xref: /linux/drivers/gpu/drm/xe/xe_guc_ct.c (revision 9b043680446067358913edc2e9dd71bf8ffae208)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #include "xe_guc_ct.h"
7 
8 #include <linux/bitfield.h>
9 #include <linux/circ_buf.h>
10 #include <linux/delay.h>
11 #include <linux/fault-inject.h>
12 
13 #include <kunit/static_stub.h>
14 
15 #include <drm/drm_managed.h>
16 
17 #include "abi/guc_actions_abi.h"
18 #include "abi/guc_actions_sriov_abi.h"
19 #include "abi/guc_klvs_abi.h"
20 #include "xe_bo.h"
21 #include "xe_devcoredump.h"
22 #include "xe_device.h"
23 #include "xe_gt.h"
24 #include "xe_gt_printk.h"
25 #include "xe_gt_sriov_pf_control.h"
26 #include "xe_gt_sriov_pf_monitor.h"
27 #include "xe_guc.h"
28 #include "xe_guc_log.h"
29 #include "xe_guc_pagefault.h"
30 #include "xe_guc_relay.h"
31 #include "xe_guc_submit.h"
32 #include "xe_guc_tlb_inval.h"
33 #include "xe_map.h"
34 #include "xe_pm.h"
35 #include "xe_sriov_vf.h"
36 #include "xe_trace_guc.h"
37 
38 static void receive_g2h(struct xe_guc_ct *ct);
39 static void g2h_worker_func(struct work_struct *w);
40 static void safe_mode_worker_func(struct work_struct *w);
41 static void ct_exit_safe_mode(struct xe_guc_ct *ct);
42 static void guc_ct_change_state(struct xe_guc_ct *ct,
43 				enum xe_guc_ct_state state);
44 
45 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
46 enum {
47 	/* Internal states, not error conditions */
48 	CT_DEAD_STATE_REARM,			/* 0x0001 */
49 	CT_DEAD_STATE_CAPTURE,			/* 0x0002 */
50 
51 	/* Error conditions */
52 	CT_DEAD_SETUP,				/* 0x0004 */
53 	CT_DEAD_H2G_WRITE,			/* 0x0008 */
54 	CT_DEAD_H2G_HAS_ROOM,			/* 0x0010 */
55 	CT_DEAD_G2H_READ,			/* 0x0020 */
56 	CT_DEAD_G2H_RECV,			/* 0x0040 */
57 	CT_DEAD_G2H_RELEASE,			/* 0x0080 */
58 	CT_DEAD_DEADLOCK,			/* 0x0100 */
59 	CT_DEAD_PROCESS_FAILED,			/* 0x0200 */
60 	CT_DEAD_FAST_G2H,			/* 0x0400 */
61 	CT_DEAD_PARSE_G2H_RESPONSE,		/* 0x0800 */
62 	CT_DEAD_PARSE_G2H_UNKNOWN,		/* 0x1000 */
63 	CT_DEAD_PARSE_G2H_ORIGIN,		/* 0x2000 */
64 	CT_DEAD_PARSE_G2H_TYPE,			/* 0x4000 */
65 	CT_DEAD_CRASH,				/* 0x8000 */
66 };
67 
68 static void ct_dead_worker_func(struct work_struct *w);
69 static void ct_dead_capture(struct xe_guc_ct *ct, struct guc_ctb *ctb, u32 reason_code);
70 
71 #define CT_DEAD(ct, ctb, reason_code)		ct_dead_capture((ct), (ctb), CT_DEAD_##reason_code)
72 #else
73 #define CT_DEAD(ct, ctb, reason)			\
74 	do {						\
75 		struct guc_ctb *_ctb = (ctb);		\
76 		if (_ctb)				\
77 			_ctb->info.broken = true;	\
78 	} while (0)
79 #endif
80 
81 /* Used when a CT send wants to block and / or receive data */
82 struct g2h_fence {
83 	u32 *response_buffer;
84 	u32 seqno;
85 	u32 response_data;
86 	u16 response_len;
87 	u16 error;
88 	u16 hint;
89 	u16 reason;
90 	bool cancel;
91 	bool retry;
92 	bool fail;
93 	bool done;
94 };
95 
g2h_fence_init(struct g2h_fence * g2h_fence,u32 * response_buffer)96 static void g2h_fence_init(struct g2h_fence *g2h_fence, u32 *response_buffer)
97 {
98 	memset(g2h_fence, 0, sizeof(*g2h_fence));
99 	g2h_fence->response_buffer = response_buffer;
100 	g2h_fence->seqno = ~0x0;
101 }
102 
g2h_fence_cancel(struct g2h_fence * g2h_fence)103 static void g2h_fence_cancel(struct g2h_fence *g2h_fence)
104 {
105 	g2h_fence->cancel = true;
106 	g2h_fence->fail = true;
107 
108 	/* WRITE_ONCE pairs with READ_ONCEs in guc_ct_send_recv. */
109 	WRITE_ONCE(g2h_fence->done, true);
110 }
111 
g2h_fence_needs_alloc(struct g2h_fence * g2h_fence)112 static bool g2h_fence_needs_alloc(struct g2h_fence *g2h_fence)
