| 1 | /* Target-dependent code for ROCm. |
| 2 | |
| 3 | Copyright (C) 2019-2020 Free Software Foundation, Inc. |
| 4 | Copyright (C) 2019-2020 Advanced Micro Devices, Inc. All rights reserved. |
| 5 | |
| 6 | This file is part of GDB. |
| 7 | |
| 8 | This program is free software; you can redistribute it and/or modify |
| 9 | it under the terms of the GNU General Public License as published by |
| 10 | the Free Software Foundation; either version 3 of the License, or |
| 11 | (at your option) any later version. |
| 12 | |
| 13 | This program is distributed in the hope that it will be useful, |
| 14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | GNU General Public License for more details. |
| 17 | |
| 18 | You should have received a copy of the GNU General Public License |
| 19 | along with this program. If not, see <http://www.gnu.org/licenses/>. */ |
| 20 | |
| 21 | #include "defs.h" |
| 22 | |
| 23 | #include "amdgcn-rocm-tdep.h" |
| 24 | #include "arch-utils.h" |
| 25 | #include "cli/cli-style.h" |
| 26 | #include "environ.h" |
| 27 | #include "event-loop.h" |
| 28 | #include "filenames.h" |
| 29 | #include "gdbcmd.h" |
| 30 | #include "gdbcore.h" |
| 31 | #include "gdbsupport/filestuff.h" |
| 32 | #include "gdbsupport/scoped_fd.h" |
| 33 | #include "gdbthread.h" |
| 34 | #include "hashtab.h" |
| 35 | #include "inf-loop.h" |
| 36 | #include "inferior.h" |
| 37 | #include "location.h" |
| 38 | #include "objfiles.h" |
| 39 | #include "observable.h" |
| 40 | #include "progspace-and-thread.h" |
| 41 | #include "regcache.h" |
| 42 | #include "rocm-tdep.h" |
| 43 | #include "solib.h" |
| 44 | #include "solist.h" |
| 45 | #include "symfile.h" |
| 46 | |
| 47 | #include <dlfcn.h> |
| 48 | #include <list> |
| 49 | #include <set> |
| 50 | #include <signal.h> |
| 51 | #include <stdarg.h> |
| 52 | #include <unordered_map> |
| 53 | |
| 54 | #include <amd-dbgapi.h> |
| 55 | |
| 56 | /* Big enough to hold the size of the largest register in bytes. */ |
| 57 | #define AMDGCN_MAX_REGISTER_SIZE 256 |
| 58 | |
| 59 | #define DEFINE_OBSERVABLE(name) decltype (name) name (#name) |
| 60 | |
| 61 | DEFINE_OBSERVABLE (amd_dbgapi_activated); |
| 62 | DEFINE_OBSERVABLE (amd_dbgapi_deactivated); |
| 63 | DEFINE_OBSERVABLE (amd_dbgapi_code_object_list_updated); |
| 64 | |
| 65 | #undef DEFINE_OBSERVABLE |
| 66 | |
| 67 | struct rocm_notify_shared_library_info |
| 68 | { |
| 69 | std::string compare; /* Compare loaded library names with this string. */ |
| 70 | struct so_list *solib; |
| 71 | bool is_loaded; |
| 72 | }; |
| 73 | |
| 74 | /* ROCm-specific inferior data. */ |
| 75 | |
| 76 | struct rocm_inferior_info |
| 77 | { |
| 78 | /* The amd_dbgapi_process_id for this inferior. */ |
| 79 | amd_dbgapi_process_id_t process_id{ AMD_DBGAPI_PROCESS_NONE }; |
| 80 | |
| 81 | /* The amd_dbgapi_notifier_t for this inferior. */ |
| 82 | amd_dbgapi_notifier_t notifier{ -1 }; |
| 83 | |
| 84 | /* True if commit_resume should all-start the GPU queues. */ |
| 85 | bool commit_resume_all_start; |
| 86 | |
| 87 | std::unordered_map<decltype (amd_dbgapi_breakpoint_id_t::handle), |
| 88 | struct breakpoint *> |
| 89 | breakpoint_map; |
| 90 | |
| 91 | /* List of pending events the rocm target retrieved from the dbgapi. */ |
| 92 | std::list<std::pair<amd_dbgapi_wave_id_t, amd_dbgapi_wave_stop_reason_t>> |
| 93 | wave_stop_events; |
| 94 | |
| 95 | /* Map of rocm_notify_shared_library_info's for libraries that have been |
| 96 | registered to receive notifications when loading/unloading. */ |
| 97 | std::unordered_map<decltype (amd_dbgapi_shared_library_id_t::handle), |
| 98 | struct rocm_notify_shared_library_info> |
| 99 | notify_solib_map; |
| 100 | }; |
| 101 | |
| 102 | static amd_dbgapi_event_id_t |
| 103 | rocm_process_event_queue (amd_dbgapi_event_kind_t until_event_kind |
| 104 | = AMD_DBGAPI_EVENT_KIND_NONE); |
| 105 | |
| 106 | /* Return the inferior's rocm_inferior_info struct. */ |
| 107 | static struct rocm_inferior_info * |
| 108 | get_rocm_inferior_info (struct inferior *inferior = nullptr); |
| 109 | |
| 110 | static const target_info rocm_ops_info |
| 111 | = { "rocm", N_ ("ROCm GPU debugging support"), |
| 112 | N_ ("ROCm GPU debugging support") }; |
| 113 | |
| 114 | static amd_dbgapi_log_level_t get_debug_amd_dbgapi_log_level (); |
| 115 | |
| 116 | struct rocm_target_ops final : public target_ops |
| 117 | { |
| 118 | const target_info & |
| 119 | info () const override |
| 120 | { |
| 121 | return rocm_ops_info; |
| 122 | } |
| 123 | strata |
| 124 | stratum () const override |
| 125 | { |
| 126 | return arch_stratum; |
| 127 | } |
| 128 | |
| 129 | void mourn_inferior () override; |
| 130 | |
| 131 | void async (int enable) override; |
| 132 | |
| 133 | ptid_t wait (ptid_t, struct target_waitstatus *, int) override; |
| 134 | void resume (ptid_t, int, enum gdb_signal) override; |
| 135 | void commit_resume () override; |
| 136 | void stop (ptid_t ptid) override; |
| 137 | |
| 138 | void fetch_registers (struct regcache *, int) override; |
| 139 | void store_registers (struct regcache *, int) override; |
| 140 | |
| 141 | void update_thread_list () override; |
| 142 | |
| 143 | struct gdbarch *thread_architecture (ptid_t) override; |
| 144 | |
| 145 | std::string pid_to_str (ptid_t ptid) override; |
| 146 | |
| 147 | const char *thread_name (thread_info *tp) override; |
| 148 | |
| 149 | const char *extra_thread_info (thread_info *tp) override; |
| 150 | |
| 151 | bool thread_alive (ptid_t ptid) override; |
| 152 | |
| 153 | enum target_xfer_status xfer_partial (enum target_object object, |
| 154 | const char *annex, gdb_byte *readbuf, |
| 155 | const gdb_byte *writebuf, |
| 156 | ULONGEST offset, ULONGEST len, |
| 157 | ULONGEST *xfered_len) override; |
| 158 | |
| 159 | bool |
| 160 | stopped_by_watchpoint () override |
| 161 | { |
| 162 | return !ptid_is_gpu (inferior_ptid) |
| 163 | && beneath ()->stopped_by_watchpoint (); |
| 164 | } |
| 165 | |
| 166 | bool |
| 167 | stopped_data_address (CORE_ADDR *addr_p) override |
| 168 | { |
| 169 | return !ptid_is_gpu (inferior_ptid) |
| 170 | && beneath ()->stopped_data_address (addr_p); |
| 171 | } |
| 172 | |
| 173 | bool |
| 174 | supports_stopped_by_sw_breakpoint () override |
| 175 | { |
| 176 | return true; |
| 177 | } |
| 178 | |
| 179 | bool stopped_by_sw_breakpoint () override; |
| 180 | |
| 181 | bool |
| 182 | stopped_by_hw_breakpoint () override |
| 183 | { |
| 184 | return !ptid_is_gpu (inferior_ptid) |
| 185 | && beneath ()->stopped_by_hw_breakpoint (); |
| 186 | } |
| 187 | }; |
| 188 | |
| 189 | /* ROCm's target vector. */ |
| 190 | static struct rocm_target_ops rocm_ops; |
| 191 | |
| 192 | /* ROCm breakpoint ops. */ |
| 193 | static struct breakpoint_ops rocm_breakpoint_ops; |
| 194 | |
| 195 | /* Per-inferior data key. */ |
| 196 | static const struct inferior_key<rocm_inferior_info> rocm_inferior_data; |
| 197 | |
| 198 | /* The read/write ends of the pipe registered as waitable file in the |
| 199 | event loop. */ |
| 200 | static int rocm_event_pipe[2] = { -1, -1 }; |
| 201 | |
| 202 | /* Flush the event pipe. */ |
| 203 | |
| 204 | static void |
| 205 | async_file_flush (void) |
| 206 | { |
| 207 | int ret; |
| 208 | char buf; |
| 209 | |
| 210 | do |
| 211 | { |
| 212 | ret = read (rocm_event_pipe[0], &buf, 1); |
| 213 | } |
| 214 | while (ret >= 0 || (ret == -1 && errno == EINTR)); |
| 215 | } |
| 216 | |
| 217 | /* Put something (anything, doesn't matter what, or how much) in event |
| 218 | pipe, so that the select/poll in the event-loop realizes we have |
| 219 | something to process. */ |
| 220 | |
| 221 | static void |
| 222 | async_file_mark (void) |
| 223 | { |
| 224 | int ret; |
| 225 | |
| 226 | /* It doesn't really matter what the pipe contains, as long we end |
| 227 | up with something in it. Might as well flush the previous |
| 228 | left-overs. */ |
| 229 | async_file_flush (); |
| 230 | |
| 231 | do |
| 232 | { |
| 233 | ret = write (rocm_event_pipe[1], "+", 1); |
| 234 | } |
| 235 | while (ret == -1 && errno == EINTR); |
| 236 | |
| 237 | /* Ignore EAGAIN. If the pipe is full, the event loop will already |
| 238 | be awakened anyway. */ |
| 239 | } |
| 240 | |
| 241 | /* Fetch the rocm_inferior_info data for the given inferior. */ |
| 242 | |
| 243 | static struct rocm_inferior_info * |
| 244 | get_rocm_inferior_info (struct inferior *inferior) |
| 245 | { |
| 246 | if (!inferior) |
