1 /* SPDX-License-Identifier: GPL-2.0-only */ <<
2 /* 1 /*
3 * Copyright (c) 2012-2021, Arm Limited. !! 2 * This file is subject to the terms and conditions of the GNU General Public
>> 3 * License. See the file "COPYING" in the main directory of this archive
>> 4 * for more details.
4 * 5 *
5 * Adapted from the original at: !! 6 * Unified implementation of memcpy, memmove and the __copy_user backend.
6 * https://github.com/ARM-software/optimized-r !! 7 *
>> 8 * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org)
>> 9 * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc.
>> 10 * Copyright (C) 2002 Broadcom, Inc.
>> 11 * memcpy/copy_user author: Mark Vandevoorde
>> 12 * Copyright (C) 2007 Maciej W. Rozycki
>> 13 * Copyright (C) 2014 Imagination Technologies Ltd.
>> 14 *
>> 15 * Mnemonic names for arguments to memcpy/__copy_user
7 */ 16 */
8 17
9 #include <linux/linkage.h> !! 18 /*
10 #include <asm/assembler.h> !! 19 * Hack to resolve longstanding prefetch issue
11 !! 20 *
12 /* Assumptions: !! 21 * Prefetching may be fatal on some systems if we're prefetching beyond the
13 * !! 22 * end of memory on some systems. It's also a seriously bad idea on non
14 * ARMv8-a, AArch64, unaligned accesses. !! 23 * dma-coherent systems.
15 * !! 24 */
16 */ !! 25 #ifdef CONFIG_DMA_NONCOHERENT
17 !! 26 #undef CONFIG_CPU_HAS_PREFETCH
18 #define L(label) .L ## label !! 27 #endif
19 !! 28 #ifdef CONFIG_MIPS_MALTA
20 #define dstin x0 !! 29 #undef CONFIG_CPU_HAS_PREFETCH
21 #define src x1 !! 30 #endif
22 #define count x2 !! 31 #ifdef CONFIG_CPU_MIPSR6
23 #define dst x3 !! 32 #undef CONFIG_CPU_HAS_PREFETCH
24 #define srcend x4 !! 33 #endif
25 #define dstend x5 !! 34
26 #define A_l x6 !! 35 #include <linux/export.h>
27 #define A_lw w6 !! 36 #include <asm/asm.h>
28 #define A_h x7 !! 37 #include <asm/asm-offsets.h>
29 #define B_l x8 !! 38 #include <asm/regdef.h>
30 #define B_lw w8 !! 39
31 #define B_h x9 !! 40 #define dst a0
32 #define C_l x10 !! 41 #define src a1
33 #define C_lw w10 !! 42 #define len a2
34 #define C_h x11 !! 43
35 #define D_l x12 !! 44 /*
36 #define D_h x13 !! 45 * Spec
37 #define E_l x14 !! 46 *
38 #define E_h x15 !! 47 * memcpy copies len bytes from src to dst and sets v0 to dst.
39 #define F_l x16 !! 48 * It assumes that
40 #define F_h x17 !! 49 * - src and dst don't overlap
41 #define G_l count !! 50 * - src is readable
42 #define G_h dst !! 51 * - dst is writable
43 #define H_l src !! 52 * memcpy uses the standard calling convention
44 #define H_h srcend !! 53 *
45 #define tmp1 x14 !! 54 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
46 !! 55 * the number of uncopied bytes due to an exception caused by a read or write.
47 /* This implementation handles overlaps and su !! 56 * __copy_user assumes that src and dst don't overlap, and that the call is
48 from a single entry point. It uses unalign !! 57 * implementing one of the following:
49 sequences to keep the code small, simple an !! 58 * copy_to_user
50 !! 59 * - src is readable (no exceptions when reading src)
51 Copies are split into 3 main cases: small c !! 60 * copy_from_user
52 copies of up to 128 bytes, and large copies !! 61 * - dst is writable (no exceptions when writing dst)
53 check is negligible since it is only requir !! 62 * __copy_user uses a non-standard calling convention; see
54 !! 63 * include/asm-mips/uaccess.h
55 Large copies use a software pipelined loop !! 64 *
56 The destination pointer is 16-byte aligned !! 65 * When an exception happens on a load, the handler must
57 The loop tail is handled by always copying !! 66 # ensure that all of the destination buffer is overwritten to prevent
58 */ !! 67 * leaking information to user mode programs.
