2
* Lockless get_user_pages_fast for sparc, cribbed from powerpc
4
* Copyright (C) 2008 Nick Piggin
5
* Copyright (C) 2008 Novell Inc.
8
#include <linux/sched.h>
10
#include <linux/vmstat.h>
11
#include <linux/pagemap.h>
12
#include <linux/rwsem.h>
13
#include <asm/pgtable.h>
16
* The performance critical leaf functions are made noinline otherwise gcc
17
* inlines everything into a single function which results in too much
20
static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
21
unsigned long end, int write, struct page **pages, int *nr)
23
unsigned long mask, result;
26
if (tlb_type == hypervisor) {
27
result = _PAGE_PRESENT_4V|_PAGE_P_4V;
29
result |= _PAGE_WRITE_4V;
31
result = _PAGE_PRESENT_4U|_PAGE_P_4U;
33
result |= _PAGE_WRITE_4U;
35
mask = result | _PAGE_SPECIAL;
37
ptep = pte_offset_kernel(&pmd, addr);
39
struct page *page, *head;
42
if ((pte_val(pte) & mask) != result)
44
VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
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.
52
head = compound_head(page);
53
if (!page_cache_get_speculative(head))
55
if (unlikely(pte_val(pte) != pte_val(*ptep))) {
60
get_huge_page_tail(page);
64
} while (ptep++, addr += PAGE_SIZE, addr != end);
69
static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
70
int write, struct page **pages, int *nr)
75
pmdp = pmd_offset(&pud, addr);
79
next = pmd_addr_end(addr, end);
82
if (!gup_pte_range(pmd, addr, next, write, pages, nr))
84
} while (pmdp++, addr = next, addr != end);
89
static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
90
int write, struct page **pages, int *nr)
95
pudp = pud_offset(&pgd, addr);
99
next = pud_addr_end(addr, end);
102
if (!gup_pmd_range(pud, addr, next, write, pages, nr))
104
} while (pudp++, addr = next, addr != end);
109
int get_user_pages_fast(unsigned long start, int nr_pages, int write,
112
struct mm_struct *mm = current->mm;
113
unsigned long addr, len, end;
120
len = (unsigned long) nr_pages << PAGE_SHIFT;
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.
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.
134
* This doesn't prevent pagetable teardown, but does prevent
135
* the pagetables from being freed on sparc.
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.
142
pgdp = pgd_offset(mm, addr);
146
next = pgd_addr_end(addr, end);
149
if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
151
} while (pgdp++, addr = next, addr != end);
155
VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
164
/* Try to get the remaining pages with get_user_pages */
165
start += nr << PAGE_SHIFT;
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);
173
/* Have to be a bit careful with return values */