| 1 | // SPDX-License-Identifier: LGPL-2.1-or-later |
| 2 | /* |
| 3 | * Copyright (C) 2015-2021 Paul Cercueil <paul@crapouillou.net> |
| 4 | */ |
| 5 | |
| 6 | #include "blockcache.h" |
| 7 | #include "debug.h" |
| 8 | #include "lightrec-private.h" |
| 9 | #include "memmanager.h" |
| 10 | #include "reaper.h" |
| 11 | #include "recompiler.h" |
| 12 | |
| 13 | #include <stdbool.h> |
| 14 | #include <stdlib.h> |
| 15 | #include <string.h> |
| 16 | |
| 17 | /* Must be power of two */ |
| 18 | #define LUT_SIZE 0x4000 |
| 19 | |
| 20 | struct blockcache { |
| 21 | struct lightrec_state *state; |
| 22 | struct block * lut[LUT_SIZE]; |
| 23 | }; |
| 24 | |
| 25 | u16 lightrec_get_lut_entry(const struct block *block) |
| 26 | { |
| 27 | return (kunseg(block->pc) >> 2) & (LUT_SIZE - 1); |
| 28 | } |
| 29 | |
| 30 | struct block * lightrec_find_block(struct blockcache *cache, u32 pc) |
| 31 | { |
| 32 | struct block *block; |
| 33 | |
| 34 | pc = kunseg(pc); |
| 35 | |
| 36 | for (block = cache->lut[(pc >> 2) & (LUT_SIZE - 1)]; |
| 37 | block; block = block->next) |
| 38 | if (kunseg(block->pc) == pc) |
| 39 | return block; |
| 40 | |
| 41 | return NULL; |
| 42 | } |
| 43 | |
| 44 | struct block * lightrec_find_block_from_lut(struct blockcache *cache, |
| 45 | u16 lut_entry, u32 addr_in_block) |
| 46 | { |
| 47 | struct block *block; |
| 48 | u32 pc; |
| 49 | |
| 50 | addr_in_block = kunseg(addr_in_block); |
| 51 | |
| 52 | for (block = cache->lut[lut_entry]; block; block = block->next) { |
| 53 | pc = kunseg(block->pc); |
| 54 | if (addr_in_block >= pc && |
| 55 | addr_in_block < pc + (block->nb_ops << 2)) |
| 56 | return block; |
| 57 | } |
| 58 | |
| 59 | return NULL; |
| 60 | } |
| 61 | |
| 62 | void remove_from_code_lut(struct blockcache *cache, struct block *block) |
| 63 | { |
| 64 | struct lightrec_state *state = cache->state; |
| 65 | u32 offset = lut_offset(block->pc); |
| 66 | |
| 67 | if (block->function) { |
| 68 | memset(lut_address(state, offset), 0, |
| 69 | block->nb_ops * lut_elm_size(state)); |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | void lightrec_register_block(struct blockcache *cache, struct block *block) |
| 74 | { |
| 75 | u32 pc = kunseg(block->pc); |
| 76 | struct block *old; |
| 77 | |
| 78 | old = cache->lut[(pc >> 2) & (LUT_SIZE - 1)]; |
| 79 | if (old) |
| 80 | block->next = old; |
| 81 | |
| 82 | cache->lut[(pc >> 2) & (LUT_SIZE - 1)] = block; |
| 83 | |
| 84 | remove_from_code_lut(cache, block); |
| 85 | } |
| 86 | |
| 87 | void lightrec_unregister_block(struct blockcache *cache, struct block *block) |
| 88 | { |
| 89 | u32 pc = kunseg(block->pc); |
| 90 | struct block *old = cache->lut[(pc >> 2) & (LUT_SIZE - 1)]; |
| 91 | |
| 92 | if (old == block) { |
| 93 | cache->lut[(pc >> 2) & (LUT_SIZE - 1)] = old->next; |
| 94 | return; |
| 95 | } |
| 96 | |
| 97 | for (; old; old = old->next) { |
| 98 | if (old->next == block) { |
| 99 | old->next = block->next; |
| 100 | return; |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | pr_err("Block at "PC_FMT" is not in cache\n", block->pc); |
| 105 | } |
| 106 | |
| 107 | static bool lightrec_block_is_old(const struct lightrec_state *state, |
| 108 | const struct block *block) |
| 109 | { |
| 110 | u32 diff = state->current_cycle - block->precompile_date; |
| 111 | |
| 112 | return diff > (1 << 27); /* About 4 seconds */ |
| 113 | } |
| 114 | |
| 115 | static void lightrec_free_blocks(struct blockcache *cache, |
| 116 | const struct block *except, bool all) |
| 117 | { |
| 118 | struct lightrec_state *state = cache->state; |
| 119 | struct block *block, *next; |
| 120 | bool outdated = all; |
| 121 | unsigned int i; |
| 122 | u8 old_flags; |
| 123 | |
| 124 | for (i = 0; i < LUT_SIZE; i++) { |
| 125 | for (block = cache->lut[i]; block; block = next) { |
| 126 | next = block->next; |
| 127 | |
| 128 | if (except && block == except) |
| 129 | continue; |
| 130 | |