113 {
114 	return g2h_fence->seqno == ~0x0;
115 }
116 
117 static struct xe_guc *
ct_to_guc(struct xe_guc_ct * ct)118 ct_to_guc(struct xe_guc_ct *ct)
119 {
120 	return container_of(ct, struct xe_guc, ct);
121 }
122 
123 static struct xe_gt *
ct_to_gt(struct xe_guc_ct * ct)124 ct_to_gt(struct xe_guc_ct *ct)
125 {
126 	return container_of(ct, struct xe_gt, uc.guc.ct);
127 }
128 
129 static struct xe_device *
ct_to_xe(struct xe_guc_ct * ct)130 ct_to_xe(struct xe_guc_ct *ct)
131 {
132 	return gt_to_xe(ct_to_gt(ct));
133 }
134 
135 /**
136  * DOC: GuC CTB Blob
137  *
138  * We allocate single blob to hold both CTB descriptors and buffers:
139  *
140  *      +--------+-----------------------------------------------+------+
141  *      | offset | contents                                      | size |
142  *      +========+===============================================+======+
143  *      | 0x0000 | H2G CTB Descriptor (send)                     |      |
144  *      +--------+-----------------------------------------------+  4K  |
145  *      | 0x0800 | G2H CTB Descriptor (g2h)                      |      |
146  *      +--------+-----------------------------------------------+------+
147  *      | 0x1000 | H2G CT Buffer (send)                          | n*4K |
148  *      |        |                                               |      |
149  *      +--------+-----------------------------------------------+------+
150  *      | 0x1000 | G2H CT Buffer (g2h)                           | m*4K |
151  *      | + n*4K |                                               |      |
152  *      +--------+-----------------------------------------------+------+
153  *
154  * Size of each ``CT Buffer`` must be multiple of 4K.
155  * We don't expect too many messages in flight at any time, unless we are
156  * using the GuC submission. In that case each request requires a minimum
157  * 2 dwords which gives us a maximum 256 queue'd requests. Hopefully this
158  * enough space to avoid backpressure on the driver. We increase the size
159  * of the receive buffer (relative to the send) to ensure a G2H response
160  * CTB has a landing spot.
161  *
162  * In addition to submissions, the G2H buffer needs to be able to hold
163  * enough space for recoverable page fault notifications. The number of
164  * page faults is interrupt driven and can be as much as the number of
165  * compute resources available. However, most of the actual work for these
166  * is in a separate page fault worker thread. Therefore we only need to
167  * make sure the queue has enough space to handle all of the submissions
168  * and responses and an extra buffer for incoming page faults.
169  */
170 
171 #define CTB_DESC_SIZE		ALIGN(sizeof(struct guc_ct_buffer_desc), SZ_2K)
172 #define CTB_H2G_BUFFER_OFFSET	(CTB_DESC_SIZE * 2)
173 #define CTB_H2G_BUFFER_SIZE	(SZ_4K)
174 #define CTB_G2H_BUFFER_SIZE	(SZ_128K)
175 #define G2H_ROOM_BUFFER_SIZE	(CTB_G2H_BUFFER_SIZE / 2)
176 
177 /**
178  * xe_guc_ct_queue_proc_time_jiffies - Return maximum time to process a full
179  * CT command queue
180  * @ct: the &xe_guc_ct. Unused at this moment but will be used in the future.
181  *
182  * Observation is that a 4KiB buffer full of commands takes a little over a
183  * second to process. Use that to calculate maximum time to process a full CT
184  * command queue.
185  *
186  * Return: Maximum time to process a full CT queue in jiffies.
