages[*nr] = page;
 63		(*nr)++;
 64	} while (ptep++, addr += PAGE_SIZE, addr != end);
 65
 66	return 1;
 67}
 68
 69static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
 70		int write, struct page **pages, int *nr)
 71{
 72	unsigned long next;
 73	pmd_t *pmdp;
 74
 75	pmdp = pmd_offset(&pud, addr);
 76	do {
 77		pmd_t pmd = *pmdp;
 78
 79		next = pmd_addr_end(addr, end);
 80		if (pmd_none(pmd))
 81			return 0;
 82		if (!gup_pte_range(pmd, addr, next, write, pages, nr))
 83			return 0;
 84	} while (pmdp++, addr = next, addr != end);
 85
 86	return 1;
 87}
 88
 89static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
 90		int write, struct page **pages, int *nr)
 91{
 92	unsigned long next;
 93	pud_t *pudp;
 94
 95	pudp = pud_offset(&pgd, addr);
 96	do {
 97		pud_t pud = *pudp;
 98
 99		next = pud_addr_end(addr, end);
100		if (pud_none(pud))
101			return 0;
102		if (!gup_pmd_range(pud, addr, next, write, pages, nr))
103			return 0;
104	} while (pudp++, addr = next, addr != end);
105
106	return 1;
107}
108
109int get_user_pages_fast(unsigned long start, int nr_pages, int write,
110			struct page **pages)
111{
112	struct mm_struct *mm = current->mm;
113	unsigned long addr, len, end;
114	unsigned long next;
115	pgd_t *pgdp;
116	int nr = 0;
117
118	start &= PAGE_MASK;
119	addr = start;
120	len = (unsigned long) nr_pages << PAGE_SHIFT;
121	end = start + len;
122
123	/*
124	 * XXX: batch / limit 'nr', to avoid large irq off latency
125	 * needs some instrumenting to determine the common sizes used by
126	 * important workloads (eg. DB2), and whether limiting the batch size
127	 * will decrease performance.
128	 *
129	 * It seems like we're in the clear for the moment. Direct-IO is
130	 * the main guy that batches up lots of get_user_pages, and even
131	 * they are limited to 64-at-a-time which is not so many.
132	 */
133	/*
134	 * This doesn't prevent pagetable teardown, but does prevent
135	 * the pagetables from being freed on sparc.
136	 *
137	 * So long as we atomically load page table pointers versus teardown,
138	 * we can follow the address down to the the page and take a ref on it.
139	 */
140	local_irq_disable();
141
142	pgdp = pgd_offset(mm, addr);
143	do {
144		pgd_t pgd = *pgdp;
145
146		next = pgd_addr_end(addr, end);
147		if (pgd_none(pgd))
148			goto slow;
149		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
150			goto slow;
151	} while (pgdp++, addr = next, addr != end);
152
153	local_irq_enable();
154
155	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
156	return nr;
157
158	{
159		int ret;
160
161slow:
162		local_irq_enable();
163
164		/* Try to get the remaining pages with get_user_pages */
165		start += nr << PAGE_SHIFT;
166		pages += nr;
167
168		down_read(&mm->mmap_sem);
169		ret = get_user_pages(current, mm, start,
170			(end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
171		up_read(&mm->mmap_sem);
172
173		/* Have to be a bit careful with return values */
174		if (nr > 0) {
175			if (ret < 0)
176				ret = nr;
177			else
178				ret += nr;
179		}
180
181		return ret;
182	}
183}