| 1 | /* |
|---|
| 2 | * hal_remote.c - implementation of Generic Remote Access API for TSAR-MIPS32 |
|---|
| 3 | * |
|---|
| 4 | * Authors : Mohammed Karaoui (2015) |
|---|
| 5 | * Alain Greiner (2016) |
|---|
| 6 | * |
|---|
| 7 | * Copyright (c) UPMC Sorbonne Universites |
|---|
| 8 | * |
|---|
| 9 | * This file is part of ALMOS-MKH.. |
|---|
| 10 | * |
|---|
| 11 | * ALMOS-MKH. is free software; you can redistribute it and/or modify it |
|---|
| 12 | * under the terms of the GNU General Public License as published by |
|---|
| 13 | * the Free Software Foundation; version 2.0 of the License. |
|---|
| 14 | * |
|---|
| 15 | * ALMOS-MKH. is distributed in the hope that it will be useful, but |
|---|
| 16 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
|---|
| 17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
|---|
| 18 | * General Public License for more details. |
|---|
| 19 | * |
|---|
| 20 | * You should have received a copy of the GNU General Public License |
|---|
| 21 | * along with ALMOS-MKH.; if not, write to the Free Software Foundation, |
|---|
| 22 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
|---|
| 23 | */ |
|---|
| 24 | |
|---|
| 25 | #include <hal_types.h> |
|---|
| 26 | |
|---|
| 27 | //////////////////////////////// |
|---|
| 28 | void hal_remote_sb( xptr_t xp, |
|---|
| 29 | char data ) |
|---|
| 30 | { |
|---|
| 31 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 32 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 33 | |
|---|
| 34 | asm volatile( |
|---|
| 35 | ".set noreorder \n" |
|---|
| 36 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 37 | "mtc2 %2, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 38 | "sb %0, 0(%1) \n" /* *paddr <= value */ |
|---|
| 39 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 40 | "sync \n" |
|---|
| 41 | ".set reorder \n" |
|---|
| 42 | : : "r" (data), "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 43 | } |
|---|
| 44 | |
|---|
| 45 | ///////////////////////////////// |
|---|
| 46 | void hal_remote_sw( xptr_t xp, |
|---|
| 47 | uint32_t data ) |
|---|
| 48 | { |
|---|
| 49 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 50 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 51 | |
|---|
| 52 | asm volatile( "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 53 | ".set noreorder \n" |
|---|
| 54 | "mtc2 %2, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 55 | "sw %0, 0(%1) \n" /* *paddr <= value */ |
|---|
| 56 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 57 | "sync \n" |
|---|
| 58 | ".set reorder \n" |
|---|
| 59 | : : "r" (data), "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 60 | } |
|---|
| 61 | |
|---|
| 62 | ////////////////////////////////// |
|---|
| 63 | void hal_remote_swd( xptr_t xp, |
|---|
| 64 | uint64_t data ) |
|---|
| 65 | { |
|---|
| 66 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 67 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 68 | |
|---|
| 69 | uint32_t data_lsb = (uint32_t)data; |
|---|
| 70 | uint32_t data_msb = (uint32_t)(data>>32); |
|---|
| 71 | |
|---|
| 72 | asm volatile( |
|---|
| 73 | ".set noreorder \n" |
|---|
| 74 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 75 | "mtc2 %3, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 76 | "sw %0, 0(%2) \n" /* *paddr <= lsb */ |
|---|
| 77 | "sw %1, 4(%2) \n" /* *(paddr+4) <= msb */ |
|---|
| 78 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 79 | "sync \n" |
|---|
| 80 | ".set reorder \n" |
|---|
| 81 | : : "r" (data_lsb), "r" (data_msb), "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 82 | } |
|---|
| 83 | |
|---|
| 84 | //////////////////////////////////// |
|---|
| 85 | void hal_remote_spt( xptr_t xp, |
|---|
| 86 | void * pt ) |
|---|
| 87 | { |
|---|
