xref: /linux/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c (revision 8934827db5403eae57d4537114a9ff88b0a8460f)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2016-2019 Intel Corporation
4  */
5 
6 #include <linux/circ_buf.h>
7 #include <linux/ktime.h>
8 #include <linux/string_helpers.h>
9 #include <linux/time64.h>
10 #include <linux/timekeeping.h>
11 
12 #include "i915_drv.h"
13 #include "i915_wait_util.h"
14 #include "intel_guc_ct.h"
15 #include "intel_guc_print.h"
16 
17 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG)
18 enum {
19 	CT_DEAD_ALIVE = 0,
20 	CT_DEAD_SETUP,
21 	CT_DEAD_WRITE,
22 	CT_DEAD_DEADLOCK,
23 	CT_DEAD_H2G_HAS_ROOM,
24 	CT_DEAD_READ,
25 	CT_DEAD_PROCESS_FAILED,
26 };
27 
28 static void ct_dead_ct_worker_func(struct work_struct *w);
29 
30 #define CT_DEAD(ct, reason)	\
31 	do { \
32 		if (!(ct)->dead_ct_reported) { \
33 			(ct)->dead_ct_reason |= 1 << CT_DEAD_##reason; \
34 			queue_work(system_unbound_wq, &(ct)->dead_ct_worker); \
35 		} \
36 	} while (0)
37 #else
38 #define CT_DEAD(ct, reason)	do { } while (0)
39 #endif
40 
ct_to_guc(struct intel_guc_ct * ct)41 static inline struct intel_guc *ct_to_guc(struct intel_guc_ct *ct)
42 {
43 	return container_of(ct, struct intel_guc, ct);
44 }
45 
46 #define CT_ERROR(_ct, _fmt, ...) \
47 	guc_err(ct_to_guc(_ct), "CT: " _fmt, ##__VA_ARGS__)
48 #ifdef CONFIG_DRM_I915_DEBUG_GUC
49 #define CT_DEBUG(_ct, _fmt, ...) \
50 	guc_dbg(ct_to_guc(_ct), "CT: " _fmt, ##__VA_ARGS__)
51 #else
52 #define CT_DEBUG(...)	do { } while (0)
53 #endif
54 #define CT_PROBE_ERROR(_ct, _fmt, ...) \
55 	guc_probe_error(ct_to_guc(ct), "CT: " _fmt, ##__VA_ARGS__)
56 
57 /**
58  * DOC: CTB Blob
59  *
60  * We allocate single blob to hold both CTB descriptors and buffers:
61  *
62  *      +--------+-----------------------------------------------+------+
63  *      | offset | contents                                      | size |
64  *      +========+===============================================+======+
65  *      | 0x0000 | H2G `CTB Descriptor`_ (send)                  |      |
66  *      +--------+-----------------------------------------------+  4K  |
67  *      | 0x0800 | G2H `CTB Descriptor`_ (recv)                  |      |
68  *      +--------+-----------------------------------------------+------+
69  *      | 0x1000 | H2G `CT Buffer`_ (send)                       | n*4K |
70  *      |        |                                               |      |
71  *      +--------+-----------------------------------------------+------+
72  *      | 0x1000 | G2H `CT Buffer`_ (recv)                       | m*4K |
73  *      | + n*4K |                                               |      |
74  *      +--------+-----------------------------------------------+------+
75  *
76  * Size of each `CT Buffer`_ must be multiple of 4K.
77  * We don't expect too many messages in flight at any time, unless we are
78  * using the GuC submission. In that case each request requires a minimum
79  * 2 dwords which gives us a maximum 256 queue'd requests. Hopefully this
80  * enough space to avoid backpressure on the driver. We increase the size
81  * of the receive buffer (relative to the send) to ensure a G2H response
82  * CTB has a landing spot.
83  */
84 #define CTB_DESC_SIZE		ALIGN(sizeof(struct guc_ct_buffer_desc), SZ_2K)
85 #define CTB_H2G_BUFFER_SIZE	(SZ_4K)
86 #define CTB_G2H_BUFFER_SIZE	(4 * CTB_H2G_BUFFER_SIZE)
87 #define G2H_ROOM_BUFFER_SIZE	(CTB_G2H_BUFFER_SIZE / 4)
88 
89 struct ct_request {
90 	struct list_head link;
91 	u32 fence;
92 	u32 status;
93 	u32 response_len;
94 	u32 *response_buf;
95 };
96 
97 struct ct_incoming_msg {
98 	struct list_head link;
99 	u32 size;
100 	u32 msg[] __counted_by(size);
101 };
102 
103 enum { CTB_SEND = 0, CTB_RECV = 1 };
104 
105 enum { CTB_OWNER_HOST = 0 };
106 
107 /*
108  * Some H2G commands involve a synchronous response that the driver needs
109  * to wait for. In such cases, a timeout is required to prevent the driver
110  * from waiting forever in the case of an error (either no error response
111  * is defined in the protocol or something has died and requires a reset).
112  * The specific command may be defined as having a time bound response but
113  * the CT is a queue and that time guarantee only starts from the point
114  * when the command reaches the head of the queue and is processed by GuC.
115  *
116  * Ideally there would be a helper to report the progress of a given
117  * command through the CT. However, that would require a significant
118  * amount of work in the CT layer. In the meantime, provide a reasonable
119  * estimation of the worst case latency it should take for the entire
120  * queue to drain. And therefore, how long a caller should wait before
121  * giving up on their request. The current estimate is based on empirical
122  * measurement of a test that fills the buffer with context creation and
123  * destruction requests as they seem to be the slowest operation.
124  */
intel_guc_ct_max_queue_time_jiffies(void)125 long intel_guc_ct_max_queue_time_jiffies(void)
126 {
127 	/*
128 	 * A 4KB buffer full of context destroy commands takes a little
129 	 * over a second to process so bump that to 2s to be super safe.
