Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1// SPDX-License-Identifier: MIT
2/*
3 * Copyright © 2023 Intel Corporation
4 */
5
6#include "xe_devcoredump.h"
7#include "xe_devcoredump_types.h"
8
9#include <linux/ascii85.h>
10#include <linux/devcoredump.h>
11#include <generated/utsrelease.h>
12
13#include <drm/drm_managed.h>
14
15#include "xe_device.h"
16#include "xe_exec_queue.h"
17#include "xe_force_wake.h"
18#include "xe_gt_printk.h"
19#include "xe_gt_types.h"
20#include "xe_guc_capture.h"
21#include "xe_guc_ct.h"
22#include "xe_guc_log.h"
23#include "xe_guc_submit.h"
24#include "xe_hw_engine.h"
25#include "xe_pm.h"
26#include "xe_sched_job.h"
27#include "xe_vm.h"
28
29/**
30 * DOC: Xe device coredump
31 *
32 * Xe uses dev_coredump infrastructure for exposing the crash errors in a
33 * standardized way. Once a crash occurs, devcoredump exposes a temporary
34 * node under ``/sys/class/devcoredump/devcd<m>/``. The same node is also
35 * accessible in ``/sys/class/drm/card<n>/device/devcoredump/``. The
36 * ``failing_device`` symlink points to the device that crashed and created the
37 * coredump.
38 *
39 * The following characteristics are observed by xe when creating a device
40 * coredump:
41 *
42 * **Snapshot at hang**:
43 * The 'data' file contains a snapshot of the HW and driver states at the time
44 * the hang happened. Due to the driver recovering from resets/crashes, it may
45 * not correspond to the state of the system when the file is read by
46 * userspace.
47 *
48 * **Coredump release**:
49 * After a coredump is generated, it stays in kernel memory until released by
50 * userspace by writing anything to it, or after an internal timer expires. The
51 * exact timeout may vary and should not be relied upon. Example to release
52 * a coredump:
53 *
54 * .. code-block:: shell
55 *
56 * $ > /sys/class/drm/card0/device/devcoredump/data
57 *
58 * **First failure only**:
59 * In general, the first hang is the most critical one since the following
60 * hangs can be a consequence of the initial hang. For this reason a snapshot
61 * is taken only for the first failure. Until the devcoredump is released by
62 * userspace or kernel, all subsequent hangs do not override the snapshot nor
63 * create new ones. Devcoredump has a delayed work queue that will eventually
64 * delete the file node and free all the dump information.
65 */
66
67#ifdef CONFIG_DEV_COREDUMP
68
69/* 1 hour timeout */
70#define XE_COREDUMP_TIMEOUT_JIFFIES (60 * 60 * HZ)
71
72static struct xe_device *coredump_to_xe(const struct xe_devcoredump *coredump)
73{
74 return container_of(coredump, struct xe_device, devcoredump);
75}
76
77static struct xe_guc *exec_queue_to_guc(struct xe_exec_queue *q)
78{
79 return &q->gt->uc.guc;
80}
81
82static ssize_t __xe_devcoredump_read(char *buffer, ssize_t count,
83 ssize_t start,
84 struct xe_devcoredump *coredump)
85{
86 struct xe_device *xe;
87 struct xe_devcoredump_snapshot *ss;
88 struct drm_printer p;
89 struct drm_print_iterator iter;
90 struct timespec64 ts;
91 int i;
