| 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | /* |
| 3 | * spu_restore.c |
| 4 | * |
| 5 | * (C) Copyright IBM Corp. 2005 |
| 6 | * |
| 7 | * SPU-side context restore sequence outlined in |
| 8 | * Synergistic Processor Element Book IV |
| 9 | * |
| 10 | * Author: Mark Nutter <mnutter@us.ibm.com> |
| 11 | */ |
| 12 | |
| 13 | |
| 14 | #ifndef LS_SIZE |
| 15 | #define LS_SIZE 0x40000 /* 256K (in bytes) */ |
| 16 | #endif |
| 17 | |
| 18 | typedef unsigned int u32; |
| 19 | typedef unsigned long long u64; |
| 20 | |
| 21 | #include <spu_intrinsics.h> |
| 22 | #include <asm/spu_csa.h> |
| 23 | #include "spu_utils.h" |
| 24 | |
| 25 | #define BR_INSTR 0x327fff80 /* br -4 */ |
| 26 | #define NOP_INSTR 0x40200000 /* nop */ |
| 27 | #define HEQ_INSTR 0x7b000000 /* heq $0, $0 */ |
| 28 | #define STOP_INSTR 0x00000000 /* stop 0x0 */ |
| 29 | #define ILLEGAL_INSTR 0x00800000 /* illegal instr */ |
| 30 | #define RESTORE_COMPLETE 0x00003ffc /* stop 0x3ffc */ |
| 31 | |
| 32 | static inline void fetch_regs_from_mem(addr64 lscsa_ea) |
| 33 | { |
| 34 | unsigned int ls = (unsigned int)®s_spill[0]; |
| 35 | unsigned int size = sizeof(regs_spill); |
| 36 | unsigned int tag_id = 0; |
| 37 | unsigned int cmd = 0x40; /* GET */ |
| 38 | |
| 39 | spu_writech(MFC_LSA, ls); |
| 40 | spu_writech(MFC_EAH, lscsa_ea.ui[0]); |
| 41 | spu_writech(MFC_EAL, lscsa_ea.ui[1]); |
| 42 | spu_writech(MFC_Size, size); |
| 43 | spu_writech(MFC_TagID, tag_id); |
| 44 | spu_writech(MFC_Cmd, cmd); |
| 45 | } |
| 46 | |
| 47 | static inline void restore_upper_240kb(addr64 lscsa_ea) |
| 48 | { |
| 49 | unsigned int ls = 16384; |
| 50 | unsigned int list = (unsigned int)&dma_list[0]; |
| 51 | unsigned int size = sizeof(dma_list); |
| 52 | unsigned int tag_id = 0; |
| 53 | unsigned int cmd = 0x44; /* GETL */ |
| 54 | |
| 55 | /* Restore, Step 4: |
| 56 | * Enqueue the GETL command (tag 0) to the MFC SPU command |
| 57 | * queue to transfer the upper 240 kb of LS from CSA. |
| 58 | */ |
| 59 | spu_writech(MFC_LSA, ls); |
| 60 | spu_writech(MFC_EAH, lscsa_ea.ui[0]); |
| 61 | spu_writech(MFC_EAL, list); |
| 62 | spu_writech(MFC_Size, size); |
| 63 | spu_writech(MFC_TagID, tag_id); |
| 64 | spu_writech(MFC_Cmd, cmd); |
| 65 | } |
| 66 | |
| 67 | static inline void restore_decr(void) |
| 68 | { |
| 69 | unsigned int offset; |
| 70 | unsigned int decr_running; |
| 71 | unsigned int decr; |
| 72 | |
| 73 | /* Restore, Step 6(moved): |
| 74 | * If the LSCSA "decrementer running" flag is set |
| 75 | * then write the SPU_WrDec channel with the |
| 76 | * decrementer value from LSCSA. |
| 77 | */ |
| 78 | offset = LSCSA_QW_OFFSET(decr_status); |
| 79 | decr_running = regs_spill[offset].slot[0] & SPU_DECR_STATUS_RUNNING; |
| 80 | if (decr_running) { |
| 81 | offset = LSCSA_QW_OFFSET(decr); |