130 	 */
131 	return (CTB_H2G_BUFFER_SIZE * HZ) / SZ_2K;
132 }
133 
134 static void ct_receive_tasklet_func(struct tasklet_struct *t);
135 static void ct_incoming_request_worker_func(struct work_struct *w);
136 
137 /**
138  * intel_guc_ct_init_early - Initialize CT state without requiring device access
139  * @ct: pointer to CT struct
140  */
intel_guc_ct_init_early(struct intel_guc_ct * ct)141 void intel_guc_ct_init_early(struct intel_guc_ct *ct)
142 {
143 	spin_lock_init(&ct->ctbs.send.lock);
144 	spin_lock_init(&ct->ctbs.recv.lock);
145 	spin_lock_init(&ct->requests.lock);
146 	INIT_LIST_HEAD(&ct->requests.pending);
147 	INIT_LIST_HEAD(&ct->requests.incoming);
148 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG)
149 	INIT_WORK(&ct->dead_ct_worker, ct_dead_ct_worker_func);
150 #endif
151 	INIT_WORK(&ct->requests.worker, ct_incoming_request_worker_func);
152 	tasklet_setup(&ct->receive_tasklet, ct_receive_tasklet_func);
153 	init_waitqueue_head(&ct->wq);
154 }
155 
guc_ct_buffer_desc_init(struct guc_ct_buffer_desc * desc)156 static void guc_ct_buffer_desc_init(struct guc_ct_buffer_desc *desc)
157 {
158 	memset(desc, 0, sizeof(*desc));
159 }
160 
guc_ct_buffer_reset(struct intel_guc_ct_buffer * ctb)161 static void guc_ct_buffer_reset(struct intel_guc_ct_buffer *ctb)
162 {
163 	u32 space;
164 
165 	ctb->broken = false;
166 	ctb->tail = 0;
167 	ctb->head = 0;
168 	space = CIRC_SPACE(ctb->tail, ctb->head, ctb->size) - ctb->resv_space;
169 	atomic_set(&ctb->space, space);
170 
171 	guc_ct_buffer_desc_init(ctb->desc);
172 }
173 
guc_ct_buffer_init(struct intel_guc_ct_buffer * ctb,struct guc_ct_buffer_desc * desc,u32 * cmds,u32 size_in_bytes,u32 resv_space)174 static void guc_ct_buffer_init(struct intel_guc_ct_buffer *ctb,
175 			       struct guc_ct_buffer_desc *desc,
176 			       u32 *cmds, u32 size_in_bytes, u32 resv_space)
177 {
178 	GEM_BUG_ON(size_in_bytes % 4);
179 
180 	ctb->desc = desc;
181 	ctb->cmds = cmds;
182 	ctb->size = size_in_bytes / 4;
183 	ctb->resv_space = resv_space / 4;
184 
185 	guc_ct_buffer_reset(ctb);
186 }
187 
guc_action_control_ctb(struct intel_guc * guc,u32 control)188 static int guc_action_control_ctb(struct intel_guc *guc, u32 control)
189 {
190 	u32 request[HOST2GUC_CONTROL_CTB_REQUEST_MSG_LEN] = {
191 		FIELD_PREP(GUC_HXG_MSG_0_ORIGIN, GUC_HXG_ORIGIN_HOST) |
192 		FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_REQUEST) |
193 		FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION, GUC_ACTION_HOST2GUC_CONTROL_CTB),
194 		FIELD_PREP(HOST2GUC_CONTROL_CTB_REQUEST_MSG_1_CONTROL, control),
195 	};
196 	int ret;
197 
198 	GEM_BUG_ON(control != GUC_CTB_CONTROL_DISABLE && control != GUC_CTB_CONTROL_ENABLE);
199 
200 	/* CT control must go over MMIO */
201 	ret = intel_guc_send_mmio(guc, request, ARRAY_SIZE(request), NULL, 0);
202 
203 	return ret > 0 ? -EPROTO : ret;
204 }
205 
ct_control_enable(struct intel_guc_ct * ct,bool enable)206 static int ct_control_enable(struct intel_guc_ct *ct, bool enable)
207 {
208 	int err;
209 
210 	err = guc_action_control_ctb(ct_to_guc(ct), enable ?
211 				     GUC_CTB_CONTROL_ENABLE : GUC_CTB_CONTROL_DISABLE);
212 	if (unlikely(err))
213 		CT_PROBE_ERROR(ct, "Failed to control/%s CTB (%pe)\n",
214 			       str_enable_disable(enable), ERR_PTR(err));
215 
216 	return err;
217 }
218 
ct_register_buffer(struct intel_guc_ct * ct,bool send,u32 desc_addr,u32 buff_addr,u32 size)219 static int ct_register_buffer(struct intel_guc_ct *ct, bool send,
220 			      u32 desc_addr, u32 buff_addr, u32 size)
221 {
222 	int err;
223 
224 	err = intel_guc_self_cfg64(ct_to_guc(ct), send ?
225 				   GUC_KLV_SELF_CFG_H2G_CTB_DESCRIPTOR_ADDR_KEY :
226 				   GUC_KLV_SELF_CFG_G2H_CTB_DESCRIPTOR_ADDR_KEY,
227 				   desc_addr);
228 	if (unlikely(err))
229 		goto failed;
230 
231 	err = intel_guc_self_cfg64(ct_to_guc(ct), send ?
232 				   GUC_KLV_SELF_CFG_H2G_CTB_ADDR_KEY :
233 				   GUC_KLV_SELF_CFG_G2H_CTB_ADDR_KEY,
234 				   buff_addr);
235 	if (unlikely(err))
236 		goto failed;
237 
238 	err = intel_guc_self_cfg32(ct_to_guc(ct), send ?
239 				   GUC_KLV_SELF_CFG_H2G_CTB_SIZE_KEY :
240 				   GUC_KLV_SELF_CFG_G2H_CTB_SIZE_KEY,
241 				   size);
242 	if (unlikely(err))
243 failed:
244 		CT_PROBE_ERROR(ct, "Failed to register %s buffer (%pe)\n",
245 			       send ? "SEND" : "RECV", ERR_PTR(err));
246 
247 	return err;
248 }
249 
250 /**
251  * intel_guc_ct_init - Init buffer-based communication
252  * @ct: pointer to CT struct
253  *
254  * Allocate memory required for buffer-based communication.
255  *
256  * Return: 0 on success, a negative errno code on failure.
257  */
intel_guc_ct_init(struct intel_guc_ct * ct)258 int intel_guc_ct_init(struct intel_guc_ct *ct)
259 {
260 	struct intel_guc *guc = ct_to_guc(ct);
261 	struct guc_ct_buffer_desc *desc;
262 	u32 blob_size;
263 	u32 cmds_size;
264 	u32 resv_space;
265 	void *blob;
266 	u32 *cmds;
267 	int err;
268 
269 	GEM_BUG_ON(ct->vma);
270 
271 	blob_size = 2 * CTB_DESC_SIZE + CTB_H2G_BUFFER_SIZE + CTB_G2H_BUFFER_SIZE;
272 	err = intel_guc_allocate_and_map_vma(guc, blob_size, &ct->vma, &blob);
273 	if (unlikely(err)) {
274 		CT_PROBE_ERROR(ct, "Failed to allocate %u for CTB data (%pe)\n",
275 			       blob_size, ERR_PTR(err));
276 		return err;
277 	}
278 
279 	CT_DEBUG(ct, "base=%#x size=%u\n", intel_guc_ggtt_offset(guc, ct->vma), blob_size);
280 
281 	/* store pointers to desc and cmds for send ctb */
282 	desc = blob;
283 	cmds = blob + 2 * CTB_DESC_SIZE;
284 	cmds_size = CTB_H2G_BUFFER_SIZE;
285 	resv_space = 0;
286 	CT_DEBUG(ct, "%s desc %#tx cmds %#tx size %u/%u\n", "send",
287 		 ptrdiff(desc, blob), ptrdiff(cmds, blob), cmds_size,
288 		 resv_space);
289 
290 	guc_ct_buffer_init(&ct->ctbs.send, desc, cmds, cmds_size, resv_space);
291 
292 	/* store pointers to desc and cmds for recv ctb */
293 	desc = blob + CTB_DESC_SIZE;
294 	cmds = blob + 2 * CTB_DESC_SIZE + CTB_H2G_BUFFER_SIZE;
295 	cmds_size = CTB_G2H_BUFFER_SIZE;
296 	resv_space = G2H_ROOM_BUFFER_SIZE;
297 	CT_DEBUG(ct, "%s desc %#tx cmds %#tx size %u/%u\n", "recv",
298 		 ptrdiff(desc, blob), ptrdiff(cmds, blob), cmds_size,
299 		 resv_space);
300 
301 	guc_ct_buffer_init(&ct->ctbs.recv, desc, cmds, cmds_size, resv_space);
302 
303 	return 0;
304 }
305 ALLOW_ERROR_INJECTION(intel_guc_ct_init, ERRNO);
306 
307 /**
308  * intel_guc_ct_fini - Fini buffer-based communication
309  * @ct: pointer to CT struct
310  *
311  * Deallocate memory required for buffer-based communication.
312  */
intel_guc_ct_fini(struct intel_guc_ct * ct)313 void intel_guc_ct_fini(struct intel_guc_ct *ct)
314 {
315 	GEM_BUG_ON(ct->enabled);
316 
317 	tasklet_kill(&ct->receive_tasklet);
318 	i915_vma_unpin_and_release(&ct->vma, I915_VMA_RELEASE_MAP);
319 	memset(ct, 0, sizeof(*ct));
320 }
321 
322 /**
323  * intel_guc_ct_enable - Enable buffer based command transport.
324  * @ct: pointer to CT struct
325  *
326  * Return: 0 on success, a negative errno code on failure.
