Commit | Line | Data |
---|---|---|
5922befa LM |
1 | /* Target-dependent code for ROCm. |
2 | ||
3 | Copyright (C) 2019 Free Software Foundation, Inc. | |
4 | Copyright (C) 2019 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; | |
80 | ||
81 | /* The amd_dbgapi_notifier_t for this inferior. */ | |
82 | amd_dbgapi_notifier_t notifier; | |
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 | 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 | inferior *inf = find_inferior_ptid (ptid); | |
354 | if (!inf) | |
355 | return false; | |
356 | ||
357 | amd_dbgapi_wave_state_t state; | |
358 | return amd_dbgapi_wave_get_info ( | |
359 | get_amd_dbgapi_process_id (inf), get_amd_dbgapi_wave_id (ptid), | |
360 | AMD_DBGAPI_WAVE_INFO_STATE, sizeof (state), &state) | |
361 | == AMD_DBGAPI_STATUS_SUCCESS; | |
362 | } | |
363 | ||
364 | const char * | |
365 | rocm_target_ops::thread_name (thread_info *tp) | |
366 | { | |
367 | if (!ptid_is_gpu (tp->ptid)) | |
368 | return beneath ()->thread_name (tp); | |
369 | ||
370 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (tp->inf); | |
371 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (tp->ptid); | |
372 | amd_dbgapi_dispatch_id_t dispatch_id; | |
373 | amd_dbgapi_global_address_t kernel_addr; | |
374 | ||
375 | if (amd_dbgapi_wave_get_info (process_id, wave_id, | |
376 | AMD_DBGAPI_WAVE_INFO_DISPATCH, | |
377 | sizeof (dispatch_id), &dispatch_id) | |
378 | != AMD_DBGAPI_STATUS_SUCCESS | |
379 | || amd_dbgapi_dispatch_get_info ( | |
380 | process_id, dispatch_id, | |
381 | AMD_DBGAPI_DISPATCH_INFO_KERNEL_ENTRY_ADDRESS, | |
382 | sizeof (kernel_addr), &kernel_addr) | |
383 | != AMD_DBGAPI_STATUS_SUCCESS) | |
384 | return NULL; | |
385 | ||
386 | struct bound_minimal_symbol msymbol | |
387 | = lookup_minimal_symbol_by_pc_section (kernel_addr, nullptr); | |
388 | ||
389 | if (msymbol.minsym != NULL) | |
390 | { | |
391 | static char buf[256]; | |
392 | char *endp; | |
393 | ||
394 | xsnprintf (buf, sizeof (buf), "%s", msymbol.minsym->print_name ()); | |
395 | ||
396 | /* Strip the arguments from the demangled function name. */ | |
397 | if ((endp = strchr (buf, '('))) | |
398 | *endp = '\0'; | |
399 | ||
400 | return buf; | |
401 | } | |
402 | ||
403 | return NULL; | |
404 | } | |
405 | ||
406 | std::string | |
407 | rocm_target_ops::pid_to_str (ptid_t ptid) | |
408 | { | |
409 | if (!ptid_is_gpu (ptid)) | |
410 | { | |
411 | return beneath ()->pid_to_str (ptid); | |
412 | } | |
413 | ||
414 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); | |
415 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (ptid); | |
416 | amd_dbgapi_dispatch_id_t dispatch_id; | |
417 | uint32_t group_ids[3], wave_in_group; | |
418 | ||
419 | if (amd_dbgapi_wave_get_info (process_id, wave_id, | |
420 | AMD_DBGAPI_WAVE_INFO_DISPATCH, | |
421 | sizeof (dispatch_id), &dispatch_id) | |
422 | != AMD_DBGAPI_STATUS_SUCCESS | |
423 | || amd_dbgapi_wave_get_info (process_id, wave_id, | |
424 | AMD_DBGAPI_WAVE_INFO_WORK_GROUP_COORD, | |
425 | sizeof (group_ids), &group_ids) | |
426 | != AMD_DBGAPI_STATUS_SUCCESS | |
427 | || amd_dbgapi_wave_get_info ( | |
428 | process_id, wave_id, | |
429 | AMD_DBGAPI_WAVE_INFO_WAVE_NUMBER_IN_WORK_GROUP, | |
430 | sizeof (wave_in_group), &wave_in_group) | |
431 | != AMD_DBGAPI_STATUS_SUCCESS) | |
66388092 | 432 | return std::string ("AMDGPU Thread"); |
5922befa | 433 | else |
66388092 | 434 | return string_printf ("AMDGPU Thread %ld.%ld (%d,%d,%d)/%d", |
5922befa LM |
435 | dispatch_id.handle, wave_id.handle, group_ids[2], |
436 | group_ids[1], group_ids[0], wave_in_group); | |
437 | } | |
438 | ||
439 | const char * | |
440 | rocm_target_ops::extra_thread_info (thread_info *tp) | |
441 | { | |
442 | if (!ptid_is_gpu (tp->ptid)) | |
443 | beneath ()->extra_thread_info (tp); | |
444 | ||
445 | return NULL; | |
446 | } | |
447 | ||
448 | enum target_xfer_status | |
449 | rocm_target_ops::xfer_partial (enum target_object object, const char *annex, | |
450 | gdb_byte *readbuf, const gdb_byte *writebuf, | |
451 | ULONGEST offset, ULONGEST requested_len, | |
452 | ULONGEST *xfered_len) | |
453 | { | |
454 | gdb::optional<scoped_restore_current_thread> maybe_restore_thread; | |
455 | ||
456 | if (ptid_is_gpu (inferior_ptid)) | |
457 | { | |
458 | gdb_assert (requested_len && xfered_len && "checking invariants"); | |
459 | ||
460 | if (object != TARGET_OBJECT_MEMORY) | |
461 | return TARGET_XFER_E_IO; | |
462 | ||
463 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); | |
464 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (inferior_ptid); | |
465 | ||
466 | amd_dbgapi_architecture_id_t architecture_id; | |
467 | amd_dbgapi_address_space_id_t address_space_id; | |
468 | ||
469 | if (amd_dbgapi_wave_get_info (process_id, wave_id, | |
470 | AMD_DBGAPI_WAVE_INFO_ARCHITECTURE, | |
471 | sizeof (architecture_id), &architecture_id) | |
472 | != AMD_DBGAPI_STATUS_SUCCESS | |
473 | || amd_dbgapi_architecture_get_info ( | |
474 | architecture_id, | |
475 | AMD_DBGAPI_ARCHITECTURE_INFO_DEFAULT_GLOBAL_ADDRESS_SPACE, | |
476 | sizeof (address_space_id), &address_space_id) | |
477 | != AMD_DBGAPI_STATUS_SUCCESS) | |
478 | error (_ ("Couldn't get default global address space")); | |
479 | ||
480 | size_t len = requested_len; | |
481 | amd_dbgapi_status_t status; | |
482 | ||
483 | if (readbuf) | |
484 | status = amd_dbgapi_read_memory ( | |
485 | process_id, wave_id, 0, address_space_id, offset, &len, readbuf); | |
486 | else | |
487 | status = amd_dbgapi_write_memory ( | |
488 | process_id, wave_id, 0, address_space_id, offset, &len, writebuf); | |
489 | ||
490 | if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
491 | { | |
492 | if (status != AMD_DBGAPI_STATUS_ERROR_MEMORY_ACCESS) | |
493 | error (_ ("amd_dbgapi_%s_memory failed (rc=%d"), | |
494 | readbuf ? "read" : "write", status); | |
495 | return TARGET_XFER_EOF; | |
496 | } | |
497 | ||
498 | *xfered_len = len; | |
499 | return TARGET_XFER_OK; | |
500 | } | |
501 | else | |
502 | return beneath ()->xfer_partial (object, annex, readbuf, writebuf, offset, | |
503 | requested_len, xfered_len); | |
504 | } | |
505 | ||
506 | void | |
507 | rocm_target_ops::resume (ptid_t ptid, int step, enum gdb_signal signo) | |
508 | { | |
509 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
510 | ||
511 | if (debug_infrun) | |
512 | fprintf_unfiltered ( | |
513 | gdb_stdlog, | |
514 | "\e[1;34minfrun: rocm_target_ops::resume ([%d,%ld,%ld])\e[0m\n", | |
515 | ptid.pid (), ptid.lwp (), ptid.tid ()); | |
516 | ||
517 | /* Check if the thread focus is on the GPU device. */ | |
518 | if (ptid == minus_one_ptid || !ptid_is_gpu (ptid)) | |
519 | { | |
520 | beneath ()->resume (ptid, step, signo); | |
521 | if (ptid != minus_one_ptid) | |
522 | return; | |
523 | } | |
524 | ||
525 | /* A specific PTID means `step only this process id'. */ | |
