xref: /linux/drivers/gpu/drm/xe/xe_devcoredump.c (revision 569d7db70e5dcf13fbf072f10e9096577ac1e565)
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43  * information is outdated since the resets might have happened in between.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 static struct xe_device *coredump_to_xe(const struct xe_devcoredump *coredump)
57 {
58 	return container_of(coredump, struct xe_device, devcoredump);
59 }
60 
61 static struct xe_guc *exec_queue_to_guc(struct xe_exec_queue *q)
62 {
63 	return &q->gt->uc.guc;
64 }
65 
66 static void xe_devcoredump_deferred_snap_work(struct work_struct *work)
67 {
68 	struct xe_devcoredump_snapshot *ss = container_of(work, typeof(*ss), work);
69 
70 	/* keep going if fw fails as we still want to save the memory and SW data */
71 	if (xe_force_wake_get(gt_to_fw(ss->gt), XE_FORCEWAKE_ALL))
72 		xe_gt_info(ss->gt, "failed to get forcewake for coredump capture\n");
73 	xe_vm_snapshot_capture_delayed(ss->vm);
74 	xe_guc_exec_queue_snapshot_capture_delayed(ss->ge);
75 	xe_force_wake_put(gt_to_fw(ss->gt), XE_FORCEWAKE_ALL);
76 }
77 
78 static ssize_t xe_devcoredump_read(char *buffer, loff_t offset,
79 				   size_t count, void *data, size_t datalen)
80 {
81 	struct xe_devcoredump *coredump = data;
82 	struct xe_device *xe;
83 	struct xe_devcoredump_snapshot *ss;
84 	struct drm_printer p;
85 	struct drm_print_iterator iter;
86 	struct timespec64 ts;
87 	int i;
88 
89 	if (!coredump)
90 		return -ENODEV;
91 
92 	xe = coredump_to_xe(coredump);
93 	ss = &coredump->snapshot;
94 
95 	/* Ensure delayed work is captured before continuing */
96 	flush_work(&ss->work);
97 
98 	iter.data = buffer;
99 	iter.offset = 0;
100 	iter.start = offset;
101 	iter.remain = count;
102 
103 	p = drm_coredump_printer(&iter);
104 
105 	drm_printf(&p, "**** Xe Device Coredump ****\n");
106 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
107 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
108 
109 	ts = ktime_to_timespec64(ss->snapshot_time);
110 	drm_printf(&p, "Snapshot time: %lld.%09ld\n", ts.tv_sec, ts.tv_nsec);
111 	ts = ktime_to_timespec64(ss->boot_time);
112 	drm_printf(&p, "Uptime: %lld.%09ld\n", ts.tv_sec, ts.tv_nsec);
113 	drm_printf(&p, "Process: %s\n", ss->process_name);
114 	xe_device_snapshot_print(xe, &p);
115 
116 	drm_printf(&p, "\n**** GuC CT ****\n");
117 	xe_guc_ct_snapshot_print(coredump->snapshot.ct, &p);
118 	xe_guc_exec_queue_snapshot_print(coredump->snapshot.ge, &p);
119 
120 	drm_printf(&p, "\n**** Job ****\n");
121 	xe_sched_job_snapshot_print(coredump->snapshot.job, &p);
122 
123 	drm_printf(&p, "\n**** HW Engines ****\n");
124 	for (i = 0; i < XE_NUM_HW_ENGINES; i++)
125 		if (coredump->snapshot.hwe[i])
126 			xe_hw_engine_snapshot_print(coredump->snapshot.hwe[i],
127 						    &p);
128 	drm_printf(&p, "\n**** VM state ****\n");
129 	xe_vm_snapshot_print(coredump->snapshot.vm, &p);
130 
131 	return count - iter.remain;
132 }
133 
134 static void xe_devcoredump_free(void *data)
135 {
136 	struct xe_devcoredump *coredump = data;
137 	int i;
138 
139 	/* Our device is gone. Nothing to do... */
140 	if (!data || !coredump_to_xe(coredump))
141 		return;
142 
143 	cancel_work_sync(&coredump->snapshot.work);
144 
145 	xe_guc_ct_snapshot_free(coredump->snapshot.ct);
146 	xe_guc_exec_queue_snapshot_free(coredump->snapshot.ge);
147 	xe_sched_job_snapshot_free(coredump->snapshot.job);
148 	for (i = 0; i < XE_NUM_HW_ENGINES; i++)
149 		if (coredump->snapshot.hwe[i])
150 			xe_hw_engine_snapshot_free(coredump->snapshot.hwe[i]);
151 	xe_vm_snapshot_free(coredump->snapshot.vm);
152 
153 	/* To prevent stale data on next snapshot, clear everything */
154 	memset(&coredump->snapshot, 0, sizeof(coredump->snapshot));
