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 <linux/highmem.h>
21#include <linux/hugetlb.h>
22#include <asm/current.h>
23#include <asm/page.h>
24
25static int
26pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
27{
28	unsigned long addr = (unsigned long)_addr;
29	pgd_t *pgd;
30	pmd_t *pmd;
31	pte_t *pte;
32	pud_t *pud;
33	spinlock_t *ptl;
34
35	pgd = pgd_offset(current->mm, addr);
36	if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd)))
37		return 0;
38
39	pud = pud_offset(pgd, addr);
40	if (unlikely(pud_none(*pud) || pud_bad(*pud)))
41		return 0;
42
43	pmd = pmd_offset(pud, addr);
44	if (unlikely(pmd_none(*pmd)))
45		return 0;
46
47	/*
48	 * A pmd can be bad if it refers to a HugeTLB or THP page.
49	 *
50	 * Both THP and HugeTLB pages have the same pmd layout
51	 * and should not be manipulated by the pte functions.
52	 *
53	 * Lock the page table for the destination and check
54	 * to see that it's still huge and whether or not we will
55	 * need to fault on write, or if we have a splitting THP.
56	 */
57	if (unlikely(pmd_thp_or_huge(*pmd))) {
58		ptl = &current->mm->page_table_lock;
59		spin_lock(ptl);
60		if (unlikely(!pmd_thp_or_huge(*pmd)
61			|| pmd_hugewillfault(*pmd)
62			|| pmd_trans_splitting(*pmd))) {
63			spin_unlock(ptl);
64			return 0;
65		}
66
67		*ptep = NULL;
68		*ptlp = ptl;
69		return 1;
70	}
71
72	if (unlikely(pmd_bad(*pmd)))
73		return 0;
74
75	pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl);
76	if (unlikely(!pte_present(*pte) || !pte_young(*pte) ||
77	    !pte_write(*pte) || !pte_dirty(*pte))) {
78		pte_unmap_unlock(pte, ptl);
79		return 0;
80	}
81
82	*ptep = pte;
83	*ptlp = ptl;
84
85	return 1;
86}
87
88static unsigned long noinline
89__copy_to_user_memcpy(void __user *to, const void *from, unsigned long n)
90{
91	unsigned long ua_flags;
92	int atomic;
93
94	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
95		memcpy((void *)to, from, n);
96		return 0;
97	}
98
99	/* the mmap semaphore is taken only if not in an atomic context */
100	atomic = faulthandler_disabled();
101
102	if (!atomic)
103		down_read(&current->mm->mmap_sem);
104	while (n) {
105		pte_t *pte;
106		spinlock_t *ptl;
107		int tocopy;
108
109		while (!pin_page_for_write(to, &pte, &ptl)) {
110			if (!atomic)
111				up_read(&current->mm->mmap_sem);
112			if (__put_user(0, (char __user *)to))
113				goto out;
114			if (!atomic)
115				down_read(&current->mm->mmap_sem);
116		}
117
118		tocopy = (~(unsigned long)to & ~PAGE_MASK) + 1;
119		if (tocopy > n)
120			tocopy = n;
121
122		ua_flags = uaccess_save_and_enable();
123		memcpy((void *)to, from, tocopy);
124		uaccess_restore(ua_flags);
125		to += tocopy;
126		from += tocopy;
127		n -= tocopy;
128
129		if (pte)
130			pte_unmap_unlock(pte, ptl);
131		else
132			spin_unlock(ptl);
133	}
134	if (!atomic)
135		up_read(&current->mm->mmap_sem);
136
137out:
138	return n;
139}
140
141unsigned long
142arm_copy_to_user(void __user *to, const void *from, unsigned long n)
143{
144	/*
145	 * This test is stubbed out of the main function above to keep
146	 * the overhead for small copies low by avoiding a large
147	 * register dump on the stack just to reload them right away.
148	 * With frame pointer disabled, tail call optimization kicks in
149	 * as well making this test almost invisible.