59 !! 68 */
60 SYM_FUNC_START(__pi_memcpy) !! 69
61 add srcend, src, count !! 70 /*
62 add dstend, dstin, count !! 71 * Implementation
63 cmp count, 128 !! 72 */
64 b.hi L(copy_long) !! 73
65 cmp count, 32 !! 74 /*
66 b.hi L(copy32_128) !! 75 * The exception handler for loads requires that:
67 !! 76 * 1- AT contain the address of the byte just past the end of the source
68 /* Small copies: 0..32 bytes. */ !! 77 * of the copy,
69 cmp count, 16 !! 78 * 2- src_entry <= src < AT, and
70 b.lo L(copy16) !! 79 * 3- (dst - src) == (dst_entry - src_entry),
71 ldp A_l, A_h, [src] !! 80 * The _entry suffix denotes values when __copy_user was called.
72 ldp D_l, D_h, [srcend, -16] !! 81 *
73 stp A_l, A_h, [dstin] !! 82 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
74 stp D_l, D_h, [dstend, -16] !! 83 * (2) is met by incrementing src by the number of bytes copied
75 ret !! 84 * (3) is met by not doing loads between a pair of increments of dst and src
76 !! 85 *
77 /* Copy 8-15 bytes. */ !! 86 * The exception handlers for stores adjust len (if necessary) and return.
78 L(copy16): !! 87 * These handlers do not need to overwrite any data.
79 tbz count, 3, L(copy8) !! 88 *
80 ldr A_l, [src] !! 89 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
81 ldr A_h, [srcend, -8] !! 90 * they're not protected.
82 str A_l, [dstin] !! 91 */
83 str A_h, [dstend, -8] !! 92
84 ret !! 93 /* Instruction type */
85 !! 94 #define LD_INSN 1
86 .p2align 3 !! 95 #define ST_INSN 2
87 /* Copy 4-7 bytes. */ !! 96 /* Pretech type */
88 L(copy8): !! 97 #define SRC_PREFETCH 1
89 tbz count, 2, L(copy4) !! 98 #define DST_PREFETCH 2
90 ldr A_lw, [src] !! 99 #define LEGACY_MODE 1
91 ldr B_lw, [srcend, -4] !! 100 #define EVA_MODE 2
92 str A_lw, [dstin] !! 101 #define USEROP 1
93 str B_lw, [dstend, -4] !! 102 #define KERNELOP 2
94 ret !! 103
95 !! 104 /*
96 /* Copy 0..3 bytes using a branchless !! 105 * Wrapper to add an entry in the exception table
97 L(copy4): !! 106 * in case the insn causes a memory exception.
98 cbz count, L(copy0) !! 107 * Arguments:
99 lsr tmp1, count, 1 !! 108 * insn : Load/store instruction
100 ldrb A_lw, [src] !! 109 * type : Instruction type
101 ldrb C_lw, [srcend, -1] !! 110 * reg : Register
102 ldrb B_lw, [src, tmp1] !! 111 * addr : Address
103 strb A_lw, [dstin] !! 112 * handler : Exception handler
104 strb B_lw, [dstin, tmp1] !! 113 */
105 strb C_lw, [dstend, -1] !! 114
106 L(copy0): !! 115 #define EXC(insn, type, reg, addr, handler) \
107 ret !! 116 .if \mode == LEGACY_MODE; \
108 !! 117 9: insn reg, addr; \
109 .p2align 4 !! 118 .section __ex_table,"a"; \
110 /* Medium copies: 33..128 bytes. */ !! 119 PTR_WD 9b, handler; \
111 L(copy32_128): !! 120 .previous; \
112 ldp A_l, A_h, [src] !! 121 /* This is assembled in EVA mode */ \
113 ldp B_l, B_h, [src, 16] !! 122 .else; \
114 ldp C_l, C_h, [srcend, -32] !! 123 /* If loading from user or storing to user */ \
115 ldp D_l, D_h, [srcend, -16] !! 124 .if ((\from == USEROP) && (type == LD_INSN)) || \
116 cmp count, 64 !! 125 ((\to == USEROP) && (type == ST_INSN)); \
117 b.hi L(copy128) !! 126 9: __BUILD_EVA_INSN(insn##e, reg, addr); \
118 stp A_l, A_h, [dstin] !! 127 .section __ex_table,"a"; \
119 stp B_l, B_h, [dstin, 16] !! 128 PTR_WD 9b, handler; \
120 stp C_l, C_h, [dstend, -32] !! 129 .previous; \
121 stp D_l, D_h, [dstend, -16] !! 130 .else; \
122 ret !! 131 /* \
123 !! 132 * Still in EVA, but no need for \
124 .p2align 4 !! 133 * exception handler or EVA insn \
125 /* Copy 65..128 bytes. */ !! 134 */ \
126 L(copy128): !! 135 insn reg, addr; \
127 ldp E_l, E_h, [src, 32] !! 136 .endif; \