327  */
intel_guc_ct_enable(struct intel_guc_ct * ct)328 int intel_guc_ct_enable(struct intel_guc_ct *ct)
329 {
330 	struct intel_guc *guc = ct_to_guc(ct);
331 	u32 base, desc, cmds, size;
332 	void *blob;
333 	int err;
334 
335 	GEM_BUG_ON(ct->enabled);
336 
337 	/* vma should be already allocated and map'ed */
338 	GEM_BUG_ON(!ct->vma);
339 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(ct->vma->obj));
340 	base = intel_guc_ggtt_offset(guc, ct->vma);
341 
342 	/* blob should start with send descriptor */
343 	blob = __px_vaddr(ct->vma->obj);
344 	GEM_BUG_ON(blob != ct->ctbs.send.desc);
345 
346 	/* (re)initialize descriptors */
347 	guc_ct_buffer_reset(&ct->ctbs.send);
348 	guc_ct_buffer_reset(&ct->ctbs.recv);
349 
350 	/*
351 	 * Register both CT buffers starting with RECV buffer.
352 	 * Descriptors are in first half of the blob.
353 	 */
354 	desc = base + ptrdiff(ct->ctbs.recv.desc, blob);
355 	cmds = base + ptrdiff(ct->ctbs.recv.cmds, blob);
356 	size = ct->ctbs.recv.size * 4;
357 	err = ct_register_buffer(ct, false, desc, cmds, size);
358 	if (unlikely(err))
359 		goto err_out;
360 
361 	desc = base + ptrdiff(ct->ctbs.send.desc, blob);
362 	cmds = base + ptrdiff(ct->ctbs.send.cmds, blob);
363 	size = ct->ctbs.send.size * 4;
364 	err = ct_register_buffer(ct, true, desc, cmds, size);
365 	if (unlikely(err))
366 		goto err_out;
367 
368 	err = ct_control_enable(ct, true);
369 	if (unlikely(err))
370 		goto err_out;
371 
372 	ct->enabled = true;
373 	ct->stall_time = KTIME_MAX;
374 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG)
375 	ct->dead_ct_reported = false;
376 	ct->dead_ct_reason = CT_DEAD_ALIVE;
377 #endif
378 
379 	return 0;
380 
381 err_out:
382 	CT_PROBE_ERROR(ct, "Failed to enable CTB (%pe)\n", ERR_PTR(err));
383 	CT_DEAD(ct, SETUP);
384 	return err;
385 }
386 
387 /**
388  * intel_guc_ct_disable - Disable buffer based command transport.
389  * @ct: pointer to CT struct
390  */
intel_guc_ct_disable(struct intel_guc_ct * ct)391 void intel_guc_ct_disable(struct intel_guc_ct *ct)
392 {
393 	struct intel_guc *guc = ct_to_guc(ct);
394 
395 	GEM_BUG_ON(!ct->enabled);
396 
397 	ct->enabled = false;
398 
399 	if (intel_guc_is_fw_running(guc)) {
400 		ct_control_enable(ct, false);
401 	}
402 }
403 
404 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
ct_track_lost_and_found(struct intel_guc_ct * ct,u32 fence,u32 action)405 static void ct_track_lost_and_found(struct intel_guc_ct *ct, u32 fence, u32 action)
406 {
407 	unsigned int lost = fence % ARRAY_SIZE(ct->requests.lost_and_found);
408 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GUC)
409 	unsigned long entries[SZ_32];
410 	unsigned int n;
411 
412 	n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
413 
414 	/* May be called under spinlock, so avoid sleeping */
415 	ct->requests.lost_and_found[lost].stack = stack_depot_save(entries, n, GFP_NOWAIT);
416 #endif
417 	ct->requests.lost_and_found[lost].fence = fence;
418 	ct->requests.lost_and_found[lost].action = action;
419 }
420 #endif
421 
ct_get_next_fence(struct intel_guc_ct * ct)422 static u32 ct_get_next_fence(struct intel_guc_ct *ct)
423 {
424 	/* For now it's trivial */
425 	return ++ct->requests.last_fence;
426 }
427 
ct_write(struct intel_guc_ct * ct,const u32 * action,u32 len,u32 fence,u32 flags)428 static int ct_write(struct intel_guc_ct *ct,
429 		    const u32 *action,
430 		    u32 len /* in dwords */,
431 		    u32 fence, u32 flags)
432 {
433 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
434 	struct guc_ct_buffer_desc *desc = ctb->desc;
435 	u32 tail = ctb->tail;
436 	u32 size = ctb->size;
437 	u32 header;
438 	u32 hxg;
439 	u32 type;
440 	u32 *cmds = ctb->cmds;
441 	unsigned int i;
442 
443 	if (unlikely(desc->status))
444 		goto corrupted;
445 
446 	GEM_BUG_ON(tail > size);
447 
448 #ifdef CONFIG_DRM_I915_DEBUG_GUC
449 	if (unlikely(tail != READ_ONCE(desc->tail))) {
450 		CT_ERROR(ct, "Tail was modified %u != %u\n",
451 			 desc->tail, tail);
452 		desc->status |= GUC_CTB_STATUS_MISMATCH;
453 		goto corrupted;
454 	}
455 	if (unlikely(READ_ONCE(desc->head) >= size)) {
456 		CT_ERROR(ct, "Invalid head offset %u >= %u)\n",
457 			 desc->head, size);
458 		desc->status |= GUC_CTB_STATUS_OVERFLOW;
459 		goto corrupted;
460 	}
461 #endif
462 
463 	/*
464 	 * dw0: CT header (including fence)
465 	 * dw1: HXG header (including action code)
466 	 * dw2+: action data
467 	 */
468 	header = FIELD_PREP(GUC_CTB_MSG_0_FORMAT, GUC_CTB_FORMAT_HXG) |
469 		 FIELD_PREP(GUC_CTB_MSG_0_NUM_DWORDS, len) |
470 		 FIELD_PREP(GUC_CTB_MSG_0_FENCE, fence);
471 
472 	type = (flags & INTEL_GUC_CT_SEND_NB) ? GUC_HXG_TYPE_FAST_REQUEST :
473 		GUC_HXG_TYPE_REQUEST;
474 	hxg = FIELD_PREP(GUC_HXG_MSG_0_TYPE, type) |
475 		FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION |
476 			   GUC_HXG_REQUEST_MSG_0_DATA0, action[0]);
477 
478 	CT_DEBUG(ct, "writing (tail %u) %*ph %*ph %*ph\n",
479 		 tail, 4, &header, 4, &hxg, 4 * (len - 1), &action[1]);
480 
481 	cmds[tail] = header;
482 	tail = (tail + 1) % size;
483 
484 	cmds[tail] = hxg;
485 	tail = (tail + 1) % size;
486 
487 	for (i = 1; i < len; i++) {
488 		cmds[tail] = action[i];
489 		tail = (tail + 1) % size;
490 	}
491 	GEM_BUG_ON(tail > size);
492 
493 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
494 	ct_track_lost_and_found(ct, fence,
495 				FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, action[0]));
496 #endif
497 
498 	/*
499 	 * make sure H2G buffer update and LRC tail update (if this triggering a
500 	 * submission) are visible before updating the descriptor tail
501 	 */
502 	intel_guc_write_barrier(ct_to_guc(ct));
503 
504 	/* update local copies */
505 	ctb->tail = tail;
506 	GEM_BUG_ON(atomic_read(&ctb->space) < len + GUC_CTB_HDR_LEN);
507 	atomic_sub(len + GUC_CTB_HDR_LEN, &ctb->space);
508 
509 	/* now update descriptor */
510 	WRITE_ONCE(desc->tail, tail);
511 
512 	return 0;
513 
514 corrupted:
515 	CT_ERROR(ct, "Corrupted descriptor head=%u tail=%u status=%#x\n",
516 		 desc->head, desc->tail, desc->status);
517 	CT_DEAD(ct, WRITE);
518 	ctb->broken = true;
519 	return -EPIPE;
520 }
521 
522 /**
523  * wait_for_ct_request_update - Wait for CT request state update.
