Linux Audio

Check our new training course

Loading...
v3.15
 
  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	int atomic;
 92
 93	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
 94		memcpy((void *)to, from, n);
 95		return 0;
 96	}
 97
 98	/* the mmap semaphore is taken only if not in an atomic context */
 99	atomic = in_atomic();
100
101	if (!atomic)
102		down_read(&current->mm->mmap_sem);
103	while (n) {
104		pte_t *pte;
105		spinlock_t *ptl;
106		int tocopy;
107
108		while (!pin_page_for_write(to, &pte, &ptl)) {
109			if (!atomic)
110				up_read(&current->mm->mmap_sem);
111			if (__put_user(0, (char __user *)to))
112				goto out;
113			if (!atomic)
114				down_read(&current->mm->mmap_sem);
115		}
116
117		tocopy = (~(unsigned long)to & ~PAGE_MASK) + 1;
118		if (tocopy > n)
119			tocopy = n;
120
 
121		memcpy((void *)to, from, tocopy);
 
122		to += tocopy;
123		from += tocopy;
124		n -= tocopy;
125
126		if (pte)
127			pte_unmap_unlock(pte, ptl);
128		else
129			spin_unlock(ptl);
130	}
131	if (!atomic)
132		up_read(&current->mm->mmap_sem);
133
134out:
135	return n;
136}
137
138unsigned long
139__copy_to_user(void __user *to, const void *from, unsigned long n)
140{
141	/*
142	 * This test is stubbed out of the main function above to keep
143	 * the overhead for small copies low by avoiding a large
144	 * register dump on the stack just to reload them right away.
145	 * With frame pointer disabled, tail call optimization kicks in
146	 * as well making this test almost invisible.
147	 */
148	if (n < 64)
149		return __copy_to_user_std(to, from, n);
150	return __copy_to_user_memcpy(to, from, n);
 
 
 
 
 
 
151}
152	
153static unsigned long noinline
154__clear_user_memset(void __user *addr, unsigned long n)
155{
156	if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
 
 
157		memset((void *)addr, 0, n);
158		return 0;
159	}
160
161	down_read(&current->mm->mmap_sem);
162	while (n) {
163		pte_t *pte;
164		spinlock_t *ptl;
165		int tocopy;
166
167		while (!pin_page_for_write(addr, &pte, &ptl)) {
168			up_read(&current->mm->mmap_sem);
169			if (__put_user(0, (char __user *)addr))
170				goto out;
171			down_read(&current->mm->mmap_sem);
172		}
173
174		tocopy = (~(unsigned long)addr & ~PAGE_MASK) + 1;
175		if (tocopy > n)
176			tocopy = n;
177
 
178		memset((void *)addr, 0, tocopy);
 
179		addr += tocopy;
180		n -= tocopy;
181
182		if (pte)
183			pte_unmap_unlock(pte, ptl);
184		else
185			spin_unlock(ptl);
186	}
187	up_read(&current->mm->mmap_sem);
188
189out:
190	return n;
191}
192
193unsigned long __clear_user(void __user *addr, unsigned long n)
194{
195	/* See rational for this in __copy_to_user() above. */
196	if (n < 64)
197		return __clear_user_std(addr, n);
198	return __clear_user_memset(addr, n);
 
 
 
 
 
199}
200
201#if 0
202
203/*
204 * This code is disabled by default, but kept around in case the chosen
205 * thresholds need to be revalidated.  Some overhead (small but still)
206 * would be implied by a runtime determined variable threshold, and
207 * so far the measurement on concerned targets didn't show a worthwhile
208 * variation.
209 *
210 * Note that a fairly precise sched_clock() implementation is needed
211 * for results to make some sense.
212 */
213
214#include <linux/vmalloc.h>
215
216static int __init test_size_treshold(void)
217{
218	struct page *src_page, *dst_page;
219	void *user_ptr, *kernel_ptr;
220	unsigned long long t0, t1, t2;
221	int size, ret;
222
223	ret = -ENOMEM;
224	src_page = alloc_page(GFP_KERNEL);
225	if (!src_page)
226		goto no_src;
227	dst_page = alloc_page(GFP_KERNEL);
228	if (!dst_page)
229		goto no_dst;
230	kernel_ptr = page_address(src_page);
231	user_ptr = vmap(&dst_page, 1, VM_IOREMAP, __pgprot(__P010));
232	if (!user_ptr)
233		goto no_vmap;
234
235	/* warm up the src page dcache */
236	ret = __copy_to_user_memcpy(user_ptr, kernel_ptr, PAGE_SIZE);
237
238	for (size = PAGE_SIZE; size >= 4; size /= 2) {
239		t0 = sched_clock();
240		ret |= __copy_to_user_memcpy(user_ptr, kernel_ptr, size);
241		t1 = sched_clock();
242		ret |= __copy_to_user_std(user_ptr, kernel_ptr, size);
243		t2 = sched_clock();
244		printk("copy_to_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
245	}
246
247	for (size = PAGE_SIZE; size >= 4; size /= 2) {
248		t0 = sched_clock();
249		ret |= __clear_user_memset(user_ptr, size);
250		t1 = sched_clock();
251		ret |= __clear_user_std(user_ptr, size);
252		t2 = sched_clock();
253		printk("clear_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
254	}
255
256	if (ret)
257		ret = -EFAULT;
258
259	vunmap(user_ptr);
260no_vmap:
261	put_page(dst_page);
262no_dst:
263	put_page(src_page);
264no_src:
265	return ret;
266}
267
268subsys_initcall(test_size_treshold);
269
270#endif
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 *  linux/arch/arm/lib/uaccess_with_memcpy.c
  4 *
  5 *  Written by: Lennert Buytenhek and Nicolas Pitre
  6 *  Copyright (C) 2009 Marvell Semiconductor
 