| 247 | inferior = current_inferior (); |
| 248 | |
| 249 | struct rocm_inferior_info *info = rocm_inferior_data.get (inferior); |
| 250 | |
| 251 | if (!info) |
| 252 | info = rocm_inferior_data.emplace (inferior); |
| 253 | |
| 254 | return info; |
| 255 | } |
| 256 | |
| 257 | /* Fetch the amd_dbgapi_process_id for the given inferior. */ |
| 258 | |
| 259 | amd_dbgapi_process_id_t |
| 260 | get_amd_dbgapi_process_id (struct inferior *inferior) |
| 261 | { |
| 262 | return get_rocm_inferior_info (inferior)->process_id; |
| 263 | } |
| 264 | |
| 265 | static void |
| 266 | rocm_breakpoint_check_status (struct bpstats *bs) |
| 267 | { |
| 268 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 269 | amd_dbgapi_process_id_t process_id = info->process_id; |
| 270 | amd_dbgapi_status_t status; |
| 271 | |
| 272 | bs->stop = 0; |
| 273 | bs->print_it = print_it_noop; |
| 274 | |
| 275 | /* Find the address the breakpoint is set at. */ |
| 276 | auto it = std::find_if ( |
| 277 | info->breakpoint_map.begin (), info->breakpoint_map.end (), |
| 278 | [=] (const decltype (info->breakpoint_map)::value_type &value) { |
| 279 | return value.second == bs->breakpoint_at; |
| 280 | }); |
| 281 | |
| 282 | if (it == info->breakpoint_map.end ()) |
| 283 | error (_ ("Could not find breakpoint_id for breakpoint at %#lx"), |
| 284 | bs->bp_location_at->address); |
| 285 | |
| 286 | amd_dbgapi_breakpoint_id_t breakpoint_id{ it->first }; |
| 287 | amd_dbgapi_breakpoint_action_t action; |
| 288 | |
| 289 | status = amd_dbgapi_report_breakpoint_hit (process_id, breakpoint_id, |
| 290 | inferior_thread (), &action); |
| 291 | |
| 292 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 293 | error (_ ("amd_dbgapi_report_breakpoint_hit failed: breakpoint_%ld " |
| 294 | "at %#lx (rc=%d)"), |
| 295 | breakpoint_id.handle, bs->bp_location_at->address, status); |
| 296 | |
| 297 | if (action == AMD_DBGAPI_BREAKPOINT_ACTION_RESUME) |
| 298 | return; |
| 299 | |
| 300 | /* If the action is AMD_DBGAPI_BREAKPOINT_ACTION_HALT, we need to wait until |
| 301 | a breakpoint resume event for this breakpoint_id is seen. */ |
| 302 | |
| 303 | amd_dbgapi_event_id_t resume_event_id |
| 304 | = rocm_process_event_queue (AMD_DBGAPI_EVENT_KIND_BREAKPOINT_RESUME); |
| 305 | |
| 306 | /* We should always get a breakpoint_resume event after processing all |
| 307 | events generated by reporting the breakpoint was hit. */ |
| 308 | gdb_assert (resume_event_id.handle != AMD_DBGAPI_EVENT_NONE.handle); |
| 309 | |
| 310 | amd_dbgapi_breakpoint_id_t resume_breakpoint_id; |
| 311 | status = amd_dbgapi_event_get_info ( |
| 312 | process_id, resume_event_id, AMD_DBGAPI_EVENT_INFO_BREAKPOINT, |
| 313 | sizeof (resume_breakpoint_id), &resume_breakpoint_id); |
| 314 | |
| 315 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 316 | error (_ ("amd_dbgapi_event_get_info failed (rc=%d)"), status); |
| 317 | |
| 318 | /* The debugger API guarantees that [breakpoint_hit...resume_breakpoint] |
| 319 | sequences cannot interleave, so this breakpoint resume event must be |
| 320 | for our breakpoint_id. */ |
| 321 | if (resume_breakpoint_id.handle != breakpoint_id.handle) |
| 322 | error (_ ("breakpoint resume event is not for this breakpoint. " |
| 323 | "Expected breakpoint_%ld, got breakpoint_%ld"), |
| 324 | breakpoint_id.handle, resume_breakpoint_id.handle); |
| 325 | |
| 326 | amd_dbgapi_event_processed (process_id, resume_event_id); |
| 327 | } |
| 328 | |
| 329 | static void |
| 330 | rocm_target_dbgapi_activated () |
| 331 | { |
| 332 | /* FIXME: only push on the first activation. */ |
| 333 | /* Engage the ROCm target_ops and so_ops. */ |
| 334 | push_target (&rocm_ops); |
| 335 | } |
| 336 | |
| 337 | static void |
| 338 | rocm_target_dbgapi_deactivated () |
| 339 | { |
| 340 | /* FIXME: only unpush on the last activation. */ |
| 341 | /* Disengage the ROCm target_ops. */ |
| 342 | unpush_target (&rocm_ops); |
| 343 | } |
| 344 | |
| 345 | bool |
| 346 | rocm_target_ops::thread_alive (ptid_t ptid) |
| 347 | { |
| 348 | if (!ptid_is_gpu (ptid)) |
| 349 | return beneath ()->thread_alive (ptid); |
| 350 | |
| 351 | /* Check that the wave_id is valid. */ |
| 352 | |
| 353 | amd_dbgapi_wave_state_t state; |
| 354 | return amd_dbgapi_wave_get_info ( |
| 355 | get_amd_dbgapi_process_id (), get_amd_dbgapi_wave_id (ptid), |
| 356 | AMD_DBGAPI_WAVE_INFO_STATE, sizeof (state), &state) |
| 357 | == AMD_DBGAPI_STATUS_SUCCESS; |
| 358 | } |
| 359 | |
| 360 | const char * |
| 361 | rocm_target_ops::thread_name (thread_info *tp) |
| 362 | { |
| 363 | if (!ptid_is_gpu (tp->ptid)) |
| 364 | return beneath ()->thread_name (tp); |
| 365 | |
| 366 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 367 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (tp->ptid); |
| 368 | amd_dbgapi_dispatch_id_t dispatch_id; |
| 369 | amd_dbgapi_global_address_t kernel_addr; |
| 370 | |
| 371 | if (amd_dbgapi_wave_get_info (process_id, wave_id, |
| 372 | AMD_DBGAPI_WAVE_INFO_DISPATCH, |
| 373 | sizeof (dispatch_id), &dispatch_id) |
| 374 | != AMD_DBGAPI_STATUS_SUCCESS |
| 375 | || amd_dbgapi_dispatch_get_info ( |
| 376 | process_id, dispatch_id, |
| 377 | AMD_DBGAPI_DISPATCH_INFO_KERNEL_ENTRY_ADDRESS, |
| 378 | sizeof (kernel_addr), &kernel_addr) |
| 379 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 380 | return NULL; |
| 381 | |
| 382 | struct bound_minimal_symbol msymbol |
| 383 | = lookup_minimal_symbol_by_pc_section (kernel_addr, nullptr); |
| 384 | |
| 385 | if (msymbol.minsym != NULL) |
| 386 | { |
| 387 | static char buf[256]; |
| 388 | char *endp; |
| 389 | |
| 390 | xsnprintf (buf, sizeof (buf), "%s", msymbol.minsym->print_name ()); |
| 391 | |
| 392 | /* Strip the arguments from the demangled function name. */ |
| 393 | if ((endp = strchr (buf, '('))) |
| 394 | *endp = '\0'; |
| 395 | |
| 396 | return buf; |
| 397 | } |
| 398 | |
| 399 | return NULL; |
| 400 | } |
| 401 | |
| 402 | std::string |
| 403 | rocm_target_ops::pid_to_str (ptid_t ptid) |
| 404 | { |
| 405 | if (!ptid_is_gpu (ptid)) |
| 406 | { |
| 407 | return beneath ()->pid_to_str (ptid); |
| 408 | } |
| 409 | |
| 410 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 411 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (ptid); |
| 412 | amd_dbgapi_dispatch_id_t dispatch_id; |
| 413 | uint32_t group_ids[3], wave_in_group; |
| 414 | |
| 415 | if (amd_dbgapi_wave_get_info (process_id, wave_id, |
| 416 | AMD_DBGAPI_WAVE_INFO_DISPATCH, |
| 417 | sizeof (dispatch_id), &dispatch_id) |
| 418 | != AMD_DBGAPI_STATUS_SUCCESS |
| 419 | || amd_dbgapi_wave_get_info (process_id, wave_id, |
| 420 | AMD_DBGAPI_WAVE_INFO_WORK_GROUP_COORD, |
| 421 | sizeof (group_ids), &group_ids) |
| 422 | != AMD_DBGAPI_STATUS_SUCCESS |
| 423 | || amd_dbgapi_wave_get_info ( |
| 424 | process_id, wave_id, |
| 425 | AMD_DBGAPI_WAVE_INFO_WAVE_NUMBER_IN_WORK_GROUP, |
| 426 | sizeof (wave_in_group), &wave_in_group) |
| 427 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 428 | return std::string ("AMDGPU Thread"); |
| 429 | else |
| 430 | return string_printf ("AMDGPU Thread %ld.%ld (%d,%d,%d)/%d", |
| 431 | dispatch_id.handle, wave_id.handle, group_ids[2], |
| 432 | group_ids[1], group_ids[0], wave_in_group); |
| 433 | } |
| 434 | |
| 435 | const char * |
| 436 | rocm_target_ops::extra_thread_info (thread_info *tp) |
| 437 | { |
| 438 | if (!ptid_is_gpu (tp->ptid)) |
| 439 | beneath ()->extra_thread_info (tp); |
| 440 | |
| 441 | return NULL; |
| 442 | } |
| 443 | |
| 444 | enum target_xfer_status |
| 445 | rocm_target_ops::xfer_partial (enum target_object object, const char *annex, |
| 446 | gdb_byte *readbuf, const gdb_byte *writebuf, |
| 447 | ULONGEST offset, ULONGEST requested_len, |
| 448 | ULONGEST *xfered_len) |
| 449 | { |
| 450 | gdb::optional<scoped_restore_current_thread> maybe_restore_thread; |
| 451 | |
| 452 | if (ptid_is_gpu (inferior_ptid)) |
| 453 | { |
| 454 | gdb_assert (requested_len && xfered_len && "checking invariants"); |
| 455 | |
| 456 | if (object != TARGET_OBJECT_MEMORY) |
| 457 | return TARGET_XFER_E_IO; |
| 458 | |
| 459 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 460 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (inferior_ptid); |
| 461 | |
| 462 | amd_dbgapi_architecture_id_t architecture_id; |
| 463 | amd_dbgapi_address_space_id_t address_space_id; |
| 464 | |
| 465 | if (amd_dbgapi_wave_get_info (process_id, wave_id, |
| 466 | AMD_DBGAPI_WAVE_INFO_ARCHITECTURE, |