| 82 | decr = regs_spill[offset].slot[0]; |
| 83 | spu_writech(SPU_WrDec, decr); |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | static inline void write_ppu_mb(void) |
| 88 | { |
| 89 | unsigned int offset; |
| 90 | unsigned int data; |
| 91 | |
| 92 | /* Restore, Step 11: |
| 93 | * Write the MFC_WrOut_MB channel with the PPU_MB |
| 94 | * data from LSCSA. |
| 95 | */ |
| 96 | offset = LSCSA_QW_OFFSET(ppu_mb); |
| 97 | data = regs_spill[offset].slot[0]; |
| 98 | spu_writech(SPU_WrOutMbox, data); |
| 99 | } |
| 100 | |
| 101 | static inline void write_ppuint_mb(void) |
| 102 | { |
| 103 | unsigned int offset; |
| 104 | unsigned int data; |
| 105 | |
| 106 | /* Restore, Step 12: |
| 107 | * Write the MFC_WrInt_MB channel with the PPUINT_MB |
| 108 | * data from LSCSA. |
| 109 | */ |
| 110 | offset = LSCSA_QW_OFFSET(ppuint_mb); |
| 111 | data = regs_spill[offset].slot[0]; |
| 112 | spu_writech(SPU_WrOutIntrMbox, data); |
| 113 | } |
| 114 | |
| 115 | static inline void restore_fpcr(void) |
| 116 | { |
| 117 | unsigned int offset; |
| 118 | vector unsigned int fpcr; |
| 119 | |
| 120 | /* Restore, Step 13: |
| 121 | * Restore the floating-point status and control |
| 122 | * register from the LSCSA. |
| 123 | */ |
| 124 | offset = LSCSA_QW_OFFSET(fpcr); |
| 125 | fpcr = regs_spill[offset].v; |
| 126 | spu_mtfpscr(fpcr); |
| 127 | } |
| 128 | |
| 129 | static inline void restore_srr0(void) |
| 130 | { |
| 131 | unsigned int offset; |
| 132 | unsigned int srr0; |
| 133 | |
| 134 | /* Restore, Step 14: |
| 135 | * Restore the SPU SRR0 data from the LSCSA. |
| 136 | */ |
| 137 | offset = LSCSA_QW_OFFSET(srr0); |
| 138 | srr0 = regs_spill[offset].slot[0]; |
| 139 | spu_writech(SPU_WrSRR0, srr0); |
| 140 | } |
| 141 | |
| 142 | static inline void restore_event_mask(void) |
| 143 | { |
| 144 | unsigned int offset; |
| 145 | unsigned int event_mask; |
| 146 | |
| 147 | /* Restore, Step 15: |
| 148 | * Restore the SPU_RdEventMsk data from the LSCSA. |
| 149 | */ |
| 150 | offset = LSCSA_QW_OFFSET(event_mask); |
| 151 | event_mask = regs_spill[offset].slot[0]; |
| 152 | spu_writech(SPU_WrEventMask, event_mask); |
| 153 | } |
| 154 | |
| 155 | static inline void restore_tag_mask(void) |
| 156 | { |
| 157 | unsigned int offset; |
| 158 | unsigned int tag_mask; |
| 159 | |
| 160 | /* Restore, Step 16: |
| 161 | * Restore the SPU_RdTagMsk data from the LSCSA. |
| 162 | */ |
| 163 | offset = LSCSA_QW_OFFSET(tag_mask); |
| 164 | tag_mask = regs_spill[offset].slot[0]; |
| 165 | spu_writech(MFC_WrTagMask, tag_mask); |
| 166 | } |
| 167 | |
| 168 | static inline void restore_complete(void) |
| 169 | { |
| 170 | extern void exit_fini(void); |
| 171 | unsigned int *exit_instrs = (unsigned int *)exit_fini; |
| 172 | unsigned int offset; |