128 ldp F_l, F_h, [src, 48] !! 137 .endif
129 cmp count, 96 <<
130 b.ls L(copy96) <<
131 ldp G_l, G_h, [srcend, -64] <<
132 ldp H_l, H_h, [srcend, -48] <<
133 stp G_l, G_h, [dstend, -64] <<
134 stp H_l, H_h, [dstend, -48] <<
135 L(copy96): <<
136 stp A_l, A_h, [dstin] <<
137 stp B_l, B_h, [dstin, 16] <<
138 stp E_l, E_h, [dstin, 32] <<
139 stp F_l, F_h, [dstin, 48] <<
140 stp C_l, C_h, [dstend, -32] <<
141 stp D_l, D_h, [dstend, -16] <<
142 ret <<
143 <<
144 .p2align 4 <<
145 /* Copy more than 128 bytes. */ <<
146 L(copy_long): <<
147 /* Use backwards copy if there is an o <<
148 sub tmp1, dstin, src <<
149 cbz tmp1, L(copy0) <<
150 cmp tmp1, count <<
151 b.lo L(copy_long_backwards) <<
152 <<
153 /* Copy 16 bytes and then align dst to <<
154 <<
155 ldp D_l, D_h, [src] <<
156 and tmp1, dstin, 15 <<
157 bic dst, dstin, 15 <<
158 sub src, src, tmp1 <<
159 add count, count, tmp1 /* Cou <<
160 ldp A_l, A_h, [src, 16] <<
161 stp D_l, D_h, [dstin] <<
162 ldp B_l, B_h, [src, 32] <<
163 ldp C_l, C_h, [src, 48] <<
164 ldp D_l, D_h, [src, 64]! <<
165 subs count, count, 128 + 16 /* Tes <<
166 b.ls L(copy64_from_end) <<
167 <<
168 L(loop64): <<
169 stp A_l, A_h, [dst, 16] <<
170 ldp A_l, A_h, [src, 16] <<
171 stp B_l, B_h, [dst, 32] <<
172 ldp B_l, B_h, [src, 32] <<
173 stp C_l, C_h, [dst, 48] <<
174 ldp C_l, C_h, [src, 48] <<
175 stp D_l, D_h, [dst, 64]! <<
176 ldp D_l, D_h, [src, 64]! <<
177 subs count, count, 64 <<
178 b.hi L(loop64) <<
179 <<
180 /* Write the last iteration and copy 6 <<
181 L(copy64_from_end): <<
182 ldp E_l, E_h, [srcend, -64] <<
183 stp A_l, A_h, [dst, 16] <<
184 ldp A_l, A_h, [srcend, -48] <<
185 stp B_l, B_h, [dst, 32] <<
186 ldp B_l, B_h, [srcend, -32] <<
187 stp C_l, C_h, [dst, 48] <<
188 ldp C_l, C_h, [srcend, -16] <<
189 stp D_l, D_h, [dst, 64] <<
190 stp E_l, E_h, [dstend, -64] <<
191 stp A_l, A_h, [dstend, -48] <<
192 stp B_l, B_h, [dstend, -32] <<
193 stp C_l, C_h, [dstend, -16] <<
194 ret <<
195 <<
196 .p2align 4 <<
197 <<
198 /* Large backwards copy for overlappin <<
199 Copy 16 bytes and then align dst to <<
200 L(copy_long_backwards): <<
201 ldp D_l, D_h, [srcend, -16] <<
202 and tmp1, dstend, 15 <<
203 sub srcend, srcend, tmp1 <<
204 sub count, count, tmp1 <<
205 ldp A_l, A_h, [srcend, -16] <<
206 stp D_l, D_h, [dstend, -16] <<
207 ldp B_l, B_h, [srcend, -32] <<
208 ldp C_l, C_h, [srcend, -48] <<
209 ldp D_l, D_h, [srcend, -64]! <<
210 sub dstend, dstend, tmp1 <<
211 subs count, count, 128 <<
212 b.ls L(copy64_from_start) <<
213 <<
214 L(loop64_backwards): <<
215 stp A_l, A_h, [dstend, -16] <<
216 ldp A_l, A_h, [srcend, -16] <<
217 stp B_l, B_h, [dstend, -32] <<
218 ldp B_l, B_h, [srcend, -32] <<
219 stp C_l, C_h, [dstend, -48] <<
220 ldp C_l, C_h, [srcend, -48] <<
221 stp D_l, D_h, [dstend, -64]! <<
222 ldp D_l, D_h, [srcend, -64]! <<
223 subs count, count, 64 <<
224 b.hi L(loop64_backwards) <<
225 <<
226 /* Write the last iteration and copy 6 <<
227 L(copy64_from_start): <<
228 ldp G_l, G_h, [src, 48] <<
229 stp A_l, A_h, [dstend, -16] <<
230 ldp A_l, A_h, [src, 32] <<
231 stp B_l, B_h, [dstend, -32] <<
232 ldp B_l, B_h, [src, 16] <<
233 stp C_l, C_h, [dstend, -48] <<
234 ldp C_l, C_h, [src] <<
235 stp D_l, D_h, [dstend, -64] <<
236 stp G_l, G_h, [dstin, 48] <<
237 stp A_l, A_h, [dstin, 32] <<
238 stp B_l, B_h, [dstin, 16] <<
239 stp C_l, C_h, [dstin] <<
240 ret <<
241 SYM_FUNC_END(__pi_memcpy) <<
242 <<
243 SYM_FUNC_ALIAS(__memcpy, __pi_memcpy) <<
244 EXPORT_SYMBOL(__memcpy) <<
245 SYM_FUNC_ALIAS_WEAK(memcpy, __memcpy) <<
246 EXPORT_SYMBOL(memcpy) <<
247 138