150	 */
151	if (n < 64) {
152		unsigned long ua_flags = uaccess_save_and_enable();
153		n = __copy_to_user_std(to, from, n);
154		uaccess_restore(ua_flags);
155	} else {
156		n = __copy_to_user_memcpy(to, from, n);
157	}
158	return n;
159}
160
161static unsigned long noinline
162__clear_user_memset(void __user *addr, unsigned long n)
163{
164	unsigned long ua_flags;
165
166	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
167		memset((void *)addr, 0, n);
168		return 0;
169	}
170
171	down_read(&current->mm->mmap_sem);
172	while (n) {
173		pte_t *pte;
174		spinlock_t *ptl;
175		int tocopy;
176
177		while (!pin_page_for_write(addr, &pte, &ptl)) {
178			up_read(&current->mm->mmap_sem);
179			if (__put_user(0, (char __user *)addr))
180				goto out;
181			down_read(&current->mm->mmap_sem);
182		}
183
184		tocopy = (~(unsigned long)addr & ~PAGE_MASK) + 1;
185		if (tocopy > n)
186			tocopy = n;
187
188		ua_flags = uaccess_save_and_enable();
189		memset((void *)addr, 0, tocopy);
190		uaccess_restore(ua_flags);
191		addr += tocopy;
192		n -= tocopy;
193
194		if (pte)
195			pte_unmap_unlock(pte, ptl);
196		else
197			spin_unlock(ptl);
198	}
199	up_read(&current->mm->mmap_sem);
200
201out:
202	return n;
203}
204
205unsigned long arm_clear_user(void __user *addr, unsigned long n)
206{
207	/* See rational for this in __copy_to_user() above. */
208	if (n < 64) {
209		unsigned long ua_flags = uaccess_save_and_enable();
210		n = __clear_user_std(addr, n);
211		uaccess_restore(ua_flags);
212	} else {
213		n = __clear_user_memset(addr, n);
214	}
215	return n;
216}
217
218#if 0
219
220/*
221 * This code is disabled by default, but kept around in case the chosen
222 * thresholds need to be revalidated.  Some overhead (small but still)
223 * would be implied by a runtime determined variable threshold, and
224 * so far the measurement on concerned targets didn't show a worthwhile
225 * variation.
226 *
227 * Note that a fairly precise sched_clock() implementation is needed
228 * for results to make some sense.
229 */
230
231#include <linux/vmalloc.h>
232
233static int __init test_size_treshold(void)
234{
235	struct page *src_page, *dst_page;
236	void *user_ptr, *kernel_ptr;
237	unsigned long long t0, t1, t2;
238	int size, ret;
239
240	ret = -ENOMEM;
241	src_page = alloc_page(GFP_KERNEL);
242	if (!src_page)
243		goto no_src;
244	dst_page = alloc_page(GFP_KERNEL);
245	if (!dst_page)
246		goto no_dst;
247	kernel_ptr = page_address(src_page);
248	user_ptr = vmap(&dst_page, 1, VM_IOREMAP, __pgprot(__P010));
249	if (!user_ptr)
250		goto no_vmap;
251
252	/* warm up the src page dcache */
253	ret = __copy_to_user_memcpy(user_ptr, kernel_ptr, PAGE_SIZE);
254
255	for (size = PAGE_SIZE; size >= 4; size /= 2) {
256		t0 = sched_clock();
257		ret |= __copy_to_user_memcpy(user_ptr, kernel_ptr, size);
258		t1 = sched_clock();
259		ret |= __copy_to_user_std(user_ptr, kernel_ptr, size);
260		t2 = sched_clock();
261		printk("copy_to_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
262	}
263
264	for (size = PAGE_SIZE; size >= 4; size /= 2) {
265		t0 = sched_clock();
266		ret |= __clear_user_memset(user_ptr, size);
267		t1 = sched_clock();
268		ret |= __clear_user_std(user_ptr, size);
269		t2 = sched_clock();
270		printk("clear_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
271	}
272
273	if (ret)
274		ret = -EFAULT;
275
276	vunmap(user_ptr);
277no_vmap:
278	put_page(dst_page);
279no_dst:
280	put_page(src_page);
281no_src:
282	return ret;
283}
284
285subsys_initcall(test_size_treshold);
286
287#endif
288