Rev 4638 | Go to most recent revision | Show entire file | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed
Rev 4638 | Rev 4663 | ||
---|---|---|---|
Line 40... | Line 40... | ||
40 | #include <arch/cpu.h> |
40 | #include <arch/cpu.h> |
41 | #include <arch/boot/boot.h> |
41 | #include <arch/boot/boot.h> |
42 | #include <arch.h> |
42 | #include <arch.h> |
43 | #include <config.h> |
43 | #include <config.h> |
44 | #include <macros.h> |
44 | #include <macros.h> |
- | 45 | #include <func.h> |
|
45 | #include <arch/types.h> |
46 | #include <arch/types.h> |
46 | #include <synch/synch.h> |
47 | #include <synch/synch.h> |
47 | #include <synch/waitq.h> |
48 | #include <synch/waitq.h> |
48 | #include <print.h> |
49 | #include <print.h> |
49 | #include <arch/sun4v/hypercall.h> |
50 | #include <arch/sun4v/hypercall.h> |
50 | #include <arch/sun4v/md.h> |
51 | #include <arch/sun4v/md.h> |
51 | #include <arch/sun4v/ipi.h> |
52 | #include <arch/sun4v/ipi.h> |
52 | #include <time/delay.h> |
53 | #include <time/delay.h> |
53 | 54 | ||
- | 55 | /** hypervisor code of the "running" state of the CPU */ |
|
54 | #define CPU_STATE_RUNNING 2 |
56 | #define CPU_STATE_RUNNING 2 |
55 | 57 | ||
56 | extern void kernel_image_start(void); |
- | |
57 | extern void *trap_table; |
- | |
58 | - | ||
59 | /** Determine number of processors. */ |
58 | /** maximum possible number of processor cores */ |
60 | void smp_init(void) |
- | |
61 | { |
- | |
62 | md_node_t node = md_get_root(); |
59 | #define MAX_NUM_CORES 8 |
63 | count_t cpu_count = 0; |
- | |
64 | - | ||
65 | /* walk through MD, find the current CPU node & its clock-frequency */ |
- | |
66 | while(md_next_node(&node, "cpu")) { |
- | |
67 | cpu_count++; |
- | |
68 | } |
- | |
69 | - | ||
70 | config.cpu_count = cpu_count; |
- | |
71 | } |
- | |
72 | - | ||
73 | - | ||
74 | /** Wake application processors up. */ |
- | |
75 | void kmp(void *arg) |
- | |
76 | { |
- | |
77 | #if 1 |
- | |
78 | (void) arg; |
- | |
79 | 60 | ||
80 | uint64_t myid; |
61 | /** needed in the CPU_START hypercall */ |
81 | __hypercall_fast_ret1(0, 0, 0, 0, 0, CPU_MYID, &myid); |
62 | extern void kernel_image_start(void); |
82 | 63 | ||
83 | /* stop the CPUs before making them execute our code */ |
64 | /** needed in the CPU_START hypercall */ |
84 | uint64_t i; |
65 | extern void *trap_table; |
85 | for (i = 0; i < config.cpu_count; i++) { |
- | |
86 | if (i == myid) |
- | |
87 | continue; |
- | |
88 | 66 | ||
89 | if (__hypercall_fast1(CPU_STOP, i) != 0) |
67 | /** number of execution units detected */ |
90 | continue; |
68 | uint8_t exec_unit_count = 0; |
91 | 69 | ||
92 | uint64_t state; |
- | |
93 | __hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
- | |
94 | while (state == CPU_STATE_RUNNING) { |
70 | /** execution units (processor cores) */ |
95 | __hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
71 | exec_unit_t exec_units[MAX_NUM_CORES]; |
96 | } |
- | |
97 | } |
- | |
98 | 72 | ||
99 | /* wake the processors up, one by one */ |
73 | /** CPU structures */ |
100 | uint64_t state; |
74 | extern cpu_t *cpus; |
101 | for (i = 1; i < config.cpu_count; i++) { |
- | |
102 | __hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
- | |
103 | printf("Starting CPU %d, error code = %d.\n", i, __hypercall_fast4( |
- | |
104 | CPU_START, |
- | |
105 | i, |
- | |
106 | (uint64_t) KA2PA(kernel_image_start), |