92
93 xe = coredump_to_xe(coredump);
94 ss = &coredump->snapshot;
95
96 iter.data = buffer;
97 iter.start = start;
98 iter.remain = count;
99
100 p = drm_coredump_printer(&iter);
101
102 drm_puts(&p, "**** Xe Device Coredump ****\n");
103 drm_printf(&p, "Reason: %s\n", ss->reason);
104 drm_puts(&p, "kernel: " UTS_RELEASE "\n");
105 drm_puts(&p, "module: " KBUILD_MODNAME "\n");
106
107 ts = ktime_to_timespec64(ss->snapshot_time);
108 drm_printf(&p, "Snapshot time: %ptSp\n", &ts);
109 ts = ktime_to_timespec64(ss->boot_time);
110 drm_printf(&p, "Uptime: %ptSp\n", &ts);
111 drm_printf(&p, "Process: %s [%d]\n", ss->process_name, ss->pid);
112 xe_device_snapshot_print(xe, &p);
113
114 drm_printf(&p, "\n**** GT #%d ****\n", ss->gt->info.id);
115 drm_printf(&p, "\tTile: %d\n", ss->gt->tile->id);
116
117 drm_puts(&p, "\n**** GuC Log ****\n");
118 xe_guc_log_snapshot_print(ss->guc.log, &p);
119 drm_puts(&p, "\n**** GuC CT ****\n");
120 xe_guc_ct_snapshot_print(ss->guc.ct, &p);
121
122 drm_puts(&p, "\n**** Contexts ****\n");
123 xe_guc_exec_queue_snapshot_print(ss->ge, &p);
124
125 drm_puts(&p, "\n**** Job ****\n");
126 xe_sched_job_snapshot_print(ss->job, &p);
127
128 drm_puts(&p, "\n**** HW Engines ****\n");
129 for (i = 0; i < XE_NUM_HW_ENGINES; i++)
130 if (ss->hwe[i])
131 xe_engine_snapshot_print(ss->hwe[i], &p);
132
133 drm_puts(&p, "\n**** VM state ****\n");
134 xe_vm_snapshot_print(ss->vm, &p);
135
136 return count - iter.remain;
137}
138
139static void xe_devcoredump_snapshot_free(struct xe_devcoredump_snapshot *ss)
140{
141 int i;
142
143 kfree(ss->reason);
144 ss->reason = NULL;
145
146 xe_guc_log_snapshot_free(ss->guc.log);
147 ss->guc.log = NULL;
148
149 xe_guc_ct_snapshot_free(ss->guc.ct);
150 ss->guc.ct = NULL;
151
152 xe_guc_capture_put_matched_nodes(&ss->gt->uc.guc);
153 ss->matched_node = NULL;
154
155 xe_guc_exec_queue_snapshot_free(ss->ge);
156 ss->ge = NULL;
157
158 xe_sched_job_snapshot_free(ss->job);
159 ss->job = NULL;
160
161 for (i = 0; i < XE_NUM_HW_ENGINES; i++)
162 if (ss->hwe[i]) {
163 xe_hw_engine_snapshot_free(ss->hwe[i]);
164 ss->hwe[i] = NULL;
165 }
166
167 xe_vm_snapshot_free(ss->vm);
168 ss->vm = NULL;
169}
170
171#define XE_DEVCOREDUMP_CHUNK_MAX (SZ_512M + SZ_1G)
172
173/**
174 * xe_devcoredump_read() - Read data from the Xe device coredump snapshot
175 * @buffer: Destination buffer to copy the coredump data into
176 * @offset: Offset in the coredump data to start reading from
177 * @count: Number of bytes to read
178 * @data: Pointer to the xe_devcoredump structure
179 * @datalen: Length of the data (unused)
180 *
181 * Reads a chunk of the coredump snapshot data into the provided buffer.
182 * If the devcoredump is smaller than 1.5 GB (XE_DEVCOREDUMP_CHUNK_MAX),
183 * it is read directly from a pre-written buffer. For larger devcoredumps,
184 * the pre-written buffer must be periodically repopulated from the snapshot
185 * state due to kmalloc size limitations.
186 *
187 * Return: Number of bytes copied on success, or a negative error code on failure.