248 SYM_FUNC_ALIAS(__pi_memmove, __pi_memcpy) !! 139 /*
>> 140 * Only on the 64-bit kernel we can made use of 64-bit registers.
>> 141 */
>> 142 #ifdef CONFIG_64BIT
>> 143 #define USE_DOUBLE
>> 144 #endif
>> 145
>> 146 #ifdef USE_DOUBLE
>> 147
>> 148 #define LOADK ld /* No exception */
>> 149 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
>> 150 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
>> 151 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
>> 152 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
>> 153 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
>> 154 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
>> 155 #define ADD daddu
>> 156 #define SUB dsubu
>> 157 #define SRL dsrl
>> 158 #define SRA dsra
>> 159 #define SLL dsll
>> 160 #define SLLV dsllv
>> 161 #define SRLV dsrlv
>> 162 #define NBYTES 8
>> 163 #define LOG_NBYTES 3
249 164
250 SYM_FUNC_ALIAS(__memmove, __pi_memmove) !! 165 /*
251 EXPORT_SYMBOL(__memmove) !! 166 * As we are sharing code base with the mips32 tree (which use the o32 ABI
252 SYM_FUNC_ALIAS_WEAK(memmove, __memmove) !! 167 * register definitions). We need to redefine the register definitions from
>> 168 * the n64 ABI register naming to the o32 ABI register naming.
>> 169 */
>> 170 #undef t0
>> 171 #undef t1
>> 172 #undef t2
>> 173 #undef t3
>> 174 #define t0 $8
>> 175 #define t1 $9
>> 176 #define t2 $10
>> 177 #define t3 $11
>> 178 #define t4 $12
>> 179 #define t5 $13
>> 180 #define t6 $14
>> 181 #define t7 $15
>> 182
>> 183 #else
>> 184
>> 185 #define LOADK lw /* No exception */
>> 186 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
>> 187 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
>> 188 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
>> 189 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
>> 190 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
>> 191 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
>> 192 #define ADD addu
>> 193 #define SUB subu
>> 194 #define SRL srl
>> 195 #define SLL sll
>> 196 #define SRA sra
>> 197 #define SLLV sllv
>> 198 #define SRLV srlv
>> 199 #define NBYTES 4
>> 200 #define LOG_NBYTES 2
>> 201
>> 202 #endif /* USE_DOUBLE */
>> 203
>> 204 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler)
>> 205 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
>> 206
>> 207 #ifdef CONFIG_CPU_HAS_PREFETCH
>> 208 # define _PREF(hint, addr, type) \
>> 209 .if \mode == LEGACY_MODE; \
>> 210 kernel_pref(hint, addr); \
>> 211 .else; \
>> 212 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \
>> 213 ((\to == USEROP) && (type == DST_PREFETCH)); \
>> 214 /* \
>> 215 * PREFE has only 9 bits for the offset \
>> 216 * compared to PREF which has 16, so it may \
>> 217 * need to use the $at register but this \
>> 218 * register should remain intact because it's \
>> 219 * used later on. Therefore use $v1. \
>> 220 */ \
>> 221 .set at=v1; \
>> 222 user_pref(hint, addr); \
>> 223 .set noat; \
>> 224 .else; \
>> 225 kernel_pref(hint, addr); \
>> 226 .endif; \
>> 227 .endif
>> 228 #else
>> 229 # define _PREF(hint, addr, type)
>> 230 #endif
>> 231
>> 232 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
>> 233 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
>> 234
>> 235 #ifdef CONFIG_CPU_LITTLE_ENDIAN
>> 236 #define LDFIRST LOADR
>> 237 #define LDREST LOADL
>> 238 #define STFIRST STORER
>> 239 #define STREST STOREL
>> 240 #define SHIFT_DISCARD SLLV
>> 241 #else
>> 242 #define LDFIRST LOADL
>> 243 #define LDREST LOADR
>> 244 #define STFIRST STOREL
>> 245 #define STREST STORER
>> 246 #define SHIFT_DISCARD SRLV
>> 247 #endif