| 467 | sizeof (architecture_id), &architecture_id) |
| 468 | != AMD_DBGAPI_STATUS_SUCCESS |
| 469 | || amd_dbgapi_architecture_get_info ( |
| 470 | architecture_id, |
| 471 | AMD_DBGAPI_ARCHITECTURE_INFO_DEFAULT_GLOBAL_ADDRESS_SPACE, |
| 472 | sizeof (address_space_id), &address_space_id) |
| 473 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 474 | error (_ ("Couldn't get default global address space")); |
| 475 | |
| 476 | size_t len = requested_len; |
| 477 | amd_dbgapi_status_t status; |
| 478 | |
| 479 | if (readbuf) |
| 480 | status = amd_dbgapi_read_memory ( |
| 481 | process_id, wave_id, 0, address_space_id, offset, &len, readbuf); |
| 482 | else |
| 483 | status = amd_dbgapi_write_memory ( |
| 484 | process_id, wave_id, 0, address_space_id, offset, &len, writebuf); |
| 485 | |
| 486 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 487 | { |
| 488 | if (status != AMD_DBGAPI_STATUS_ERROR_MEMORY_ACCESS) |
| 489 | error (_ ("amd_dbgapi_%s_memory failed (rc=%d"), |
| 490 | readbuf ? "read" : "write", status); |
| 491 | return TARGET_XFER_EOF; |
| 492 | } |
| 493 | |
| 494 | *xfered_len = len; |
| 495 | return TARGET_XFER_OK; |
| 496 | } |
| 497 | else |
| 498 | return beneath ()->xfer_partial (object, annex, readbuf, writebuf, offset, |
| 499 | requested_len, xfered_len); |
| 500 | } |
| 501 | |
| 502 | void |
| 503 | rocm_target_ops::resume (ptid_t ptid, int step, enum gdb_signal signo) |
| 504 | { |
| 505 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 506 | |
| 507 | if (debug_infrun) |
| 508 | fprintf_unfiltered ( |
| 509 | gdb_stdlog, |
| 510 | "\e[1;34minfrun: rocm_target_ops::resume ([%d,%ld,%ld])\e[0m\n", |
| 511 | ptid.pid (), ptid.lwp (), ptid.tid ()); |
| 512 | |
| 513 | /* Check if the thread focus is on the GPU device. */ |
| 514 | if (ptid == minus_one_ptid || !ptid_is_gpu (ptid)) |
| 515 | { |
| 516 | beneath ()->resume (ptid, step, signo); |
| 517 | if (ptid != minus_one_ptid) |
| 518 | return; |
| 519 | } |
| 520 | |
| 521 | /* A specific PTID means `step only this process id'. */ |
| 522 | bool resume_one = ptid != minus_one_ptid && !ptid.is_pid (); |
| 523 | gdb_assert (resume_one || !step); |
| 524 | |
| 525 | if (!resume_one) |
| 526 | error (_ ("internal error - unimplemented ")); |
| 527 | |
| 528 | amd_dbgapi_process_set_progress (info->process_id, |
| 529 | AMD_DBGAPI_PROGRESS_NO_FORWARD); |
| 530 | |
| 531 | amd_dbgapi_status_t status = amd_dbgapi_wave_resume ( |
| 532 | info->process_id, get_amd_dbgapi_wave_id (ptid), |
| 533 | step ? AMD_DBGAPI_RESUME_MODE_SINGLE_STEP |
| 534 | : AMD_DBGAPI_RESUME_MODE_NORMAL); |
| 535 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 536 | warning (_ ("Could not resume %s (rc=%d)"), |
| 537 | target_pid_to_str (ptid).c_str (), status); |
| 538 | |
| 539 | info->commit_resume_all_start = true; |
| 540 | } |
| 541 | |
| 542 | void |
| 543 | rocm_target_ops::commit_resume () |
| 544 | { |
| 545 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 546 | |
| 547 | if (debug_infrun) |
| 548 | fprintf_unfiltered ( |
| 549 | gdb_stdlog, |
| 550 | "\e[1;34minfrun: rocm_target_ops::commit_resume ()\e[0m\n"); |
| 551 | |
| 552 | beneath ()->commit_resume (); |
| 553 | |
| 554 | if (info->commit_resume_all_start) |
| 555 | { |
| 556 | amd_dbgapi_process_set_progress (info->process_id, |
| 557 | AMD_DBGAPI_PROGRESS_NORMAL); |
| 558 | info->commit_resume_all_start = false; |
| 559 | } |
| 560 | |
| 561 | if (target_can_async_p ()) |
| 562 | target_async (1); |
| 563 | } |
| 564 | |
| 565 | static void |
| 566 | rocm_target_stop_one_wave (ptid_t ptid) |
| 567 | { |
| 568 | amd_dbgapi_status_t status; |
| 569 | |
| 570 | status = amd_dbgapi_wave_stop (get_amd_dbgapi_process_id (), |
| 571 | get_amd_dbgapi_wave_id (ptid)); |
| 572 | |
| 573 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID) |
| 574 | { |
| 575 | /* the wave must have exited, set the thread status to reflect that. */ |
| 576 | auto *tp = find_thread_ptid (current_inferior (), ptid); |
| 577 | gdb_assert (tp); |
| 578 | |
| 579 | tp->state = THREAD_EXITED; |
| 580 | } |
| 581 | else if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 582 | warning (_ ("Could not stop %s (rc=%d)"), |
| 583 | target_pid_to_str (ptid).c_str (), status); |
| 584 | } |
| 585 | |
| 586 | void |
| 587 | rocm_target_ops::stop (ptid_t ptid) |
| 588 | { |
| 589 | if (debug_infrun) |
| 590 | fprintf_unfiltered ( |
| 591 | gdb_stdlog, |
| 592 | "\e[1;34minfrun: rocm_target_ops::stop ([%d,%ld,%ld])\e[0m\n", |
| 593 | ptid.pid (), ptid.lwp (), ptid.tid ()); |
| 594 | |
| 595 | if (ptid == minus_one_ptid || !ptid_is_gpu (ptid)) |
| 596 | { |
| 597 | beneath ()->stop (ptid); |
| 598 | if (ptid != minus_one_ptid) |
| 599 | return; |
| 600 | } |
| 601 | |
| 602 | if (ptid == minus_one_ptid) |
| 603 | error (_ ("internal error - unimplemented ")); |
| 604 | |
| 605 | rocm_target_stop_one_wave (ptid); |
| 606 | } |
| 607 | |
| 608 | static void |
| 609 | handle_target_event (int error, gdb_client_data client_data) |
| 610 | { |
| 611 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 612 | amd_dbgapi_process_id_t process_id = info->process_id; |
| 613 | |
| 614 | amd_dbgapi_process_set_progress (process_id, AMD_DBGAPI_PROGRESS_NO_FORWARD); |
| 615 | |
| 616 | /* Flush the async file first. */ |
| 617 | if (target_is_async_p ()) |
| 618 | async_file_flush (); |
| 619 | |
| 620 | rocm_process_event_queue (); |
| 621 | |
| 622 | /* In all-stop mode, unless the event queue is empty (spurious wake-up), |
| 623 | we can keep the process in progress_no_forward mode. The infrun loop |
| 624 | will enable forward progress when a thread is resumed. */ |
| 625 | if (non_stop || info->wave_stop_events.empty ()) |
| 626 | amd_dbgapi_process_set_progress (process_id, AMD_DBGAPI_PROGRESS_NORMAL); |
| 627 | |
| 628 | if (!info->wave_stop_events.empty ()) |
| 629 | inferior_event_handler (INF_REG_EVENT, nullptr); |
| 630 | } |
| 631 | |
| 632 | void |
| 633 | rocm_target_ops::async (int enable) |
| 634 | { |
| 635 | beneath ()->async (enable); |
| 636 | |
| 637 | if (enable) |
| 638 | { |
| 639 | if (rocm_event_pipe[0] != -1) |
| 640 | return; |
| 641 | |
| 642 | if (gdb_pipe_cloexec (rocm_event_pipe) == -1) |
| 643 | internal_error (__FILE__, __LINE__, "creating event pipe failed."); |
| 644 | |
| 645 | ::fcntl (rocm_event_pipe[0], F_SETFL, O_NONBLOCK); |
| 646 | ::fcntl (rocm_event_pipe[1], F_SETFL, O_NONBLOCK); |
| 647 | |
| 648 | add_file_handler (rocm_event_pipe[0], handle_target_event, nullptr); |
| 649 | |
| 650 | /* There may be pending events to handle. Tell the event loop |
| 651 | to poll them. */ |
| 652 | async_file_mark (); |
| 653 | } |
| 654 | else |
| 655 | { |
| 656 | delete_file_handler (rocm_event_pipe[0]); |
| 657 | |
| 658 | if (rocm_event_pipe[0] == -1) |
| 659 | return; |
| 660 | |
| 661 | ::close (rocm_event_pipe[0]); |
| 662 | ::close (rocm_event_pipe[1]); |
| 663 | rocm_event_pipe[0] = -1; |
| 664 | rocm_event_pipe[1] = -1; |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | static void |
| 669 | rocm_process_one_event (amd_dbgapi_event_id_t event_id, |
| 670 | amd_dbgapi_event_kind_t event_kind) |
| 671 | { |
| 672 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 673 | amd_dbgapi_process_id_t process_id = info->process_id; |
| 674 | amd_dbgapi_status_t status; |
| 675 | |
| 676 | switch (event_kind) |
| 677 | { |
| 678 | case AMD_DBGAPI_EVENT_KIND_WAVE_STOP: |
| 679 | { |
| 680 | amd_dbgapi_wave_id_t wave_id; |
| 681 | if ((status = amd_dbgapi_event_get_info (process_id, event_id, |
| 682 | AMD_DBGAPI_EVENT_INFO_WAVE, |
| 683 | sizeof (wave_id), &wave_id)) |
| 684 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 685 | error (_ ("event_get_info for event_%ld failed (rc=%d)"), |
| 686 | event_id.handle, status); |
| 687 | |
| 688 | amd_dbgapi_wave_stop_reason_t stop_reason; |
| 689 | status = amd_dbgapi_wave_get_info (process_id, wave_id, |
| 690 | AMD_DBGAPI_WAVE_INFO_STOP_REASON, |
| 691 | sizeof (stop_reason), &stop_reason); |
| 692 | |
| 693 | if (status != AMD_DBGAPI_STATUS_SUCCESS |
| 694 | && status != AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID) |
| 695 | error (_ ("wave_get_info for wave_%ld failed (rc=%d)"), |
| 696 | wave_id.handle, status); |
| 697 | |
| 698 | /* The wave may have exited, or the queue went into an error |
| 699 | state. In such cases, we will see another wave command |
