2 * Copyright 2008 Vitaly Mayatskikh <vmayatsk@redhat.com>
3 * Copyright 2002 Andi Kleen, SuSE Labs.
4 * Subject to the GNU Public License v2.
6 * Functions to copy from and to user space.
9 #include <linux/linkage.h>
10 #include <asm/dwarf2.h>
12 #define FIX_ALIGNMENT 1
14 #include <asm/current.h>
15 #include <asm/asm-offsets.h>
16 #include <asm/thread_info.h>
17 #include <asm/cpufeatures.h>
18 #include <asm/alternative-asm.h>
23 * By placing feature2 after feature1 in altinstructions section, we logically
25 * If CPU has feature2, jmp to alt2 is used
26 * else if CPU has feature1, jmp to alt1 is used
27 * else jmp to orig is used.
29 .macro ALTERNATIVE_JUMP feature1,feature2,orig,alt1,alt2
33 .section .altinstr_replacement,"ax"
40 .section .altinstructions,"a"
41 altinstruction_entry 0b,2b,\feature1,5,5,0
42 altinstruction_entry 0b,3b,\feature2,5,5,0
46 .macro ALIGN_DESTINATION
48 /* check for bad alignment of destination */
51 jz 102f /* already aligned */
63 103: addl %ecx,%edx /* ecx is zerorest also */
64 jmp copy_user_handle_tail
67 _ASM_EXTABLE(100b,103b)
68 _ASM_EXTABLE(101b,103b)
72 /* Standard copy_to_user with segment limit checking */
79 cmpq TI_addr_limit(%rax),%rcx
81 ALTERNATIVE_JUMP X86_FEATURE_REP_GOOD,X86_FEATURE_ERMS, \
82 copy_user_generic_unrolled,copy_user_generic_string, \
83 copy_user_enhanced_fast_string
85 ENDPROC(_copy_to_user)
87 /* Standard copy_from_user with segment limit checking */
88 ENTRY(_copy_from_user)
94 cmpq TI_addr_limit(%rax),%rcx
96 ALTERNATIVE_JUMP X86_FEATURE_REP_GOOD,X86_FEATURE_ERMS, \
97 copy_user_generic_unrolled,copy_user_generic_string, \
98 copy_user_enhanced_fast_string
100 ENDPROC(_copy_from_user)
115 ENDPROC(bad_from_user)
119 * copy_user_generic_unrolled - memory copy with exception handling.
120 * This version is for CPUs like P4 that don't have efficient micro
129 * eax uncopied bytes or 0 if successful.
131 ENTRY(copy_user_generic_unrolled)
135 jb 20f /* less then 8 bytes, go to byte copy loop */
142 2: movq 1*8(%rsi),%r9
143 3: movq 2*8(%rsi),%r10
144 4: movq 3*8(%rsi),%r11
146 6: movq %r9,1*8(%rdi)
147 7: movq %r10,2*8(%rdi)
148 8: movq %r11,3*8(%rdi)
149 9: movq 4*8(%rsi),%r8
150 10: movq 5*8(%rsi),%r9
151 11: movq 6*8(%rsi),%r10
152 12: movq 7*8(%rsi),%r11
153 13: movq %r8,4*8(%rdi)
154 14: movq %r9,5*8(%rdi)
155 15: movq %r10,6*8(%rdi)
156 16: movq %r11,7*8(%rdi)
188 40: leal (%rdx,%rcx,8),%edx
191 60: jmp copy_user_handle_tail /* ecx is zerorest also */
203 _ASM_EXTABLE(10b,30b)
204 _ASM_EXTABLE(11b,30b)
205 _ASM_EXTABLE(12b,30b)
206 _ASM_EXTABLE(13b,30b)
207 _ASM_EXTABLE(14b,30b)
208 _ASM_EXTABLE(15b,30b)
209 _ASM_EXTABLE(16b,30b)
210 _ASM_EXTABLE(18b,40b)
211 _ASM_EXTABLE(19b,40b)
212 _ASM_EXTABLE(21b,50b)
213 _ASM_EXTABLE(22b,50b)
215 ENDPROC(copy_user_generic_unrolled)
217 /* Some CPUs run faster using the string copy instructions.
218 * This is also a lot simpler. Use them when possible.
220 * Only 4GB of copy is supported. This shouldn't be a problem
221 * because the kernel normally only writes from/to page sized chunks
222 * even if user space passed a longer buffer.
223 * And more would be dangerous because both Intel and AMD have
224 * errata with rep movsq > 4GB. If someone feels the need to fix
225 * this please consider this.
233 * eax uncopied bytes or 0 if successful.
235 ENTRY(copy_user_generic_string)
239 jb 2f /* less than 8 bytes, go to byte copy loop */
254 11: leal (%rdx,%rcx,8),%ecx
255 12: movl %ecx,%edx /* ecx is zerorest also */
256 jmp copy_user_handle_tail
262 ENDPROC(copy_user_generic_string)
265 * Some CPUs are adding enhanced REP MOVSB/STOSB instructions.
266 * It's recommended to use enhanced REP MOVSB/STOSB if it's enabled.
274 * eax uncopied bytes or 0 if successful.
276 ENTRY(copy_user_enhanced_fast_string)
287 12: movl %ecx,%edx /* ecx is zerorest also */
288 jmp copy_user_handle_tail
293 ENDPROC(copy_user_enhanced_fast_string)