524  * @ct:		pointer to CT
525  * @req:	pointer to pending request
526  * @status:	placeholder for status
527  *
528  * For each sent request, GuC shall send back CT response message.
529  * Our message handler will update status of tracked request once
530  * response message with given fence is received. Wait here and
531  * check for valid response status value.
532  *
533  * Return:
534  * *	0 response received (status is valid)
535  * *	-ETIMEDOUT no response within hardcoded timeout
536  */
wait_for_ct_request_update(struct intel_guc_ct * ct,struct ct_request * req,u32 * status)537 static int wait_for_ct_request_update(struct intel_guc_ct *ct, struct ct_request *req, u32 *status)
538 {
539 	int err;
540 	bool ct_enabled;
541 
542 	/*
543 	 * Fast commands should complete in less than 10us, so sample quickly
544 	 * up to that length of time, then switch to a slower sleep-wait loop.
545 	 * No GuC command should ever take longer than 10ms but many GuC
546 	 * commands can be inflight at time, so use a 1s timeout on the slower
547 	 * sleep-wait loop.
548 	 */
549 #define GUC_CTB_RESPONSE_TIMEOUT_SHORT_MS 10
550 #define GUC_CTB_RESPONSE_TIMEOUT_LONG_MS 1000
551 #define done \
552 	(!(ct_enabled = intel_guc_ct_enabled(ct)) || \
553 	 FIELD_GET(GUC_HXG_MSG_0_ORIGIN, READ_ONCE(req->status)) == \
554 	 GUC_HXG_ORIGIN_GUC)
555 	err = wait_for_us(done, GUC_CTB_RESPONSE_TIMEOUT_SHORT_MS);
556 	if (err)
557 		err = wait_for(done, GUC_CTB_RESPONSE_TIMEOUT_LONG_MS);
558 #undef done
559 	if (!ct_enabled)
560 		err = -ENODEV;
561 
562 	*status = req->status;
563 	return err;
564 }
565 
566 #define GUC_CTB_TIMEOUT_MS	1500
ct_deadlocked(struct intel_guc_ct * ct)567 static inline bool ct_deadlocked(struct intel_guc_ct *ct)
568 {
569 	long timeout = GUC_CTB_TIMEOUT_MS;
570 	bool ret = ktime_ms_delta(ktime_get(), ct->stall_time) > timeout;
571 
572 	if (unlikely(ret)) {
573 		struct guc_ct_buffer_desc *send = ct->ctbs.send.desc;
574 		struct guc_ct_buffer_desc *recv = ct->ctbs.send.desc;
575 
576 		CT_ERROR(ct, "Communication stalled for %lld ms, desc status=%#x,%#x\n",
577 			 ktime_ms_delta(ktime_get(), ct->stall_time),
578 			 send->status, recv->status);
579 		CT_ERROR(ct, "H2G Space: %u (Bytes)\n",
580 			 atomic_read(&ct->ctbs.send.space) * 4);
581 		CT_ERROR(ct, "Head: %u (Dwords)\n", ct->ctbs.send.desc->head);
582 		CT_ERROR(ct, "Tail: %u (Dwords)\n", ct->ctbs.send.desc->tail);
583 		CT_ERROR(ct, "G2H Space: %u (Bytes)\n",
584 			 atomic_read(&ct->ctbs.recv.space) * 4);
585 		CT_ERROR(ct, "Head: %u\n (Dwords)", ct->ctbs.recv.desc->head);
586 		CT_ERROR(ct, "Tail: %u\n (Dwords)", ct->ctbs.recv.desc->tail);
587 
588 		CT_DEAD(ct, DEADLOCK);
589 		ct->ctbs.send.broken = true;
590 	}
591 
592 	return ret;
593 }
594 
g2h_has_room(struct intel_guc_ct * ct,u32 g2h_len_dw)595 static inline bool g2h_has_room(struct intel_guc_ct *ct, u32 g2h_len_dw)
596 {
597 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.recv;
598 
599 	/*
600 	 * We leave a certain amount of space in the G2H CTB buffer for
601 	 * unexpected G2H CTBs (e.g. logging, engine hang, etc...)
602 	 */
603 	return !g2h_len_dw || atomic_read(&ctb->space) >= g2h_len_dw;
604 }
605 
g2h_reserve_space(struct intel_guc_ct * ct,u32 g2h_len_dw)606 static inline void g2h_reserve_space(struct intel_guc_ct *ct, u32 g2h_len_dw)
607 {
608 	lockdep_assert_held(&ct->ctbs.send.lock);
609 
610 	GEM_BUG_ON(!g2h_has_room(ct, g2h_len_dw));
611 
612 	if (g2h_len_dw)
613 		atomic_sub(g2h_len_dw, &ct->ctbs.recv.space);
614 }
615 
g2h_release_space(struct intel_guc_ct * ct,u32 g2h_len_dw)616 static inline void g2h_release_space(struct intel_guc_ct *ct, u32 g2h_len_dw)
617 {
618 	atomic_add(g2h_len_dw, &ct->ctbs.recv.space);
619 }
620 
h2g_has_room(struct intel_guc_ct * ct,u32 len_dw)621 static inline bool h2g_has_room(struct intel_guc_ct *ct, u32 len_dw)
622 {
623 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
624 	struct guc_ct_buffer_desc *desc = ctb->desc;
625 	u32 head;
626 	u32 space;
627 
628 	if (atomic_read(&ctb->space) >= len_dw)
629 		return true;
630 
631 	head = READ_ONCE(desc->head);
632 	if (unlikely(head > ctb->size)) {
633 		CT_ERROR(ct, "Invalid head offset %u >= %u)\n",
634 			 head, ctb->size);
635 		desc->status |= GUC_CTB_STATUS_OVERFLOW;
636 		ctb->broken = true;
637 		CT_DEAD(ct, H2G_HAS_ROOM);
638 		return false;
639 	}
640 
641 	space = CIRC_SPACE(ctb->tail, head, ctb->size);
642 	atomic_set(&ctb->space, space);
643 
644 	return space >= len_dw;
645 }
646 
has_room_nb(struct intel_guc_ct * ct,u32 h2g_dw,u32 g2h_dw)647 static int has_room_nb(struct intel_guc_ct *ct, u32 h2g_dw, u32 g2h_dw)
648 {
649 	bool h2g = h2g_has_room(ct, h2g_dw);
650 	bool g2h = g2h_has_room(ct, g2h_dw);
651 
652 	lockdep_assert_held(&ct->ctbs.send.lock);
653 
654 	if (unlikely(!h2g || !g2h)) {
655 		if (ct->stall_time == KTIME_MAX)
656 			ct->stall_time = ktime_get();
657 
658 		/* Be paranoid and kick G2H tasklet to free credits */
659 		if (!g2h)
660 			tasklet_hi_schedule(&ct->receive_tasklet);
661 
662 		if (unlikely(ct_deadlocked(ct)))
663 			return -EPIPE;
664 		else
665 			return -EBUSY;
666 	}
667 
668 	ct->stall_time = KTIME_MAX;
669 	return 0;
670 }
671 
672 #define G2H_LEN_DW(f) ({ \
673 	typeof(f) f_ = (f); \
674 	FIELD_GET(INTEL_GUC_CT_SEND_G2H_DW_MASK, f_) ? \
675 	FIELD_GET(INTEL_GUC_CT_SEND_G2H_DW_MASK, f_) + \
676 	GUC_CTB_HXG_MSG_MIN_LEN : 0; \
677 })
ct_send_nb(struct intel_guc_ct * ct,const u32 * action,u32 len,u32 flags)678 static int ct_send_nb(struct intel_guc_ct *ct,
679 		      const u32 *action,
680 		      u32 len,
681 		      u32 flags)
682 {
683 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
684 	unsigned long spin_flags;
685 	u32 g2h_len_dw = G2H_LEN_DW(flags);
686 	u32 fence;
687 	int ret;
688 
689 	spin_lock_irqsave(&ctb->lock, spin_flags);
690 
691 	ret = has_room_nb(ct, len + GUC_CTB_HDR_LEN, g2h_len_dw);
692 	if (unlikely(ret))
693 		goto out;
694 
695 	fence = ct_get_next_fence(ct);
696 	ret = ct_write(ct, action, len, fence, flags);
697 	if (unlikely(ret))
698 		goto out;
699 
700 	g2h_reserve_space(ct, g2h_len_dw);
701 	intel_guc_notify(ct_to_guc(ct));
702 
703 out:
704 	spin_unlock_irqrestore(&ctb->lock, spin_flags);
705 
706 	return ret;
707 }
708 
ct_send(struct intel_guc_ct * ct,const u32 * action,u32 len,u32 * response_buf,u32 response_buf_size,u32 * status)709 static int ct_send(struct intel_guc_ct *ct,
710 		   const u32 *action,
711 		   u32 len,
712 		   u32 *response_buf,
713 		   u32 response_buf_size,
714 		   u32 *status)
715 {
716 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
717 	struct ct_request request;
718 	unsigned long flags;
719 	unsigned int sleep_period_ms = 1;
720 	bool send_again;
721 	u32 fence;
722 	int err;
723 
724 	GEM_BUG_ON(!ct->enabled);
725 	GEM_BUG_ON(!len);
726 	GEM_BUG_ON(len > GUC_CTB_HXG_MSG_MAX_LEN - GUC_CTB_HDR_LEN);
727 	GEM_BUG_ON(!response_buf && response_buf_size);
728 	might_sleep();
729 
730 resend:
731 	send_again = false;
732 
733 	/*
734 	 * We use a lazy spin wait loop here as we believe that if the CT
735 	 * buffers are sized correctly the flow control condition should be
736 	 * rare. Reserving the maximum size in the G2H credits as we don't know
737 	 * how big the response is going to be.