526 | bool resume_one = ptid != minus_one_ptid && !ptid.is_pid (); | |
527 | gdb_assert (resume_one || !step); | |
528 | ||
529 | if (!resume_one) | |
530 | error (_ ("internal error - unimplemented ")); | |
531 | ||
532 | amd_dbgapi_process_set_progress (info->process_id, | |
533 | AMD_DBGAPI_PROGRESS_NO_FORWARD); | |
534 | ||
535 | amd_dbgapi_status_t status = amd_dbgapi_wave_resume ( | |
536 | info->process_id, get_amd_dbgapi_wave_id (ptid), | |
537 | step ? AMD_DBGAPI_RESUME_MODE_SINGLE_STEP | |
538 | : AMD_DBGAPI_RESUME_MODE_NORMAL); | |
539 | if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
540 | warning (_ ("Could not resume %s (rc=%d)"), | |
541 | target_pid_to_str (ptid).c_str (), status); | |
542 | ||
543 | info->commit_resume_all_start = true; | |
544 | } | |
545 | ||
546 | void | |
547 | rocm_target_ops::commit_resume () | |
548 | { | |
549 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
550 | ||
551 | if (debug_infrun) | |
552 | fprintf_unfiltered ( | |
553 | gdb_stdlog, | |
554 | "\e[1;34minfrun: rocm_target_ops::commit_resume ()\e[0m\n"); | |
555 | ||
556 | beneath ()->commit_resume (); | |
557 | ||
558 | if (info->commit_resume_all_start) | |
559 | { | |
560 | amd_dbgapi_process_set_progress (info->process_id, | |
561 | AMD_DBGAPI_PROGRESS_NORMAL); | |
562 | info->commit_resume_all_start = false; | |
563 | } | |
564 | ||
565 | if (target_can_async_p ()) | |
566 | target_async (1); | |
567 | } | |
568 | ||
569 | static void | |
570 | rocm_target_stop_one_wave (ptid_t ptid) | |
571 | { | |
572 | amd_dbgapi_status_t status; | |
573 | ||
574 | status = amd_dbgapi_wave_stop (get_amd_dbgapi_process_id (), | |
575 | get_amd_dbgapi_wave_id (ptid)); | |
576 | ||
577 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID) | |
578 | { | |
579 | /* the wave must have exited, set the thread status to reflect that. */ | |
580 | auto *tp = find_thread_ptid (ptid); | |
581 | gdb_assert (tp); | |
582 | ||
583 | tp->state = THREAD_EXITED; | |
584 | } | |
585 | else if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
586 | warning (_ ("Could not stop %s (rc=%d)"), | |
587 | target_pid_to_str (ptid).c_str (), status); | |
588 | } | |
589 | ||
590 | void | |
591 | rocm_target_ops::stop (ptid_t ptid) | |
592 | { | |
593 | if (debug_infrun) | |
594 | fprintf_unfiltered ( | |
595 | gdb_stdlog, | |
596 | "\e[1;34minfrun: rocm_target_ops::stop ([%d,%ld,%ld])\e[0m\n", | |
597 | ptid.pid (), ptid.lwp (), ptid.tid ()); | |
598 | ||
599 | if (ptid == minus_one_ptid || !ptid_is_gpu (ptid)) | |
600 | { | |
601 | beneath ()->stop (ptid); | |
602 | if (ptid != minus_one_ptid) | |
603 | return; | |
604 | } | |
605 | ||
606 | if (ptid == minus_one_ptid) | |
607 | error (_ ("internal error - unimplemented ")); | |
608 | ||
609 | rocm_target_stop_one_wave (ptid); | |
610 | } | |
611 | ||
612 | static void | |
613 | handle_target_event (int error, gdb_client_data client_data) | |
614 | { | |
615 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
616 | amd_dbgapi_process_id_t process_id = info->process_id; | |
617 | ||
618 | amd_dbgapi_process_set_progress (process_id, AMD_DBGAPI_PROGRESS_NO_FORWARD); | |
619 | ||
620 | /* Flush the async file first. */ | |
621 | if (target_is_async_p ()) | |
622 | async_file_flush (); | |
623 | ||
624 | rocm_process_event_queue (); | |
625 | ||
626 | /* In all-stop mode, unless the event queue is empty (spurious wake-up), | |
627 | we can keep the process in progress_no_forward mode. The infrun loop | |
628 | will enable forward progress when a thread is resumed. */ | |
629 | if (non_stop || info->wave_stop_events.empty ()) | |
630 | amd_dbgapi_process_set_progress (process_id, AMD_DBGAPI_PROGRESS_NORMAL); | |
631 | ||
632 | if (!info->wave_stop_events.empty ()) | |
633 | inferior_event_handler (INF_REG_EVENT, nullptr); | |
634 | } | |
635 | ||
636 | void | |
637 | rocm_target_ops::async (int enable) | |
638 | { | |
639 | beneath ()->async (enable); | |
640 | ||
641 | if (enable) | |
642 | { | |
643 | if (rocm_event_pipe[0] != -1) | |
644 | return; | |
645 | ||
646 | if (gdb_pipe_cloexec (rocm_event_pipe) == -1) | |
647 | internal_error (__FILE__, __LINE__, "creating event pipe failed."); | |
648 | ||
649 | ::fcntl (rocm_event_pipe[0], F_SETFL, O_NONBLOCK); | |
650 | ::fcntl (rocm_event_pipe[1], F_SETFL, O_NONBLOCK); | |
651 | ||
652 | add_file_handler (rocm_event_pipe[0], handle_target_event, nullptr); | |
653 | ||
654 | /* There may be pending events to handle. Tell the event loop | |
655 | to poll them. */ | |
656 | async_file_mark (); | |
657 | } | |
658 | else | |
659 | { | |
660 | delete_file_handler (rocm_event_pipe[0]); | |
661 | ||
662 | if (rocm_event_pipe[0] == -1) | |
663 | return; | |
664 | ||
665 | ::close (rocm_event_pipe[0]); | |
666 | ::close (rocm_event_pipe[1]); | |
667 | rocm_event_pipe[0] = -1; | |
668 | rocm_event_pipe[1] = -1; | |
669 | } | |
670 | } | |
671 | ||
672 | static void | |
673 | rocm_process_one_event (amd_dbgapi_event_id_t event_id, | |
674 | amd_dbgapi_event_kind_t event_kind) | |
675 | { | |
676 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
677 | amd_dbgapi_process_id_t process_id = info->process_id; | |
678 | amd_dbgapi_status_t status; | |
679 | ||
680 | switch (event_kind) | |
681 | { | |
682 | case AMD_DBGAPI_EVENT_KIND_WAVE_STOP: | |
683 | { | |
684 | amd_dbgapi_wave_id_t wave_id; | |
685 | if ((status = amd_dbgapi_event_get_info (process_id, event_id, | |
686 | AMD_DBGAPI_EVENT_INFO_WAVE, | |
687 | sizeof (wave_id), &wave_id)) | |
688 | != AMD_DBGAPI_STATUS_SUCCESS) | |
689 | error (_ ("event_get_info for event_%ld failed (rc=%d)"), | |
690 | event_id.handle, status); | |
691 | ||
692 | amd_dbgapi_wave_stop_reason_t stop_reason; | |
693 | status = amd_dbgapi_wave_get_info (process_id, wave_id, | |
694 | AMD_DBGAPI_WAVE_INFO_STOP_REASON, | |
695 | sizeof (stop_reason), &stop_reason); | |
696 | ||
697 | if (status != AMD_DBGAPI_STATUS_SUCCESS | |
698 | && status != AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID) | |
699 | error (_ ("wave_get_info for wave_%ld failed (rc=%d)"), | |
700 | wave_id.handle, status); | |
701 | ||
702 | /* The wave may have exited, or the queue went into an error | |
703 | state. In such cases, we will see another wave command | |
704 | terminated event, and handle the wave termination then. */ | |
705 | ||
706 | if (status == AMD_DBGAPI_STATUS_SUCCESS) | |
707 | info->wave_stop_events.emplace_back ( | |
708 | std::make_pair (wave_id, stop_reason)); | |
709 | } | |
710 | break; | |
711 | ||
712 | case AMD_DBGAPI_EVENT_KIND_CODE_OBJECT_LIST_UPDATED: | |
713 | amd_dbgapi_code_object_list_updated.notify (); | |
714 | break; | |
715 | ||
716 | case AMD_DBGAPI_EVENT_KIND_BREAKPOINT_RESUME: | |
717 | /* Breakpoint resume events should be handled by the breakpoint | |
718 | action, and this code should not reach this. */ | |
719 | gdb_assert_not_reached (_ ("unhandled event kind")); | |