| 88 | hal_remote_sw ( xp , (uint32_t)pt ); |
|---|
| 89 | } |
|---|
| 90 | |
|---|
| 91 | //////////////////////////////// |
|---|
| 92 | char hal_remote_lb( xptr_t xp ) |
|---|
| 93 | { |
|---|
| 94 | char data; |
|---|
| 95 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 96 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 97 | |
|---|
| 98 | asm volatile( |
|---|
| 99 | ".set noreorder \n" |
|---|
| 100 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 101 | "mtc2 %2, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 102 | "lb %0, 0(%1) \n" /* data <= *paddr */ |
|---|
| 103 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 104 | ".set reorder \n" |
|---|
| 105 | : "=r" (data) : "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 106 | |
|---|
| 107 | return ( data ); |
|---|
| 108 | } |
|---|
| 109 | |
|---|
| 110 | //////////////////////////////////// |
|---|
| 111 | uint32_t hal_remote_lw( xptr_t xp ) |
|---|
| 112 | { |
|---|
| 113 | uint32_t data; |
|---|
| 114 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 115 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 116 | |
|---|
| 117 | asm volatile( |
|---|
| 118 | ".set noreorder \n" |
|---|
| 119 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 120 | "mtc2 %2, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 121 | "lw %0, 0(%1) \n" /* data <= *paddr */ |
|---|
| 122 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 123 | ".set reorder \n" |
|---|
| 124 | : "=r" (data) : "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 125 | |
|---|
| 126 | return ( data ); |
|---|
| 127 | } |
|---|
| 128 | |
|---|
| 129 | ///////////////////////////////////// |
|---|
| 130 | uint64_t hal_remote_lwd( xptr_t xp ) |
|---|
| 131 | { |
|---|
| 132 | uint32_t data_lsb; |
|---|
| 133 | uint32_t data_msb; |
|---|
| 134 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 135 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 136 | |
|---|
| 137 | asm volatile( |
|---|
| 138 | ".set noreorder \n" |
|---|
| 139 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 140 | "mtc2 %3, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 141 | "lw %0, 0(%2) \n" /* data_lsb <= *paddr */ |
|---|
| 142 | "lw %1, 4(%2) \n" /* data_msb <= *paddr+4 */ |
|---|
| 143 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 144 | ".set reorder \n" |
|---|
| 145 | : "=r" (data_lsb), "=r"(data_msb) : "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 146 | |
|---|
| 147 | return ( (((uint64_t)data_msb)<<32) + (((uint64_t)data_lsb)) ); |
|---|
| 148 | } |
|---|
| 149 | |
|---|
| 150 | //////////////////////////////////// |
|---|
| 151 | void * hal_remote_lpt( xptr_t xp, |
|---|
| 152 | void * pt ) |
|---|
| 153 | { |
|---|
| 154 | return (void *)hal_remote_lw ( xp ); |
|---|
| 155 | } |
|---|
| 156 | |
|---|
| 157 | //////////////////////////////////////// |
|---|
| 158 | uint32_t hal_remote_lw_unc( xptr_t xp ) |
|---|
| 159 | { |
|---|
| 160 | uint32_t data; |
|---|
| 161 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 162 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 163 | |
|---|
| 164 | asm volatile( |
|---|
| 165 | ".set noreorder \n" |
|---|
| 166 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 167 | "mtc2 %2, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 168 | "ll %0, 0(%1) \n" /* data <= *paddr */ |
|---|
| 169 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 170 | ".set reorder \n" |
|---|
| 171 | : "=r" (data) : "r" (ptr), "r" (cxy) : "$15" ); |
|---|
| 172 | |
|---|
| 173 | return ( data ); |
|---|
| 174 | } |
|---|
| 175 | |
|---|
| 176 | /////////////////////////////////////////// |
|---|
| 177 | bool_t hal_remote_atomic_cas( xptr_t xp, |
|---|
| 178 | uint32_t old, |
|---|
| 179 | uint32_t new ) |
|---|