 
 
 
  7 */
  8
  9#include <linux/kernel.h>
 10#include <linux/ctype.h>
 11#include <linux/uaccess.h>
 12#include <linux/rwsem.h>
 13#include <linux/mm.h>
 14#include <linux/sched.h>
 15#include <linux/hardirq.h> /* for in_atomic() */
 16#include <linux/gfp.h>
 17#include <linux/highmem.h>
 18#include <linux/hugetlb.h>
 19#include <asm/current.h>
 20#include <asm/page.h>
 21
 22static int
 23pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
 24{
 25	unsigned long addr = (unsigned long)_addr;
 26	pgd_t *pgd;
 27	pmd_t *pmd;
 28	pte_t *pte;
 29	pud_t *pud;
 30	spinlock_t *ptl;
 31
 32	pgd = pgd_offset(current->mm, addr);
 33	if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd)))
 34		return 0;
 35
 36	pud = pud_offset(pgd, addr);
 37	if (unlikely(pud_none(*pud) || pud_bad(*pud)))
 38		return 0;
 39
 40	pmd = pmd_offset(pud, addr);
 41	if (unlikely(pmd_none(*pmd)))
 42		return 0;
 43
 44	/*
 45	 * A pmd can be bad if it refers to a HugeTLB or THP page.
 46	 *
 47	 * Both THP and HugeTLB pages have the same pmd layout
 48	 * and should not be manipulated by the pte functions.
 49	 *
 50	 * Lock the page table for the destination and check
 51	 * to see that it's still huge and whether or not we will
 52	 * need to fault on write.
 53	 */
 54	if (unlikely(pmd_thp_or_huge(*pmd))) {
 55		ptl = &current->mm->page_table_lock;
 56		spin_lock(ptl);
 57		if (unlikely(!pmd_thp_or_huge(*pmd)
 58			|| pmd_hugewillfault(*pmd))) {
 