720 | break; | |
721 | ||
722 | case AMD_DBGAPI_EVENT_KIND_RUNTIME: | |
723 | { | |
724 | amd_dbgapi_runtime_state_t runtime_state; | |
725 | ||
726 | if ((status = amd_dbgapi_event_get_info ( | |
727 | process_id, event_id, AMD_DBGAPI_EVENT_INFO_RUNTIME_STATE, | |
728 | sizeof (runtime_state), &runtime_state)) | |
729 | != AMD_DBGAPI_STATUS_SUCCESS) | |
730 | error (_ ("event_get_info for event_%ld failed (rc=%d)"), | |
731 | event_id.handle, status); | |
732 | ||
733 | switch (runtime_state) | |
734 | { | |
735 | case AMD_DBGAPI_RUNTIME_STATE_LOADED_SUPPORTED: | |
736 | amd_dbgapi_activated.notify (); | |
737 | break; | |
738 | case AMD_DBGAPI_RUNTIME_STATE_LOADED_UNSUPPORTED: | |
739 | warning (_ ("ROCm-GDB: low-level runtime version not supported")); | |
740 | break; | |
741 | ||
742 | case AMD_DBGAPI_RUNTIME_STATE_UNLOADED: | |
743 | amd_dbgapi_deactivated.notify (); | |
744 | break; | |
745 | } | |
746 | } | |
747 | break; | |
748 | ||
749 | default: | |
750 | error (_ ("event kind (%d) not supported"), event_kind); | |
751 | } | |
752 | ||
753 | amd_dbgapi_event_processed (process_id, event_id); | |
754 | } | |
755 | ||
756 | /* Drain the amd_dbgapi event queue until an event of the given type is seen. | |
757 | If no particular event kind is specified (AMD_DBGAPI_EVENT_KIND_NONE), the | |
758 | event queue is completely drained. Wave stop events that are not returned | |
759 | are re-queued into the current's process pending wave events. */ | |
760 | static amd_dbgapi_event_id_t | |
761 | rocm_process_event_queue (amd_dbgapi_event_kind_t until_event_kind) | |
762 | { | |
763 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
764 | ||
765 | while (true) | |
766 | { | |
767 | amd_dbgapi_event_id_t event_id; | |
768 | amd_dbgapi_event_kind_t event_kind; | |
769 | ||
770 | amd_dbgapi_status_t status = amd_dbgapi_next_pending_event ( | |
771 | info->process_id, &event_id, &event_kind); | |
772 | ||
773 | if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
774 | error (_ ("next_pending_event failed (rc=%d)"), status); | |
775 | ||
776 | if (event_id.handle == AMD_DBGAPI_EVENT_NONE.handle | |
777 | || event_kind == until_event_kind) | |
778 | return event_id; | |
779 | ||
780 | rocm_process_one_event (event_id, event_kind); | |
781 | } | |
782 | } | |
783 | ||
784 | ptid_t | |
785 | rocm_target_ops::wait (ptid_t ptid, struct target_waitstatus *ws, | |
786 | int target_options) | |
787 | { | |
788 | if (debug_infrun) | |
789 | fprintf_unfiltered (gdb_stdlog, | |
790 | "\e[1;34minfrun: rocm_target_ops::wait\e[0m\n"); | |
791 | ||
792 | if (!ptid_is_gpu (ptid)) | |
793 | { | |
794 | ptid_t event_ptid = beneath ()->wait (ptid, ws, target_options); | |
795 | if (event_ptid != minus_one_ptid) | |
796 | return event_ptid; | |
797 | } | |
798 | ||
799 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
800 | amd_dbgapi_process_id_t process_id = info->process_id; | |
801 | ||
802 | /* Drain all the events from the amd_dbgapi, and preserve the ordering. */ | |
803 | if (info->wave_stop_events.empty ()) | |
804 | { | |
805 | amd_dbgapi_process_set_progress (process_id, | |
806 | AMD_DBGAPI_PROGRESS_NO_FORWARD); | |
807 | ||
808 | /* Flush the async file first. */ | |
809 | if (target_is_async_p ()) | |
810 | async_file_flush (); | |
811 | ||
812 | rocm_process_event_queue (); | |
813 | ||
814 | /* In all-stop mode, unless the event queue is empty (spurious wake-up), | |
815 | we can keep the process in progress_no_forward mode. The infrun loop | |
816 | will enable forward progress when a thread is resumed. */ | |
817 | if (non_stop || info->wave_stop_events.empty ()) | |
818 | amd_dbgapi_process_set_progress (process_id, | |
819 | AMD_DBGAPI_PROGRESS_NORMAL); | |
820 | } | |
821 | ||
822 | if (info->wave_stop_events.empty ()) | |
823 | return minus_one_ptid; | |
824 | ||
825 | amd_dbgapi_wave_id_t event_wave_id; | |
826 | amd_dbgapi_wave_stop_reason_t stop_reason; | |
827 | ||
828 | std::tie (event_wave_id, stop_reason) = info->wave_stop_events.front (); | |
829 | info->wave_stop_events.pop_front (); | |
830 | ||
831 | ptid_t event_ptid (current_inferior ()->pid, 1, event_wave_id.handle); | |
832 | ||
833 | if (!find_thread_ptid (event_ptid)) | |
834 | { | |
835 | add_thread_silent (event_ptid); | |
836 | set_running (event_ptid, 1); | |
837 | set_executing (event_ptid, 1); | |
838 | } | |
839 | ||
840 | /* Since we are manipulating the register cache for the event thread, | |
841 | make sure it is the current thread. */ | |
842 | switch_to_thread (event_ptid); | |
843 | ||
844 | /* By caching the PC now, we avoid having to suspend/resume the queue | |
845 | later when we need to access it. */ | |
846 | amd_dbgapi_global_address_t stop_pc; | |
847 | if (amd_dbgapi_wave_get_info (process_id, event_wave_id, | |
848 | AMD_DBGAPI_WAVE_INFO_PC, sizeof (stop_pc), | |
849 | &stop_pc) | |
850 | == AMD_DBGAPI_STATUS_SUCCESS) | |
851 | { | |
852 | struct regcache *regcache = get_thread_regcache (event_ptid); | |
853 | regcache->raw_supply (gdbarch_pc_regnum (regcache->arch ()), &stop_pc); | |
854 | } | |
855 | ws->kind = TARGET_WAITKIND_STOPPED; | |
856 | ||
857 | if (stop_reason | |
858 | & (AMD_DBGAPI_WAVE_STOP_REASON_BREAKPOINT | |
859 | | AMD_DBGAPI_WAVE_STOP_REASON_SINGLE_STEP)) | |
860 | ws->value.sig = GDB_SIGNAL_TRAP; | |
861 | else if (stop_reason & AMD_DBGAPI_WAVE_STOP_REASON_MEMORY_VIOLATION) | |
862 | ws->value.sig = GDB_SIGNAL_SEGV; | |
863 | else if (stop_reason | |
864 | & (AMD_DBGAPI_WAVE_STOP_REASON_FP_INPUT_DENORMAL | |
865 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_DIVIDE_BY_0 | |
866 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_OVERFLOW | |
867 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_UNDERFLOW | |
868 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_INEXACT | |
869 | | AMD_DBGAPI_WAVE_STOP_REASON_FP_INVALID_OPERATION | |
870 | | AMD_DBGAPI_WAVE_STOP_REASON_INT_DIVIDE_BY_0)) | |
871 | ws->value.sig = GDB_SIGNAL_FPE; | |
872 | else | |
873 | ws->value.sig = GDB_SIGNAL_0; | |
874 | ||
875 | /* If there are more events in the list, mark the async file so that | |
876 | rocm_target_ops::wait gets called again. */ | |
877 | if (target_is_async_p () && !info->wave_stop_events.empty ()) | |
878 | async_file_mark (); | |
879 | ||
880 | return event_ptid; | |
881 | } | |
882 | ||
883 | bool | |
884 | rocm_target_ops::stopped_by_sw_breakpoint () | |
885 | { | |
886 | if (!ptid_is_gpu (inferior_ptid)) | |
887 | return beneath ()->supports_stopped_by_sw_breakpoint () | |
888 | && beneath ()->stopped_by_sw_breakpoint (); | |
889 | ||
890 | /* FIXME: we should check that the wave is not single-stepping. */ | |
891 | ||
892 | struct regcache *regcache = get_thread_regcache (inferior_ptid); | |
893 | ||
894 | CORE_ADDR bkpt_pc = regcache_read_pc (regcache) | |
