xref: /linux/arch/arm/lib/uaccess_with_memcpy.c (revision cff4fa8415a3224a5abdd2b1dd7f431e4ea49366)
1 /*
2  *  linux/arch/arm/lib/uaccess_with_memcpy.c
3  *
4  *  Written by: Lennert Buytenhek and Nicolas Pitre
5  *  Copyright (C) 2009 Marvell Semiconductor
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/ctype.h>
14 #include <linux/uaccess.h>
15 #include <linux/rwsem.h>
16 #include <linux/mm.h>
17 #include <linux/sched.h>
18 #include <linux/hardirq.h> /* for in_atomic() */
19 #include <linux/gfp.h>
20 #include <asm/current.h>
21 #include <asm/page.h>
22 
23 static int
24 pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
25 {
26 	unsigned long addr = (unsigned long)_addr;
27 	pgd_t *pgd;
28 	pmd_t *pmd;
29 	pte_t *pte;
30 	pud_t *pud;
31 	spinlock_t *ptl;
32 
33 	pgd = pgd_offset(current->mm, addr);
34 	if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd)))
35 		return 0;
36 
37 	pud = pud_offset(pgd, addr);
38 	if (unlikely(pud_none(*pud) || pud_bad(*pud)))
39 		return 0;
40 
41 	pmd = pmd_offset(pud, addr);
42 	if (unlikely(pmd_none(*pmd) || pmd_bad(*pmd)))
43 		return 0;
44 
45 	pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl);
46 	if (unlikely(!pte_present(*pte) || !pte_young(*pte) ||
47 	    !pte_write(*pte) || !pte_dirty(*pte))) {
48 		pte_unmap_unlock(pte, ptl);
49 		return 0;
50 	}
51 
52 	*ptep = pte;
53 	*ptlp = ptl;
54 
55 	return 1;
56 }
57 
58 static unsigned long noinline
59 __copy_to_user_memcpy(void __user *to, const void *from, unsigned long n)
60 {
61 	int atomic;
62 
63 	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
64 		memcpy((void *)to, from, n);
65 		return 0;
66 	}
67 
68 	/* the mmap semaphore is taken only if not in an atomic context */
69 	atomic = in_atomic();
70 
71 	if (!atomic)
72 		down_read(&current->mm->mmap_sem);
73 	while (n) {
74 		pte_t *pte;
75 		spinlock_t *ptl;
76 		int tocopy;
77 
78 		while (!pin_page_for_write(to, &pte, &ptl)) {
79 			if (!atomic)
80 				up_read(&current->mm->mmap_sem);
81 			if (__put_user(0, (char __user *)to))
82 				goto out;
83 			if (!atomic)
84 				down_read(&current->mm->mmap_sem);
85 		}
86 
87 		tocopy = (~(unsigned long)to & ~PAGE_MASK) + 1;
88 		if (tocopy > n)
89 			tocopy = n;
90 
91 		memcpy((void *)to, from, tocopy);
92 		to += tocopy;
93 		from += tocopy;
94 		n -= tocopy;
95 
96 		pte_unmap_unlock(pte, ptl);
97 	}
98 	if (!atomic)
99 		up_read(&current->mm->mmap_sem);
100 
101 out:
102 	return n;
103 }
104 
105 unsigned long
106 __copy_to_user(void __user *to, const void *from, unsigned long n)
107 {
108 	/*
109 	 * This test is stubbed out of the main function above to keep
110 	 * the overhead for small copies low by avoiding a large
111 	 * register dump on the stack just to reload them right away.
112 	 * With frame pointer disabled, tail call optimization kicks in
113 	 * as well making this test almost invisible.
114 	 */
115 	if (n < 64)
116 		return __copy_to_user_std(to, from, n);
117 	return __copy_to_user_memcpy(to, from, n);
118 }
119 
120 static unsigned long noinline
121 __clear_user_memset(void __user *addr, unsigned long n)
122 {
123 	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
124 		memset((void *)addr, 0, n);
125 		return 0;
126 	}
127 
128 	down_read(&current->mm->mmap_sem);
129 	while (n) {
130 		pte_t *pte;
131 		spinlock_t *ptl;
132 		int tocopy;
133 
134 		while (!pin_page_for_write(addr, &pte, &ptl)) {
135 			up_read(&current->mm->mmap_sem);
136 			if (__put_user(0, (char __user *)addr))
137 				goto out;
138 			down_read(&current->mm->mmap_sem);
139 		}
140 
141 		tocopy = (~(unsigned long)addr & ~PAGE_MASK) + 1;
142 		if (tocopy > n)
143 			tocopy = n;
144 
145 		memset((void *)addr, 0, tocopy);
146 		addr += tocopy;
147 		n -= tocopy;
148 
149 		pte_unmap_unlock(pte, ptl);
150 	}
151 	up_read(&current->mm->mmap_sem);
152 
153 out:
154 	return n;
155 }
156 
157 unsigned long __clear_user(void __user *addr, unsigned long n)
158 {
159 	/* See rational for this in __copy_to_user() above. */
160 	if (n < 64)
161 		return __clear_user_std(addr, n);
162 	return __clear_user_memset(addr, n);
163 }
164 
165 #if 0
166 
167 /*
168  * This code is disabled by default, but kept around in case the chosen
169  * thresholds need to be revalidated.  Some overhead (small but still)
170  * would be implied by a runtime determined variable threshold, and
171  * so far the measurement on concerned targets didn't show a worthwhile
172  * variation.
173  *
174  * Note that a fairly precise sched_clock() implementation is needed
175  * for results to make some sense.
176  */
177 
178 #include <linux/vmalloc.h>
179 
180 static int __init test_size_treshold(void)
181 {
182 	struct page *src_page, *dst_page;
183 	void *user_ptr, *kernel_ptr;
184 	unsigned long long t0, t1, t2;
185 	int size, ret;
186 
187 	ret = -ENOMEM;
188 	src_page = alloc_page(GFP_KERNEL);
189 	if (!src_page)
190 		goto no_src;
191 	dst_page = alloc_page(GFP_KERNEL);
192 	if (!dst_page)
193 		goto no_dst;
194 	kernel_ptr = page_address(src_page);
195 	user_ptr = vmap(&dst_page, 1, VM_IOREMAP, __pgprot(__P010));
196 	if (!user_ptr)
197 		goto no_vmap;
198 
199 	/* warm up the src page dcache */
200 	ret = __copy_to_user_memcpy(user_ptr, kernel_ptr, PAGE_SIZE);
201 
202 	for (size = PAGE_SIZE; size >= 4; size /= 2) {
203 		t0 = sched_clock();
204 		ret |= __copy_to_user_memcpy(user_ptr, kernel_ptr, size);
205 		t1 = sched_clock();
206 		ret |= __copy_to_user_std(user_ptr, kernel_ptr, size);
207 		t2 = sched_clock();
208 		printk("copy_to_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
209 	}
210 
211 	for (size = PAGE_SIZE; size >= 4; size /= 2) {
212 		t0 = sched_clock();
213 		ret |= __clear_user_memset(user_ptr, size);
214 		t1 = sched_clock();
215 		ret |= __clear_user_std(user_ptr, size);
216 		t2 = sched_clock();
217 		printk("clear_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
218 	}
219 
220 	if (ret)
221 		ret = -EFAULT;
222 
223 	vunmap(user_ptr);
224 no_vmap:
225 	put_page(dst_page);
226 no_dst:
227 	put_page(src_page);
228 no_src:
229 	return ret;
230 }
231 
232 subsys_initcall(test_size_treshold);
233 
234 #endif
235