- | |
107 | KA2PA(trap_table), |
- | |
108 | bootinfo.physmem_start |
- | |
109 | )); |
- | |
110 | - | ||
111 | if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) == |
- | |
112 | ESYNCH_TIMEOUT) |
- | |
113 | printf("%s: waiting for processor (cpuid = %" PRIu32 |
- | |
114 | ") timed out\n", __func__, i); |
- | |
115 | - | ||
116 | } |
- | |
117 | #else |
- | |
118 | 75 | ||
- | 76 | /** maximum number of strands per a physical core detected */ |
|
- | 77 | unsigned int max_core_strands = 0; |
|
- | 78 | ||
- | 79 | #ifdef CONFIG_SIMICS_SMP_HACK |
|
- | 80 | /** |
|
- | 81 | * Copies a piece of HelenOS code to the place where OBP had its IPI handler. |
|
- | 82 | * By sending an IPI by the BSP to the AP the code will be executed. |
|
- | 83 | * The code will jump to the first instruction of the kernel. This is |
|
- | 84 | * a workaround how to make APs execute HelenOS code on Simics. |
|
- | 85 | */ |
|
- | 86 | static void simics_smp_hack_init(void) { |
|
119 | asm volatile ( |
87 | asm volatile ( |
120 | "setx temp_cpu_mondo_handler, %g4, %g6 \n" |
88 | "setx temp_cpu_mondo_handler, %g4, %g6 \n" |
121 | //"setx 0x80246ad8, %g4, %g7 \n" |
- | |
122 | "setx 0x80200f80, %g4, %g7 \n" |
89 | "setx 0x80200f80, %g4, %g7 \n" |
123 | 90 | ||
124 | "ldx [%g6], %g4 \n" |
91 | "ldx [%g6], %g4 \n" |
125 | "stxa %g4, [%g7] 0x14 \n" |
92 | "stxa %g4, [%g7] 0x14 \n" |
126 | "membar #Sync \n" |
93 | "membar #Sync \n" |
Line 166... | Line 133... | ||
166 | "membar #Sync \n" |
133 | "membar #Sync \n" |
167 | 134 | ||
168 | "flush %i7" |
135 | "flush %i7" |
169 | 136 | ||
170 | ); |
137 | ); |
- | 138 | } |
|
- | 139 | #endif |
|
- | 140 | ||
- | 141 | /** |
|
- | 142 | * Finds out which execution units belong to particular CPUs. By execution unit |
|
- | 143 | * we mean the physical core the logical processor is backed by. Since each |
|
- | 144 | * Niagara physical core has just one integer execution unit and we will |
|
- | 145 | * ignore other execution units than the integer ones, we will use the terms |
|
- | 146 | * "integer execution unit", "execution unit" and "physical core" |
|
- | 147 | * interchangeably. |
|
- | 148 | * |
|
- | 149 | * The physical cores are detected by browsing the children of the CPU node |
|
- | 150 | * in the machine description and looking for a node representing an integer |
|
- | 151 | * execution unit. Once the integer execution unit of a particular CPU is |
|
- | 152 | * known, the ID of the CPU is added to the list of cpuids of the corresponding |
|
- | 153 | * execution unit structure (exec_unit_t). If an execution unit is encountered |
|
- | 154 | * for the first time, a new execution unit structure (exec_unit_t) must be |
|
- | 155 | * created first and added to the execution units array (exec_units). |
|
- | 156 | * |
|
- | 157 | * If the function fails to find an execution unit for a CPU (this may happen |
|
- | 158 | * on machines with older firmware or on Simics), it performs a fallback code |
|
- | 159 | * which pretends there exists just one execution unit and all CPUs belong to |
|
- | 160 | * it. |
|
- | 161 | * |
|
- | 162 | * Finally, the array of all execution units is reordered such that its element |
|
- | 163 | * which represents the physical core of the the bootstrap CPU is at index 0. |
|
- | 164 | * Moreover, the array of CPU IDs within the BSP's physical core structure is |