 59			spin_unlock(ptl);
 60			return 0;
 61		}
 62
 63		*ptep = NULL;
 64		*ptlp = ptl;
 65		return 1;
 66	}
 67
 68	if (unlikely(pmd_bad(*pmd)))
 69		return 0;
 70
 71	pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl);
 72	if (unlikely(!pte_present(*pte) || !pte_young(*pte) ||
 73	    !pte_write(*pte) || !pte_dirty(*pte))) {
 74		pte_unmap_unlock(pte, ptl);
 75		return 0;
 76	}
 77
 78	*ptep = pte;
 79	*ptlp = ptl;
 80
 81	return 1;
 82}
 83
 84static unsigned long noinline
 85__copy_to_user_memcpy(void __user *to, const void *from, unsigned long n)
 86{
 87	unsigned long ua_flags;
 88	int atomic;
 89
 90	if (uaccess_kernel()) {
 91		memcpy((void *)to, from, n);
 92		return 0;
 93	}
 94
 95	/* the mmap semaphore is taken only if not in an atomic context */
 96	atomic = faulthandler_disabled();
 97
 98	if (!atomic)
 99		down_read(&current->mm->mmap_sem);
100	while (n) {
101		pte_t *pte;
102		spinlock_t *ptl;
103		int tocopy;
104
105		while (!pin_page_for_write(to, &pte, &ptl)) {
106			if (!atomic)
107				up_read(&current->mm->mmap_sem);
108			if (__put_user(0, (char __user *)to))
109				goto out;
110			if (!atomic)
111				down_read(&current->mm->mmap_sem);
112		}
113
114		tocopy = (~(unsigned long)to & ~PAGE_MASK) + 1;
115		if (tocopy > n)
116			tocopy = n;
117
118		ua_flags = uaccess_save_and_enable();
119		memcpy((void *)to, from, tocopy);
120		uaccess_restore(ua_flags);
121		to += tocopy;
122		from += tocopy;
123		n -= tocopy;
124
125		if (pte)
126			pte_unmap_unlock(pte, ptl);
127		else
128			spin_unlock(ptl);
129	}
130	if (!atomic)
131		up_read(&current->mm->mmap_sem);
132
133out:
134	return n;
135}
136
137unsigned long
138arm_copy_to_user(void __user *to, const void *from, unsigned long n)
139{
140	/*
141	 * This test is stubbed out of the main function above to keep
142	 * the overhead for small copies low by avoiding a large
143	 * register dump on the stack just to reload them right away.
144	 * With frame pointer disabled, tail call optimization kicks in
145	 * as well making this test almost invisible.
146	 */
147	if (n < 64) {
148		unsigned long ua_flags = uaccess_save_and_enable();
149		n = __copy_to_user_std(to, from, n);
150		uaccess_restore(ua_flags);
151	} else {
152		n = __copy_to_user_memcpy(uaccess_mask_range_ptr(to, n),
153					  from, n);
154	}
155	return n;
156}
157	
158static unsigned long noinline
159__clear_user_memset(void __user *addr, unsigned long n)
160{
161	unsigned long ua_flags;
162
163	if (uaccess_kernel()) {
164		memset((void *)addr, 0, n);
165		return 0;
166	}
167
168	down_read(&current->mm->mmap_sem);
169	while (n) {
170		pte_t *pte;
171		spinlock_t *ptl;
172		int tocopy;
173
174		while (!pin_page_for_write(addr, &pte, &ptl)) {
175			up_read(&current->mm->mmap_sem);
176			if (__put_user(0, (char __user *)addr))
177				goto out;
178			down_read(&current->mm->mmap_sem);
179		}
180
181		tocopy = (~(unsigned long)addr & ~PAGE_MASK) + 1;
182		if (tocopy > n)
183			tocopy = n;
184
185		ua_flags = uaccess_save_and_enable();
186		memset((void *)addr, 0, tocopy);
187		uaccess_restore(ua_flags);
188		addr += tocopy;
189		n -= tocopy;
190
191		if (pte)
192			pte_unmap_unlock(pte, ptl);
193		else
194			spin_unlock(ptl);
195	}
196	up_read(&current->mm->mmap_sem);
197
198out:
199	return n;
200}
201
202unsigned long arm_clear_user(void __user *addr, unsigned long n)
203{
204	/* See rational for this in __copy_to_user() above. */
205	if (n < 64) {
206		unsigned long ua_flags = uaccess_save_and_enable();
207		n = __clear_user_std(addr, n);
208		uaccess_restore(ua_flags);
209	} else {
210		n = __clear_user_memset(addr, n);
211	}
212	return n;
213}
214
215#if 0
216
217/*
218 * This code is disabled by default, but kept around in case the chosen
219 * thresholds need to be revalidated.  Some overhead (small but still)
220 * would be implied by a runtime determined variable threshold, and
221 * so far the measurement on concerned targets didn't show a worthwhile
222 * variation.
223 *
224 * Note that a fairly precise sched_clock() implementation is needed
225 * for results to make some sense.
226 */
227
228#include <linux/vmalloc.h>
229
230static int __init test_size_treshold(void)
231{
232	struct page *src_page, *dst_page;
233	void *user_ptr, *kernel_ptr;
234	unsigned long long t0, t1, t2;
235	int size, ret;
236
237	ret = -ENOMEM;
238	src_page = alloc_page(GFP_KERNEL);
239	if (!src_page)
240		goto no_src;
241	dst_page = alloc_page(GFP_KERNEL);
242	if (!dst_page)
243		goto no_dst;
244	kernel_ptr = page_address(src_page);
245	user_ptr = vmap(&dst_page, 1, VM_IOREMAP, __pgprot(__P010));
246	if (!user_ptr)
247		goto no_vmap;
248
249	/* warm up the src page dcache */
250	ret = __copy_to_user_memcpy(user_ptr, kernel_ptr, PAGE_SIZE);
251
252	for (size = PAGE_SIZE; size >= 4; size /= 2) {
253		t0 = sched_clock();
254		ret |= __copy_to_user_memcpy(user_ptr, kernel_ptr, size);
255		t1 = sched_clock();
256		ret |= __copy_to_user_std(user_ptr, kernel_ptr, size);
257		t2 = sched_clock();
258		printk("copy_to_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
259	}
260
261	for (size = PAGE_SIZE; size >= 4; size /= 2) {
262		t0 = sched_clock();
263		ret |= __clear_user_memset(user_ptr, size);
264		t1 = sched_clock();
265		ret |= __clear_user_std(user_ptr, size);
266		t2 = sched_clock();
267		printk("clear_user: %d %llu %llu\n", size, t1 - t0, t2 - t1);
268	}
269
270	if (ret)
271		ret = -EFAULT;
272
273	vunmap(user_ptr);
274no_vmap:
275	put_page(dst_page);
276no_dst:
277	put_page(src_page);
278no_src:
279	return ret;
280}
281
282subsys_initcall(test_size_treshold);
283
284#endif