xref: /linux/drivers/gpu/drm/xe/xe_devcoredump.c (revision 06a130e42a5bfc84795464bff023bff4c16f58c5)
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/devcoredump.h>
10 #include <generated/utsrelease.h>
11 
12 #include <drm/drm_managed.h>
13 
14 #include "xe_device.h"
15 #include "xe_exec_queue.h"
16 #include "xe_force_wake.h"
17 #include "xe_gt.h"
18 #include "xe_gt_printk.h"
19 #include "xe_guc_ct.h"
20 #include "xe_guc_submit.h"
21 #include "xe_hw_engine.h"
22 #include "xe_sched_job.h"
23 #include "xe_vm.h"
24 
25 /**
26  * DOC: Xe device coredump
27  *
28  * Devices overview:
29  * Xe uses dev_coredump infrastructure for exposing the crash errors in a
30  * standardized way.
31  * devcoredump exposes a temporary device under /sys/class/devcoredump/
32  * which is linked with our card device directly.
33  * The core dump can be accessed either from
34  * /sys/class/drm/card<n>/device/devcoredump/ or from
35  * /sys/class/devcoredump/devcd<m> where
36  * /sys/class/devcoredump/devcd<m>/failing_device is a link to
37  * /sys/class/drm/card<n>/device/.
38  *
39  * Snapshot at hang:
40  * The 'data' file is printed with a drm_printer pointer at devcoredump read
41  * time. For this reason, we need to take snapshots from when the hang has
42  * happened, and not only when the user is reading the file. Otherwise the
43  * information is outdated since the resets might have happened in between.
44  *
45  * 'First' failure snapshot:
46  * In general, the first hang is the most critical one since the following hangs
47  * can be a consequence of the initial hang. For this reason we only take the
48  * snapshot of the 'first' failure and ignore subsequent calls of this function,
49  * at least while the coredump device is alive. Dev_coredump has a delayed work
50  * queue that will eventually delete the device and free all the dump
51  * information.
52  */
53 
54 #ifdef CONFIG_DEV_COREDUMP
55 
56 /* 1 hour timeout */
57 #define XE_COREDUMP_TIMEOUT_JIFFIES (60 * 60 * HZ)
58 
59 static struct xe_device *coredump_to_xe(const struct xe_devcoredump *coredump)
60 {
61 	return container_of(coredump, struct xe_device, devcoredump);
62 }
63 
64 static struct xe_guc *exec_queue_to_guc(struct xe_exec_queue *q)
65 {
66 	return &q->gt->uc.guc;
67 }
68 
69 static ssize_t __xe_devcoredump_read(char *buffer, size_t count,
70 				     struct xe_devcoredump *coredump)
71 {
72 	struct xe_device *xe;
73 	struct xe_devcoredump_snapshot *ss;
74 	struct drm_printer p;
75 	struct drm_print_iterator iter;
76 	struct timespec64 ts;
77 	int i;
78 
79 	xe = coredump_to_xe(coredump);
80 	ss = &coredump->snapshot;
81 
82 	iter.data = buffer;
83 	iter.start = 0;
84 	iter.remain = count;
85 
86 	p = drm_coredump_printer(&iter);
87 
88 	drm_printf(&p, "**** Xe Device Coredump ****\n");
89 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
90 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
91 
92 	ts = ktime_to_timespec64(ss->snapshot_time);
93 	drm_printf(&p, "Snapshot time: %lld.%09ld\n", ts.tv_sec, ts.tv_nsec);
94 	ts = ktime_to_timespec64(ss->boot_time);
95 	drm_printf(&p, "Uptime: %lld.%09ld\n", ts.tv_sec, ts.tv_nsec);
96 	drm_printf(&p, "Process: %s\n", ss->process_name);
97 	xe_device_snapshot_print(xe, &p);
98 
99 	drm_printf(&p, "\n**** GuC CT ****\n");
100 	xe_guc_ct_snapshot_print(coredump->snapshot.ct, &p);
101 	xe_guc_exec_queue_snapshot_print(coredump->snapshot.ge, &p);
102 