>> 248
>> 249 #define FIRST(unit) ((unit)*NBYTES)
>> 250 #define REST(unit) (FIRST(unit)+NBYTES-1)
>> 251 #define UNIT(unit) FIRST(unit)
>> 252
>> 253 #define ADDRMASK (NBYTES-1)
>> 254
>> 255 .text
>> 256 .set noreorder
>> 257 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
>> 258 .set noat
>> 259 #else
>> 260 .set at=v1
>> 261 #endif
>> 262
>> 263 .align 5
>> 264
>> 265 /*
>> 266 * Macro to build the __copy_user common code
>> 267 * Arguments:
>> 268 * mode : LEGACY_MODE or EVA_MODE
>> 269 * from : Source operand. USEROP or KERNELOP
>> 270 * to : Destination operand. USEROP or KERNELOP
>> 271 */
>> 272 .macro __BUILD_COPY_USER mode, from, to
>> 273
>> 274 /* initialize __memcpy if this the first time we execute this macro */
>> 275 .ifnotdef __memcpy
>> 276 .set __memcpy, 1
>> 277 .hidden __memcpy /* make sure it does not leak */
>> 278 .endif
>> 279
>> 280 /*
>> 281 * Note: dst & src may be unaligned, len may be 0
>> 282 * Temps
>> 283 */
>> 284 #define rem t8
>> 285
>> 286 R10KCBARRIER(0(ra))
>> 287 /*
>> 288 * The "issue break"s below are very approximate.
>> 289 * Issue delays for dcache fills will perturb the schedule, as will
>> 290 * load queue full replay traps, etc.
>> 291 *
>> 292 * If len < NBYTES use byte operations.
>> 293 */
>> 294 PREFS( 0, 0(src) )
>> 295 PREFD( 1, 0(dst) )
>> 296 sltu t2, len, NBYTES
>> 297 and t1, dst, ADDRMASK
>> 298 PREFS( 0, 1*32(src) )
>> 299 PREFD( 1, 1*32(dst) )
>> 300 bnez t2, .Lcopy_bytes_checklen\@
>> 301 and t0, src, ADDRMASK
>> 302 PREFS( 0, 2*32(src) )
>> 303 PREFD( 1, 2*32(dst) )
>> 304 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
>> 305 bnez t1, .Ldst_unaligned\@
>> 306 nop
>> 307 bnez t0, .Lsrc_unaligned_dst_aligned\@
>> 308 #else /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 309 or t0, t0, t1
>> 310 bnez t0, .Lcopy_unaligned_bytes\@
>> 311 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 312 /*
>> 313 * use delay slot for fall-through
>> 314 * src and dst are aligned; need to compute rem
>> 315 */
>> 316 .Lboth_aligned\@:
>> 317 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
>> 318 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
>> 319 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES)
>> 320 PREFS( 0, 3*32(src) )
>> 321 PREFD( 1, 3*32(dst) )
>> 322 .align 4
>> 323 1:
>> 324 R10KCBARRIER(0(ra))
>> 325 LOAD(t0, UNIT(0)(src), .Ll_exc\@)
>> 326 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
>> 327 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
>> 328 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
>> 329 SUB len, len, 8*NBYTES
>> 330 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
>> 331 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
>> 332 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@)
>> 333 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@)
>> 334 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
>> 335 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
>> 336 ADD src, src, 8*NBYTES
>> 337 ADD dst, dst, 8*NBYTES
>> 338 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
>> 339 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
>> 340 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
>> 341 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
>> 342 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
>> 343 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
>> 344 PREFS( 0, 8*32(src) )
>> 345 PREFD( 1, 8*32(dst) )
>> 346 bne len, rem, 1b
>> 347 nop
>> 348
>> 349 /*
>> 350 * len == rem == the number of bytes left to copy < 8*NBYTES
>> 351 */
