xref: /linux/drivers/gpu/drm/i915/gt/uc/selftest_guc.c (revision 8a79db5e83a5d52c74e6f3c40d6f312cf899213e)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2017 Intel Corporation
4  */
5 
6 #include "i915_selftest.h"
7 #include "gem/i915_gem_pm.h"
8 
9 /* max doorbell number + negative test for each client type */
10 #define ATTEMPTS (GUC_NUM_DOORBELLS + GUC_CLIENT_PRIORITY_NUM)
11 
12 static struct intel_guc_client *clients[ATTEMPTS];
13 
14 static bool available_dbs(struct intel_guc *guc, u32 priority)
15 {
16 	unsigned long offset;
17 	unsigned long end;
18 	u16 id;
19 
20 	/* first half is used for normal priority, second half for high */
21 	offset = 0;
22 	end = GUC_NUM_DOORBELLS / 2;
23 	if (priority <= GUC_CLIENT_PRIORITY_HIGH) {
24 		offset = end;
25 		end += offset;
26 	}
27 
28 	id = find_next_zero_bit(guc->doorbell_bitmap, end, offset);
29 	if (id < end)
30 		return true;
31 
32 	return false;
33 }
34 
35 static int check_all_doorbells(struct intel_guc *guc)
36 {
37 	u16 db_id;
38 
39 	pr_info_once("Max number of doorbells: %d", GUC_NUM_DOORBELLS);
40 	for (db_id = 0; db_id < GUC_NUM_DOORBELLS; ++db_id) {
41 		if (!doorbell_ok(guc, db_id)) {
42 			pr_err("doorbell %d, not ok\n", db_id);
43 			return -EIO;
44 		}
45 	}
46 
47 	return 0;
48 }
49 
50 static int ring_doorbell_nop(struct intel_guc_client *client)
51 {
52 	struct guc_process_desc *desc = __get_process_desc(client);
53 	int err;
54 
55 	client->use_nop_wqi = true;
56 
57 	spin_lock_irq(&client->wq_lock);
58 
59 	guc_wq_item_append(client, 0, 0, 0, 0);
60 	guc_ring_doorbell(client);
61 
62 	spin_unlock_irq(&client->wq_lock);
63 
64 	client->use_nop_wqi = false;
65 
66 	/* if there are no issues GuC will update the WQ head and keep the
67 	 * WQ in active status
68 	 */
69 	err = wait_for(READ_ONCE(desc->head) == READ_ONCE(desc->tail), 10);
70 	if (err) {
71 		pr_err("doorbell %u ring failed!\n", client->doorbell_id);
72 		return -EIO;
73 	}
74 
75 	if (desc->wq_status != WQ_STATUS_ACTIVE) {
76 		pr_err("doorbell %u ring put WQ in bad state (%u)!\n",
77 		       client->doorbell_id, desc->wq_status);
78 		return -EIO;
79 	}
80 
81 	return 0;
82 }
83 
84 /*
85  * Basic client sanity check, handy to validate create_clients.
86  */
87 static int validate_client(struct intel_guc_client *client, int client_priority)
88 {
89 	if (client->priority != client_priority ||
90 	    client->doorbell_id == GUC_DOORBELL_INVALID)
91 		return -EINVAL;
92 	else
93 		return 0;
94 }
95 
96 static bool client_doorbell_in_sync(struct intel_guc_client *client)
97 {
98 	return !client || doorbell_ok(client->guc, client->doorbell_id);
99 }
100 
101 /*
102  * Check that we're able to synchronize guc_clients with their doorbells
103  *
104  * We're creating clients and reserving doorbells once, at module load. During
105  * module lifetime, GuC, doorbell HW, and i915 state may go out of sync due to
106  * GuC being reset. In other words - GuC clients are still around, but the
107  * status of their doorbells may be incorrect. This is the reason behind
108  * validating that the doorbells status expected by the driver matches what the
109  * GuC/HW have.
110  */
111 static int igt_guc_clients(void *arg)
112 {
113 	struct intel_gt *gt = arg;
114 	struct intel_guc *guc = &gt->uc.guc;
115 	intel_wakeref_t wakeref;
116 	int err = 0;
117 
118 	GEM_BUG_ON(!HAS_GT_UC(gt->i915));
119 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
120 
121 	err = check_all_doorbells(guc);
122 	if (err)
123 		goto unlock;
124 
125 	/*
126 	 * Get rid of clients created during driver load because the test will
127 	 * recreate them.
128 	 */
129 	guc_clients_disable(guc);
130 	guc_clients_destroy(guc);
131 	if (guc->execbuf_client) {
132 		pr_err("guc_clients_destroy lied!\n");
133 		err = -EINVAL;
134 		goto unlock;
135 	}
136 
137 	err = guc_clients_create(guc);
138 	if (err) {
139 		pr_err("Failed to create clients\n");
140 		goto unlock;
141 	}
142 	GEM_BUG_ON(!guc->execbuf_client);
143 
144 	err = validate_client(guc->execbuf_client,
145 			      GUC_CLIENT_PRIORITY_KMD_NORMAL);
146 	if (err) {
147 		pr_err("execbug client validation failed\n");
148 		goto out;
149 	}
150 
151 	/* the client should now have reserved a doorbell */
152 	if (!has_doorbell(guc->execbuf_client)) {
153 		pr_err("guc_clients_create didn't reserve doorbells\n");
154 		err = -EINVAL;
155 		goto out;
156 	}
157 
158 	/* Now enable the clients */
159 	guc_clients_enable(guc);
160 
161 	/* each client should now have received a doorbell */
162 	if (!client_doorbell_in_sync(guc->execbuf_client)) {
163 		pr_err("failed to initialize the doorbells\n");
164 		err = -EINVAL;
165 		goto out;
166 	}
167 
168 	/*
169 	 * Basic test - an attempt to reallocate a valid doorbell to the
170 	 * client it is currently assigned should not cause a failure.