895 | - gdbarch_decr_pc_after_break (regcache->arch ()); | |
896 | ||
897 | return software_breakpoint_inserted_here_p (regcache->aspace (), bkpt_pc); | |
898 | } | |
899 | ||
900 | void | |
901 | rocm_target_ops::mourn_inferior () | |
902 | { | |
903 | /* FIXME: only unpush on the last activation. */ | |
904 | /* Disengage the ROCm target_ops. */ | |
905 | unpush_target (&rocm_ops); | |
906 | ||
907 | beneath ()->mourn_inferior (); | |
908 | } | |
909 | ||
910 | void | |
911 | rocm_target_ops::fetch_registers (struct regcache *regcache, int regno) | |
912 | { | |
913 | struct gdbarch *gdbarch = regcache->arch (); | |
914 | struct gdbarch_tdep *tdep = gdbarch_tdep (gdbarch); | |
915 | ||
916 | /* delegate to the host routines when not on the device */ | |
917 | ||
918 | if (!rocm_is_amdgcn_gdbarch (gdbarch)) | |
919 | { | |
920 | beneath ()->fetch_registers (regcache, regno); | |
921 | return; | |
922 | } | |
923 | ||
924 | inferior *inf = find_inferior_ptid (regcache->ptid ()); | |
925 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (inf); | |
926 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (regcache->ptid ()); | |
927 | ||
928 | gdb_byte raw[AMDGCN_MAX_REGISTER_SIZE]; | |
929 | ||
930 | amd_dbgapi_status_t status = amd_dbgapi_read_register ( | |
931 | process_id, wave_id, tdep->register_ids[regno], 0, | |
932 | TYPE_LENGTH (register_type (gdbarch, regno)), raw); | |
933 | ||
934 | if (status == AMD_DBGAPI_STATUS_SUCCESS) | |
935 | { | |
936 | regcache->raw_supply (regno, raw); | |
937 | } | |
938 | else if (status != AMD_DBGAPI_STATUS_ERROR_INVALID_REGISTER_ID) | |
939 | { | |
940 | warning (_ ("Couldn't read register %s (#%d)."), | |
941 | gdbarch_register_name (gdbarch, regno), regno); | |
942 | } | |
943 | } | |
944 | ||
945 | void | |
946 | rocm_target_ops::store_registers (struct regcache *regcache, int regno) | |
947 | { | |
948 | struct gdbarch *gdbarch = regcache->arch (); | |
949 | struct gdbarch_tdep *tdep = gdbarch_tdep (gdbarch); | |
950 | gdb_byte raw[AMDGCN_MAX_REGISTER_SIZE]; | |
951 | ||
952 | if (!rocm_is_amdgcn_gdbarch (gdbarch)) | |
953 | { | |
954 | beneath ()->store_registers (regcache, regno); | |
955 | return; | |
956 | } | |
957 | ||
958 | inferior *inf = find_inferior_ptid (regcache->ptid ()); | |
959 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (inf); | |
960 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (regcache->ptid ()); | |
961 | ||
962 | regcache->raw_collect (regno, &raw); | |
963 | ||
964 | amd_dbgapi_status_t status = amd_dbgapi_write_register ( | |
965 | process_id, wave_id, tdep->register_ids[regno], 0, | |
966 | TYPE_LENGTH (register_type (gdbarch, regno)), raw); | |
967 | ||
968 | if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
969 | { | |
970 | warning (_ ("Couldn't write register %s (#%d)."), | |
971 | gdbarch_register_name (gdbarch, regno), regno); | |
972 | } | |
973 | } | |
974 | ||
975 | /* Fix breakpoints created with an address location while the | |
976 | architecture was set to the host (could be fixed in core GDB). */ | |
977 | ||
978 | static void | |
979 | rocm_target_breakpoint_fixup (struct breakpoint *b) | |
980 | { | |
981 | if (b->location.get () | |
982 | && event_location_type (b->location.get ()) == ADDRESS_LOCATION | |
983 | && gdbarch_bfd_arch_info (b->loc->gdbarch)->arch == bfd_arch_amdgcn | |
984 | && gdbarch_bfd_arch_info (b->gdbarch)->arch != bfd_arch_amdgcn) | |
985 | { | |
986 | b->gdbarch = b->loc->gdbarch; | |
987 | } | |
988 | } | |
989 | ||
990 | struct gdbarch * | |
991 | rocm_target_ops::thread_architecture (ptid_t ptid) | |
992 | { | |
993 | static std::result_of<decltype (&ptid_t::tid) (ptid_t)>::type last_tid = 0; | |
994 | static struct gdbarch *cached_arch = nullptr; | |
995 | ||
996 | if (!ptid_is_gpu (ptid)) | |
997 | return beneath ()->thread_architecture (ptid); | |
998 | ||
999 | /* We can cache the gdbarch for a given wave_id (ptid::tid) because | |
1000 | wave IDs are unique, and aren't reused. */ | |
1001 | if (ptid.tid () == last_tid) | |
1002 | return cached_arch; | |
1003 | ||
1004 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); | |
1005 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (ptid); | |
1006 | amd_dbgapi_architecture_id_t architecture_id; | |
1007 | ||
1008 | if (amd_dbgapi_wave_get_info (process_id, wave_id, | |
1009 | AMD_DBGAPI_WAVE_INFO_ARCHITECTURE, | |
1010 | sizeof (architecture_id), &architecture_id) | |
1011 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1012 | error (_ ("Couldn't get architecture for wave_%ld"), ptid.tid ()); | |
1013 | ||
1014 | uint32_t elf_amdgpu_machine; | |
1015 | if (amd_dbgapi_architecture_get_info ( | |
1016 | architecture_id, AMD_DBGAPI_ARCHITECTURE_INFO_ELF_AMDGPU_MACHINE, | |
1017 | sizeof (elf_amdgpu_machine), &elf_amdgpu_machine) | |
1018 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1019 | error (_ ("Couldn't get elf_amdgpu_machine for architecture_%ld"), | |
1020 | architecture_id.handle); | |
1021 | ||
1022 | struct gdbarch_info info; | |
1023 | gdbarch_info_init (&info); | |
1024 | ||
1025 | info.bfd_arch_info = bfd_lookup_arch (bfd_arch_amdgcn, elf_amdgpu_machine); | |
1026 | info.byte_order = BFD_ENDIAN_LITTLE; | |
1027 | info.osabi = GDB_OSABI_AMDGPU_HSA; | |
1028 | ||
1029 | last_tid = ptid.tid (); | |
1030 | if (!(cached_arch = gdbarch_find_by_info (info))) | |
1031 | error (_ ("Couldn't get elf_amdgpu_machine (%#x)"), elf_amdgpu_machine); | |
1032 | ||
1033 | return cached_arch; | |
1034 | } | |
1035 | ||
1036 | void | |
1037 | rocm_target_ops::update_thread_list () | |
1038 | { | |
1039 | for (inferior *inf : all_inferiors ()) | |
1040 | { | |
1041 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (inf); | |
1042 | amd_dbgapi_wave_id_t *wave_list; | |
1043 | size_t count; | |
1044 | ||
1045 | amd_dbgapi_changed_t changed; | |
1046 | amd_dbgapi_status_t status; | |
1047 | if ((status | |
1048 | = amd_dbgapi_wave_list (process_id, &count, &wave_list, &changed)) | |
1049 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1050 | error (_ ("amd_dbgapi_wave_list failed (rc=%d"), status); | |
1051 | ||
1052 | if (changed == AMD_DBGAPI_CHANGED_NO) | |
1053 | continue; | |
1054 | ||
1055 | /* Create a set and free the wave list. */ | |
1056 | std::set<std::result_of<decltype (&ptid_t::tid) (ptid_t)>::type> threads; | |
1057 | for (size_t i = 0; i < count; ++i) | |
1058 | threads.emplace (wave_list[i].handle); | |
1059 | xfree (wave_list); | |
1060 | ||
1061 | /* Then prune the wave_ids that already have a thread_info. */ | |
1062 | for (thread_info *tp : inf->non_exited_threads ()) | |
1063 | if (ptid_is_gpu (tp->ptid)) | |
1064 | threads.erase (tp->ptid.tid ()); | |
1065 | ||
1066 | /* The wave_ids that are left require a new thread_info. */ | |
1067 | for (auto &&tid : threads) | |
1068 | { | |
1069 | ptid_t wave_ptid (inf->pid, 1, tid); | |
1070 | amd_dbgapi_wave_state_t state; | |