| 700 | terminated event, and handle the wave termination then. */ |
| 701 | |
| 702 | if (status == AMD_DBGAPI_STATUS_SUCCESS) |
| 703 | info->wave_stop_events.emplace_back ( |
| 704 | std::make_pair (wave_id, stop_reason)); |
| 705 | } |
| 706 | break; |
| 707 | |
| 708 | case AMD_DBGAPI_EVENT_KIND_CODE_OBJECT_LIST_UPDATED: |
| 709 | amd_dbgapi_code_object_list_updated.notify (); |
| 710 | break; |
| 711 | |
| 712 | case AMD_DBGAPI_EVENT_KIND_BREAKPOINT_RESUME: |
| 713 | /* Breakpoint resume events should be handled by the breakpoint |
| 714 | action, and this code should not reach this. */ |
| 715 | gdb_assert_not_reached (_ ("unhandled event kind")); |
| 716 | break; |
| 717 | |
| 718 | case AMD_DBGAPI_EVENT_KIND_RUNTIME: |
| 719 | { |
| 720 | amd_dbgapi_runtime_state_t runtime_state; |
| 721 | |
| 722 | if ((status = amd_dbgapi_event_get_info ( |
| 723 | process_id, event_id, AMD_DBGAPI_EVENT_INFO_RUNTIME_STATE, |
| 724 | sizeof (runtime_state), &runtime_state)) |
| 725 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 726 | error (_ ("event_get_info for event_%ld failed (rc=%d)"), |
| 727 | event_id.handle, status); |
| 728 | |
| 729 | switch (runtime_state) |
| 730 | { |
| 731 | case AMD_DBGAPI_RUNTIME_STATE_LOADED_SUPPORTED: |
| 732 | amd_dbgapi_activated.notify (); |
| 733 | break; |
| 734 | case AMD_DBGAPI_RUNTIME_STATE_LOADED_UNSUPPORTED: |
| 735 | warning (_ ("ROCm-GDB: low-level runtime version not supported")); |
| 736 | break; |
| 737 | |
| 738 | case AMD_DBGAPI_RUNTIME_STATE_UNLOADED: |
| 739 | amd_dbgapi_deactivated.notify (); |
| 740 | break; |
| 741 | } |
| 742 | } |
| 743 | break; |
| 744 | |
| 745 | default: |
| 746 | error (_ ("event kind (%d) not supported"), event_kind); |
| 747 | } |
| 748 | |
| 749 | amd_dbgapi_event_processed (process_id, event_id); |
| 750 | } |
| 751 | |
| 752 | /* Drain the amd_dbgapi event queue until an event of the given type is seen. |
| 753 | If no particular event kind is specified (AMD_DBGAPI_EVENT_KIND_NONE), the |
| 754 | event queue is completely drained. Wave stop events that are not returned |
| 755 | are re-queued into the current's process pending wave events. */ |
| 756 | static amd_dbgapi_event_id_t |
| 757 | rocm_process_event_queue (amd_dbgapi_event_kind_t until_event_kind) |
| 758 | { |
| 759 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 760 | |
| 761 | while (true) |
| 762 | { |
| 763 | amd_dbgapi_event_id_t event_id; |
| 764 | amd_dbgapi_event_kind_t event_kind; |
| 765 | |
| 766 | amd_dbgapi_status_t status = amd_dbgapi_next_pending_event ( |
| 767 | info->process_id, &event_id, &event_kind); |
| 768 | |
| 769 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 770 | error (_ ("next_pending_event failed (rc=%d)"), status); |
| 771 | |
| 772 | if (event_id.handle == AMD_DBGAPI_EVENT_NONE.handle |
| 773 | || event_kind == until_event_kind) |
| 774 | return event_id; |
| 775 | |
| 776 | rocm_process_one_event (event_id, event_kind); |
| 777 | } |
| 778 | } |
| 779 | |
| 780 | ptid_t |
| 781 | rocm_target_ops::wait (ptid_t ptid, struct target_waitstatus *ws, |
| 782 | int target_options) |
| 783 | { |
| 784 | if (debug_infrun) |
| 785 | fprintf_unfiltered (gdb_stdlog, |
| 786 | "\e[1;34minfrun: rocm_target_ops::wait\e[0m\n"); |
| 787 | |
| 788 | if (!ptid_is_gpu (ptid)) |
| 789 | { |
| 790 | ptid_t event_ptid = beneath ()->wait (ptid, ws, target_options); |
| 791 | if (event_ptid != minus_one_ptid) |
| 792 | return event_ptid; |
| 793 | } |
| 794 | |
| 795 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 796 | amd_dbgapi_process_id_t process_id = info->process_id; |
| 797 | |
| 798 | /* Drain all the events from the amd_dbgapi, and preserve the ordering. */ |
| 799 | if (info->wave_stop_events.empty ()) |
| 800 | { |
| 801 | amd_dbgapi_process_set_progress (process_id, |
| 802 | AMD_DBGAPI_PROGRESS_NO_FORWARD); |
| 803 | |
| 804 | /* Flush the async file first. */ |
| 805 | if (target_is_async_p ()) |
| 806 | async_file_flush (); |
| 807 | |
| 808 | rocm_process_event_queue (); |
| 809 | |
| 810 | /* In all-stop mode, unless the event queue is empty (spurious wake-up), |
| 811 | we can keep the process in progress_no_forward mode. The infrun loop |
| 812 | will enable forward progress when a thread is resumed. */ |
| 813 | if (non_stop || info->wave_stop_events.empty ()) |
| 814 | amd_dbgapi_process_set_progress (process_id, |
| 815 | AMD_DBGAPI_PROGRESS_NORMAL); |
| 816 | } |
| 817 | |
| 818 | if (info->wave_stop_events.empty ()) |
| 819 | return minus_one_ptid; |
| 820 | |
| 821 | amd_dbgapi_wave_id_t event_wave_id; |
| 822 | amd_dbgapi_wave_stop_reason_t stop_reason; |
| 823 | |
| 824 | std::tie (event_wave_id, stop_reason) = info->wave_stop_events.front (); |
| 825 | info->wave_stop_events.pop_front (); |
| 826 | |
| 827 | struct inferior *inf = current_inferior (); |
| 828 | ptid_t event_ptid (inf->pid, 1, event_wave_id.handle); |
| 829 | |
| 830 | if (!find_thread_ptid (inf->process_target (), event_ptid)) |
| 831 | { |
| 832 | add_thread_silent (inf->process_target (), event_ptid); |
| 833 | set_running (inf->process_target (), event_ptid, 1); |
| 834 | set_executing (inf->process_target (), event_ptid, 1); |
| 835 | } |
| 836 | |
| 837 | /* Since we are manipulating the register cache for the event thread, |
| 838 | make sure it is the current thread. */ |
| 839 | switch_to_thread (inf->process_target (), event_ptid); |
| 840 | |
| 841 | /* By caching the PC now, we avoid having to suspend/resume the queue |
| 842 | later when we need to access it. */ |
| 843 | amd_dbgapi_global_address_t stop_pc; |
| 844 | if (amd_dbgapi_wave_get_info (process_id, event_wave_id, |
| 845 | AMD_DBGAPI_WAVE_INFO_PC, sizeof (stop_pc), |
| 846 | &stop_pc) |
| 847 | == AMD_DBGAPI_STATUS_SUCCESS) |
| 848 | { |
| 849 | struct regcache *regcache = get_thread_regcache_for_ptid (event_ptid); |
| 850 | regcache->raw_supply (gdbarch_pc_regnum (regcache->arch ()), &stop_pc); |
| 851 | } |
| 852 | ws->kind = TARGET_WAITKIND_STOPPED; |
| 853 | |
| 854 | if (stop_reason |
| 855 | & (AMD_DBGAPI_WAVE_STOP_REASON_BREAKPOINT |
| 856 | | AMD_DBGAPI_WAVE_STOP_REASON_SINGLE_STEP)) |
| 857 | ws->value.sig = GDB_SIGNAL_TRAP; |
| 858 | else if (stop_reason & AMD_DBGAPI_WAVE_STOP_REASON_MEMORY_VIOLATION) |
| 859 | ws->value.sig = GDB_SIGNAL_SEGV; |
| 860 | else if (stop_reason |
| 861 | & (AMD_DBGAPI_WAVE_STOP_REASON_FP_INPUT_DENORMAL |
| 862 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_DIVIDE_BY_0 |
| 863 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_OVERFLOW |
| 864 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_UNDERFLOW |
| 865 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_INEXACT |
| 866 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_INVALID_OPERATION |
| 867 | | AMD_DBGAPI_WAVE_STOP_REASON_INT_DIVIDE_BY_0)) |
| 868 | ws->value.sig = GDB_SIGNAL_FPE; |
| 869 | else |
| 870 | ws->value.sig = GDB_SIGNAL_0; |
| 871 | |
| 872 | /* If there are more events in the list, mark the async file so that |
| 873 | rocm_target_ops::wait gets called again. */ |
| 874 | if (target_is_async_p () && !info->wave_stop_events.empty ()) |
| 875 | async_file_mark (); |
| 876 | |
| 877 | return event_ptid; |
| 878 | } |
| 879 | |
| 880 | bool |
| 881 | rocm_target_ops::stopped_by_sw_breakpoint () |
| 882 | { |
| 883 | if (!ptid_is_gpu (inferior_ptid)) |
| 884 | return beneath ()->supports_stopped_by_sw_breakpoint () |
| 885 | && beneath ()->stopped_by_sw_breakpoint (); |
| 886 | |
| 887 | /* FIXME: we should check that the wave is not single-stepping. */ |
| 888 | |
| 889 | struct regcache *regcache = get_thread_regcache_for_ptid (inferior_ptid); |
| 890 | |
| 891 | CORE_ADDR bkpt_pc = regcache_read_pc (regcache) |
| 892 | - gdbarch_decr_pc_after_break (regcache->arch ()); |
| 893 | |
| 894 | return software_breakpoint_inserted_here_p (regcache->aspace (), bkpt_pc); |
| 895 | } |
| 896 | |
| 897 | void |
| 898 | rocm_target_ops::mourn_inferior () |
| 899 | { |
| 900 | /* FIXME: only unpush on the last activation. */ |
| 901 | /* Disengage the ROCm target_ops. */ |
| 902 | unpush_target (&rocm_ops); |
| 903 | |
| 904 | beneath ()->mourn_inferior (); |
| 905 | } |
| 906 | |
| 907 | void |
| 908 | rocm_target_ops::fetch_registers (struct regcache *regcache, int regno) |
| 909 | { |
| 910 | struct gdbarch *gdbarch = regcache->arch (); |
| 911 | struct gdbarch_tdep *tdep = gdbarch_tdep (gdbarch); |
| 912 | |