>> 352 .Lcleanup_both_aligned\@:
>> 353 beqz len, .Ldone\@
>> 354 sltu t0, len, 4*NBYTES
>> 355 bnez t0, .Lless_than_4units\@
>> 356 and rem, len, (NBYTES-1) # rem = len % NBYTES
>> 357 /*
>> 358 * len >= 4*NBYTES
>> 359 */
>> 360 LOAD( t0, UNIT(0)(src), .Ll_exc\@)
>> 361 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@)
>> 362 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@)
>> 363 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@)
>> 364 SUB len, len, 4*NBYTES
>> 365 ADD src, src, 4*NBYTES
>> 366 R10KCBARRIER(0(ra))
>> 367 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
>> 368 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
>> 369 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
>> 370 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
>> 371 .set reorder /* DADDI_WAR */
>> 372 ADD dst, dst, 4*NBYTES
>> 373 beqz len, .Ldone\@
>> 374 .set noreorder
>> 375 .Lless_than_4units\@:
>> 376 /*
>> 377 * rem = len % NBYTES
>> 378 */
>> 379 beq rem, len, .Lcopy_bytes\@
>> 380 nop
>> 381 1:
>> 382 R10KCBARRIER(0(ra))
>> 383 LOAD(t0, 0(src), .Ll_exc\@)
>> 384 ADD src, src, NBYTES
>> 385 SUB len, len, NBYTES
>> 386 STORE(t0, 0(dst), .Ls_exc_p1u\@)
>> 387 .set reorder /* DADDI_WAR */
>> 388 ADD dst, dst, NBYTES
>> 389 bne rem, len, 1b
>> 390 .set noreorder
>> 391
>> 392 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
>> 393 /*
>> 394 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
>> 395 * A loop would do only a byte at a time with possible branch
>> 396 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
>> 397 * because can't assume read-access to dst. Instead, use
>> 398 * STREST dst, which doesn't require read access to dst.
>> 399 *
>> 400 * This code should perform better than a simple loop on modern,
>> 401 * wide-issue mips processors because the code has fewer branches and
>> 402 * more instruction-level parallelism.
>> 403 */
>> 404 #define bits t2
>> 405 beqz len, .Ldone\@
>> 406 ADD t1, dst, len # t1 is just past last byte of dst
>> 407 li bits, 8*NBYTES
>> 408 SLL rem, len, 3 # rem = number of bits to keep
>> 409 LOAD(t0, 0(src), .Ll_exc\@)
>> 410 SUB bits, bits, rem # bits = number of bits to discard
>> 411 SHIFT_DISCARD t0, t0, bits
>> 412 STREST(t0, -1(t1), .Ls_exc\@)
>> 413 jr ra
>> 414 move len, zero
>> 415 .Ldst_unaligned\@:
>> 416 /*
>> 417 * dst is unaligned
>> 418 * t0 = src & ADDRMASK
>> 419 * t1 = dst & ADDRMASK; T1 > 0
>> 420 * len >= NBYTES
>> 421 *
>> 422 * Copy enough bytes to align dst
>> 423 * Set match = (src and dst have same alignment)
>> 424 */
>> 425 #define match rem
>> 426 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
>> 427 ADD t2, zero, NBYTES
>> 428 LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
>> 429 SUB t2, t2, t1 # t2 = number of bytes copied
>> 430 xor match, t0, t1
>> 431 R10KCBARRIER(0(ra))
>> 432 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
>> 433 beq len, t2, .Ldone\@
>> 434 SUB len, len, t2
>> 435 ADD dst, dst, t2
>> 436 beqz match, .Lboth_aligned\@
>> 437 ADD src, src, t2
>> 438
>> 439 .Lsrc_unaligned_dst_aligned\@:
>> 440 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
>> 441 PREFS( 0, 3*32(src) )
>> 442 beqz t0, .Lcleanup_src_unaligned\@
>> 443 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
>> 444 PREFD( 1, 3*32(dst) )
>> 445 1:
>> 446 /*
>> 447 * Avoid consecutive LD*'s to the same register since some mips
>> 448 * implementations can't issue them in the same cycle.
>> 449 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
>> 450 * are to the same unit (unless src is aligned, but it's not).