155 	coredump->captured = false;
156 	drm_info(&coredump_to_xe(coredump)->drm,
157 		 "Xe device coredump has been deleted.\n");
158 }
159 
160 static void devcoredump_snapshot(struct xe_devcoredump *coredump,
161 				 struct xe_sched_job *job)
162 {
163 	struct xe_devcoredump_snapshot *ss = &coredump->snapshot;
164 	struct xe_exec_queue *q = job->q;
165 	struct xe_guc *guc = exec_queue_to_guc(q);
166 	struct xe_hw_engine *hwe;
167 	enum xe_hw_engine_id id;
168 	u32 adj_logical_mask = q->logical_mask;
169 	u32 width_mask = (0x1 << q->width) - 1;
170 	const char *process_name = "no process";
171 	struct task_struct *task = NULL;
172 
173 	int i;
174 	bool cookie;
175 
176 	ss->snapshot_time = ktime_get_real();
177 	ss->boot_time = ktime_get_boottime();
178 
179 	if (q->vm && q->vm->xef) {
180 		task = get_pid_task(q->vm->xef->drm->pid, PIDTYPE_PID);
181 		if (task)
182 			process_name = task->comm;
183 	}
184 	strscpy(ss->process_name, process_name);
185 	if (task)
186 		put_task_struct(task);
187 
188 	ss->gt = q->gt;
189 	INIT_WORK(&ss->work, xe_devcoredump_deferred_snap_work);
190 
191 	cookie = dma_fence_begin_signalling();
192 	for (i = 0; q->width > 1 && i < XE_HW_ENGINE_MAX_INSTANCE;) {
193 		if (adj_logical_mask & BIT(i)) {
194 			adj_logical_mask |= width_mask << i;
195 			i += q->width;
196 		} else {
197 			++i;
198 		}
199 	}
200 
201 	/* keep going if fw fails as we still want to save the memory and SW data */
202 	if (xe_force_wake_get(gt_to_fw(q->gt), XE_FORCEWAKE_ALL))
203 		xe_gt_info(ss->gt, "failed to get forcewake for coredump capture\n");
204 
205 	coredump->snapshot.ct = xe_guc_ct_snapshot_capture(&guc->ct, true);
206 	coredump->snapshot.ge = xe_guc_exec_queue_snapshot_capture(q);
207 	coredump->snapshot.job = xe_sched_job_snapshot_capture(job);
208 	coredump->snapshot.vm = xe_vm_snapshot_capture(q->vm);
209 
210 	for_each_hw_engine(hwe, q->gt, id) {
211 		if (hwe->class != q->hwe->class ||
212 		    !(BIT(hwe->logical_instance) & adj_logical_mask)) {
213 			coredump->snapshot.hwe[id] = NULL;
214 			continue;
215 		}
216 		coredump->snapshot.hwe[id] = xe_hw_engine_snapshot_capture(hwe);
217 	}
218 
219 	queue_work(system_unbound_wq, &ss->work);
220 
221 	xe_force_wake_put(gt_to_fw(q->gt), XE_FORCEWAKE_ALL);
222 	dma_fence_end_signalling(cookie);
223 }
224 
225 /**
226  * xe_devcoredump - Take the required snapshots and initialize coredump device.
227  * @job: The faulty xe_sched_job, where the issue was detected.
228  *
229  * This function should be called at the crash time within the serialized
230  * gt_reset. It is skipped if we still have the core dump device available
231  * with the information of the 'first' snapshot.
232  */
233 void xe_devcoredump(struct xe_sched_job *job)
234 {
235 	struct xe_device *xe = gt_to_xe(job->q->gt);
236 	struct xe_devcoredump *coredump = &xe->devcoredump;
237 
238 	if (coredump->captured) {
239 		drm_dbg(&xe->drm, "Multiple hangs are occurring, but only the first snapshot was taken\n");
240 		return;
241 	}
242 
243 	coredump->captured = true;
244 	devcoredump_snapshot(coredump, job);
245 
246 	drm_info(&xe->drm, "Xe device coredump has been created\n");
247 	drm_info(&xe->drm, "Check your /sys/class/drm/card%d/device/devcoredump/data\n",
248 		 xe->drm.primary->index);
249 
250 	dev_coredumpm(xe->drm.dev, THIS_MODULE, coredump, 0, GFP_KERNEL,
251 		      xe_devcoredump_read, xe_devcoredump_free);
252 }
253 
254 static void xe_driver_devcoredump_fini(void *arg)
255 {
256 	struct drm_device *drm = arg;
257 
258 	dev_coredump_put(drm->dev);
259 }
260 
261 int xe_devcoredump_init(struct xe_device *xe)
262 {
263 	return devm_add_action_or_reset(xe->drm.dev, xe_driver_devcoredump_fini, &xe->drm);
264 }
265 #endif
266