| 913 | /* delegate to the host routines when not on the device */ |
| 914 | |
| 915 | if (!rocm_is_amdgcn_gdbarch (gdbarch)) |
| 916 | { |
| 917 | beneath ()->fetch_registers (regcache, regno); |
| 918 | return; |
| 919 | } |
| 920 | |
| 921 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 922 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (regcache->ptid ()); |
| 923 | |
| 924 | gdb_byte raw[AMDGCN_MAX_REGISTER_SIZE]; |
| 925 | |
| 926 | amd_dbgapi_status_t status = amd_dbgapi_read_register ( |
| 927 | process_id, wave_id, tdep->register_ids[regno], 0, |
| 928 | TYPE_LENGTH (register_type (gdbarch, regno)), raw); |
| 929 | |
| 930 | if (status == AMD_DBGAPI_STATUS_SUCCESS) |
| 931 | { |
| 932 | regcache->raw_supply (regno, raw); |
| 933 | } |
| 934 | else if (status != AMD_DBGAPI_STATUS_ERROR_INVALID_REGISTER_ID) |
| 935 | { |
| 936 | warning (_ ("Couldn't read register %s (#%d)."), |
| 937 | gdbarch_register_name (gdbarch, regno), regno); |
| 938 | } |
| 939 | } |
| 940 | |
| 941 | void |
| 942 | rocm_target_ops::store_registers (struct regcache *regcache, int regno) |
| 943 | { |
| 944 | struct gdbarch *gdbarch = regcache->arch (); |
| 945 | struct gdbarch_tdep *tdep = gdbarch_tdep (gdbarch); |
| 946 | gdb_byte raw[AMDGCN_MAX_REGISTER_SIZE]; |
| 947 | |
| 948 | if (!rocm_is_amdgcn_gdbarch (gdbarch)) |
| 949 | { |
| 950 | beneath ()->store_registers (regcache, regno); |
| 951 | return; |
| 952 | } |
| 953 | |
| 954 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 955 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (regcache->ptid ()); |
| 956 | |
| 957 | regcache->raw_collect (regno, &raw); |
| 958 | |
| 959 | amd_dbgapi_status_t status = amd_dbgapi_write_register ( |
| 960 | process_id, wave_id, tdep->register_ids[regno], 0, |
| 961 | TYPE_LENGTH (register_type (gdbarch, regno)), raw); |
| 962 | |
| 963 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 964 | { |
| 965 | warning (_ ("Couldn't write register %s (#%d)."), |
| 966 | gdbarch_register_name (gdbarch, regno), regno); |
| 967 | } |
| 968 | } |
| 969 | |
| 970 | /* Fix breakpoints created with an address location while the |
| 971 | architecture was set to the host (could be fixed in core GDB). */ |
| 972 | |
| 973 | static void |
| 974 | rocm_target_breakpoint_fixup (struct breakpoint *b) |
| 975 | { |
| 976 | if (b->location.get () |
| 977 | && event_location_type (b->location.get ()) == ADDRESS_LOCATION |
| 978 | && gdbarch_bfd_arch_info (b->loc->gdbarch)->arch == bfd_arch_amdgcn |
| 979 | && gdbarch_bfd_arch_info (b->gdbarch)->arch != bfd_arch_amdgcn) |
| 980 | { |
| 981 | b->gdbarch = b->loc->gdbarch; |
| 982 | } |
| 983 | } |
| 984 | |
| 985 | struct gdbarch * |
| 986 | rocm_target_ops::thread_architecture (ptid_t ptid) |
| 987 | { |
| 988 | static std::result_of<decltype (&ptid_t::tid) (ptid_t)>::type last_tid = 0; |
| 989 | static struct gdbarch *cached_arch = nullptr; |
| 990 | |
| 991 | if (!ptid_is_gpu (ptid)) |
| 992 | return beneath ()->thread_architecture (ptid); |
| 993 | |
| 994 | /* We can cache the gdbarch for a given wave_id (ptid::tid) because |
| 995 | wave IDs are unique, and aren't reused. */ |
| 996 | if (ptid.tid () == last_tid) |
| 997 | return cached_arch; |
| 998 | |
| 999 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 1000 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (ptid); |
| 1001 | amd_dbgapi_architecture_id_t architecture_id; |
| 1002 | |
| 1003 | if (amd_dbgapi_wave_get_info (process_id, wave_id, |
| 1004 | AMD_DBGAPI_WAVE_INFO_ARCHITECTURE, |
| 1005 | sizeof (architecture_id), &architecture_id) |
| 1006 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1007 | error (_ ("Couldn't get architecture for wave_%ld"), ptid.tid ()); |
| 1008 | |
| 1009 | uint32_t elf_amdgpu_machine; |
| 1010 | if (amd_dbgapi_architecture_get_info ( |
| 1011 | architecture_id, AMD_DBGAPI_ARCHITECTURE_INFO_ELF_AMDGPU_MACHINE, |
| 1012 | sizeof (elf_amdgpu_machine), &elf_amdgpu_machine) |
| 1013 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1014 | error (_ ("Couldn't get elf_amdgpu_machine for architecture_%ld"), |
| 1015 | architecture_id.handle); |
| 1016 | |
| 1017 | struct gdbarch_info info; |
| 1018 | gdbarch_info_init (&info); |
| 1019 | |
| 1020 | info.bfd_arch_info = bfd_lookup_arch (bfd_arch_amdgcn, elf_amdgpu_machine); |
| 1021 | info.byte_order = BFD_ENDIAN_LITTLE; |
| 1022 | info.osabi = GDB_OSABI_AMDGPU_HSA; |
| 1023 | |
| 1024 | last_tid = ptid.tid (); |
| 1025 | if (!(cached_arch = gdbarch_find_by_info (info))) |
| 1026 | error (_ ("Couldn't get elf_amdgpu_machine (%#x)"), elf_amdgpu_machine); |
| 1027 | |
| 1028 | return cached_arch; |
| 1029 | } |
| 1030 | |
| 1031 | void |
| 1032 | rocm_target_ops::update_thread_list () |
| 1033 | { |
| 1034 | for (inferior *inf : all_inferiors ()) |
| 1035 | { |
| 1036 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (inf); |
| 1037 | amd_dbgapi_wave_id_t *wave_list; |
| 1038 | size_t count; |
| 1039 | |
| 1040 | amd_dbgapi_changed_t changed; |
| 1041 | amd_dbgapi_status_t status; |
| 1042 | if ((status |
| 1043 | = amd_dbgapi_wave_list (process_id, &count, &wave_list, &changed)) |
| 1044 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1045 | error (_ ("amd_dbgapi_wave_list failed (rc=%d"), status); |
| 1046 | |
| 1047 | if (changed == AMD_DBGAPI_CHANGED_NO) |
| 1048 | continue; |
| 1049 | |
| 1050 | /* Create a set and free the wave list. */ |
| 1051 | std::set<std::result_of<decltype (&ptid_t::tid) (ptid_t)>::type> threads; |
| 1052 | for (size_t i = 0; i < count; ++i) |
| 1053 | threads.emplace (wave_list[i].handle); |
| 1054 | xfree (wave_list); |
| 1055 | |
| 1056 | /* Then prune the wave_ids that already have a thread_info. */ |
| 1057 | for (thread_info *tp : inf->non_exited_threads ()) |
| 1058 | if (ptid_is_gpu (tp->ptid)) |
| 1059 | threads.erase (tp->ptid.tid ()); |
| 1060 | |
| 1061 | /* The wave_ids that are left require a new thread_info. */ |
| 1062 | for (auto &&tid : threads) |
| 1063 | { |
| 1064 | ptid_t wave_ptid (inf->pid, 1, tid); |
| 1065 | amd_dbgapi_wave_state_t state; |
| 1066 | |
| 1067 | if (amd_dbgapi_wave_get_info ( |
| 1068 | process_id, get_amd_dbgapi_wave_id (wave_ptid), |
| 1069 | AMD_DBGAPI_WAVE_INFO_STATE, sizeof (state), &state) |
| 1070 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1071 | continue; |
| 1072 | |
| 1073 | add_thread_silent (inf->process_target (), wave_ptid); |
| 1074 | set_running (inf->process_target (), wave_ptid, 1); |
| 1075 | set_executing (inf->process_target (), wave_ptid, 1); |
| 1076 | } |
| 1077 | } |
| 1078 | |
| 1079 | /* Give the beneath target a chance to do extra processing. */ |
| 1080 | this->beneath ()->update_thread_list (); |
| 1081 | } |
| 1082 | |
| 1083 | static void |
| 1084 | rocm_target_solib_loaded (struct so_list *solib) |
| 1085 | { |
| 1086 | /* Notify the amd_dbgapi that a shared library has been loaded. */ |
| 1087 | for (auto &&value : get_rocm_inferior_info ()->notify_solib_map) |
| 1088 | /* TODO: If we want to support file name wildcards, change this code. */ |
| 1089 | if (::strstr (solib->so_original_name, value.second.compare.c_str ()) |
| 1090 | && !value.second.is_loaded) |
| 1091 | { |
| 1092 | value.second.solib = solib; |
| 1093 | value.second.is_loaded = true; |
| 1094 | |
| 1095 | amd_dbgapi_report_shared_library ( |
| 1096 | get_amd_dbgapi_process_id (), |
| 1097 | amd_dbgapi_shared_library_id_t{ value.first }, |
| 1098 | AMD_DBGAPI_SHARED_LIBRARY_STATE_LOADED); |
| 1099 | } |
| 1100 | } |
| 1101 | |
| 1102 | static void |
| 1103 | rocm_target_solib_unloaded (struct so_list *solib) |
| 1104 | { |
| 1105 | /* Notify the amd_dbgapi that a shared library will unload. */ |
| 1106 | for (auto &&value : get_rocm_inferior_info ()->notify_solib_map) |
| 1107 | /* TODO: If we want to support file name wildcards, change this code. */ |
| 1108 | if (::strstr (solib->so_original_name, value.second.compare.c_str ()) |
| 1109 | && value.second.is_loaded) |
| 1110 | { |
| 1111 | struct rocm_inferior_info *info = get_rocm_inferior_info (); |
| 1112 | |
| 1113 | amd_dbgapi_report_shared_library ( |
| 1114 | info->process_id, amd_dbgapi_shared_library_id_t{ value.first }, |
| 1115 | AMD_DBGAPI_SHARED_LIBRARY_STATE_UNLOADED); |
| 1116 | |
| 1117 | /* Delete breakpoints that were left inserted in this shared library. |
| 1118 | */ |
| 1119 | for (auto it = info->breakpoint_map.begin (); |