| 173 | unsigned int stopped_status; |
| 174 | unsigned int stopped_code; |
| 175 | |
| 176 | /* Restore, Step 18: |
| 177 | * Issue a stop-and-signal instruction with |
| 178 | * "good context restore" signal value. |
| 179 | * |
| 180 | * Restore, Step 19: |
| 181 | * There may be additional instructions placed |
| 182 | * here by the PPE Sequence for SPU Context |
| 183 | * Restore in order to restore the correct |
| 184 | * "stopped state". |
| 185 | * |
| 186 | * This step is handled here by analyzing the |
| 187 | * LSCSA.stopped_status and then modifying the |
| 188 | * exit() function to behave appropriately. |
| 189 | */ |
| 190 | |
| 191 | offset = LSCSA_QW_OFFSET(stopped_status); |
| 192 | stopped_status = regs_spill[offset].slot[0]; |
| 193 | stopped_code = regs_spill[offset].slot[1]; |
| 194 | |
| 195 | switch (stopped_status) { |
| 196 | case SPU_STOPPED_STATUS_P_I: |
| 197 | /* SPU_Status[P,I]=1. Add illegal instruction |
| 198 | * followed by stop-and-signal instruction after |
| 199 | * end of restore code. |
| 200 | */ |
| 201 | exit_instrs[0] = RESTORE_COMPLETE; |
| 202 | exit_instrs[1] = ILLEGAL_INSTR; |
| 203 | exit_instrs[2] = STOP_INSTR | stopped_code; |
| 204 | break; |
| 205 | case SPU_STOPPED_STATUS_P_H: |
| 206 | /* SPU_Status[P,H]=1. Add 'heq $0, $0' followed |
| 207 | * by stop-and-signal instruction after end of |
| 208 | * restore code. |
| 209 | */ |
| 210 | exit_instrs[0] = RESTORE_COMPLETE; |
| 211 | exit_instrs[1] = HEQ_INSTR; |
| 212 | exit_instrs[2] = STOP_INSTR | stopped_code; |
| 213 | break; |
| 214 | case SPU_STOPPED_STATUS_S_P: |
| 215 | /* SPU_Status[S,P]=1. Add nop instruction |
| 216 | * followed by 'br -4' after end of restore |
| 217 | * code. |
| 218 | */ |
| 219 | exit_instrs[0] = RESTORE_COMPLETE; |
| 220 | exit_instrs[1] = STOP_INSTR | stopped_code; |
| 221 | exit_instrs[2] = NOP_INSTR; |
| 222 | exit_instrs[3] = BR_INSTR; |
| 223 | break; |
| 224 | case SPU_STOPPED_STATUS_S_I: |
| 225 | /* SPU_Status[S,I]=1. Add illegal instruction |
| 226 | * followed by 'br -4' after end of restore code. |
| 227 | */ |
| 228 | exit_instrs[0] = RESTORE_COMPLETE; |
| 229 | exit_instrs[1] = ILLEGAL_INSTR; |
| 230 | exit_instrs[2] = NOP_INSTR; |
| 231 | exit_instrs[3] = BR_INSTR; |
| 232 | break; |
| 233 | case SPU_STOPPED_STATUS_I: |
| 234 | /* SPU_Status[I]=1. Add illegal instruction followed |
| 235 | * by infinite loop after end of restore sequence. |
| 236 | */ |
| 237 | exit_instrs[0] = RESTORE_COMPLETE; |
| 238 | exit_instrs[1] = ILLEGAL_INSTR; |
| 239 | exit_instrs[2] = NOP_INSTR; |
| 240 | exit_instrs[3] = BR_INSTR; |
| 241 | break; |
| 242 | case SPU_STOPPED_STATUS_S: |
| 243 | /* SPU_Status[S]=1. Add two 'nop' instructions. */ |
| 244 | exit_instrs[0] = RESTORE_COMPLETE; |
| 245 | exit_instrs[1] = NOP_INSTR; |