| 180 | { |
|---|
| 181 | bool_t isAtomic; |
|---|
| 182 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 183 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 184 | |
|---|
| 185 | asm volatile( |
|---|
| 186 | ".set noreorder \n" |
|---|
| 187 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 188 | "mtc2 %4, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 189 | "or $8, $0, %3 \n" /* $8 <= new */ |
|---|
| 190 | "ll $3, 0(%1) \n" /* $3 <= *paddr */ |
|---|
| 191 | "bne $3, %2, 1f \n" /* if ($3 != old) */ |
|---|
| 192 | "li $7, 0 \n" /* $7 <= 0 */ |
|---|
| 193 | "sc $8, (%1) \n" /* *paddr <= new */ |
|---|
| 194 | "or $7, $8, $0 \n" /* $7 <= atomic */ |
|---|
| 195 | "sync \n" |
|---|
| 196 | "1: \n" |
|---|
| 197 | "or %0, $7, $0 \n" /* isAtomic <= $7 */ |
|---|
| 198 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 199 | ".set reorder \n" |
|---|
| 200 | : "=&r" (isAtomic) : "r" (ptr), "r" (old) , "r" (new), "r" (cxy) |
|---|
| 201 | : "$3", "$7", "$8", "$15" ); |
|---|
| 202 | |
|---|
| 203 | return isAtomic; |
|---|
| 204 | } |
|---|
| 205 | |
|---|
| 206 | //////////////////////////////////////////// |
|---|
| 207 | uint32_t hal_remote_atomic_add( xptr_t xp, |
|---|
| 208 | uint32_t incr ) |
|---|
| 209 | { |
|---|
| 210 | uint32_t current; |
|---|
| 211 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 212 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 213 | |
|---|
| 214 | asm volatile( |
|---|
| 215 | ".set noreorder \n" |
|---|
| 216 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 217 | "mtc2 %3, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 218 | "1: \n" |
|---|
| 219 | "ll %0, (%1) \n" /* current <= *paddr */ |
|---|
| 220 | "addu $3, %0, %2 \n" /* $3 <= current + incr */ |
|---|
| 221 | "sc $3, (%1) \n" /* *paddr <= $3 */ |
|---|
| 222 | "beq $3, $0, 1b \n" /* retry if failure */ |
|---|
| 223 | "nop \n" |
|---|
| 224 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 225 | ".set reorder \n" |
|---|
| 226 | : "=&r" (current) : "r" (ptr), "r" (incr), "r" (cxy) : "$3", "$15" ); |
|---|
| 227 | |
|---|
| 228 | return current; |
|---|
| 229 | } |
|---|
| 230 | |
|---|
| 231 | //////////////////////////////////////////// |
|---|
| 232 | uint32_t hal_remote_atomic_and( xptr_t xp, |
|---|
| 233 | uint32_t mask ) |
|---|
| 234 | { |
|---|
| 235 | uint32_t current; |
|---|
| 236 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 237 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 238 | |
|---|
| 239 | asm volatile( |
|---|
| 240 | ".set noreorder \n" |
|---|
| 241 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 242 | "mtc2 %3, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 243 | "1: \n" |
|---|
| 244 | "ll %0, (%1) \n" /* current <= *paddr */ |
|---|
| 245 | "and $3, %0, %2 \n" /* $3 <= current & mask */ |
|---|
| 246 | "sc $3, (%1) \n" /* *paddr <= $3 */ |
|---|
| 247 | "beq $3, $0, 1b \n" /* retry if failure */ |
|---|
| 248 | "nop \n" |
|---|
| 249 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 250 | ".set reorder \n" |
|---|
| 251 | : "=&r" (current) : "r" (ptr), "r" (mask), "r" (cxy) : "$3", "$15" ); |
|---|
| 252 | |
|---|
| 253 | return current; |
|---|
| 254 | } |
|---|
| 255 | |
|---|
| 256 | //////////////////////////////////////////// |
|---|
| 257 | uint32_t hal_remote_atomic_or( xptr_t xp, |
|---|
| 258 | uint32_t mask ) |
|---|
| 259 | { |
|---|
| 260 | uint32_t current; |
|---|
| 261 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 262 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 263 | |
|---|
| 264 | asm volatile( |
|---|
| 265 | ".set noreorder \n" |
|---|
| 266 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 267 | "mtc2 %3, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 268 | "1: \n" |
|---|