738 	 */
739 retry:
740 	spin_lock_irqsave(&ctb->lock, flags);
741 	if (unlikely(!h2g_has_room(ct, len + GUC_CTB_HDR_LEN) ||
742 		     !g2h_has_room(ct, GUC_CTB_HXG_MSG_MAX_LEN))) {
743 		if (ct->stall_time == KTIME_MAX)
744 			ct->stall_time = ktime_get();
745 		spin_unlock_irqrestore(&ctb->lock, flags);
746 
747 		if (unlikely(ct_deadlocked(ct)))
748 			return -EPIPE;
749 
750 		if (msleep_interruptible(sleep_period_ms))
751 			return -EINTR;
752 		sleep_period_ms = sleep_period_ms << 1;
753 
754 		goto retry;
755 	}
756 
757 	ct->stall_time = KTIME_MAX;
758 
759 	fence = ct_get_next_fence(ct);
760 	request.fence = fence;
761 	request.status = 0;
762 	request.response_len = response_buf_size;
763 	request.response_buf = response_buf;
764 
765 	spin_lock(&ct->requests.lock);
766 	list_add_tail(&request.link, &ct->requests.pending);
767 	spin_unlock(&ct->requests.lock);
768 
769 	err = ct_write(ct, action, len, fence, 0);
770 	g2h_reserve_space(ct, GUC_CTB_HXG_MSG_MAX_LEN);
771 
772 	spin_unlock_irqrestore(&ctb->lock, flags);
773 
774 	if (unlikely(err))
775 		goto unlink;
776 
777 	intel_guc_notify(ct_to_guc(ct));
778 
779 	err = wait_for_ct_request_update(ct, &request, status);
780 	g2h_release_space(ct, GUC_CTB_HXG_MSG_MAX_LEN);
781 	if (unlikely(err)) {
782 		if (err == -ENODEV)
783 			/* wait_for_ct_request_update returns -ENODEV on reset/suspend in progress.
784 			 * In this case, output is debug rather than error info
785 			 */
786 			CT_DEBUG(ct, "Request %#x (fence %u) cancelled as CTB is disabled\n",
787 				 action[0], request.fence);
788 		else
789 			CT_ERROR(ct, "No response for request %#x (fence %u)\n",
790 				 action[0], request.fence);
791 		goto unlink;
792 	}
793 
794 	if (FIELD_GET(GUC_HXG_MSG_0_TYPE, *status) == GUC_HXG_TYPE_NO_RESPONSE_RETRY) {
795 		CT_DEBUG(ct, "retrying request %#x (%u)\n", *action,
796 			 FIELD_GET(GUC_HXG_RETRY_MSG_0_REASON, *status));
797 		send_again = true;
798 		goto unlink;
799 	}
800 
801 	if (FIELD_GET(GUC_HXG_MSG_0_TYPE, *status) != GUC_HXG_TYPE_RESPONSE_SUCCESS) {
802 		err = -EIO;
803 		goto unlink;
804 	}
805 
806 	if (response_buf) {
807 		/* There shall be no data in the status */
808 		WARN_ON(FIELD_GET(GUC_HXG_RESPONSE_MSG_0_DATA0, request.status));
809 		/* Return actual response len */
810 		err = request.response_len;
811 	} else {
812 		/* There shall be no response payload */
813 		WARN_ON(request.response_len);
814 		/* Return data decoded from the status dword */
815 		err = FIELD_GET(GUC_HXG_RESPONSE_MSG_0_DATA0, *status);
816 	}
817 
818 unlink:
819 	spin_lock_irqsave(&ct->requests.lock, flags);
820 	list_del(&request.link);
821 	spin_unlock_irqrestore(&ct->requests.lock, flags);
822 
823 	if (unlikely(send_again))
824 		goto resend;
825 
826 	return err;
827 }
828 
829 /*
830  * Command Transport (CT) buffer based GuC send function.
831  */
intel_guc_ct_send(struct intel_guc_ct * ct,const u32 * action,u32 len,u32 * response_buf,u32 response_buf_size,u32 flags)832 int intel_guc_ct_send(struct intel_guc_ct *ct, const u32 *action, u32 len,
833 		      u32 *response_buf, u32 response_buf_size, u32 flags)
834 {
835 	u32 status = ~0; /* undefined */
836 	int ret;
837 
838 	if (unlikely(!ct->enabled)) {
839 		struct intel_guc *guc = ct_to_guc(ct);
840 		struct intel_uc *uc = container_of(guc, struct intel_uc, guc);
841 
842 		WARN(!uc->reset_in_progress, "Unexpected send: action=%#x\n", *action);
843 		return -ENODEV;
844 	}
845 
846 	if (unlikely(ct->ctbs.send.broken))
847 		return -EPIPE;
848 
849 	if (flags & INTEL_GUC_CT_SEND_NB)
850 		return ct_send_nb(ct, action, len, flags);
851 
852 	ret = ct_send(ct, action, len, response_buf, response_buf_size, &status);
853 	if (unlikely(ret < 0)) {
854 		if (ret != -ENODEV)
855 			CT_ERROR(ct, "Sending action %#x failed (%pe) status=%#X\n",
856 				 action[0], ERR_PTR(ret), status);
857 	} else if (unlikely(ret)) {
858 		CT_DEBUG(ct, "send action %#x returned %d (%#x)\n",
859 			 action[0], ret, ret);
860 	}
861 
862 	return ret;
863 }
864 
ct_alloc_msg(u32 num_dwords)865 static struct ct_incoming_msg *ct_alloc_msg(u32 num_dwords)
866 {
867 	struct ct_incoming_msg *msg;
868 
869 	msg = kmalloc_flex(*msg, msg, num_dwords, GFP_ATOMIC);
870 	if (msg)
871 		msg->size = num_dwords;
872 	return msg;
873 }
874 
ct_free_msg(struct ct_incoming_msg * msg)875 static void ct_free_msg(struct ct_incoming_msg *msg)
876 {
877 	kfree(msg);
878 }
879 
880 /*
881  * Return: number available remaining dwords to read (0 if empty)
882  *         or a negative error code on failure
883  */
ct_read(struct intel_guc_ct * ct,struct ct_incoming_msg ** msg)884 static int ct_read(struct intel_guc_ct *ct, struct ct_incoming_msg **msg)
885 {
886 	struct intel_guc_ct_buffer *ctb = &ct->ctbs.recv;
887 	struct guc_ct_buffer_desc *desc = ctb->desc;
888 	u32 head = ctb->head;
889 	u32 tail = READ_ONCE(desc->tail);
890 	u32 size = ctb->size;
891 	u32 *cmds = ctb->cmds;
892 	s32 available;
893 	unsigned int len;
894 	unsigned int i;
895 	u32 header;
896 
897 	if (unlikely(ctb->broken))
898 		return -EPIPE;
899 
900 	if (unlikely(desc->status)) {
901 		u32 status = desc->status;
902 
903 		if (status & GUC_CTB_STATUS_UNUSED) {
904 			/*
905 			 * Potentially valid if a CLIENT_RESET request resulted in
906 			 * contexts/engines being reset. But should never happen as
907 			 * no contexts should be active when CLIENT_RESET is sent.