|
- | 165 | * reordered such that the element which represents the ID of the BSP is at |
|
- | 166 | * index 0. This is done because we would like the CPUs to be woken up |
|
- | 167 | * such that the 0-index CPU of the 0-index execution unit is |
|
- | 168 | * woken up first. And since the BSP is already woken up, we would like it to be |
|
- | 169 | * at 0-th position of the 0-th execution unit structure. |
|
- | 170 | * |
|
- | 171 | * Apart from that, the code also counts the total number of CPUs and stores |
|
- | 172 | * it to the global config.cpu_count variable. |
|
- | 173 | */ |
|
- | 174 | static void detect_execution_units(void) |
|
- | 175 | { |
|
- | 176 | /* ID of the bootstrap processor */ |
|
- | 177 | uint64_t myid; |
|
- | 178 | ||
- | 179 | /* total number of CPUs detected */ |
|
- | 180 | count_t cpu_count = 0; |
|
- | 181 | ||
- | 182 | /* will be set to 1 if detecting the physical cores fails */ |
|
- | 183 | bool exec_unit_assign_error = 0; |
|
- | 184 | ||
- | 185 | /* index of the bootstrap physical core in the array of cores */ |
|
- | 186 | unsigned int bsp_exec_unit_index = 0; |
|
- | 187 | ||
- | 188 | /* index of the BSP ID inside the array of bootstrap core's cpuids */ |
|
- | 189 | unsigned int bsp_core_strand_index = 0; |
|
- | 190 | ||
- | 191 | __hypercall_fast_ret1(0, 0, 0, 0, 0, CPU_MYID, &myid); |
|
- | 192 | md_node_t node = md_get_root(); |
|
- | 193 | ||
- | 194 | /* walk through all the CPU nodes in the MD*/ |
|
- | 195 | while (md_next_node(&node, "cpu")) { |
|
- | 196 | ||
- | 197 | uint64_t cpuid; |
|
- | 198 | md_get_integer_property(node, "id", &cpuid); |
|
- | 199 | cpu_count++; |
|
- | 200 | ||
- | 201 | /* |
|
- | 202 | * if failed in previous CPUs, don't try |
|
- | 203 | * to detect physical cores any more |
|
- | 204 | */ |
|
- | 205 | if (exec_unit_assign_error) |
|
- | 206 | continue; |
|
- | 207 | ||
- | 208 | /* detect exec. unit for the CPU represented by current node */ |
|
- | 209 | uint64_t exec_unit_id = 0; |
|
- | 210 | md_child_iter_t it = md_get_child_iterator(node); |
|
- | 211 | ||
- | 212 | while (md_next_child(&it)) { |
|
- | 213 | md_node_t child = md_get_child_node(it); |
|
- | 214 | const char *exec_unit_type; |
|
- | 215 | md_get_string_property(child, "type", &exec_unit_type); |
|
- | 216 | ||
- | 217 | /* each physical core has just 1 integer exec. unit */ |
|
- | 218 | if (strcmp(exec_unit_type, "integer") == 0) { |
|
- | 219 | exec_unit_id = child; |
|
- | 220 | break; |
|
- | 221 | } |
|
- | 222 | } |
|
- | 223 | ||
- | 224 | /* execution unit detected successfully */ |
|
- | 225 | if (exec_unit_id != 0) { |
|
- | 226 | ||
- | 227 | /* find the exec. unit in array of existing units */ |
|
- | 228 | unsigned int i = 0; |
|
- | 229 | for (i = 0; i < exec_unit_count; i++) { |
|
- | 230 | if (exec_units[i].exec_unit_id == exec_unit_id) |
|
- | 231 | break; |
|
- | 232 | } |
|
- | 233 | ||
- | 234 | /* |
|
- | 235 | * execution unit just met has not been met before, so |
|
- | 236 | * create a new entry in array of all execution units |
|
- | 237 | */ |
|
- | 238 | if (i == exec_unit_count) { |
|
- | 239 | exec_units[i].exec_unit_id = exec_unit_id; |
|
- | 240 | exec_units[i].strand_count = 0; |
|
- | 241 | exec_unit_count++; |
|
- | 242 | } |
|
- | 243 | ||
- | 244 | /* |
|
- | 245 | * remember the exec. unit and strand of the BSP |