| 269 | "ll %0, (%1) \n" /* current <= *paddr */ |
|---|
| 270 | "or $3, %0, %2 \n" /* $3 <= current | mask */ |
|---|
| 271 | "sc $3, (%1) \n" /* *paddr <= $3 */ |
|---|
| 272 | "beq $3, $0, 1b \n" /* retry if failure */ |
|---|
| 273 | "nop \n" |
|---|
| 274 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 275 | ".set reorder \n" |
|---|
| 276 | : "=&r" (current) : "r" (ptr), "r" (mask), "r" (cxy) : "$3", "$15" ); |
|---|
| 277 | |
|---|
| 278 | return current; |
|---|
| 279 | } |
|---|
| 280 | |
|---|
| 281 | ///////////////////////////////////////////////// |
|---|
| 282 | error_t hal_remote_atomic_try_add( xptr_t xp, |
|---|
| 283 | uint32_t incr, |
|---|
| 284 | uint32_t * old ) |
|---|
| 285 | { |
|---|
| 286 | uint32_t current; |
|---|
| 287 | error_t error; |
|---|
| 288 | uint32_t ptr = (uint32_t)GET_PTR( xp ); |
|---|
| 289 | uint32_t cxy = (uint32_t)GET_CXY( xp ); |
|---|
| 290 | |
|---|
| 291 | asm volatile( |
|---|
| 292 | ".set noreorder \n" |
|---|
| 293 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 294 | "mtc2 %4, $24 \n" /* PADDR_EXT <= cxy */ |
|---|
| 295 | "ll %0, (%2) \n" /* current <= *paddr */ |
|---|
| 296 | "addu $3, %0, %3 \n" /* $3 <= current + incr */ |
|---|
| 297 | "sc $3, (%2) \n" /* *paddr <= $3 */ |
|---|
| 298 | "beq $3, $0, 1f \n" /* exit if failure */ |
|---|
| 299 | "ori %1, $0, 1 \n" /* fail: ret <= 1 */ |
|---|
| 300 | "and %1, $0, $0 \n" /* success: ret <= 0 */ |
|---|
| 301 | "1: \n" |
|---|
| 302 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 303 | ".set reorder \n" |
|---|
| 304 | : "=&r" (current), "=&r" (error) : "r" (ptr), "r" (incr), "r" (cxy) : "$3", "$15" ); |
|---|
| 305 | |
|---|
| 306 | *old = current; |
|---|
| 307 | |
|---|
| 308 | return error; |
|---|
| 309 | } |
|---|
| 310 | |
|---|
| 311 | ///////////////////////////////////// |
|---|
| 312 | void hal_remote_memcpy( xptr_t dst, |
|---|
| 313 | xptr_t src, |
|---|
| 314 | uint32_t size ) |
|---|
| 315 | { |
|---|
| 316 | uint32_t i; |
|---|
| 317 | uint32_t wsize; |
|---|
| 318 | uint32_t dptr = (uint32_t)GET_PTR( dst ); |
|---|
| 319 | uint32_t dcxy = (uint32_t)GET_CXY( dst ); |
|---|
| 320 | uint32_t sptr = (uint32_t)GET_PTR( src ); |
|---|
| 321 | uint32_t scxy = (uint32_t)GET_CXY( src ); |
|---|
| 322 | |
|---|
| 323 | if( (dptr & 0x3) || (sptr & 0x3) ) wsize = 0; // do it all in bytes |
|---|
| 324 | else wsize = size >> 2; |
|---|
| 325 | |
|---|
| 326 | for( i = 0 ; i < wsize ; i++ ) // transfer one word per iteration |
|---|
| 327 | { |
|---|
| 328 | asm volatile( |
|---|
| 329 | ".set noreorder \n" |
|---|
| 330 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 331 | "mtc2 %0, $24 \n" /* PADDR_EXT <= scxy */ |
|---|
| 332 | "lw $3, 0(%1) \n" /* $3 <= *src */ |
|---|
| 333 | "mtc2 %2, $24 \n" /* PADDR_EXT <= dcxy */ |
|---|
| 334 | "sw $3, 0(%3) \n" /* *dst <= $3 */ |
|---|
| 335 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 336 | ".set reorder \n" |
|---|
| 337 | : : "r"(scxy), "r" (sptr+(i<<2)), "r"(dcxy), "r" (dptr+(i<<2)) : "$3", "$15" ); |
|---|
| 338 | } |
|---|
| 339 | |
|---|
| 340 | for( i = wsize << 2 ; i < size ; i++ ) // transfer one byte per iteration |
|---|
| 341 | { |
|---|
| 342 | asm volatile( |
|---|
| 343 | ".set noreorder \n" |
|---|
| 344 | "mfc2 $15, $24 \n" /* $15 <= PADDR_EXT */ |
|---|
| 345 | "mtc2 %0, $24 \n" /* PADDR_EXT <= scxy */ |
|---|
| 346 | "lb $3, 0(%1) \n" /* $3 <= *src */ |
|---|
| 347 | "mtc2 %2, $24 \n" /* PADDR_EXT <= dcxy */ |
|---|
| 348 | "sb $3, 0(%3) \n" /* *dst <= $3 */ |
|---|
| 349 | "mtc2 $15, $24 \n" /* PADDR_EXT <= $15 */ |
|---|
| 350 | ".set reorder \n" |
|---|
| 351 | : : "r"(scxy), "r" (sptr+i), "r"(dcxy), "r" (dptr+i) : "$3", "$15" ); |
|---|
| 352 | } |
|---|
| 353 | } |
|---|
| 354 | |
|---|
| 355 | |
|---|