188 */
189static ssize_t xe_devcoredump_read(char *buffer, loff_t offset,
190 size_t count, void *data, size_t datalen)
191{
192 struct xe_devcoredump *coredump = data;
193 struct xe_devcoredump_snapshot *ss;
194 ssize_t byte_copied = 0;
195 u32 chunk_offset;
196 ssize_t new_chunk_position;
197 bool pm_needed = false;
198 int ret = 0;
199
200 if (!coredump)
201 return -ENODEV;
202
203 ss = &coredump->snapshot;
204
205 /* Ensure delayed work is captured before continuing */
206 flush_work(&ss->work);
207
208 pm_needed = ss->read.size > XE_DEVCOREDUMP_CHUNK_MAX;
209 if (pm_needed)
210 xe_pm_runtime_get(gt_to_xe(ss->gt));
211
212 mutex_lock(&coredump->lock);
213
214 if (!ss->read.buffer) {
215 ret = -ENODEV;
216 goto unlock;
217 }
218
219 if (offset >= ss->read.size)
220 goto unlock;
221
222 new_chunk_position = div_u64_rem(offset,
223 XE_DEVCOREDUMP_CHUNK_MAX,
224 &chunk_offset);
225
226 if (offset >= ss->read.chunk_position + XE_DEVCOREDUMP_CHUNK_MAX ||
227 offset < ss->read.chunk_position) {
228 ss->read.chunk_position = new_chunk_position *
229 XE_DEVCOREDUMP_CHUNK_MAX;
230
231 __xe_devcoredump_read(ss->read.buffer,
232 XE_DEVCOREDUMP_CHUNK_MAX,
233 ss->read.chunk_position, coredump);
234 }
235
236 byte_copied = count < ss->read.size - offset ? count :
237 ss->read.size - offset;
238 memcpy(buffer, ss->read.buffer + chunk_offset, byte_copied);
239
240unlock:
241 mutex_unlock(&coredump->lock);
242
243 if (pm_needed)
244 xe_pm_runtime_put(gt_to_xe(ss->gt));
245
246 return byte_copied ? byte_copied : ret;
247}
248
249static void xe_devcoredump_free(void *data)
250{
251 struct xe_devcoredump *coredump = data;
252
253 /* Our device is gone. Nothing to do... */
254 if (!data || !coredump_to_xe(coredump))
255 return;
256
257 cancel_work_sync(&coredump->snapshot.work);
258
259 mutex_lock(&coredump->lock);
260
261 xe_devcoredump_snapshot_free(&coredump->snapshot);
262 kvfree(coredump->snapshot.read.buffer);
263
264 /* To prevent stale data on next snapshot, clear everything */
265 memset(&coredump->snapshot, 0, sizeof(coredump->snapshot));
266 coredump->captured = false;
267 drm_info(&coredump_to_xe(coredump)->drm,
268 "Xe device coredump has been deleted.\n");
269
270 mutex_unlock(&coredump->lock);
271}
272
273static void xe_devcoredump_deferred_snap_work(struct work_struct *work)
274{
275 struct xe_devcoredump_snapshot *ss = container_of(work, typeof(*ss), work);
276 struct xe_devcoredump *coredump = container_of(ss, typeof(*coredump), snapshot);
277 struct xe_device *xe = coredump_to_xe(coredump);
278
279 /*
280 * NB: Despite passing a GFP_ flags parameter here, more allocations are done
281 * internally using GFP_KERNEL explicitly. Hence this call must be in the worker
282 * thread and not in the initial capture call.