1071 | ||
1072 | if (amd_dbgapi_wave_get_info ( | |
1073 | process_id, get_amd_dbgapi_wave_id (wave_ptid), | |
1074 | AMD_DBGAPI_WAVE_INFO_STATE, sizeof (state), &state) | |
1075 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1076 | continue; | |
1077 | ||
1078 | add_thread_silent (wave_ptid); | |
1079 | set_running (wave_ptid, 1); | |
1080 | set_executing (wave_ptid, 1); | |
1081 | } | |
1082 | } | |
1083 | ||
1084 | /* Give the beneath target a chance to do extra processing. */ | |
1085 | this->beneath ()->update_thread_list (); | |
1086 | } | |
1087 | ||
1088 | static void | |
1089 | rocm_target_solib_loaded (struct so_list *solib) | |
1090 | { | |
1091 | /* Notify the amd_dbgapi that a shared library has been loaded. */ | |
1092 | for (auto &&value : get_rocm_inferior_info ()->notify_solib_map) | |
1093 | /* TODO: If we want to support file name wildcards, change this code. */ | |
1094 | if (::strstr (solib->so_original_name, value.second.compare.c_str ()) | |
1095 | && !value.second.is_loaded) | |
1096 | { | |
1097 | value.second.solib = solib; | |
1098 | value.second.is_loaded = true; | |
1099 | ||
1100 | amd_dbgapi_report_shared_library ( | |
1101 | get_amd_dbgapi_process_id (), | |
1102 | amd_dbgapi_shared_library_id_t{ value.first }, | |
1103 | AMD_DBGAPI_SHARED_LIBRARY_STATE_LOADED); | |
1104 | } | |
1105 | } | |
1106 | ||
1107 | static void | |
1108 | rocm_target_solib_unloaded (struct so_list *solib) | |
1109 | { | |
1110 | /* Notify the amd_dbgapi that a shared library will unload. */ | |
1111 | for (auto &&value : get_rocm_inferior_info ()->notify_solib_map) | |
1112 | /* TODO: If we want to support file name wildcards, change this code. */ | |
1113 | if (::strstr (solib->so_original_name, value.second.compare.c_str ()) | |
1114 | && value.second.is_loaded) | |
1115 | { | |
1116 | struct rocm_inferior_info *info = get_rocm_inferior_info (); | |
1117 | ||
1118 | amd_dbgapi_report_shared_library ( | |
1119 | info->process_id, amd_dbgapi_shared_library_id_t{ value.first }, | |
1120 | AMD_DBGAPI_SHARED_LIBRARY_STATE_UNLOADED); | |
1121 | ||
1122 | /* Delete breakpoints that were left inserted in this shared library. | |
1123 | */ | |
1124 | for (auto it = info->breakpoint_map.begin (); | |
1125 | it != info->breakpoint_map.end ();) | |
1126 | if (solib_contains_address_p (solib, it->second->loc->address)) | |
1127 | { | |
1128 | warning (_ ("breakpoint_%ld is still inserted after " | |
1129 | "shared_library_%ld was unloaded"), | |
1130 | it->first, value.first); | |
1131 | delete_breakpoint (it->second); | |
1132 | it = info->breakpoint_map.erase (it); | |
1133 | } | |
1134 | else | |
1135 | ++it; | |
1136 | ||
1137 | value.second.solib = nullptr; | |
1138 | value.second.is_loaded = false; | |
1139 | } | |
1140 | } | |
1141 | ||
1142 | static void | |
1143 | rocm_target_inferior_created (struct target_ops *target, int from_tty) | |
1144 | { | |
1145 | struct inferior *inf = current_inferior (); | |
1146 | auto *info = get_rocm_inferior_info (inf); | |
1147 | amd_dbgapi_status_t status; | |
1148 | ||
1149 | if (!target_can_async_p ()) | |
1150 | { | |
1151 | warning ( | |
1152 | _ ("ROCm-GDB requires target-async, GPU debugging is disabled")); | |
1153 | return; | |
1154 | } | |
1155 | ||
1156 | gdb_assert (info->wave_stop_events.empty ()); | |
1157 | ||
1158 | status = amd_dbgapi_process_attach (inf, &info->process_id); | |
1159 | ||
1160 | if (status == AMD_DBGAPI_STATUS_ERROR_VERSION_MISMATCH) | |
1161 | warning (_ ("The version of the kernel driver does not match the version " | |
1162 | "required by the ROCm debugger library")); | |
1163 | ||
1164 | if (status != AMD_DBGAPI_STATUS_SUCCESS) | |
1165 | { | |
1166 | warning (_ ("Could not attach to process %d"), inf->pid); | |
1167 | return; | |
1168 | } | |
1169 | ||
1170 | if (amd_dbgapi_process_get_info (info->process_id, | |
1171 | AMD_DBGAPI_PROCESS_INFO_NOTIFIER, | |
1172 | sizeof (info->notifier), &info->notifier) | |
1173 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1174 | { | |
1175 | warning (_ ("Could not retrieve process %d's notifier"), inf->pid); | |
1176 | amd_dbgapi_process_detach (info->process_id); | |
1177 | return; | |
1178 | } | |
1179 | ||
1180 | /* We add a file handler for events returned by the debugger api. We'll use | |
1181 | this handler to signal our async handler that events are available. */ | |
1182 | add_file_handler ( | |
1183 | info->notifier, | |
1184 | [] (int error, gdb_client_data client_data) { | |
1185 | auto info_ = static_cast<struct rocm_inferior_info *> (client_data); | |
1186 | int ret; | |
1187 | ||
1188 | /* Drain the notifier pipe. */ | |
1189 | do | |
1190 | { | |
1191 | char buf; | |
1192 | ret = read (info_->notifier, &buf, 1); | |
1193 | } | |
1194 | while (ret >= 0 || (ret == -1 && errno == EINTR)); | |
1195 | ||
1196 | /* Signal our async handler. */ | |
1197 | async_file_mark (); | |
1198 | }, | |
1199 | info); | |
1200 | ||
1201 | /* Attaching to the inferior may have generated runtime events, process | |
1202 | them now. */ | |
1203 | rocm_process_event_queue (); | |
1204 | } | |
1205 | ||
1206 | static void | |
1207 | rocm_target_inferior_exit (struct inferior *inf) | |
1208 | { | |
1209 | auto *info = get_rocm_inferior_info (inf); | |
1210 | ||
1211 | amd_dbgapi_deactivated.notify (); | |
1212 | ||
1213 | delete_file_handler (info->notifier); | |
1214 | ||
1215 | amd_dbgapi_process_detach (info->process_id); | |
1216 | ||
1217 | /* Delete the breakpoints that are still active. */ | |
1218 | for (auto &&value : info->breakpoint_map) | |
1219 | delete_breakpoint (value.second); | |
1220 | ||
1221 | rocm_inferior_data.clear (inf); | |
1222 | } | |
1223 | ||
1224 | static cli_style_option warning_style ("rocm_warning", ui_file_style::RED); | |
1225 | static cli_style_option info_style ("rocm_info", ui_file_style::GREEN); | |
1226 | static cli_style_option verbose_style ("rocm_verbose", ui_file_style::BLUE); | |
1227 | ||
1228 | static amd_dbgapi_callbacks_t dbgapi_callbacks = { | |
1229 | /* allocate_memory. */ | |
1230 | .allocate_memory = xmalloc, | |
1231 | ||
1232 | /* deallocate_memory. */ | |
1233 | .deallocate_memory = xfree, | |
1234 | ||
1235 | /* get_os_pid. */ | |
1236 | .get_os_pid = [] (amd_dbgapi_client_process_id_t client_process_id, | |
1237 | pid_t *pid) -> amd_dbgapi_status_t { | |
1238 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1239 | ||
1240 | *pid = inf->pid; | |
1241 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1242 | }, | |
1243 | ||
1244 | /* enable_notify_shared_library callback. */ | |
1245 | .enable_notify_shared_library | |
1246 | = [] (amd_dbgapi_client_process_id_t client_process_id, | |
1247 | const char *library_name, amd_dbgapi_shared_library_id_t library_id, | |
1248 | amd_dbgapi_shared_library_state_t *library_state) | |
1249 | -> amd_dbgapi_status_t { | |