|
- | 246 | */ |
|
- | 247 | if (cpuid == myid) { |
|
- | 248 | bsp_exec_unit_index = i; |
|
- | 249 | bsp_core_strand_index = exec_units[i].strand_count; |
|
- | 250 | } |
|
- | 251 | ||
- | 252 | /* add the CPU just met to the exec. unit's list */ |
|
- | 253 | exec_units[i].cpuids[exec_units[i].strand_count] = cpuid; |
|
- | 254 | exec_units[i].strand_count++; |
|
- | 255 | max_core_strands = |
|
- | 256 | exec_units[i].strand_count > max_core_strands ? |
|
- | 257 | exec_units[i].strand_count : max_core_strands; |
|
- | 258 | ||
- | 259 | /* detecting execution unit failed */ |
|
- | 260 | } else { |
|
- | 261 | exec_unit_assign_error = 1; |
|
- | 262 | } |
|
- | 263 | } |
|
- | 264 | ||
- | 265 | /* save the number of CPUs to a globally accessible variable */ |
|
- | 266 | config.cpu_count = cpu_count; |
|
- | 267 | ||
- | 268 | /* |
|
- | 269 | * A fallback code which will be executed if finding out which |
|
- | 270 | * execution units belong to particular CPUs fails. Pretend there |
|
- | 271 | * exists just one execution unit and all CPUs belong to it. |
|
- | 272 | */ |
|
- | 273 | if (exec_unit_assign_error) { |
|
- | 274 | bsp_exec_unit_index = 0; |
|
- | 275 | exec_unit_count = 1; |
|
- | 276 | exec_units[0].strand_count = cpu_count; |
|
- | 277 | exec_units[0].exec_unit_id = 1; |
|
- | 278 | max_core_strands = cpu_count; |
|
- | 279 | ||
- | 280 | /* browse CPUs again, assign them the fictional exec. unit */ |
|
- | 281 | node = md_get_root(); |
|
- | 282 | unsigned int i = 0; |
|
- | 283 | ||
- | 284 | while (md_next_node(&node, "cpu")) { |
|
- | 285 | uint64_t cpuid; |
|
- | 286 | md_get_integer_property(node, "id", &cpuid); |
|
- | 287 | if (cpuid == myid) { |
|
- | 288 | bsp_core_strand_index = i; |
|
- | 289 | } |
|
- | 290 | exec_units[0].cpuids[i++] = cpuid; |
|
- | 291 | } |
|
- | 292 | } |
|
- | 293 | ||
- | 294 | /* |
|
- | 295 | * Reorder the execution units array elements and the cpuid array |
|
- | 296 | * elements so that the BSP will always be the very first CPU of |
|
- | 297 | * the very first execution unit. |
|
- | 298 | */ |
|
- | 299 | exec_unit_t temp_exec_unit = exec_units[0]; |
|
- | 300 | exec_units[0] = exec_units[bsp_exec_unit_index]; |
|
- | 301 | exec_units[bsp_exec_unit_index] = temp_exec_unit; |
|
- | 302 | ||
- | 303 | uint64_t temp_cpuid = exec_units[0].cpuids[0]; |
|
- | 304 | exec_units[0].cpuids[0] = exec_units[0].cpuids[bsp_exec_unit_index]; |
|
- | 305 | exec_units[0].cpuids[bsp_core_strand_index] = temp_cpuid; |
|
- | 306 | ||
- | 307 | } |
|
- | 308 | ||
- | 309 | /** |
|
- | 310 | * Determine number of processors and detect physical cores. On Simics |
|
- | 311 | * copy the code which will be executed by the AP when the BSP sends an |
|
- | 312 | * IPI to it in order to make it execute HelenOS code. |
|
- | 313 | */ |
|
- | 314 | void smp_init(void) |
|
- | 315 | { |
|
- | 316 | detect_execution_units(); |
|
- | 317 | #ifdef CONFIG_SIMICS_SMP_HACK |
|
- | 318 | simics_smp_hack_init(); |
|
- | 319 | #endif |
|
- | 320 | } |
|
- | 321 | ||
- | 322 | /** |
|
- | 323 | * For each CPU sets the value of cpus[i].arch.id, where i is the |
|
- | 324 | * index of the CPU in the cpus variable, to the cpuid of the i-th processor |
|
- | 325 | * to be run. The CPUs are run such that the CPU represented by cpus[0] |
|