283 */
284 dev_coredumpm_timeout(gt_to_xe(ss->gt)->drm.dev, THIS_MODULE, coredump, 0, GFP_KERNEL,
285 xe_devcoredump_read, xe_devcoredump_free,
286 XE_COREDUMP_TIMEOUT_JIFFIES);
287
288 guard(xe_pm_runtime)(xe);
289
290 /* keep going if fw fails as we still want to save the memory and SW data */
291 xe_with_force_wake(fw_ref, gt_to_fw(ss->gt), XE_FORCEWAKE_ALL) {
292 if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FORCEWAKE_ALL))
293 xe_gt_info(ss->gt, "failed to get forcewake for coredump capture\n");
294 xe_vm_snapshot_capture_delayed(ss->vm);
295 xe_guc_exec_queue_snapshot_capture_delayed(ss->ge);
296 }
297
298 ss->read.chunk_position = 0;
299
300 /* Calculate devcoredump size */
301 ss->read.size = __xe_devcoredump_read(NULL, LONG_MAX, 0, coredump);
302
303 if (ss->read.size > XE_DEVCOREDUMP_CHUNK_MAX) {
304 ss->read.buffer = kvmalloc(XE_DEVCOREDUMP_CHUNK_MAX,
305 GFP_USER);
306 if (!ss->read.buffer)
307 return;
308
309 __xe_devcoredump_read(ss->read.buffer,
310 XE_DEVCOREDUMP_CHUNK_MAX,
311 0, coredump);
312 } else {
313 ss->read.buffer = kvmalloc(ss->read.size, GFP_USER);
314 if (!ss->read.buffer)
315 return;
316
317 __xe_devcoredump_read(ss->read.buffer, ss->read.size, 0,
318 coredump);
319 xe_devcoredump_snapshot_free(ss);
320 }
321}
322
323static void devcoredump_snapshot(struct xe_devcoredump *coredump,
324 struct xe_exec_queue *q,
325 struct xe_sched_job *job)
326{
327 struct xe_devcoredump_snapshot *ss = &coredump->snapshot;
328 struct xe_guc *guc = exec_queue_to_guc(q);
329 const char *process_name = "no process";
330 bool cookie;
331
332 ss->snapshot_time = ktime_get_real();
333 ss->boot_time = ktime_get_boottime();
334
335 if (q->vm && q->vm->xef) {
336 process_name = q->vm->xef->process_name;
337 ss->pid = q->vm->xef->pid;
338 }
339
340 strscpy(ss->process_name, process_name);
341
342 ss->gt = q->gt;
343 INIT_WORK(&ss->work, xe_devcoredump_deferred_snap_work);
344
345 /* keep going if fw fails as we still want to save the memory and SW data */
346 CLASS(xe_force_wake, fw_ref)(gt_to_fw(q->gt), XE_FORCEWAKE_ALL);
347
348 cookie = dma_fence_begin_signalling();
349
350 ss->guc.log = xe_guc_log_snapshot_capture(&guc->log, true);
351 ss->guc.ct = xe_guc_ct_snapshot_capture(&guc->ct);
352 ss->ge = xe_guc_exec_queue_snapshot_capture(q);
353 if (job)
354 ss->job = xe_sched_job_snapshot_capture(job);
355 ss->vm = xe_vm_snapshot_capture(q->vm);
356
357 xe_engine_snapshot_capture_for_queue(q);
358
359 queue_work(system_dfl_wq, &ss->work);
360
361 dma_fence_end_signalling(cookie);
362}
363
364/**
365 * xe_devcoredump - Take the required snapshots and initialize coredump device.
366 * @q: The faulty xe_exec_queue, where the issue was detected.
367 * @job: The faulty xe_sched_job, where the issue was detected.
368 * @fmt: Printf format + args to describe the reason for the core dump
369 *
370 * This function should be called at the crash time within the serialized
371 * gt_reset. It is skipped if we still have the core dump device available
372 * with the information of the 'first' snapshot.
373 */
374__printf(3, 4)
375void xe_devcoredump(struct xe_exec_queue *q, struct xe_sched_job *job, const char *fmt, ...)