1250 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1251 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1252 | ||
1253 | if (!library_name || !library_state) | |
1254 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; | |
1255 | ||
1256 | if (info->notify_solib_map.find (library_id.handle) | |
1257 | != info->notify_solib_map.end ()) | |
1258 | { | |
1259 | /* This library id is already registered. */ | |
1260 | return AMD_DBGAPI_STATUS_ERROR; | |
1261 | } | |
1262 | ||
1263 | /* Check whether the library is already loaded. */ | |
1264 | bool is_loaded = false; | |
1265 | struct so_list *solib; | |
1266 | for (solib = inf->pspace->so_list; solib; solib = solib->next) | |
1267 | if (::strstr (solib->so_original_name, library_name)) | |
1268 | { | |
1269 | is_loaded = true; | |
1270 | break; | |
1271 | } | |
1272 | ||
1273 | /* Check that the library_name is valid. If must not be empty, and | |
1274 | should not have wildcard characters. */ | |
1275 | if (*library_name == '\0' | |
1276 | || std::string (library_name).find_first_of ("*?[]") | |
1277 | != std::string::npos) | |
1278 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; | |
1279 | ||
1280 | /* Add a new entry in the notify_solib_map. */ | |
1281 | if (!info->notify_solib_map | |
1282 | .emplace (std::piecewise_construct, | |
1283 | std::forward_as_tuple (library_id.handle), | |
1284 | std::forward_as_tuple (rocm_notify_shared_library_info{ | |
1285 | library_name, solib, is_loaded })) | |
1286 | .second) | |
1287 | return AMD_DBGAPI_STATUS_ERROR; | |
1288 | ||
1289 | *library_state = is_loaded ? AMD_DBGAPI_SHARED_LIBRARY_STATE_LOADED | |
1290 | : AMD_DBGAPI_SHARED_LIBRARY_STATE_UNLOADED; | |
1291 | ||
1292 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1293 | }, | |
1294 | ||
1295 | /* disable_notify_shared_library callback. */ | |
1296 | .disable_notify_shared_library | |
1297 | = [] (amd_dbgapi_client_process_id_t client_process_id, | |
1298 | amd_dbgapi_shared_library_id_t library_id) -> amd_dbgapi_status_t { | |
1299 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1300 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1301 | ||
1302 | auto it = info->notify_solib_map.find (library_id.handle); | |
1303 | if (it == info->notify_solib_map.end ()) | |
1304 | return AMD_DBGAPI_STATUS_ERROR_INVALID_SHARED_LIBRARY_ID; | |
1305 | ||
1306 | info->notify_solib_map.erase (it); | |
1307 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1308 | }, | |
1309 | ||
1310 | /* get_symbol_address callback. */ | |
1311 | .get_symbol_address = | |
1312 | [] (amd_dbgapi_client_process_id_t client_process_id, | |
1313 | amd_dbgapi_shared_library_id_t library_id, const char *symbol_name, | |
1314 | amd_dbgapi_global_address_t *address) { | |
1315 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1316 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1317 | ||
1318 | auto it = info->notify_solib_map.find (library_id.handle); | |
1319 | if (it == info->notify_solib_map.end ()) | |
1320 | return AMD_DBGAPI_STATUS_ERROR_INVALID_SHARED_LIBRARY_ID; | |
1321 | ||
1322 | struct so_list *solib = it->second.solib; | |
1323 | if (!solib) | |
1324 | return AMD_DBGAPI_STATUS_ERROR_LIBRARY_NOT_LOADED; | |
1325 | ||
1326 | solib_read_symbols (solib, 0); | |
1327 | gdb_assert (solib->objfile); | |
1328 | ||
1329 | struct bound_minimal_symbol msymbol | |
1330 | = lookup_minimal_symbol (symbol_name, NULL, solib->objfile); | |
1331 | ||
1332 | if (!msymbol.minsym || BMSYMBOL_VALUE_ADDRESS (msymbol) == 0) | |
1333 | return AMD_DBGAPI_STATUS_ERROR_SYMBOL_NOT_FOUND; | |
1334 | ||
1335 | *address = BMSYMBOL_VALUE_ADDRESS (msymbol); | |
1336 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1337 | }, | |
1338 | ||
1339 | /* set_breakpoint callback. */ | |
1340 | .add_breakpoint = | |
1341 | [] (amd_dbgapi_client_process_id_t client_process_id, | |
1342 | amd_dbgapi_shared_library_id_t shared_library_id, | |
1343 | amd_dbgapi_global_address_t address, | |
1344 | amd_dbgapi_breakpoint_id_t breakpoint_id) { | |
1345 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1346 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1347 | ||
1348 | /* Initialize the breakpoint ops lazily since we depend on | |
1349 | bkpt_breakpoint_ops and we can't control the order in which | |
1350 | initializers are called. */ | |
1351 | if (rocm_breakpoint_ops.check_status == NULL) | |
1352 | { | |
1353 | rocm_breakpoint_ops = bkpt_breakpoint_ops; | |
1354 | rocm_breakpoint_ops.check_status = rocm_breakpoint_check_status; | |
1355 | } | |
1356 | ||
1357 | auto it = info->breakpoint_map.find (breakpoint_id.handle); | |
1358 | if (it != info->breakpoint_map.end ()) | |
1359 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; | |
1360 | ||
1361 | /* Create a new breakpoint. */ | |
1362 | struct obj_section *section = find_pc_section (address); | |
1363 | if (!section || !section->objfile) | |
1364 | return AMD_DBGAPI_STATUS_ERROR; | |
1365 | ||
1366 | event_location_up location | |
1367 | = new_address_location (address, nullptr, 0); | |
1368 | if (!create_breakpoint ( | |
1369 | get_objfile_arch (section->objfile), location.get (), | |
1370 | /*cond_string*/ NULL, /*thread*/ -1, /*extra_sring*/ NULL, | |
1371 | /*parse_extra*/ 0, /*tempflag*/ 0, /*bptype*/ bp_breakpoint, | |
1372 | /*ignore_count*/ 0, /*pending_break*/ AUTO_BOOLEAN_FALSE, | |
1373 | /*ops*/ &rocm_breakpoint_ops, /*from_tty*/ 0, | |
1374 | /*enabled*/ 1, /*internal*/ 1, /*flags*/ 0)) | |
1375 | return AMD_DBGAPI_STATUS_ERROR; | |
1376 | ||
1377 | /* Find our breakpoint in the breakpoint list. */ | |
1378 | auto bp_loc = std::make_pair (inf->aspace, address); | |
1379 | auto bp = breakpoint_find_if ( | |
1380 | [] (struct breakpoint *b, void *data) { | |
1381 | auto *arg = static_cast<decltype (&bp_loc)> (data); | |
1382 | if (b->ops == &rocm_breakpoint_ops && b->loc | |
1383 | && b->loc->pspace->aspace == arg->first | |
1384 | && b->loc->address == arg->second) | |
1385 | return 1; | |
1386 | return 0; | |
1387 | }, | |
1388 | reinterpret_cast<void *> (&bp_loc)); | |
1389 | ||
1390 | if (!bp) | |
1391 | error (_ ("Could not find breakpoint")); | |
1392 | ||
1393 | info->breakpoint_map.emplace (breakpoint_id.handle, bp); | |
1394 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1395 | }, | |
1396 | ||
1397 | /* remove_breakpoint callback. */ | |
1398 | .remove_breakpoint = | |
1399 | [] (amd_dbgapi_client_process_id_t client_process_id, | |
1400 | amd_dbgapi_breakpoint_id_t breakpoint_id) { | |
1401 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1402 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1403 | ||
1404 | auto it = info->breakpoint_map.find (breakpoint_id.handle); | |
1405 | if (it == info->breakpoint_map.end ()) | |
1406 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; | |
1407 | ||
1408 | delete_breakpoint (it->second); | |