| 246 | exit_instrs[2] = NOP_INSTR; |
| 247 | exit_instrs[3] = BR_INSTR; |
| 248 | break; |
| 249 | case SPU_STOPPED_STATUS_H: |
| 250 | /* SPU_Status[H]=1. Add 'heq $0, $0' instruction |
| 251 | * after end of restore code. |
| 252 | */ |
| 253 | exit_instrs[0] = RESTORE_COMPLETE; |
| 254 | exit_instrs[1] = HEQ_INSTR; |
| 255 | exit_instrs[2] = NOP_INSTR; |
| 256 | exit_instrs[3] = BR_INSTR; |
| 257 | break; |
| 258 | case SPU_STOPPED_STATUS_P: |
| 259 | /* SPU_Status[P]=1. Add stop-and-signal instruction |
| 260 | * after end of restore code. |
| 261 | */ |
| 262 | exit_instrs[0] = RESTORE_COMPLETE; |
| 263 | exit_instrs[1] = STOP_INSTR | stopped_code; |
| 264 | break; |
| 265 | case SPU_STOPPED_STATUS_R: |
| 266 | /* SPU_Status[I,S,H,P,R]=0. Add infinite loop. */ |
| 267 | exit_instrs[0] = RESTORE_COMPLETE; |
| 268 | exit_instrs[1] = NOP_INSTR; |
| 269 | exit_instrs[2] = NOP_INSTR; |
| 270 | exit_instrs[3] = BR_INSTR; |
| 271 | break; |
| 272 | default: |
| 273 | /* SPU_Status[R]=1. No additional instructions. */ |
| 274 | break; |
| 275 | } |
| 276 | spu_sync(); |
| 277 | } |
| 278 | |
| 279 | /** |
| 280 | * main - entry point for SPU-side context restore. |
| 281 | * |
| 282 | * This code deviates from the documented sequence in the |
| 283 | * following aspects: |
| 284 | * |
| 285 | * 1. The EA for LSCSA is passed from PPE in the |
| 286 | * signal notification channels. |
| 287 | * 2. The register spill area is pulled by SPU |
| 288 | * into LS, rather than pushed by PPE. |
| 289 | * 3. All 128 registers are restored by exit(). |
| 290 | * 4. The exit() function is modified at run |
| 291 | * time in order to properly restore the |
| 292 | * SPU_Status register. |
| 293 | */ |
| 294 | int main() |
| 295 | { |
| 296 | addr64 lscsa_ea; |
| 297 | |
| 298 | lscsa_ea.ui[0] = spu_readch(SPU_RdSigNotify1); |
| 299 | lscsa_ea.ui[1] = spu_readch(SPU_RdSigNotify2); |
| 300 | fetch_regs_from_mem(lscsa_ea); |
| 301 | |
| 302 | set_event_mask(); /* Step 1. */ |
| 303 | set_tag_mask(); /* Step 2. */ |
| 304 | build_dma_list(lscsa_ea); /* Step 3. */ |
| 305 | restore_upper_240kb(lscsa_ea); /* Step 4. */ |
| 306 | /* Step 5: done by 'exit'. */ |
| 307 | enqueue_putllc(lscsa_ea); /* Step 7. */ |
| 308 | set_tag_update(); /* Step 8. */ |
| 309 | read_tag_status(); /* Step 9. */ |
| 310 | restore_decr(); /* moved Step 6. */ |
| 311 | read_llar_status(); /* Step 10. */ |
| 312 | write_ppu_mb(); /* Step 11. */ |
| 313 | write_ppuint_mb(); /* Step 12. */ |
| 314 | restore_fpcr(); /* Step 13. */ |
| 315 | restore_srr0(); /* Step 14. */ |
| 316 | restore_event_mask(); /* Step 15. */ |
| 317 | restore_tag_mask(); /* Step 16. */ |
| 318 | /* Step 17. done by 'exit'. */ |
| 319 | restore_complete(); /* Step 18. */ |
| 320 | |
| 321 | return 0; |
| 322 | } |
| 323 | |