171 	 */
172 	err = create_doorbell(guc->execbuf_client);
173 
174 out:
175 	/*
176 	 * Leave clean state for other test, plus the driver always destroy the
177 	 * clients during unload.
178 	 */
179 	guc_clients_disable(guc);
180 	guc_clients_destroy(guc);
181 	guc_clients_create(guc);
182 	guc_clients_enable(guc);
183 unlock:
184 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
185 	return err;
186 }
187 
188 /*
189  * Create as many clients as number of doorbells. Note that there's already
190  * client(s)/doorbell(s) created during driver load, but this test creates
191  * its own and do not interact with the existing ones.
192  */
193 static int igt_guc_doorbells(void *arg)
194 {
195 	struct intel_gt *gt = arg;
196 	struct intel_guc *guc = &gt->uc.guc;
197 	intel_wakeref_t wakeref;
198 	int i, err = 0;
199 	u16 db_id;
200 
201 	GEM_BUG_ON(!HAS_GT_UC(gt->i915));
202 	wakeref = intel_runtime_pm_get(gt->uncore->rpm);
203 
204 	err = check_all_doorbells(guc);
205 	if (err)
206 		goto unlock;
207 
208 	for (i = 0; i < ATTEMPTS; i++) {
209 		clients[i] = guc_client_alloc(guc, i % GUC_CLIENT_PRIORITY_NUM);
210 
211 		if (!clients[i]) {
212 			pr_err("[%d] No guc client\n", i);
213 			err = -EINVAL;
214 			goto out;
215 		}
216 
217 		if (IS_ERR(clients[i])) {
218 			if (PTR_ERR(clients[i]) != -ENOSPC) {
219 				pr_err("[%d] unexpected error\n", i);
220 				err = PTR_ERR(clients[i]);
221 				goto out;
222 			}
223 
224 			if (available_dbs(guc, i % GUC_CLIENT_PRIORITY_NUM)) {
225 				pr_err("[%d] non-db related alloc fail\n", i);
226 				err = -EINVAL;
227 				goto out;
228 			}
229 
230 			/* expected, ran out of dbs for this client type */
231 			continue;
232 		}
233 
234 		/*
235 		 * The check below is only valid because we keep a doorbell
236 		 * assigned during the whole life of the client.
237 		 */
238 		if (clients[i]->stage_id >= GUC_NUM_DOORBELLS) {
239 			pr_err("[%d] more clients than doorbells (%d >= %d)\n",
240 			       i, clients[i]->stage_id, GUC_NUM_DOORBELLS);
241 			err = -EINVAL;
242 			goto out;
243 		}
244 
245 		err = validate_client(clients[i], i % GUC_CLIENT_PRIORITY_NUM);
246 		if (err) {
247 			pr_err("[%d] client_alloc sanity check failed!\n", i);
248 			err = -EINVAL;
249 			goto out;
250 		}
251 
252 		db_id = clients[i]->doorbell_id;
253 
254 		err = __guc_client_enable(clients[i]);
255 		if (err) {
256 			pr_err("[%d] Failed to create a doorbell\n", i);
257 			goto out;
258 		}
259 
260 		/* doorbell id shouldn't change, we are holding the mutex */
261 		if (db_id != clients[i]->doorbell_id) {
262 			pr_err("[%d] doorbell id changed (%d != %d)\n",
263 			       i, db_id, clients[i]->doorbell_id);
264 			err = -EINVAL;
265 			goto out;
266 		}
267 
268 		err = check_all_doorbells(guc);
269 		if (err)
270 			goto out;
271 
272 		err = ring_doorbell_nop(clients[i]);
273 		if (err)
274 			goto out;
275 	}
276 
277 out:
278 	for (i = 0; i < ATTEMPTS; i++)
279 		if (!IS_ERR_OR_NULL(clients[i])) {
280 			__guc_client_disable(clients[i]);
281 			guc_client_free(clients[i]);
282 		}
283 unlock:
284 	intel_runtime_pm_put(gt->uncore->rpm, wakeref);
285 	return err;
286 }
287 
288 int intel_guc_live_selftest(struct drm_i915_private *i915)
289 {
290 	static const struct i915_subtest tests[] = {
291 		SUBTEST(igt_guc_clients),
292 		SUBTEST(igt_guc_doorbells),
293 	};
294 
295 	if (!USES_GUC_SUBMISSION(i915))
296 		return 0;
297 
298 	return intel_gt_live_subtests(tests, &i915->gt);
299 }
300