1409 | info->breakpoint_map.erase (it); | |
1410 | ||
1411 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1412 | }, | |
1413 | ||
1414 | /* set_breakpoint_state callback. */ | |
1415 | .set_breakpoint_state = | |
1416 | [] (amd_dbgapi_client_process_id_t client_process_id, | |
1417 | amd_dbgapi_breakpoint_id_t breakpoint_id, | |
1418 | amd_dbgapi_breakpoint_state_t breakpoint_state) { | |
1419 | inferior *inf = static_cast<inferior *> (client_process_id); | |
1420 | struct rocm_inferior_info *info = get_rocm_inferior_info (inf); | |
1421 | ||
1422 | auto it = info->breakpoint_map.find (breakpoint_id.handle); | |
1423 | if (it == info->breakpoint_map.end ()) | |
1424 | return AMD_DBGAPI_STATUS_ERROR_INVALID_BREAKPOINT_ID; | |
1425 | ||
1426 | if (breakpoint_state == AMD_DBGAPI_BREAKPOINT_STATE_ENABLE) | |
1427 | it->second->enable_state = bp_enabled; | |
1428 | else if (breakpoint_state == AMD_DBGAPI_BREAKPOINT_STATE_DISABLE) | |
1429 | it->second->enable_state = bp_disabled; | |
1430 | else | |
1431 | return AMD_DBGAPI_STATUS_ERROR_INVALID_ARGUMENT; | |
1432 | ||
1433 | return AMD_DBGAPI_STATUS_SUCCESS; | |
1434 | }, | |
1435 | ||
1436 | .log_message | |
1437 | = [] (amd_dbgapi_log_level_t level, const char *message) -> void { | |
1438 | gdb::optional<target_terminal::scoped_restore_terminal_state> tstate; | |
1439 | ||
1440 | if (level > get_debug_amd_dbgapi_log_level ()) | |
1441 | return; | |
1442 | ||
1443 | if (target_supports_terminal_ours ()) | |
1444 | { | |
1445 | tstate.emplace (); | |
1446 | target_terminal::ours_for_output (); | |
1447 | } | |
1448 | ||
1449 | if (filtered_printing_initialized ()) | |
1450 | wrap_here (""); | |
1451 | ||
1452 | struct ui_file *out_file | |
1453 | = (level >= AMD_DBGAPI_LOG_LEVEL_INFO) ? gdb_stdlog : gdb_stderr; | |
1454 | ||
1455 | switch (level) | |
1456 | { | |
1457 | case AMD_DBGAPI_LOG_LEVEL_FATAL_ERROR: | |
1458 | fputs_unfiltered ("[amd-dbgapi]: ", out_file); | |
1459 | break; | |
1460 | case AMD_DBGAPI_LOG_LEVEL_WARNING: | |
1461 | fputs_styled ("[amd-dbgapi]: ", warning_style.style (), out_file); | |
1462 | break; | |
1463 | case AMD_DBGAPI_LOG_LEVEL_INFO: | |
1464 | fputs_styled ("[amd-dbgapi]: ", info_style.style (), out_file); | |
1465 | break; | |
1466 | case AMD_DBGAPI_LOG_LEVEL_VERBOSE: | |
1467 | fputs_styled ("[amd-dbgapi]: ", verbose_style.style (), out_file); | |
1468 | break; | |
1469 | } | |
1470 | ||
1471 | fputs_unfiltered (message, out_file); | |
1472 | fputs_unfiltered ("\n", out_file); | |
1473 | } | |
1474 | }; | |
1475 | ||
1476 | /* Implementation of `_wave_id' variable. */ | |
1477 | ||
1478 | static struct value * | |
1479 | rocm_wave_id_make_value (struct gdbarch *gdbarch, struct internalvar *var, | |
1480 | void *ignore) | |
1481 | { | |
1482 | if (ptid_is_gpu (inferior_ptid)) | |
1483 | { | |
1484 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); | |
1485 | amd_dbgapi_wave_id_t wave_id = get_amd_dbgapi_wave_id (inferior_ptid); | |
1486 | uint32_t group_ids[3], wave_in_group; | |
1487 | ||
1488 | if (amd_dbgapi_wave_get_info (process_id, wave_id, | |
1489 | AMD_DBGAPI_WAVE_INFO_WORK_GROUP_COORD, | |
1490 | sizeof (group_ids), &group_ids) | |
1491 | == AMD_DBGAPI_STATUS_SUCCESS | |
1492 | && amd_dbgapi_wave_get_info ( | |
1493 | process_id, wave_id, | |
1494 | AMD_DBGAPI_WAVE_INFO_WAVE_NUMBER_IN_WORK_GROUP, | |
1495 | sizeof (wave_in_group), &wave_in_group) | |
1496 | == AMD_DBGAPI_STATUS_SUCCESS) | |
1497 | { | |
1498 | std::string wave_id_str | |
1499 | = string_printf ("(%d,%d,%d)/%d", group_ids[2], group_ids[1], | |
1500 | group_ids[0], wave_in_group); | |
1501 | ||
1502 | return value_cstring (wave_id_str.data (), wave_id_str.length () + 1, | |
1503 | builtin_type (gdbarch)->builtin_char); | |
1504 | } | |
1505 | } | |
1506 | ||
1507 | return allocate_value (builtin_type (gdbarch)->builtin_void); | |
1508 | } | |
1509 | ||
1510 | static const struct internalvar_funcs rocm_wave_id_funcs | |
1511 | = { rocm_wave_id_make_value, NULL, NULL }; | |
1512 | ||
1513 | /* List of set/show debug amd_dbgapi commands. */ | |
1514 | struct cmd_list_element *set_debug_amd_dbgapi_list; | |
1515 | struct cmd_list_element *show_debug_amd_dbgapi_list; | |
1516 | ||
1517 | static void | |
1518 | set_debug_amd_dbgapi (const char *arg, int from_tty) | |
1519 | { | |
1520 | help_list (set_debug_amd_dbgapi_list, "set debug amd-dbgapi ", | |
1521 | (enum command_class) - 1, gdb_stdout); | |
1522 | } | |
1523 | ||
1524 | static void | |
1525 | show_debug_amd_dbgapi (const char *args, int from_tty) | |
1526 | { | |
1527 | cmd_show_list (show_debug_amd_dbgapi_list, from_tty, ""); | |
1528 | } | |
1529 | ||
1530 | constexpr char amd_dbgapi_log_level_off[] = "off"; | |
1531 | constexpr char amd_dbgapi_log_level_error[] = "error"; | |
1532 | constexpr char amd_dbgapi_log_level_warning[] = "warning"; | |
1533 | constexpr char amd_dbgapi_log_level_info[] = "info"; | |
1534 | constexpr char amd_dbgapi_log_level_verbose[] = "verbose"; | |
1535 | ||
1536 | constexpr const char *debug_amd_dbgapi_log_level_enums[] | |
1537 | = { [AMD_DBGAPI_LOG_LEVEL_NONE] = amd_dbgapi_log_level_off, | |
1538 | [AMD_DBGAPI_LOG_LEVEL_FATAL_ERROR] = amd_dbgapi_log_level_error, | |
1539 | [AMD_DBGAPI_LOG_LEVEL_WARNING] = amd_dbgapi_log_level_warning, | |
1540 | [AMD_DBGAPI_LOG_LEVEL_INFO] = amd_dbgapi_log_level_info, | |
1541 | [AMD_DBGAPI_LOG_LEVEL_VERBOSE] = amd_dbgapi_log_level_verbose, | |
1542 | nullptr }; | |
1543 | ||
1544 | static const char *debug_amd_dbgapi_log_level = amd_dbgapi_log_level_error; | |
1545 | ||
1546 | static amd_dbgapi_log_level_t | |
1547 | get_debug_amd_dbgapi_log_level () | |
1548 | { | |
1549 | size_t pos; | |
1550 | for (pos = 0; debug_amd_dbgapi_log_level_enums[pos]; ++pos) | |
1551 | if (debug_amd_dbgapi_log_level == debug_amd_dbgapi_log_level_enums[pos]) | |
1552 | break; | |
1553 | ||
1554 | gdb_assert (debug_amd_dbgapi_log_level_enums[pos]); | |
1555 | return static_cast<amd_dbgapi_log_level_t> (pos); | |
1556 | } | |
1557 | ||
1558 | static void | |
1559 | set_debug_amd_dbgapi_log_level (const char *args, int from_tty, | |
1560 | struct cmd_list_element *c) | |
1561 | { | |
1562 | amd_dbgapi_set_log_level (get_debug_amd_dbgapi_log_level ()); | |
1563 | } | |
1564 | ||
1565 | static void | |
1566 | show_debug_amd_dbgapi_log_level (struct ui_file *file, int from_tty, | |
1567 | struct cmd_list_element *c, | |
1568 | const char *value) | |
1569 | { | |
1570 | fprintf_filtered (file, _ ("The amd-dbgapi log level is %s.\n"), value); | |
1571 | } | |
1572 | ||
1573 | static void | |
1574 | info_agents_command (const char *args, int from_tty) | |
1575 | { | |
1576 | amd_dbgapi_process_id_t process_id = get_amd_dbgapi_process_id (); | |
1577 | struct ui_out *uiout = current_uiout; | |
1578 | amd_dbgapi_status_t status; | |
1579 | ||
1580 | amd_dbgapi_agent_id_t *agent_list; | |