| 131 | if (!all) { |
| 132 | outdated = lightrec_block_is_old(state, block) || |
| 133 | lightrec_block_is_outdated(state, block); |
| 134 | } |
| 135 | |
| 136 | if (!outdated) |
| 137 | continue; |
| 138 | |
| 139 | old_flags = block_set_flags(block, BLOCK_IS_DEAD); |
| 140 | |
| 141 | if (!(old_flags & BLOCK_IS_DEAD)) { |
| 142 | if (ENABLE_THREADED_COMPILER) |
| 143 | lightrec_recompiler_remove(state->rec, block); |
| 144 | |
| 145 | pr_debug("Freeing outdated block at "PC_FMT"\n", block->pc); |
| 146 | remove_from_code_lut(cache, block); |
| 147 | lightrec_unregister_block(cache, block); |
| 148 | lightrec_free_block(state, block); |
| 149 | } |
| 150 | } |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | void lightrec_remove_outdated_blocks(struct blockcache *cache, |
| 155 | const struct block *except) |
| 156 | { |
| 157 | pr_info("Running out of code space. Cleaning block cache...\n"); |
| 158 | |
| 159 | lightrec_free_blocks(cache, except, false); |
| 160 | } |
| 161 | |
| 162 | void lightrec_free_block_cache(struct blockcache *cache) |
| 163 | { |
| 164 | lightrec_free_blocks(cache, NULL, true); |
| 165 | lightrec_free(cache->state, MEM_FOR_LIGHTREC, sizeof(*cache), cache); |
| 166 | } |
| 167 | |
| 168 | struct blockcache * lightrec_blockcache_init(struct lightrec_state *state) |
| 169 | { |
| 170 | struct blockcache *cache; |
| 171 | |
| 172 | cache = lightrec_calloc(state, MEM_FOR_LIGHTREC, sizeof(*cache)); |
| 173 | if (!cache) |
| 174 | return NULL; |
| 175 | |
| 176 | cache->state = state; |
| 177 | |
| 178 | return cache; |
| 179 | } |
| 180 | |
| 181 | u32 lightrec_calculate_block_hash(const struct block *block) |
| 182 | { |
| 183 | const u32 *code = block->code; |
| 184 | u32 hash = 0xffffffff; |
| 185 | unsigned int i; |
| 186 | |
| 187 | /* Jenkins one-at-a-time hash algorithm */ |
| 188 | for (i = 0; i < block->nb_ops; i++) { |
| 189 | hash += *code++; |
| 190 | hash += (hash << 10); |
| 191 | hash ^= (hash >> 6); |
| 192 | } |
| 193 | |
| 194 | hash += (hash << 3); |
| 195 | hash ^= (hash >> 11); |
| 196 | hash += (hash << 15); |
| 197 | |
| 198 | return hash; |
| 199 | } |
| 200 | |
| 201 | static void lightrec_reset_lut_offset(struct lightrec_state *state, void *d) |
| 202 | { |
| 203 | u32 pc = (u32)(uintptr_t) d; |
| 204 | struct block *block; |
| 205 | void *addr; |
| 206 | |
| 207 | block = lightrec_find_block(state->block_cache, pc); |
| 208 | if (!block) |
| 209 | return; |
| 210 | |
| 211 | if (block_has_flag(block, BLOCK_IS_DEAD)) |
| 212 | return; |
| 213 | |
| 214 | addr = block->function ?: state->get_next_block; |
| 215 | lut_write(state, lut_offset(pc), addr); |
| 216 | } |
| 217 | |
| 218 | bool lightrec_block_is_outdated(struct lightrec_state *state, struct block *block) |
| 219 | { |
| 220 | u32 offset = lut_offset(block->pc); |
| 221 | bool outdated; |
| 222 | |
| 223 | if (lut_read(state, offset)) |
| 224 | return false; |
| 225 | |
| 226 | outdated = block->hash != lightrec_calculate_block_hash(block); |
| 227 | if (likely(!outdated)) { |
| 228 | /* The block was marked as outdated, but the content is still |
| 229 | * the same */ |
| 230 | |
| 231 | if (ENABLE_THREADED_COMPILER) { |
| 232 | /* |
| 233 | * When compiling a block that covers ours, the threaded |
| 234 | * compiler will set the LUT entries of the various |
| 235 | * entry points. Therefore we cannot write the LUT here, |
| 236 | * as we would risk overwriting the new entry points. |
| 237 | * Leave it to the reaper to re-install the LUT entries. |
| 238 | */ |
| 239 | |
| 240 | lightrec_reaper_add(state->reaper, |
| 241 | lightrec_reset_lut_offset, |
| 242 | (void *)(uintptr_t) block->pc); |
| 243 | } else if (block->function) { |
| 244 | lut_write(state, offset, block->function); |
| 245 | } else { |
| 246 | lut_write(state, offset, state->get_next_block); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | return outdated; |
| 251 | } |