>> 451 */
>> 452 R10KCBARRIER(0(ra))
>> 453 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
>> 454 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
>> 455 SUB len, len, 4*NBYTES
>> 456 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
>> 457 LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
>> 458 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
>> 459 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
>> 460 LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
>> 461 LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
>> 462 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed)
>> 463 ADD src, src, 4*NBYTES
>> 464 #ifdef CONFIG_CPU_SB1
>> 465 nop # improves slotting
>> 466 #endif
>> 467 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
>> 468 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
>> 469 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
>> 470 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
>> 471 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed)
>> 472 .set reorder /* DADDI_WAR */
>> 473 ADD dst, dst, 4*NBYTES
>> 474 bne len, rem, 1b
>> 475 .set noreorder
>> 476
>> 477 .Lcleanup_src_unaligned\@:
>> 478 beqz len, .Ldone\@
>> 479 and rem, len, NBYTES-1 # rem = len % NBYTES
>> 480 beq rem, len, .Lcopy_bytes\@
>> 481 nop
>> 482 1:
>> 483 R10KCBARRIER(0(ra))
>> 484 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
>> 485 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
>> 486 ADD src, src, NBYTES
>> 487 SUB len, len, NBYTES
>> 488 STORE(t0, 0(dst), .Ls_exc_p1u\@)
>> 489 .set reorder /* DADDI_WAR */
>> 490 ADD dst, dst, NBYTES
>> 491 bne len, rem, 1b
>> 492 .set noreorder
>> 493
>> 494 #endif /* !CONFIG_CPU_NO_LOAD_STORE_LR */
>> 495 .Lcopy_bytes_checklen\@:
>> 496 beqz len, .Ldone\@
>> 497 nop
>> 498 .Lcopy_bytes\@:
>> 499 /* 0 < len < NBYTES */
>> 500 R10KCBARRIER(0(ra))
>> 501 #define COPY_BYTE(N) \
>> 502 LOADB(t0, N(src), .Ll_exc\@); \
>> 503 SUB len, len, 1; \
>> 504 beqz len, .Ldone\@; \
>> 505 STOREB(t0, N(dst), .Ls_exc_p1\@)
>> 506
>> 507 COPY_BYTE(0)
>> 508 COPY_BYTE(1)
>> 509 #ifdef USE_DOUBLE
>> 510 COPY_BYTE(2)
>> 511 COPY_BYTE(3)
>> 512 COPY_BYTE(4)
>> 513 COPY_BYTE(5)
>> 514 #endif
>> 515 LOADB(t0, NBYTES-2(src), .Ll_exc\@)
>> 516 SUB len, len, 1
>> 517 jr ra
>> 518 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
>> 519 .Ldone\@:
>> 520 jr ra
>> 521 nop
>> 522
>> 523 #ifdef CONFIG_CPU_NO_LOAD_STORE_LR
>> 524 .Lcopy_unaligned_bytes\@:
>> 525 1:
>> 526 COPY_BYTE(0)
>> 527 COPY_BYTE(1)
>> 528 COPY_BYTE(2)
>> 529 COPY_BYTE(3)
>> 530 COPY_BYTE(4)
>> 531 COPY_BYTE(5)
>> 532 COPY_BYTE(6)
>> 533 COPY_BYTE(7)
>> 534 ADD src, src, 8
>> 535 b 1b
>> 536 ADD dst, dst, 8
>> 537 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 538 .if __memcpy == 1
>> 539 END(memcpy)
>> 540 .set __memcpy, 0
>> 541 .hidden __memcpy
>> 542 .endif
>> 543
>> 544 .Ll_exc_copy\@:
>> 545 /*
>> 546 * Copy bytes from src until faulting load address (or until a
>> 547 * lb faults)
>> 548 *
>> 549 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
>> 550 * may be more than a byte beyond the last address.
>> 551 * Hence, the lb below may get an exception.