| 1120 | it != info->breakpoint_map.end ();) |
| 1121 | if (solib_contains_address_p (solib, it->second->loc->address)) |
| 1122 | { |
| 1123 | warning (_ ("breakpoint_%ld is still inserted after " |
| 1124 | "shared_library_%ld was unloaded"), |
| 1125 | it->first, value.first); |
| 1126 | delete_breakpoint (it->second); |
| 1127 | it = info->breakpoint_map.erase (it); |
| 1128 | } |
| 1129 | else |
| 1130 | ++it; |
| 1131 | |
| 1132 | value.second.solib = nullptr; |
| 1133 | value.second.is_loaded = false; |
| 1134 | } |
| 1135 | } |
| 1136 | |
| 1137 | static void |
| 1138 | rocm_target_inferior_created (struct target_ops *target, int from_tty) |
| 1139 | { |
| 1140 | struct inferior *inf = current_inferior (); |
| 1141 | auto *info = get_rocm_inferior_info (inf); |
| 1142 | amd_dbgapi_status_t status; |
| 1143 | |
| 1144 | if (!target_can_async_p ()) |
| 1145 | { |
| 1146 | warning ( |
| 1147 | _ ("ROCm-GDB requires target-async, GPU debugging is disabled")); |
| 1148 | return; |
| 1149 | } |
| 1150 | |
| 1151 | gdb_assert (info->wave_stop_events.empty ()); |
| 1152 | |
| 1153 | status = amd_dbgapi_process_attach (inf, &info->process_id); |
| 1154 | |
| 1155 | if (status == AMD_DBGAPI_STATUS_ERROR_VERSION_MISMATCH) |
| 1156 | warning (_ ("The version of the kernel driver does not match the version " |
| 1157 | "required by the ROCm debugger library")); |
| 1158 | |
| 1159 | if (status != AMD_DBGAPI_STATUS_SUCCESS) |
| 1160 | { |
| 1161 | warning (_ ("Could not attach to process %d"), inf->pid); |
| 1162 | return; |
| 1163 | } |
| 1164 | |
| 1165 | if (amd_dbgapi_process_get_info (info->process_id, |
| 1166 | AMD_DBGAPI_PROCESS_INFO_NOTIFIER, |
| 1167 | sizeof (info->notifier), &info->notifier) |
| 1168 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1169 | { |
| 1170 | warning (_ ("Could not retrieve process %d's notifier"), inf->pid); |
| 1171 | amd_dbgapi_process_detach (info->process_id); |
| 1172 | return; |
| 1173 | } |
| 1174 | |
| 1175 | /* We add a file handler for events returned by the debugger api. We'll use |
| 1176 | this handler to signal our async handler that events are available. */ |
| 1177 | add_file_handler ( |
| 1178 | info->notifier, |
| 1179 | [] (int error, gdb_client_data client_data) { |
| 1180 | auto info_ = static_cast<struct rocm_inferior_info *> (client_data); |
| 1181 | int ret; |
| 1182 | |
| 1183 | /* Drain the notifier pipe. */ |
| 1184 | do |
| 1185 | { |
| 1186 | char buf; |
| 1187 | ret = read (info_->notifier, &buf, 1); |
| 1188 | } |
| 1189 | while (ret >= 0 || (ret == -1 && errno == EINTR)); |
| 1190 | |
| 1191 | /* Signal our async handler. */ |
| 1192 | async_file_mark (); |
| 1193 | }, |
| 1194 | info); |
| 1195 | |
| 1196 | /* Attaching to the inferior may have generated runtime events, process |
| 1197 | them now. */ |
| 1198 | rocm_process_event_queue (); |
| 1199 | } |
| 1200 | |
| 1201 | static void |
| 1202 | rocm_target_inferior_exit (struct inferior *inf) |
| 1203 | { |
| 1204 | auto *info = get_rocm_inferior_info (inf); |
| 1205 | |
| 1206 | amd_dbgapi_deactivated.notify (); |
| 1207 | |
| 1208 | if (info->notifier != -1) |
| 1209 | delete_file_handler (info->notifier); |
| 1210 | |
| 1211 | amd_dbgapi_process_detach (info->process_id); |
| 1212 | |
| 1213 | /* Delete the breakpoints that are still active. */ |
| 1214 | for (auto &&value : info->breakpoint_map) |
| 1215 | delete_breakpoint (value.second); |
| 1216 | |
| 1217 | rocm_inferior_data.clear (inf); |
| 1218 | } |
| 1219 | |
| 1220 | static cli_style_option warning_style ("rocm_warning", ui_file_style::RED); |
| 1221 | static cli_style_option info_style ("rocm_info", ui_file_style::GREEN); |
| 1222 | static cli_style_option verbose_style ("rocm_verbose", ui_file_style::BLUE); |
| 1223 | |
| 1224 | static amd_dbgapi_callbacks_t dbgapi_callbacks = { |
| 1225 | /* allocate_memory. */ |
| 1226 | .allocate_memory = xmalloc, |
| 1227 | |
| 1228 | /* deallocate_memory. */ |
| 1229 | .deallocate_memory = xfree, |
| 1230 | |
| 1231 | /* get_os_pid. */ |
| 1232 | .get_os_pid = [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1233 | pid_t *pid) -> amd_dbgapi_status_t { |
| 1234 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1235 | |
| 1236 | *pid = inf->pid; |
| 1237 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1238 | }, |
| 1239 | |
| 1240 | /* enable_notify_shared_library callback. */ |
| 1241 | .enable_notify_shared_library |
| 1242 | = [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1243 | const char *library_name, amd_dbgapi_shared_library_id_t library_id, |
| 1244 | amd_dbgapi_shared_library_state_t *library_state) |
| 1245 | -> amd_dbgapi_status_t { |
| 1246 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1247 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1248 | |
| 1249 | if (!library_name || !library_state) |
| 1250 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; |
| 1251 | |
| 1252 | if (info->notify_solib_map.find (library_id.handle) |
| 1253 | != info->notify_solib_map.end ()) |
| 1254 | { |
| 1255 | /* This library id is already registered. */ |
| 1256 | return AMD_DBGAPI_STATUS_ERROR; |
| 1257 | } |
| 1258 | |
| 1259 | /* Check whether the library is already loaded. */ |
| 1260 | bool is_loaded = false; |
| 1261 | struct so_list *solib; |
| 1262 | for (solib = inf->pspace->so_list; solib; solib = solib->next) |
| 1263 | if (::strstr (solib->so_original_name, library_name)) |
| 1264 | { |
| 1265 | is_loaded = true; |
| 1266 | break; |
| 1267 | } |
| 1268 | |
| 1269 | /* Check that the library_name is valid. If must not be empty, and |
| 1270 | should not have wildcard characters. */ |
| 1271 | if (*library_name == '\0' |
| 1272 | || std::string (library_name).find_first_of ("*?[]") |
| 1273 | != std::string::npos) |
| 1274 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; |
| 1275 | |
| 1276 | /* Add a new entry in the notify_solib_map. */ |
| 1277 | if (!info->notify_solib_map |
| 1278 | .emplace (std::piecewise_construct, |
| 1279 | std::forward_as_tuple (library_id.handle), |
| 1280 | std::forward_as_tuple (rocm_notify_shared_library_info{ |
| 1281 | library_name, solib, is_loaded })) |
| 1282 | .second) |
| 1283 | return AMD_DBGAPI_STATUS_ERROR; |
| 1284 | |
| 1285 | *library_state = is_loaded ? AMD_DBGAPI_SHARED_LIBRARY_STATE_LOADED |
| 1286 | : AMD_DBGAPI_SHARED_LIBRARY_STATE_UNLOADED; |
| 1287 | |
| 1288 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1289 | }, |
| 1290 | |
| 1291 | /* disable_notify_shared_library callback. */ |
| 1292 | .disable_notify_shared_library |
| 1293 | = [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1294 | amd_dbgapi_shared_library_id_t library_id) -> amd_dbgapi_status_t { |
| 1295 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1296 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1297 | |
| 1298 | auto it = info->notify_solib_map.find (library_id.handle); |
| 1299 | if (it == info->notify_solib_map.end ()) |
| 1300 | return AMD_DBGAPI_STATUS_ERROR_INVALID_SHARED_LIBRARY_ID; |
| 1301 | |
| 1302 | info->notify_solib_map.erase (it); |
| 1303 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1304 | }, |
| 1305 | |
| 1306 | /* get_symbol_address callback. */ |
| 1307 | .get_symbol_address = |
| 1308 | [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1309 | amd_dbgapi_shared_library_id_t library_id, const char *symbol_name, |
| 1310 | amd_dbgapi_global_address_t *address) { |
| 1311 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1312 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1313 | |
| 1314 | auto it = info->notify_solib_map.find (library_id.handle); |
| 1315 | if (it == info->notify_solib_map.end ()) |
| 1316 | return AMD_DBGAPI_STATUS_ERROR_INVALID_SHARED_LIBRARY_ID; |
| 1317 | |
| 1318 | struct so_list *solib = it->second.solib; |
| 1319 | if (!solib) |
| 1320 | return AMD_DBGAPI_STATUS_ERROR_LIBRARY_NOT_LOADED; |
| 1321 | |
| 1322 | solib_read_symbols (solib, 0); |
| 1323 | gdb_assert (solib->objfile); |
| 1324 | |
| 1325 | struct bound_minimal_symbol msymbol |
| 1326 | = lookup_minimal_symbol (symbol_name, NULL, solib->objfile); |
| 1327 | |
| 1328 | if (!msymbol.minsym || BMSYMBOL_VALUE_ADDRESS (msymbol) == 0) |
| 1329 | return AMD_DBGAPI_STATUS_ERROR_SYMBOL_NOT_FOUND; |
| 1330 | |
| 1331 | *address = BMSYMBOL_VALUE_ADDRESS (msymbol); |
| 1332 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1333 | }, |
| 1334 | |