- | 326 | * is run first, cpus[1] is run after it, and cpus[cpu_count - 1] is run as the |
|
- | 327 | * last one. |
|
- | 328 | * |
|
- | 329 | * The CPU IDs are set such that during waking the CPUs up the |
|
- | 330 | * processor cores will be alternated, i.e. first one CPU from the first core |
|
- | 331 | * will be run, after that one CPU from the second CPU core will be run,... |
|
- | 332 | * then one CPU from the last core will be run, after that another CPU |
|
- | 333 | * from the first core will be run, then another CPU from the second core |
|
- | 334 | * will be run,... then another CPU from the last core will be run, and so on. |
|
- | 335 | */ |
|
- | 336 | static void init_cpuids(void) |
|
- | 337 | { |
|
- | 338 | unsigned int cur_core_strand; |
|
- | 339 | unsigned int cur_core; |
|
- | 340 | unsigned int cur_cpu = 0; |
|
- | 341 | ||
- | 342 | for (cur_core_strand = 0; cur_core_strand < max_core_strands; cur_core_strand++) { |
|
- | 343 | for (cur_core = 0; cur_core < exec_unit_count; cur_core++) { |
|
- | 344 | if (cur_core_strand > exec_units[cur_core].strand_count) |
|
171 | delay(1000); |
345 | continue; |
- | 346 | ||
- | 347 | cpus[cur_cpu++].arch.id = exec_units[cur_core].cpuids[cur_core_strand]; |
|
- | 348 | } |
|
- | 349 | } |
|
- | 350 | } |
|
- | 351 | ||
- | 352 | /** |
|
- | 353 | * Wakes up a single CPU. |
|
- | 354 | * |
|
- | 355 | * @param cpuid ID of the CPU to be woken up |
|
- | 356 | */ |
|
- | 357 | static bool wake_cpu(uint64_t cpuid) |
|
- | 358 | { |
|
- | 359 | ||
- | 360 | #ifdef CONFIG_SIMICS_SMP_HACK |
|
172 | printf("Result: %d\n", ipi_unicast_to((void (*)(void)) 1234, 1)); |
361 | ipi_unicast_to((void (*)(void)) 1234, cpuid); |
- | 362 | #else |
|
- | 363 | /* stop the CPU before making it execute our code */ |
|
173 | if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) == |
364 | if (__hypercall_fast1(CPU_STOP, cpuid) != EOK) |
- | 365 | return false; |
|
- | 366 | ||
- | 367 | /* wait for the CPU to stop */ |
|
- | 368 | uint64_t state; |
|
- | 369 | __hypercall_fast_ret1(cpuid, 0, 0, 0, 0, |
|
174 | ESYNCH_TIMEOUT) |
370 | CPU_STATE, &state); |
- | 371 | while (state == CPU_STATE_RUNNING) { |
|
175 | printf("%s: waiting for processor (cpuid = %" PRIu32 |
372 | __hypercall_fast_ret1(cpuid, 0, 0, 0, 0, |
- | 373 | CPU_STATE, &state); |
|
- | 374 | } |
|
- | 375 | ||
- | 376 | /* make the CPU run again and execute HelenOS code */ |
|
- | 377 | if (__hypercall_fast4( |
|
- | 378 | CPU_START, cpuid, |
|
176 | ") timed out\n", __func__, 1); |
379 | (uint64_t) KA2PA(kernel_image_start), |
- | 380 | KA2PA(trap_table), bootinfo.physmem_start |
|
- | 381 | ) != EOK) |
|
- | 382 | return false; |
|
177 | #endif |
383 | #endif |
- | 384 | ||
- | 385 | if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) == |
|
- | 386 | ESYNCH_TIMEOUT) |
|
- | 387 | printf("%s: waiting for processor (cpuid = %" PRIu32 |
|
- | 388 | ") timed out\n", __func__, cpuid); |
|
- | 389 | ||
- | 390 | return true; |
|
- | 391 | } |
|
- | 392 | ||
- | 393 | /** Wake application processors up. */ |
|
- | 394 | void kmp(void *arg) |
|
- | 395 | { |
|
- | 396 | init_cpuids(); |
|
- | 397 | ||
- | 398 | unsigned int i; |
|
- | 399 | ||
- | 400 | for (i = 1; i < config.cpu_count; i++) { |
|
- | 401 | wake_cpu(cpus[i].arch.id); |
|
- | 402 | } |
|
178 | } |
403 | } |
179 | 404 | ||
180 | /** @} |
405 | /** @} |
181 | */ |
406 | */ |