>> 552 *
>> 553 * Assumes src < THREAD_BUADDR($28)
>> 554 */
>> 555 LOADK t0, TI_TASK($28)
>> 556 nop
>> 557 LOADK t0, THREAD_BUADDR(t0)
>> 558 1:
>> 559 LOADB(t1, 0(src), .Ll_exc\@)
>> 560 ADD src, src, 1
>> 561 sb t1, 0(dst) # can't fault -- we're copy_from_user
>> 562 .set reorder /* DADDI_WAR */
>> 563 ADD dst, dst, 1
>> 564 bne src, t0, 1b
>> 565 .set noreorder
>> 566 .Ll_exc\@:
>> 567 LOADK t0, TI_TASK($28)
>> 568 nop
>> 569 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
>> 570 nop
>> 571 SUB len, AT, t0 # len number of uncopied bytes
>> 572 jr ra
>> 573 nop
>> 574
>> 575 #define SEXC(n) \
>> 576 .set reorder; /* DADDI_WAR */ \
>> 577 .Ls_exc_p ## n ## u\@: \
>> 578 ADD len, len, n*NBYTES; \
>> 579 jr ra; \
>> 580 .set noreorder
>> 581
>> 582 SEXC(8)
>> 583 SEXC(7)
>> 584 SEXC(6)
>> 585 SEXC(5)
>> 586 SEXC(4)
>> 587 SEXC(3)
>> 588 SEXC(2)
>> 589 SEXC(1)
>> 590
>> 591 .Ls_exc_p1\@:
>> 592 .set reorder /* DADDI_WAR */
>> 593 ADD len, len, 1
>> 594 jr ra
>> 595 .set noreorder
>> 596 .Ls_exc\@:
>> 597 jr ra
>> 598 nop
>> 599 .endm
>> 600
>> 601 #ifndef CONFIG_HAVE_PLAT_MEMCPY
>> 602 .align 5
>> 603 LEAF(memmove)
253 EXPORT_SYMBOL(memmove) 604 EXPORT_SYMBOL(memmove)
>> 605 ADD t0, a0, a2
>> 606 ADD t1, a1, a2
>> 607 sltu t0, a1, t0 # dst + len <= src -> memcpy
>> 608 sltu t1, a0, t1 # dst >= src + len -> memcpy
>> 609 and t0, t1
>> 610 beqz t0, .L__memcpy
>> 611 move v0, a0 /* return value */
>> 612 beqz a2, .Lr_out
>> 613 END(memmove)
>> 614
>> 615 /* fall through to __rmemcpy */
>> 616 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
>> 617 sltu t0, a1, a0
>> 618 beqz t0, .Lr_end_bytes_up # src >= dst
>> 619 nop
>> 620 ADD a0, a2 # dst = dst + len
>> 621 ADD a1, a2 # src = src + len
>> 622
>> 623 .Lr_end_bytes:
>> 624 R10KCBARRIER(0(ra))
>> 625 lb t0, -1(a1)
>> 626 SUB a2, a2, 0x1
>> 627 sb t0, -1(a0)
>> 628 SUB a1, a1, 0x1
>> 629 .set reorder /* DADDI_WAR */
>> 630 SUB a0, a0, 0x1
>> 631 bnez a2, .Lr_end_bytes
>> 632 .set noreorder
>> 633
>> 634 .Lr_out:
>> 635 jr ra
>> 636 move a2, zero
>> 637
>> 638 .Lr_end_bytes_up:
>> 639 R10KCBARRIER(0(ra))
>> 640 lb t0, (a1)
>> 641 SUB a2, a2, 0x1
>> 642 sb t0, (a0)
>> 643 ADD a1, a1, 0x1
>> 644 .set reorder /* DADDI_WAR */
>> 645 ADD a0, a0, 0x1
>> 646 bnez a2, .Lr_end_bytes_up
>> 647 .set noreorder
>> 648
>> 649 jr ra
>> 650 move a2, zero
>> 651 END(__rmemcpy)
>> 652
>> 653 /*
>> 654 * A combined memcpy/__copy_user
>> 655 * __copy_user sets len to 0 for success; else to an upper bound of
>> 656 * the number of uncopied bytes.
>> 657 * memcpy sets v0 to dst.
>> 658 */
>> 659 .align 5
>> 660 LEAF(memcpy) /* a0=dst a1=src a2=len */
>> 661 EXPORT_SYMBOL(memcpy)
>> 662 move v0, dst /* return value */
>> 663 .L__memcpy:
>> 664 #ifndef CONFIG_EVA
>> 665 FEXPORT(__raw_copy_from_user)
>> 666 EXPORT_SYMBOL(__raw_copy_from_user)
>> 667 FEXPORT(__raw_copy_to_user)
>> 668 EXPORT_SYMBOL(__raw_copy_to_user)
>> 669 #endif
>> 670 /* Legacy Mode, user <-> user */
>> 671 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP
>> 672
>> 673 #endif
>> 674
>> 675 #ifdef CONFIG_EVA
>> 676
>> 677 /*
>> 678 * For EVA we need distinct symbols for reading and writing to user space.
>> 679 * This is because we need to use specific EVA instructions to perform the
>> 680 * virtual <-> physical translation when a virtual address is actually in user
>> 681 * space
>> 682 */
>> 683
>> 684 /*
>> 685 * __copy_from_user (EVA)
>> 686 */
>> 687
>> 688 LEAF(__raw_copy_from_user)
>> 689 EXPORT_SYMBOL(__raw_copy_from_user)
>> 690 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP
>> 691 END(__raw_copy_from_user)
>> 692
>> 693
>> 694
>> 695 /*
>> 696 * __copy_to_user (EVA)
>> 697 */
>> 698
>> 699 LEAF(__raw_copy_to_user)
>> 700 EXPORT_SYMBOL(__raw_copy_to_user)
>> 701 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP
>> 702 END(__raw_copy_to_user)
>> 703
>> 704 #endif
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.