| 1335 | /* set_breakpoint callback. */ |
| 1336 | .add_breakpoint = |
| 1337 | [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1338 | amd_dbgapi_shared_library_id_t shared_library_id, |
| 1339 | amd_dbgapi_global_address_t address, |
| 1340 | amd_dbgapi_breakpoint_id_t breakpoint_id) { |
| 1341 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1342 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1343 | |
| 1344 | /* Initialize the breakpoint ops lazily since we depend on |
| 1345 | bkpt_breakpoint_ops and we can't control the order in which |
| 1346 | initializers are called. */ |
| 1347 | if (rocm_breakpoint_ops.check_status == NULL) |
| 1348 | { |
| 1349 | rocm_breakpoint_ops = bkpt_breakpoint_ops; |
| 1350 | rocm_breakpoint_ops.check_status = rocm_breakpoint_check_status; |
| 1351 | } |
| 1352 | |
| 1353 | auto it = info->breakpoint_map.find (breakpoint_id.handle); |
| 1354 | if (it != info->breakpoint_map.end ()) |
| 1355 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; |
| 1356 | |
| 1357 | /* Create a new breakpoint. */ |
| 1358 | struct obj_section *section = find_pc_section (address); |
| 1359 | if (!section || !section->objfile) |
| 1360 | return AMD_DBGAPI_STATUS_ERROR; |
| 1361 | |
| 1362 | event_location_up location |
| 1363 | = new_address_location (address, nullptr, 0); |
| 1364 | if (!create_breakpoint ( |
| 1365 | get_objfile_arch (section->objfile), location.get (), |
| 1366 | /*cond_string*/ NULL, /*thread*/ -1, /*extra_sring*/ NULL, |
| 1367 | /*parse_extra*/ 0, /*tempflag*/ 0, /*bptype*/ bp_breakpoint, |
| 1368 | /*ignore_count*/ 0, /*pending_break*/ AUTO_BOOLEAN_FALSE, |
| 1369 | /*ops*/ &rocm_breakpoint_ops, /*from_tty*/ 0, |
| 1370 | /*enabled*/ 1, /*internal*/ 1, /*flags*/ 0)) |
| 1371 | return AMD_DBGAPI_STATUS_ERROR; |
| 1372 | |
| 1373 | /* Find our breakpoint in the breakpoint list. */ |
| 1374 | auto bp_loc = std::make_pair (inf->aspace, address); |
| 1375 | auto bp = breakpoint_find_if ( |
| 1376 | [] (struct breakpoint *b, void *data) { |
| 1377 | auto *arg = static_cast<decltype (&bp_loc)> (data); |
| 1378 | if (b->ops == &rocm_breakpoint_ops && b->loc |
| 1379 | && b->loc->pspace->aspace == arg->first |
| 1380 | && b->loc->address == arg->second) |
| 1381 | return 1; |
| 1382 | return 0; |
| 1383 | }, |
| 1384 | reinterpret_cast<void *> (&bp_loc)); |
| 1385 | |
| 1386 | if (!bp) |
| 1387 | error (_ ("Could not find breakpoint")); |
| 1388 | |
| 1389 | info->breakpoint_map.emplace (breakpoint_id.handle, bp); |
| 1390 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1391 | }, |
| 1392 | |
| 1393 | /* remove_breakpoint callback. */ |
| 1394 | .remove_breakpoint = |
| 1395 | [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1396 | amd_dbgapi_breakpoint_id_t breakpoint_id) { |
| 1397 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1398 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1399 | |
| 1400 | auto it = info->breakpoint_map.find (breakpoint_id.handle); |
| 1401 | if (it == info->breakpoint_map.end ()) |
| 1402 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; |
| 1403 | |
| 1404 | delete_breakpoint (it->second); |
| 1405 | info->breakpoint_map.erase (it); |
| 1406 | |
| 1407 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1408 | }, |
| 1409 | |
| 1410 | /* set_breakpoint_state callback. */ |
| 1411 | .set_breakpoint_state = |
| 1412 | [] (amd_dbgapi_client_process_id_t client_process_id, |
| 1413 | amd_dbgapi_breakpoint_id_t breakpoint_id, |
| 1414 | amd_dbgapi_breakpoint_state_t breakpoint_state) { |
| 1415 | inferior *inf = static_cast<inferior *> (client_process_id); |
| 1416 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); |
| 1417 | |
| 1418 | auto it = info->breakpoint_map.find (breakpoint_id.handle); |
| 1419 | if (it == info->breakpoint_map.end ()) |
| 1420 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; |
| 1421 | |
| 1422 | if (breakpoint_state == AMD_DBGAPI_BREAKPOINT_STATE_ENABLE) |
| 1423 | it->second->enable_state = bp_enabled; |
| 1424 | else if (breakpoint_state == AMD_DBGAPI_BREAKPOINT_STATE_DISABLE) |
| 1425 | it->second->enable_state = bp_disabled; |
| 1426 | else |
| 1427 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; |
| 1428 | |
| 1429 | return AMD_DBGAPI_STATUS_SUCCESS; |
| 1430 | }, |
| 1431 | |
| 1432 | .log_message |
| 1433 | = [] (amd_dbgapi_log_level_t level, const char *message) -> void { |
| 1434 | gdb::optional<target_terminal::scoped_restore_terminal_state> tstate; |
| 1435 | |
| 1436 | if (level > get_debug_amd_dbgapi_log_level ()) |
| 1437 | return; |
| 1438 | |
| 1439 | if (target_supports_terminal_ours ()) |
| 1440 | { |
| 1441 | tstate.emplace (); |
| 1442 | target_terminal::ours_for_output (); |
| 1443 | } |
| 1444 | |
| 1445 | if (filtered_printing_initialized ()) |
| 1446 | wrap_here (""); |
| 1447 | |
| 1448 | struct ui_file *out_file |
| 1449 | = (level >= AMD_DBGAPI_LOG_LEVEL_INFO) ? gdb_stdlog : gdb_stderr; |
| 1450 | |
| 1451 | switch (level) |
| 1452 | { |
| 1453 | case AMD_DBGAPI_LOG_LEVEL_FATAL_ERROR: |
| 1454 | fputs_unfiltered ("[amd-dbgapi]: ", out_file); |
| 1455 | break; |
| 1456 | case AMD_DBGAPI_LOG_LEVEL_WARNING: |
| 1457 | fputs_styled ("[amd-dbgapi]: ", warning_style.style (), out_file); |
| 1458 | break; |
| 1459 | case AMD_DBGAPI_LOG_LEVEL_INFO: |
| 1460 | fputs_styled ("[amd-dbgapi]: ", info_style.style (), out_file); |
| 1461 | break; |
| 1462 | case AMD_DBGAPI_LOG_LEVEL_VERBOSE: |
| 1463 | fputs_styled ("[amd-dbgapi]: ", verbose_style.style (), out_file); |
| 1464 | break; |
| 1465 | } |
| 1466 | |
| 1467 | fputs_unfiltered (message, out_file); |
| 1468 | fputs_unfiltered ("\n", out_file); |
| 1469 | } |
| 1470 | }; |
| 1471 | |
| 1472 | /* Implementation of `_wave_id' variable. */ |
| 1473 | |
| 1474 | static struct value * |
| 1475 | rocm_wave_id_make_value (struct gdbarch *gdbarch, struct internalvar *var, |
| 1476 | void *ignore) |
| 1477 | { |
| 1478 | if (ptid_is_gpu (inferior_ptid)) |
| 1479 | { |
| 1480 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 1481 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (inferior_ptid); |
| 1482 | uint32_t group_ids[3], wave_in_group; |
| 1483 | |
| 1484 | if (amd_dbgapi_wave_get_info (process_id, wave_id, |
| 1485 | AMD_DBGAPI_WAVE_INFO_WORK_GROUP_COORD, |
| 1486 | sizeof (group_ids), &group_ids) |
| 1487 | == AMD_DBGAPI_STATUS_SUCCESS |
| 1488 | && amd_dbgapi_wave_get_info ( |
| 1489 | process_id, wave_id, |
| 1490 | AMD_DBGAPI_WAVE_INFO_WAVE_NUMBER_IN_WORK_GROUP, |
| 1491 | sizeof (wave_in_group), &wave_in_group) |
| 1492 | == AMD_DBGAPI_STATUS_SUCCESS) |
| 1493 | { |
| 1494 | std::string wave_id_str |
| 1495 | = string_printf ("(%d,%d,%d)/%d", group_ids[2], group_ids[1], |
| 1496 | group_ids[0], wave_in_group); |
| 1497 | |
| 1498 | return value_cstring (wave_id_str.data (), wave_id_str.length () + 1, |
| 1499 | builtin_type (gdbarch)->builtin_char); |
| 1500 | } |
| 1501 | } |
| 1502 | |
| 1503 | return allocate_value (builtin_type (gdbarch)->builtin_void); |
| 1504 | } |
| 1505 | |
| 1506 | static const struct internalvar_funcs rocm_wave_id_funcs |
| 1507 | = { rocm_wave_id_make_value, NULL, NULL }; |
| 1508 | |
| 1509 | /* List of set/show debug amd_dbgapi commands. */ |
| 1510 | struct cmd_list_element *set_debug_amd_dbgapi_list; |
| 1511 | struct cmd_list_element *show_debug_amd_dbgapi_list; |
| 1512 | |
| 1513 | static void |
| 1514 | set_debug_amd_dbgapi (const char *arg, int from_tty) |
| 1515 | { |
| 1516 | help_list (set_debug_amd_dbgapi_list, "set debug amd-dbgapi ", |
| 1517 | (enum command_class) - 1, gdb_stdout); |
| 1518 | } |
| 1519 | |
| 1520 | static void |
| 1521 | show_debug_amd_dbgapi (const char *args, int from_tty) |
| 1522 | { |
| 1523 | cmd_show_list (show_debug_amd_dbgapi_list, from_tty, ""); |
| 1524 | } |
| 1525 | |
| 1526 | constexpr char amd_dbgapi_log_level_off[] = "off"; |
| 1527 | constexpr char amd_dbgapi_log_level_error[] = "error"; |
| 1528 | constexpr char amd_dbgapi_log_level_warning[] = "warning"; |
| 1529 | constexpr char amd_dbgapi_log_level_info[] = "info"; |
| 1530 | constexpr char amd_dbgapi_log_level_verbose[] = "verbose"; |
| 1531 | |
| 1532 | constexpr const char *debug_amd_dbgapi_log_level_enums[] |
| 1533 | = { [AMD_DBGAPI_LOG_LEVEL_NONE] = amd_dbgapi_log_level_off, |
| 1534 | [AMD_DBGAPI_LOG_LEVEL_FATAL_ERROR] = amd_dbgapi_log_level_error, |
| 1535 | [AMD_DBGAPI_LOG_LEVEL_WARNING] = amd_dbgapi_log_level_warning, |