103 	drm_printf(&p, "\n**** Job ****\n");
104 	xe_sched_job_snapshot_print(coredump->snapshot.job, &p);
105 
106 	drm_printf(&p, "\n**** HW Engines ****\n");
107 	for (i = 0; i < XE_NUM_HW_ENGINES; i++)
108 		if (coredump->snapshot.hwe[i])
109 			xe_hw_engine_snapshot_print(coredump->snapshot.hwe[i],
110 						    &p);
111 	drm_printf(&p, "\n**** VM state ****\n");
112 	xe_vm_snapshot_print(coredump->snapshot.vm, &p);
113 
114 	return count - iter.remain;
115 }
116 
117 static void xe_devcoredump_snapshot_free(struct xe_devcoredump_snapshot *ss)
118 {
119 	int i;
120 
121 	xe_guc_ct_snapshot_free(ss->ct);
122 	ss->ct = NULL;
123 
124 	xe_guc_exec_queue_snapshot_free(ss->ge);
125 	ss->ge = NULL;
126 
127 	xe_sched_job_snapshot_free(ss->job);
128 	ss->job = NULL;
129 
130 	for (i = 0; i < XE_NUM_HW_ENGINES; i++)
131 		if (ss->hwe[i]) {
132 			xe_hw_engine_snapshot_free(ss->hwe[i]);
133 			ss->hwe[i] = NULL;
134 		}
135 
136 	xe_vm_snapshot_free(ss->vm);
137 	ss->vm = NULL;
138 }
139 
140 static void xe_devcoredump_deferred_snap_work(struct work_struct *work)
141 {
142 	struct xe_devcoredump_snapshot *ss = container_of(work, typeof(*ss), work);
143 	struct xe_devcoredump *coredump = container_of(ss, typeof(*coredump), snapshot);
144 
145 	/* keep going if fw fails as we still want to save the memory and SW data */
146 	if (xe_force_wake_get(gt_to_fw(ss->gt), XE_FORCEWAKE_ALL))
147 		xe_gt_info(ss->gt, "failed to get forcewake for coredump capture\n");
148 	xe_vm_snapshot_capture_delayed(ss->vm);
149 	xe_guc_exec_queue_snapshot_capture_delayed(ss->ge);
150 	xe_force_wake_put(gt_to_fw(ss->gt), XE_FORCEWAKE_ALL);
151 
152 	/* Calculate devcoredump size */
153 	ss->read.size = __xe_devcoredump_read(NULL, INT_MAX, coredump);
154 
155 	ss->read.buffer = kvmalloc(ss->read.size, GFP_USER);
156 	if (!ss->read.buffer)
157 		return;
158 
159 	__xe_devcoredump_read(ss->read.buffer, ss->read.size, coredump);
160 	xe_devcoredump_snapshot_free(ss);
161 }
162 
163 static ssize_t xe_devcoredump_read(char *buffer, loff_t offset,
164 				   size_t count, void *data, size_t datalen)
165 {
166 	struct xe_devcoredump *coredump = data;
167 	struct xe_devcoredump_snapshot *ss;
168 	ssize_t byte_copied;
169 
170 	if (!coredump)
171 		return -ENODEV;
172 
173 	ss = &coredump->snapshot;
174 
175 	/* Ensure delayed work is captured before continuing */
176 	flush_work(&ss->work);
177 
178 	if (!ss->read.buffer)
179 		return -ENODEV;
180 
181 	if (offset >= ss->read.size)
182 		return 0;
183 
184 	byte_copied = count < ss->read.size - offset ? count :
185 		ss->read.size - offset;
186 	memcpy(buffer, ss->read.buffer + offset, byte_copied);
187 
188 	return byte_copied;
189 }
190 
191 static void xe_devcoredump_free(void *data)
192 {
193 	struct xe_devcoredump *coredump = data;
194 
195 	/* Our device is gone. Nothing to do... */
196 	if (!data || !coredump_to_xe(coredump))
197 		return;
198 
199 	cancel_work_sync(&coredump->snapshot.work);
200 
201 	xe_devcoredump_snapshot_free(&coredump->snapshot);
202 	kvfree(coredump->snapshot.read.buffer);
203 
204 	/* To prevent stale data on next snapshot, clear everything */
205 	memset(&coredump->snapshot, 0, sizeof(coredump->snapshot));
206 	coredump->captured = false;
207 	drm_info(&coredump_to_xe(coredump)->drm,
208 		 "Xe device coredump has been deleted.\n");
209 }
210 
211 static void devcoredump_snapshot(struct xe_devcoredump *coredump,