908 			 */
909 			CT_ERROR(ct, "Unexpected G2H after GuC has stopped!\n");
910 			status &= ~GUC_CTB_STATUS_UNUSED;
911 		}
912 
913 		if (status)
914 			goto corrupted;
915 	}
916 
917 	GEM_BUG_ON(head > size);
918 
919 #ifdef CONFIG_DRM_I915_DEBUG_GUC
920 	if (unlikely(head != READ_ONCE(desc->head))) {
921 		CT_ERROR(ct, "Head was modified %u != %u\n",
922 			 desc->head, head);
923 		desc->status |= GUC_CTB_STATUS_MISMATCH;
924 		goto corrupted;
925 	}
926 #endif
927 	if (unlikely(tail >= size)) {
928 		CT_ERROR(ct, "Invalid tail offset %u >= %u)\n",
929 			 tail, size);
930 		desc->status |= GUC_CTB_STATUS_OVERFLOW;
931 		goto corrupted;
932 	}
933 
934 	/* tail == head condition indicates empty */
935 	available = tail - head;
936 	if (unlikely(available == 0)) {
937 		*msg = NULL;
938 		return 0;
939 	}
940 
941 	/* beware of buffer wrap case */
942 	if (unlikely(available < 0))
943 		available += size;
944 	CT_DEBUG(ct, "available %d (%u:%u:%u)\n", available, head, tail, size);
945 	GEM_BUG_ON(available < 0);
946 
947 	header = cmds[head];
948 	head = (head + 1) % size;
949 
950 	/* message len with header */
951 	len = FIELD_GET(GUC_CTB_MSG_0_NUM_DWORDS, header) + GUC_CTB_MSG_MIN_LEN;
952 	if (unlikely(len > (u32)available)) {
953 		CT_ERROR(ct, "Incomplete message %*ph %*ph %*ph\n",
954 			 4, &header,
955 			 4 * (head + available - 1 > size ?
956 			      size - head : available - 1), &cmds[head],
957 			 4 * (head + available - 1 > size ?
958 			      available - 1 - size + head : 0), &cmds[0]);
959 		desc->status |= GUC_CTB_STATUS_UNDERFLOW;
960 		goto corrupted;
961 	}
962 
963 	*msg = ct_alloc_msg(len);
964 	if (!*msg) {
965 		CT_ERROR(ct, "No memory for message %*ph %*ph %*ph\n",
966 			 4, &header,
967 			 4 * (head + available - 1 > size ?
968 			      size - head : available - 1), &cmds[head],
969 			 4 * (head + available - 1 > size ?
970 			      available - 1 - size + head : 0), &cmds[0]);
971 		return available;
972 	}
973 
974 	(*msg)->msg[0] = header;
975 
976 	for (i = 1; i < len; i++) {
977 		(*msg)->msg[i] = cmds[head];
978 		head = (head + 1) % size;
979 	}
980 	CT_DEBUG(ct, "received %*ph\n", 4 * len, (*msg)->msg);
981 
982 	/* update local copies */
983 	ctb->head = head;
984 
985 	/* now update descriptor */
986 	WRITE_ONCE(desc->head, head);
987 
988 	intel_guc_write_barrier(ct_to_guc(ct));
989 
990 	return available - len;
991 
992 corrupted:
993 	CT_ERROR(ct, "Corrupted descriptor head=%u tail=%u status=%#x\n",
994 		 desc->head, desc->tail, desc->status);
995 	ctb->broken = true;
996 	CT_DEAD(ct, READ);
997 	return -EPIPE;
998 }
999 
1000 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
ct_check_lost_and_found(struct intel_guc_ct * ct,u32 fence)1001 static bool ct_check_lost_and_found(struct intel_guc_ct *ct, u32 fence)
1002 {
1003 	unsigned int n;
1004 	char *buf = NULL;
1005 	bool found = false;
1006 
1007 	lockdep_assert_held(&ct->requests.lock);
1008 
1009 	for (n = 0; n < ARRAY_SIZE(ct->requests.lost_and_found); n++) {
1010 		if (ct->requests.lost_and_found[n].fence != fence)
1011 			continue;
1012 		found = true;
1013 
1014 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GUC)
1015 		buf = kmalloc(SZ_4K, GFP_NOWAIT);
1016 		if (buf && stack_depot_snprint(ct->requests.lost_and_found[n].stack,
1017 					       buf, SZ_4K, 0)) {
1018 			CT_ERROR(ct, "Fence %u was used by action %#04x sent at\n%s",
1019 				 fence, ct->requests.lost_and_found[n].action, buf);
1020 			break;
1021 		}
1022 #endif
1023 		CT_ERROR(ct, "Fence %u was used by action %#04x\n",
1024 			 fence, ct->requests.lost_and_found[n].action);
1025 		break;
1026 	}
1027 	kfree(buf);
1028 	return found;
1029 }
1030 #else
ct_check_lost_and_found(struct intel_guc_ct * ct,u32 fence)1031 static bool ct_check_lost_and_found(struct intel_guc_ct *ct, u32 fence)
1032 {
1033 	return false;
1034 }
1035 #endif
1036 
ct_handle_response(struct intel_guc_ct * ct,struct ct_incoming_msg * response)1037 static int ct_handle_response(struct intel_guc_ct *ct, struct ct_incoming_msg *response)
1038 {
1039 	u32 len = FIELD_GET(GUC_CTB_MSG_0_NUM_DWORDS, response->msg[0]);
1040 	u32 fence = FIELD_GET(GUC_CTB_MSG_0_FENCE, response->msg[0]);
1041 	const u32 *hxg = &response->msg[GUC_CTB_MSG_MIN_LEN];
1042 	const u32 *data = &hxg[GUC_HXG_MSG_MIN_LEN];
1043 	u32 datalen = len - GUC_HXG_MSG_MIN_LEN;
1044 	struct ct_request *req;
1045 	unsigned long flags;
1046 	bool found = false;
1047 	int err = 0;
1048 
1049 	GEM_BUG_ON(len < GUC_HXG_MSG_MIN_LEN);
1050 	GEM_BUG_ON(FIELD_GET(GUC_HXG_MSG_0_ORIGIN, hxg[0]) != GUC_HXG_ORIGIN_GUC);
1051 	GEM_BUG_ON(FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_RESPONSE_SUCCESS &&
1052 		   FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_NO_RESPONSE_RETRY &&
1053 		   FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_RESPONSE_FAILURE);
1054 
1055 	CT_DEBUG(ct, "response fence %u status %#x\n", fence, hxg[0]);
1056 
1057 	spin_lock_irqsave(&ct->requests.lock, flags);
1058 	list_for_each_entry(req, &ct->requests.pending, link) {
1059 		if (unlikely(fence != req->fence)) {
1060 			CT_DEBUG(ct, "request %u awaits response\n",
1061 				 req->fence);
1062 			continue;
1063 		}
1064 		if (unlikely(datalen > req->response_len)) {
1065 			CT_ERROR(ct, "Response %u too long (datalen %u > %u)\n",
1066 				 req->fence, datalen, req->response_len);