| 1536 | [AMD_DBGAPI_LOG_LEVEL_INFO] = amd_dbgapi_log_level_info, |
| 1537 | [AMD_DBGAPI_LOG_LEVEL_VERBOSE] = amd_dbgapi_log_level_verbose, |
| 1538 | nullptr }; |
| 1539 | |
| 1540 | static const char *debug_amd_dbgapi_log_level = amd_dbgapi_log_level_error; |
| 1541 | |
| 1542 | static amd_dbgapi_log_level_t |
| 1543 | get_debug_amd_dbgapi_log_level () |
| 1544 | { |
| 1545 | size_t pos; |
| 1546 | for (pos = 0; debug_amd_dbgapi_log_level_enums[pos]; ++pos) |
| 1547 | if (debug_amd_dbgapi_log_level == debug_amd_dbgapi_log_level_enums[pos]) |
| 1548 | break; |
| 1549 | |
| 1550 | gdb_assert (debug_amd_dbgapi_log_level_enums[pos]); |
| 1551 | return static_cast<amd_dbgapi_log_level_t> (pos); |
| 1552 | } |
| 1553 | |
| 1554 | static void |
| 1555 | set_debug_amd_dbgapi_log_level (const char *args, int from_tty, |
| 1556 | struct cmd_list_element *c) |
| 1557 | { |
| 1558 | amd_dbgapi_set_log_level (get_debug_amd_dbgapi_log_level ()); |
| 1559 | } |
| 1560 | |
| 1561 | static void |
| 1562 | show_debug_amd_dbgapi_log_level (struct ui_file *file, int from_tty, |
| 1563 | struct cmd_list_element *c, const char *value) |
| 1564 | { |
| 1565 | fprintf_filtered (file, _ ("The amd-dbgapi log level is %s.\n"), value); |
| 1566 | } |
| 1567 | |
| 1568 | static void |
| 1569 | info_agents_command (const char *args, int from_tty) |
| 1570 | { |
| 1571 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); |
| 1572 | struct ui_out *uiout = current_uiout; |
| 1573 | amd_dbgapi_status_t status; |
| 1574 | |
| 1575 | amd_dbgapi_agent_id_t *agent_list; |
| 1576 | size_t count = 0; |
| 1577 | |
| 1578 | if (process_id.handle != AMD_DBGAPI_PROCESS_NONE.handle |
| 1579 | && (status |
| 1580 | = amd_dbgapi_agent_list (process_id, &count, &agent_list, nullptr)) |
| 1581 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1582 | error (_ ("amd_dbgapi_agent_list failed (rc=%d)"), status); |
| 1583 | |
| 1584 | if (!count && !uiout->is_mi_like_p ()) |
| 1585 | { |
| 1586 | uiout->field_string (NULL, _ ("No agents are currently active.\n")); |
| 1587 | return; |
| 1588 | } |
| 1589 | |
| 1590 | /* Calculate the maximum size needed to print the agents names. */ |
| 1591 | std::vector<std::string> agent_names (count); |
| 1592 | |
| 1593 | size_t max_name_len = 0; |
| 1594 | for (size_t i = 0; i < count; ++i) |
| 1595 | { |
| 1596 | char *agent_name; |
| 1597 | |
| 1598 | if ((status = amd_dbgapi_agent_get_info ( |
| 1599 | process_id, agent_list[i], AMD_DBGAPI_AGENT_INFO_NAME, |
| 1600 | sizeof (agent_name), &agent_name)) |
| 1601 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1602 | { |
| 1603 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) |
| 1604 | agent_names[i] = "N/A"; |
| 1605 | else |
| 1606 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); |
| 1607 | } |
| 1608 | else |
| 1609 | { |
| 1610 | agent_names[i] = agent_name; |
| 1611 | xfree (agent_name); |
| 1612 | } |
| 1613 | |
| 1614 | max_name_len = std::max (max_name_len, agent_names[i].size ()); |
| 1615 | } |
| 1616 | |
| 1617 | /* Header: */ |
| 1618 | ui_out_emit_table table_emmitter (uiout, 7, count, "InfoRocmDevicesTable"); |
| 1619 | |
| 1620 | uiout->table_header (2, ui_left, "agent_id", "Id"); |
| 1621 | uiout->table_header (8, ui_left, "location_id", "PCI Slot"); |
| 1622 | uiout->table_header (std::max (11ul, max_name_len), ui_left, "name", |
| 1623 | "Device Name"); |
| 1624 | uiout->table_header (14, ui_left, "num_se", "Shader Engines"); |
| 1625 | uiout->table_header (13, ui_left, "num_cu", "Compute Units"); |
| 1626 | uiout->table_header (7, ui_left, "simd", "SIMD/CU"); |
| 1627 | uiout->table_header (15, ui_left, "waves", "Wavefronts/SIMD"); |
| 1628 | uiout->table_body (); |
| 1629 | |
| 1630 | /* Rows: */ |
| 1631 | for (size_t i = 0; i < count; ++i) |
| 1632 | { |
| 1633 | ui_out_emit_tuple tuple_emitter (uiout, "InfoRocmDevicesRow"); |
| 1634 | |
| 1635 | /* agent */ |
| 1636 | uiout->field_signed ("agent_id", agent_list[i].handle); |
| 1637 | |
| 1638 | /* location */ |
| 1639 | uint32_t location_id; |
| 1640 | if ((status = amd_dbgapi_agent_get_info ( |
| 1641 | process_id, agent_list[i], AMD_DBGAPI_AGENT_INFO_PCIE_SLOT, |
| 1642 | sizeof (location_id), &location_id)) |
| 1643 | != AMD_DBGAPI_STATUS_SUCCESS) |
| 1644 | { |
| 1645 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) |
| 1646 | uiout->field_string ("location_id", "N/A"); |
| 1647 | else |
| 1648 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); |
| 1649 | } |
| 1650 | else |
| 1651 | uiout->field_string ( |
| 1652 | "location_id", |
| 1653 | string_printf ("%02x:%02x.%d", (location_id >> 8) & 0xFF, |
| 1654 | (location_id >> 3) & 0x1F, location_id & 0x7)); |
| 1655 | |
| 1656 | /* name */ |
| 1657 | uiout->field_string ("name", agent_names[i]); |
| 1658 | |
| 1659 | /* num_se, num_cu, simd, waves */ |
| 1660 | |
| 1661 | #define UIOUT_FIELD_INT(name, query) \ |
| 1662 | uint32_t name; \ |
| 1663 | if ((status = amd_dbgapi_agent_get_info (process_id, agent_list[i], query, \ |
| 1664 | sizeof (name), &name)) \ |
| 1665 | != AMD_DBGAPI_STATUS_SUCCESS) \ |
| 1666 | { \ |
| 1667 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) \ |
| 1668 | uiout->field_string (#name, "N/A"); \ |
| 1669 | else \ |
| 1670 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); \ |
| 1671 | } \ |
| 1672 | else \ |
| 1673 | uiout->field_signed (#name, name); |
| 1674 | |
| 1675 | UIOUT_FIELD_INT (num_se, AMD_DBGAPI_AGENT_INFO_SHADER_ENGINE_COUNT); |
| 1676 | UIOUT_FIELD_INT (num_cu, AMD_DBGAPI_AGENT_INFO_COMPUTE_UNIT_COUNT); |
| 1677 | UIOUT_FIELD_INT (simd, AMD_DBGAPI_AGENT_INFO_NUM_SIMD_PER_COMPUTE_UNIT); |
| 1678 | UIOUT_FIELD_INT (waves, AMD_DBGAPI_AGENT_INFO_MAX_WAVES_PER_SIMD); |
| 1679 | |
| 1680 | #undef UIOUT_FIELD_INT |
| 1681 | |
| 1682 | uiout->text ("\n"); |
| 1683 | } |
| 1684 | |
| 1685 | xfree (agent_list); |
| 1686 | gdb_flush (gdb_stdout); |
| 1687 | } |
| 1688 | |
| 1689 | /* -Wmissing-prototypes */ |
| 1690 | extern initialize_file_ftype _initialize_rocm_tdep; |
| 1691 | |
| 1692 | void |
| 1693 | _initialize_rocm_tdep (void) |
| 1694 | { |
| 1695 | /* Initialize the ROCm Debug API. */ |
| 1696 | if (amd_dbgapi_initialize (&dbgapi_callbacks) != AMD_DBGAPI_STATUS_SUCCESS) |
| 1697 | return; |
| 1698 | |
| 1699 | /* Set the initial log level. */ |
| 1700 | amd_dbgapi_set_log_level (get_debug_amd_dbgapi_log_level ()); |
| 1701 | |
| 1702 | /* Install observers. */ |
| 1703 | gdb::observers::breakpoint_created.attach (rocm_target_breakpoint_fixup); |
| 1704 | gdb::observers::solib_loaded.attach (rocm_target_solib_loaded); |
| 1705 | gdb::observers::solib_unloaded.attach (rocm_target_solib_unloaded); |
| 1706 | gdb::observers::inferior_created.attach (rocm_target_inferior_created); |
| 1707 | gdb::observers::inferior_exit.attach (rocm_target_inferior_exit); |
| 1708 | |
| 1709 | amd_dbgapi_activated.attach (rocm_target_dbgapi_activated); |
| 1710 | amd_dbgapi_deactivated.attach (rocm_target_dbgapi_deactivated); |
| 1711 | |
| 1712 | create_internalvar_type_lazy ("_wave_id", &rocm_wave_id_funcs, NULL); |
| 1713 | |
| 1714 | add_prefix_cmd ("amd-dbgapi", no_class, set_debug_amd_dbgapi, |
| 1715 | _ ("Generic command for setting amd-dbgapi debugging flags"), |
| 1716 | &set_debug_amd_dbgapi_list, "set debug amd-dbgapi ", 0, |
| 1717 | &setdebuglist); |
| 1718 | |
| 1719 | add_prefix_cmd ("amd-dbgapi", no_class, show_debug_amd_dbgapi, |
| 1720 | _ ("Generic command for showing amd-dbgapi debugging flags"), |
| 1721 | &show_debug_amd_dbgapi_list, "show debug amd-dbgapi ", 0, |
| 1722 | &showdebuglist); |
| 1723 | |
| 1724 | add_setshow_enum_cmd ( |
| 1725 | "log-level", class_maintenance, debug_amd_dbgapi_log_level_enums, |
| 1726 | &debug_amd_dbgapi_log_level, _ ("Set the amd-dbgapi log level."), |
| 1727 | _ ("Show the amd-dbgapi log level."), |
| 1728 | _ ("off == no logging is enabled\n" |
| 1729 | "error == fatal errors are reported\n" |
| 1730 | "warning == fatal errors and warnings are reported\n" |
| 1731 | "info == fatal errors, warnings, and info messages are reported\n" |
| 1732 | "verbose == all messages are reported"), |
| 1733 | set_debug_amd_dbgapi_log_level, show_debug_amd_dbgapi_log_level, |
| 1734 | &set_debug_amd_dbgapi_list, &show_debug_amd_dbgapi_list); |
| 1735 | |
| 1736 | add_cmd ("agents", class_info, info_agents_command, |
| 1737 | _ ("Info about currently active agents."), &infolist); |
| 1738 | } |