376{
377 struct xe_device *xe = gt_to_xe(q->gt);
378 struct xe_devcoredump *coredump = &xe->devcoredump;
379 va_list varg;
380
381 mutex_lock(&coredump->lock);
382
383 if (coredump->captured) {
384 drm_dbg(&xe->drm, "Multiple hangs are occurring, but only the first snapshot was taken\n");
385 mutex_unlock(&coredump->lock);
386 return;
387 }
388
389 coredump->captured = true;
390
391 va_start(varg, fmt);
392 coredump->snapshot.reason = kvasprintf(GFP_ATOMIC, fmt, varg);
393 va_end(varg);
394
395 devcoredump_snapshot(coredump, q, job);
396
397 drm_info(&xe->drm, "Xe device coredump has been created\n");
398 drm_info(&xe->drm, "Check your /sys/class/drm/card%d/device/devcoredump/data\n",
399 xe->drm.primary->index);
400
401 mutex_unlock(&coredump->lock);
402}
403
404static void xe_driver_devcoredump_fini(void *arg)
405{
406 struct drm_device *drm = arg;
407
408 dev_coredump_put(drm->dev);
409}
410
411int xe_devcoredump_init(struct xe_device *xe)
412{
413 int err;
414
415 err = drmm_mutex_init(&xe->drm, &xe->devcoredump.lock);
416 if (err)
417 return err;
418
419 if (IS_ENABLED(CONFIG_LOCKDEP)) {
420 fs_reclaim_acquire(GFP_KERNEL);
421 might_lock(&xe->devcoredump.lock);
422 fs_reclaim_release(GFP_KERNEL);
423 }
424
425 return devm_add_action_or_reset(xe->drm.dev, xe_driver_devcoredump_fini, &xe->drm);
426}
427
428#endif
429
430/**
431 * xe_print_blob_ascii85 - print a BLOB to some useful location in ASCII85
432 *
433 * The output is split into multiple calls to drm_puts() because some print
434 * targets, e.g. dmesg, cannot handle arbitrarily long lines. These targets may
435 * add newlines, as is the case with dmesg: each drm_puts() call creates a
436 * separate line.
437 *
438 * There is also a scheduler yield call to prevent the 'task has been stuck for
439 * 120s' kernel hang check feature from firing when printing to a slow target
440 * such as dmesg over a serial port.
441 *
442 * @p: the printer object to output to
443 * @prefix: optional prefix to add to output string
444 * @suffix: optional suffix to add at the end. 0 disables it and is
445 * not added to the output, which is useful when using multiple calls
446 * to dump data to @p
447 * @blob: the Binary Large OBject to dump out
448 * @offset: offset in bytes to skip from the front of the BLOB, must be a multiple of sizeof(u32)
449 * @size: the size in bytes of the BLOB, must be a multiple of sizeof(u32)
450 */
451void xe_print_blob_ascii85(struct drm_printer *p, const char *prefix, char suffix,
452 const void *blob, size_t offset, size_t size)
453{
454 const u32 *blob32 = (const u32 *)blob;
455 char buff[ASCII85_BUFSZ], *line_buff;
456 size_t line_pos = 0;
457
458#define DMESG_MAX_LINE_LEN 800
459 /* Always leave space for the suffix char and the \0 */
460#define MIN_SPACE (ASCII85_BUFSZ + 2) /* 85 + "<suffix>\0" */
461
462 if (size & 3)
463 drm_printf(p, "Size not word aligned: %zu", size);
464 if (offset & 3)
465 drm_printf(p, "Offset not word aligned: %zu", offset);
466
467 line_buff = kzalloc(DMESG_MAX_LINE_LEN, GFP_ATOMIC);
468 if (!line_buff) {
469 drm_printf(p, "Failed to allocate line buffer\n");
470 return;
471 }
472
473 blob32 += offset / sizeof(*blob32);
474 size /= sizeof(*blob32);
475
476 if (prefix) {
477 strscpy(line_buff, prefix, DMESG_MAX_LINE_LEN - MIN_SPACE - 2);
478 line_pos = strlen(line_buff);
479
480 line_buff[line_pos++] = ':';
481 line_buff[line_pos++] = ' ';
482 }
483
484 while (size--) {
485 u32 val = *(blob32++);
486
487 strscpy(line_buff + line_pos, ascii85_encode(val, buff),
488 DMESG_MAX_LINE_LEN - line_pos);
489 line_pos += strlen(line_buff + line_pos);
490
491 if ((line_pos + MIN_SPACE) >= DMESG_MAX_LINE_LEN) {
492 line_buff[line_pos++] = 0;
493
494 drm_puts(p, line_buff);
495
496 line_pos = 0;
497
498 /* Prevent 'stuck thread' time out errors */
499 cond_resched();
500 }
501 }
502
503 if (suffix)
504 line_buff[line_pos++] = suffix;
505
506 if (line_pos) {
507 line_buff[line_pos++] = 0;
508 drm_puts(p, line_buff);
509 }
510
511 kfree(line_buff);
512
513#undef MIN_SPACE
514#undef DMESG_MAX_LINE_LEN
515}