1067 			datalen = min(datalen, req->response_len);
1068 			err = -EMSGSIZE;
1069 		}
1070 		if (datalen)
1071 			memcpy(req->response_buf, data, 4 * datalen);
1072 		req->response_len = datalen;
1073 		WRITE_ONCE(req->status, hxg[0]);
1074 		found = true;
1075 		break;
1076 	}
1077 
1078 #ifdef CONFIG_DRM_I915_SELFTEST
1079 	if (!found && ct_to_guc(ct)->fast_response_selftest) {
1080 		CT_DEBUG(ct, "Assuming unsolicited response due to FAST_REQUEST selftest\n");
1081 		ct_to_guc(ct)->fast_response_selftest++;
1082 		found = true;
1083 	}
1084 #endif
1085 
1086 	if (!found) {
1087 		CT_ERROR(ct, "Unsolicited response message: len %u, data %#x (fence %u, last %u)\n",
1088 			 len, hxg[0], fence, ct->requests.last_fence);
1089 		if (!ct_check_lost_and_found(ct, fence)) {
1090 			list_for_each_entry(req, &ct->requests.pending, link)
1091 				CT_ERROR(ct, "request %u awaits response\n",
1092 					 req->fence);
1093 		}
1094 		err = -ENOKEY;
1095 	}
1096 	spin_unlock_irqrestore(&ct->requests.lock, flags);
1097 
1098 	if (unlikely(err))
1099 		return err;
1100 
1101 	ct_free_msg(response);
1102 	return 0;
1103 }
1104 
ct_process_request(struct intel_guc_ct * ct,struct ct_incoming_msg * request)1105 static int ct_process_request(struct intel_guc_ct *ct, struct ct_incoming_msg *request)
1106 {
1107 	struct intel_guc *guc = ct_to_guc(ct);
1108 	const u32 *hxg;
1109 	const u32 *payload;
1110 	u32 hxg_len, action, len;
1111 	int ret;
1112 
1113 	hxg = &request->msg[GUC_CTB_MSG_MIN_LEN];
1114 	hxg_len = request->size - GUC_CTB_MSG_MIN_LEN;
1115 	payload = &hxg[GUC_HXG_MSG_MIN_LEN];
1116 	action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1117 	len = hxg_len - GUC_HXG_MSG_MIN_LEN;
1118 
1119 	CT_DEBUG(ct, "request %x %*ph\n", action, 4 * len, payload);
1120 
1121 	switch (action) {
1122 	case INTEL_GUC_ACTION_DEFAULT:
1123 		ret = intel_guc_to_host_process_recv_msg(guc, payload, len);
1124 		break;
1125 	case INTEL_GUC_ACTION_DEREGISTER_CONTEXT_DONE:
1126 		ret = intel_guc_deregister_done_process_msg(guc, payload,
1127 							    len);
1128 		break;
1129 	case INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_DONE:
1130 		ret = intel_guc_sched_done_process_msg(guc, payload, len);
1131 		break;
1132 	case INTEL_GUC_ACTION_CONTEXT_RESET_NOTIFICATION:
1133 		ret = intel_guc_context_reset_process_msg(guc, payload, len);
1134 		break;
1135 	case INTEL_GUC_ACTION_STATE_CAPTURE_NOTIFICATION:
1136 		ret = intel_guc_error_capture_process_msg(guc, payload, len);
1137 		if (unlikely(ret))
1138 			CT_ERROR(ct, "error capture notification failed %x %*ph\n",
1139 				 action, 4 * len, payload);
1140 		break;
1141 	case INTEL_GUC_ACTION_ENGINE_FAILURE_NOTIFICATION:
1142 		ret = intel_guc_engine_failure_process_msg(guc, payload, len);
1143 		break;
1144 	case INTEL_GUC_ACTION_NOTIFY_FLUSH_LOG_BUFFER_TO_FILE:
1145 		intel_guc_log_handle_flush_event(&guc->log);
1146 		ret = 0;
1147 		break;
1148 	case INTEL_GUC_ACTION_NOTIFY_CRASH_DUMP_POSTED:
1149 	case INTEL_GUC_ACTION_NOTIFY_EXCEPTION:
1150 		ret = intel_guc_crash_process_msg(guc, action);
1151 		break;
1152 	case INTEL_GUC_ACTION_TLB_INVALIDATION_DONE:
1153 		ret = intel_guc_tlb_invalidation_done(guc, payload, len);
1154 		break;
1155 	default:
1156 		ret = -EOPNOTSUPP;
1157 		break;
1158 	}
1159 
1160 	if (unlikely(ret)) {
1161 		CT_ERROR(ct, "Failed to process request %04x (%pe)\n",
1162 			 action, ERR_PTR(ret));
1163 		return ret;
1164 	}
1165 
1166 	ct_free_msg(request);
1167 	return 0;
1168 }
1169 
ct_process_incoming_requests(struct intel_guc_ct * ct)1170 static bool ct_process_incoming_requests(struct intel_guc_ct *ct)
1171 {
1172 	unsigned long flags;
1173 	struct ct_incoming_msg *request;
1174 	bool done;
1175 	int err;
1176 
1177 	spin_lock_irqsave(&ct->requests.lock, flags);
1178 	request = list_first_entry_or_null(&ct->requests.incoming,
1179 					   struct ct_incoming_msg, link);
1180 	if (request)
1181 		list_del(&request->link);
1182 	done = !!list_empty(&ct->requests.incoming);
1183 	spin_unlock_irqrestore(&ct->requests.lock, flags);
1184 
1185 	if (!request)
1186 		return true;
1187 
1188 	err = ct_process_request(ct, request);
1189 	if (unlikely(err)) {
1190 		CT_ERROR(ct, "Failed to process CT message (%pe) %*ph\n",
1191 			 ERR_PTR(err), 4 * request->size, request->msg);
1192 		CT_DEAD(ct, PROCESS_FAILED);
1193 		ct_free_msg(request);
1194 	}
1195 
1196 	return done;
1197 }
1198 
ct_incoming_request_worker_func(struct work_struct * w)1199 static void ct_incoming_request_worker_func(struct work_struct *w)
1200 {
1201 	struct intel_guc_ct *ct =
1202 		container_of(w, struct intel_guc_ct, requests.worker);
1203 	bool done;
1204 
1205 	do {
1206 		done = ct_process_incoming_requests(ct);
1207 	} while (!done);
1208 }
1209 
ct_handle_event(struct intel_guc_ct * ct,struct ct_incoming_msg * request)1210 static int ct_handle_event(struct intel_guc_ct *ct, struct ct_incoming_msg *request)
1211 {
1212 	const u32 *hxg = &request->msg[GUC_CTB_MSG_MIN_LEN];
1213 	u32 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1214 	unsigned long flags;
1215 
1216 	GEM_BUG_ON(FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_EVENT);
1217 
1218 	/*
1219 	 * Adjusting the space must be done in IRQ or deadlock can occur as the
1220 	 * CTB processing in the below workqueue can send CTBs which creates a
1221 	 * circular dependency if the space was returned there.