1581 | size_t count = 0; | |
1582 | ||
1583 | if (process_id.handle != AMD_DBGAPI_PROCESS_NONE.handle | |
1584 | && (status | |
1585 | = amd_dbgapi_agent_list (process_id, &count, &agent_list, nullptr)) | |
1586 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1587 | error (_ ("amd_dbgapi_agent_list failed (rc=%d)"), status); | |
1588 | ||
1589 | if (!count && !uiout->is_mi_like_p ()) | |
1590 | { | |
1591 | uiout->field_string (NULL, | |
1592 | _ ("No agents are currently active.\n")); | |
1593 | return; | |
1594 | } | |
1595 | ||
1596 | /* Calculate the maximum size needed to print the agents names. */ | |
1597 | std::vector<std::string> agent_names (count); | |
1598 | ||
1599 | size_t max_name_len = 0; | |
1600 | for (size_t i = 0; i < count; ++i) | |
1601 | { | |
1602 | char *agent_name; | |
1603 | ||
1604 | if ((status = amd_dbgapi_agent_get_info ( | |
1605 | process_id, agent_list[i], AMD_DBGAPI_AGENT_INFO_NAME, | |
1606 | sizeof (agent_name), &agent_name)) | |
1607 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1608 | { | |
1609 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) | |
1610 | agent_names[i] = "N/A"; | |
1611 | else | |
1612 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); | |
1613 | } | |
1614 | else | |
1615 | { | |
1616 | agent_names[i] = agent_name; | |
1617 | xfree (agent_name); | |
1618 | } | |
1619 | ||
1620 | max_name_len = std::max (max_name_len, agent_names[i].size ()); | |
1621 | } | |
1622 | ||
1623 | /* Header: */ | |
1624 | ui_out_emit_table table_emmitter (uiout, 7, count, "InfoRocmDevicesTable"); | |
1625 | ||
1626 | uiout->table_header (2, ui_left, "agent_id", "Id"); | |
1627 | uiout->table_header (8, ui_left, "location_id", "PCI Slot"); | |
1628 | uiout->table_header (std::max (11ul, max_name_len), ui_left, "name", | |
1629 | "Device Name"); | |
1630 | uiout->table_header (14, ui_left, "num_se", "Shader Engines"); | |
1631 | uiout->table_header (13, ui_left, "num_cu", "Compute Units"); | |
1632 | uiout->table_header (7, ui_left, "simd", "SIMD/CU"); | |
1633 | uiout->table_header (15, ui_left, "waves", "Wavefronts/SIMD"); | |
1634 | uiout->table_body (); | |
1635 | ||
1636 | /* Rows: */ | |
1637 | for (size_t i = 0; i < count; ++i) | |
1638 | { | |
1639 | ui_out_emit_tuple tuple_emitter (uiout, "InfoRocmDevicesRow"); | |
1640 | ||
1641 | /* agent */ | |
1642 | uiout->field_signed ("agent_id", agent_list[i].handle); | |
1643 | ||
1644 | /* location */ | |
1645 | uint32_t location_id; | |
1646 | if ((status = amd_dbgapi_agent_get_info ( | |
1647 | process_id, agent_list[i], AMD_DBGAPI_AGENT_INFO_PCIE_SLOT, | |
1648 | sizeof (location_id), &location_id)) | |
1649 | != AMD_DBGAPI_STATUS_SUCCESS) | |
1650 | { | |
1651 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) | |
1652 | uiout->field_string ("location_id", "N/A"); | |
1653 | else | |
1654 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); | |
1655 | } | |
1656 | else | |
1657 | uiout->field_string ( | |
1658 | "location_id", | |
1659 | string_printf ("%02x:%02x.%d", (location_id >> 8) & 0xFF, | |
1660 | (location_id >> 3) & 0x1F, location_id & 0x7)); | |
1661 | ||
1662 | /* name */ | |
1663 | uiout->field_string ("name", agent_names[i]); | |
1664 | ||
1665 | /* num_se, num_cu, simd, waves */ | |
1666 | ||
1667 | #define UIOUT_FIELD_INT(name, query) \ | |
1668 | uint32_t name; \ | |
1669 | if ((status = amd_dbgapi_agent_get_info (process_id, agent_list[i], query, \ | |
1670 | sizeof (name), &name)) \ | |
1671 | != AMD_DBGAPI_STATUS_SUCCESS) \ | |
1672 | { \ | |
1673 | if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_AGENT_ID) \ | |
1674 | uiout->field_string (#name, "N/A"); \ | |
1675 | else \ | |
1676 | error (_ ("amd_dbgapi_agent_get_info failed (rc=%d"), status); \ | |
1677 | } \ | |
1678 | else \ | |
1679 | uiout->field_signed (#name, name); | |
1680 | ||
1681 | UIOUT_FIELD_INT (num_se, AMD_DBGAPI_AGENT_INFO_SHADER_ENGINE_COUNT); | |
1682 | UIOUT_FIELD_INT (num_cu, AMD_DBGAPI_AGENT_INFO_COMPUTE_UNIT_COUNT); | |
1683 | UIOUT_FIELD_INT (simd, AMD_DBGAPI_AGENT_INFO_NUM_SIMD_PER_COMPUTE_UNIT); | |
1684 | UIOUT_FIELD_INT (waves, AMD_DBGAPI_AGENT_INFO_MAX_WAVES_PER_SIMD); | |
1685 | ||
1686 | #undef UIOUT_FIELD_INT | |
1687 | ||
1688 | uiout->text ("\n"); | |
1689 | } | |
1690 | ||
1691 | xfree (agent_list); | |
1692 | gdb_flush (gdb_stdout); | |
1693 | } | |
1694 | ||
1695 | /* -Wmissing-prototypes */ | |
1696 | extern initialize_file_ftype _initialize_rocm_tdep; | |
1697 | ||
1698 | void | |
1699 | _initialize_rocm_tdep (void) | |
1700 | { | |
1701 | /* Initialize the ROCm Debug API. */ | |
1702 | if (amd_dbgapi_initialize (&dbgapi_callbacks) != AMD_DBGAPI_STATUS_SUCCESS) | |
1703 | return; | |
1704 | ||
1705 | /* Set the initial log level. */ | |
1706 | amd_dbgapi_set_log_level (get_debug_amd_dbgapi_log_level ()); | |
1707 | ||
1708 | /* Install observers. */ | |
1709 | gdb::observers::breakpoint_created.attach (rocm_target_breakpoint_fixup); | |
1710 | gdb::observers::solib_loaded.attach (rocm_target_solib_loaded); | |
1711 | gdb::observers::solib_unloaded.attach (rocm_target_solib_unloaded); | |
1712 | gdb::observers::inferior_created.attach (rocm_target_inferior_created); | |
1713 | gdb::observers::inferior_exit.attach (rocm_target_inferior_exit); | |
1714 | ||
1715 | amd_dbgapi_activated.attach (rocm_target_dbgapi_activated); | |
1716 | amd_dbgapi_deactivated.attach (rocm_target_dbgapi_deactivated); | |
1717 | ||
1718 | create_internalvar_type_lazy ("_wave_id", &rocm_wave_id_funcs, NULL); | |
1719 | ||
1720 | add_prefix_cmd ( | |
1721 | "amd-dbgapi", no_class, set_debug_amd_dbgapi, | |
1722 | _ ("Generic command for setting amd-dbgapi debugging flags"), | |
1723 | &set_debug_amd_dbgapi_list, "set debug amd-dbgapi ", 0, &setdebuglist); | |
1724 | ||
1725 | add_prefix_cmd ( | |
1726 | "amd-dbgapi", no_class, show_debug_amd_dbgapi, | |
1727 | _ ("Generic command for showing amd-dbgapi debugging flags"), | |
1728 | &show_debug_amd_dbgapi_list, "show debug amd-dbgapi ", 0, | |
1729 | &showdebuglist); | |
1730 | ||
1731 | add_setshow_enum_cmd ( | |
1732 | "log-level", class_maintenance, debug_amd_dbgapi_log_level_enums, | |
1733 | &debug_amd_dbgapi_log_level, _ ("Set the amd-dbgapi log level."), | |
1734 | _ ("Show the amd-dbgapi log level."), | |
1735 | _ ("off == no logging is enabled\n" | |
1736 | "error == fatal errors are reported\n" | |
1737 | "warning == fatal errors and warnings are reported\n" | |
1738 | "info == fatal errors, warnings, and info messages are reported\n" | |
1739 | "verbose == all messages are reported"), | |
1740 | set_debug_amd_dbgapi_log_level, show_debug_amd_dbgapi_log_level, | |
1741 | &set_debug_amd_dbgapi_list, &show_debug_amd_dbgapi_list); | |
1742 | ||
1743 | add_cmd ("agents", class_info, info_agents_command, | |
1744 | _ ("Info about currently active agents."), &infolist); | |
1745 | } |