212 				 struct xe_sched_job *job)
213 {
214 	struct xe_devcoredump_snapshot *ss = &coredump->snapshot;
215 	struct xe_exec_queue *q = job->q;
216 	struct xe_guc *guc = exec_queue_to_guc(q);
217 	struct xe_hw_engine *hwe;
218 	enum xe_hw_engine_id id;
219 	u32 adj_logical_mask = q->logical_mask;
220 	u32 width_mask = (0x1 << q->width) - 1;
221 	const char *process_name = "no process";
222 
223 	int i;
224 	bool cookie;
225 
226 	ss->snapshot_time = ktime_get_real();
227 	ss->boot_time = ktime_get_boottime();
228 
229 	if (q->vm && q->vm->xef)
230 		process_name = q->vm->xef->process_name;
231 	strscpy(ss->process_name, process_name);
232 
233 	ss->gt = q->gt;
234 	INIT_WORK(&ss->work, xe_devcoredump_deferred_snap_work);
235 
236 	cookie = dma_fence_begin_signalling();
237 	for (i = 0; q->width > 1 && i < XE_HW_ENGINE_MAX_INSTANCE;) {
238 		if (adj_logical_mask & BIT(i)) {
239 			adj_logical_mask |= width_mask << i;
240 			i += q->width;
241 		} else {
242 			++i;
243 		}
244 	}
245 
246 	/* keep going if fw fails as we still want to save the memory and SW data */
247 	if (xe_force_wake_get(gt_to_fw(q->gt), XE_FORCEWAKE_ALL))
248 		xe_gt_info(ss->gt, "failed to get forcewake for coredump capture\n");
249 
250 	coredump->snapshot.ct = xe_guc_ct_snapshot_capture(&guc->ct, true);
251 	coredump->snapshot.ge = xe_guc_exec_queue_snapshot_capture(q);
252 	coredump->snapshot.job = xe_sched_job_snapshot_capture(job);
253 	coredump->snapshot.vm = xe_vm_snapshot_capture(q->vm);
254 
255 	for_each_hw_engine(hwe, q->gt, id) {
256 		if (hwe->class != q->hwe->class ||
257 		    !(BIT(hwe->logical_instance) & adj_logical_mask)) {
258 			coredump->snapshot.hwe[id] = NULL;
259 			continue;
260 		}
261 		coredump->snapshot.hwe[id] = xe_hw_engine_snapshot_capture(hwe);
262 	}
263 
264 	queue_work(system_unbound_wq, &ss->work);
265 
266 	xe_force_wake_put(gt_to_fw(q->gt), XE_FORCEWAKE_ALL);
267 	dma_fence_end_signalling(cookie);
268 }
269 
270 /**
271  * xe_devcoredump - Take the required snapshots and initialize coredump device.
272  * @job: The faulty xe_sched_job, where the issue was detected.
273  *
274  * This function should be called at the crash time within the serialized
275  * gt_reset. It is skipped if we still have the core dump device available
276  * with the information of the 'first' snapshot.
277  */
278 void xe_devcoredump(struct xe_sched_job *job)
279 {
280 	struct xe_device *xe = gt_to_xe(job->q->gt);
281 	struct xe_devcoredump *coredump = &xe->devcoredump;
282 
283 	if (coredump->captured) {
284 		drm_dbg(&xe->drm, "Multiple hangs are occurring, but only the first snapshot was taken\n");
285 		return;
286 	}
287 
288 	coredump->captured = true;
289 	devcoredump_snapshot(coredump, job);
290 
291 	drm_info(&xe->drm, "Xe device coredump has been created\n");
292 	drm_info(&xe->drm, "Check your /sys/class/drm/card%d/device/devcoredump/data\n",
293 		 xe->drm.primary->index);
294 
295 	dev_coredumpm_timeout(xe->drm.dev, THIS_MODULE, coredump, 0, GFP_KERNEL,
296 			      xe_devcoredump_read, xe_devcoredump_free,
297 			      XE_COREDUMP_TIMEOUT_JIFFIES);
298 }
299 
300 static void xe_driver_devcoredump_fini(void *arg)
301 {
302 	struct drm_device *drm = arg;
303 
304 	dev_coredump_put(drm->dev);
305 }
306 
307 int xe_devcoredump_init(struct xe_device *xe)
308 {
309 	return devm_add_action_or_reset(xe->drm.dev, xe_driver_devcoredump_fini, &xe->drm);
310 }
311 
312 #endif
313