1222 	 */
1223 	switch (action) {
1224 	case INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_DONE:
1225 	case INTEL_GUC_ACTION_DEREGISTER_CONTEXT_DONE:
1226 	case INTEL_GUC_ACTION_TLB_INVALIDATION_DONE:
1227 		g2h_release_space(ct, request->size);
1228 	}
1229 
1230 	/*
1231 	 * TLB invalidation responses must be handled immediately as processing
1232 	 * of other G2H notifications may be blocked by an invalidation request.
1233 	 */
1234 	if (action == INTEL_GUC_ACTION_TLB_INVALIDATION_DONE)
1235 		return ct_process_request(ct, request);
1236 
1237 	spin_lock_irqsave(&ct->requests.lock, flags);
1238 	list_add_tail(&request->link, &ct->requests.incoming);
1239 	spin_unlock_irqrestore(&ct->requests.lock, flags);
1240 
1241 	queue_work(system_unbound_wq, &ct->requests.worker);
1242 	return 0;
1243 }
1244 
ct_handle_hxg(struct intel_guc_ct * ct,struct ct_incoming_msg * msg)1245 static int ct_handle_hxg(struct intel_guc_ct *ct, struct ct_incoming_msg *msg)
1246 {
1247 	u32 origin, type;
1248 	u32 *hxg;
1249 	int err;
1250 
1251 	if (unlikely(msg->size < GUC_CTB_HXG_MSG_MIN_LEN))
1252 		return -EBADMSG;
1253 
1254 	hxg = &msg->msg[GUC_CTB_MSG_MIN_LEN];
1255 
1256 	origin = FIELD_GET(GUC_HXG_MSG_0_ORIGIN, hxg[0]);
1257 	if (unlikely(origin != GUC_HXG_ORIGIN_GUC)) {
1258 		err = -EPROTO;
1259 		goto failed;
1260 	}
1261 
1262 	type = FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]);
1263 	switch (type) {
1264 	case GUC_HXG_TYPE_EVENT:
1265 		err = ct_handle_event(ct, msg);
1266 		break;
1267 	case GUC_HXG_TYPE_RESPONSE_SUCCESS:
1268 	case GUC_HXG_TYPE_RESPONSE_FAILURE:
1269 	case GUC_HXG_TYPE_NO_RESPONSE_RETRY:
1270 		err = ct_handle_response(ct, msg);
1271 		break;
1272 	default:
1273 		err = -EOPNOTSUPP;
1274 	}
1275 
1276 	if (unlikely(err)) {
1277 failed:
1278 		CT_ERROR(ct, "Failed to handle HXG message (%pe) %*ph\n",
1279 			 ERR_PTR(err), 4 * GUC_HXG_MSG_MIN_LEN, hxg);
1280 	}
1281 	return err;
1282 }
1283 
ct_handle_msg(struct intel_guc_ct * ct,struct ct_incoming_msg * msg)1284 static void ct_handle_msg(struct intel_guc_ct *ct, struct ct_incoming_msg *msg)
1285 {
1286 	u32 format = FIELD_GET(GUC_CTB_MSG_0_FORMAT, msg->msg[0]);
1287 	int err;
1288 
1289 	if (format == GUC_CTB_FORMAT_HXG)
1290 		err = ct_handle_hxg(ct, msg);
1291 	else
1292 		err = -EOPNOTSUPP;
1293 
1294 	if (unlikely(err)) {
1295 		CT_ERROR(ct, "Failed to process CT message (%pe) %*ph\n",
1296 			 ERR_PTR(err), 4 * msg->size, msg->msg);
1297 		ct_free_msg(msg);
1298 	}
1299 }
1300 
1301 /*
1302  * Return: number available remaining dwords to read (0 if empty)
1303  *         or a negative error code on failure
1304  */
ct_receive(struct intel_guc_ct * ct)1305 static int ct_receive(struct intel_guc_ct *ct)
1306 {
1307 	struct ct_incoming_msg *msg = NULL;
1308 	unsigned long flags;
1309 	int ret;
1310 
1311 	spin_lock_irqsave(&ct->ctbs.recv.lock, flags);
1312 	ret = ct_read(ct, &msg);
1313 	spin_unlock_irqrestore(&ct->ctbs.recv.lock, flags);
1314 	if (ret < 0)
1315 		return ret;
1316 
1317 	if (msg)
1318 		ct_handle_msg(ct, msg);
1319 
1320 	return ret;
1321 }
1322 
ct_try_receive_message(struct intel_guc_ct * ct)1323 static void ct_try_receive_message(struct intel_guc_ct *ct)
1324 {
1325 	struct intel_guc *guc = ct_to_guc(ct);
1326 	int ret;
1327 
1328 	if (!ct->enabled) {
1329 		GEM_WARN_ON(!guc_to_gt(guc)->uc.reset_in_progress);
1330 		return;
1331 	}
1332 
1333 	/* When interrupt disabled, message handling is not expected */
1334 	if (!guc->interrupts.enabled)
1335 		return;
1336 
1337 	ret = ct_receive(ct);
1338 	if (ret > 0)
1339 		tasklet_hi_schedule(&ct->receive_tasklet);
1340 }
1341 
ct_receive_tasklet_func(struct tasklet_struct * t)1342 static void ct_receive_tasklet_func(struct tasklet_struct *t)
1343 {
1344 	struct intel_guc_ct *ct = from_tasklet(ct, t, receive_tasklet);
1345 
1346 	ct_try_receive_message(ct);
1347 }
1348 
1349 /*
1350  * When we're communicating with the GuC over CT, GuC uses events
1351  * to notify us about new messages being posted on the RECV buffer.
1352  */
intel_guc_ct_event_handler(struct intel_guc_ct * ct)1353 void intel_guc_ct_event_handler(struct intel_guc_ct *ct)
1354 {
1355 	if (unlikely(!ct->enabled)) {
1356 		WARN(1, "Unexpected GuC event received while CT disabled!\n");
1357 		return;
1358 	}
1359 
1360 	ct_try_receive_message(ct);
1361 }
1362 
intel_guc_ct_print_info(struct intel_guc_ct * ct,struct drm_printer * p)1363 void intel_guc_ct_print_info(struct intel_guc_ct *ct,
1364 			     struct drm_printer *p)
1365 {
1366 	drm_printf(p, "CT %s\n", str_enabled_disabled(ct->enabled));
1367 
1368 	if (!ct->enabled)
1369 		return;
1370 
1371 	drm_printf(p, "H2G Space: %u\n",
1372 		   atomic_read(&ct->ctbs.send.space) * 4);
1373 	drm_printf(p, "Head: %u\n",
1374 		   ct->ctbs.send.desc->head);
1375 	drm_printf(p, "Tail: %u\n",
1376 		   ct->ctbs.send.desc->tail);
1377 	drm_printf(p, "G2H Space: %u\n",
1378 		   atomic_read(&ct->ctbs.recv.space) * 4);
1379 	drm_printf(p, "Head: %u\n",
1380 		   ct->ctbs.recv.desc->head);
1381 	drm_printf(p, "Tail: %u\n",
1382 		   ct->ctbs.recv.desc->tail);
1383 }
1384 
1385 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG)
ct_dead_ct_worker_func(struct work_struct * w)1386 static void ct_dead_ct_worker_func(struct work_struct *w)
1387 {
1388 	struct intel_guc_ct *ct = container_of(w, struct intel_guc_ct, dead_ct_worker);
1389 	struct intel_guc *guc = ct_to_guc(ct);
1390 
1391 	if (ct->dead_ct_reported)
1392 		return;
1393 
1394 	ct->dead_ct_reported = true;
1395 
1396 	guc_info(guc, "CTB is dead - reason=0x%X\n", ct->dead_ct_reason);
1397 	intel_klog_error_capture(guc_to_gt(guc), (intel_engine_mask_t)~0U);
1398 }
1399 #endif
1400