187  */
xe_guc_ct_queue_proc_time_jiffies(struct xe_guc_ct * ct)188 long xe_guc_ct_queue_proc_time_jiffies(struct xe_guc_ct *ct)
189 {
190 	BUILD_BUG_ON(!IS_ALIGNED(CTB_H2G_BUFFER_SIZE, SZ_4));
191 	return (CTB_H2G_BUFFER_SIZE / SZ_4K) * HZ;
192 }
193 
guc_ct_size(void)194 static size_t guc_ct_size(void)
195 {
196 	return CTB_H2G_BUFFER_OFFSET + CTB_H2G_BUFFER_SIZE +
197 		CTB_G2H_BUFFER_SIZE;
198 }
199 
guc_ct_fini(struct drm_device * drm,void * arg)200 static void guc_ct_fini(struct drm_device *drm, void *arg)
201 {
202 	struct xe_guc_ct *ct = arg;
203 
204 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
205 	cancel_work_sync(&ct->dead.worker);
206 #endif
207 	ct_exit_safe_mode(ct);
208 	destroy_workqueue(ct->g2h_wq);
209 	xa_destroy(&ct->fence_lookup);
210 }
211 
primelockdep(struct xe_guc_ct * ct)212 static void primelockdep(struct xe_guc_ct *ct)
213 {
214 	if (!IS_ENABLED(CONFIG_LOCKDEP))
215 		return;
216 
217 	fs_reclaim_acquire(GFP_KERNEL);
218 	might_lock(&ct->lock);
219 	fs_reclaim_release(GFP_KERNEL);
220 }
221 
xe_guc_ct_init_noalloc(struct xe_guc_ct * ct)222 int xe_guc_ct_init_noalloc(struct xe_guc_ct *ct)
223 {
224 	struct xe_device *xe = ct_to_xe(ct);
225 	struct xe_gt *gt = ct_to_gt(ct);
226 	int err;
227 
228 	xe_gt_assert(gt, !(guc_ct_size() % PAGE_SIZE));
229 
230 	err = drmm_mutex_init(&xe->drm, &ct->lock);
231 	if (err)
232 		return err;
233 
234 	primelockdep(ct);
235 
236 	ct->g2h_wq = alloc_ordered_workqueue("xe-g2h-wq", WQ_MEM_RECLAIM);
237 	if (!ct->g2h_wq)
238 		return -ENOMEM;
239 
240 	spin_lock_init(&ct->fast_lock);
241 	xa_init(&ct->fence_lookup);
242 	INIT_WORK(&ct->g2h_worker, g2h_worker_func);
243 	INIT_DELAYED_WORK(&ct->safe_mode_worker, safe_mode_worker_func);
244 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
245 	spin_lock_init(&ct->dead.lock);
246 	INIT_WORK(&ct->dead.worker, ct_dead_worker_func);
247 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC)
248 	stack_depot_init();
249 #endif
250 #endif
251 	init_waitqueue_head(&ct->wq);
252 	init_waitqueue_head(&ct->g2h_fence_wq);
253 
254 	err = drmm_add_action_or_reset(&xe->drm, guc_ct_fini, ct);
255 	if (err)
256 		return err;
257 
258 	xe_gt_assert(gt, ct->state == XE_GUC_CT_STATE_NOT_INITIALIZED);
259 	ct->state = XE_GUC_CT_STATE_DISABLED;
260 	return 0;
261 }
262 ALLOW_ERROR_INJECTION(xe_guc_ct_init_noalloc, ERRNO); /* See xe_pci_probe() */
263 
guc_action_disable_ct(void * arg)264 static void guc_action_disable_ct(void *arg)
265 {
266 	struct xe_guc_ct *ct = arg;
267 
268 	guc_ct_change_state(ct, XE_GUC_CT_STATE_DISABLED);
269 }
270 
xe_guc_ct_init(struct xe_guc_ct * ct)271 int xe_guc_ct_init(struct xe_guc_ct *ct)
272 {
273 	struct xe_device *xe = ct_to_xe(ct);
274 	struct xe_gt *gt = ct_to_gt(ct);
275 	struct xe_tile *tile = gt_to_tile(gt);
276 	struct xe_bo *bo;
277 
278 	bo = xe_managed_bo_create_pin_map(xe, tile, guc_ct_size(),
279 					  XE_BO_FLAG_SYSTEM |
280 					  XE_BO_FLAG_GGTT |
281 					  XE_BO_FLAG_GGTT_INVALIDATE |
282 					  XE_BO_FLAG_PINNED_NORESTORE);
283 	if (IS_ERR(bo))
284 		return PTR_ERR(bo);
285 
286 	ct->bo = bo;
287 
288 	return devm_add_action_or_reset(xe->drm.dev, guc_action_disable_ct, ct);
289 }
290 ALLOW_ERROR_INJECTION(xe_guc_ct_init, ERRNO); /* See xe_pci_probe() */
291 
292 /**
293  * xe_guc_ct_init_post_hwconfig - Reinitialize the GuC CTB in VRAM
294  * @ct: the &xe_guc_ct
295  *
296  * Allocate a new BO in VRAM and free the previous BO that was allocated
297  * in system memory (SMEM). Applicable only for DGFX products.
298  *
299  * Return: 0 on success, or a negative errno on failure.
300  */
xe_guc_ct_init_post_hwconfig(struct xe_guc_ct * ct)301 int xe_guc_ct_init_post_hwconfig(struct xe_guc_ct *ct)
302 {
303 	struct xe_device *xe = ct_to_xe(ct);
304 	struct xe_gt *gt = ct_to_gt(ct);
305 	struct xe_tile *tile = gt_to_tile(gt);
306 	int ret;
307 
308 	xe_assert(xe, !xe_guc_ct_enabled(ct));
309 
310 	if (IS_DGFX(xe)) {
311 		ret = xe_managed_bo_reinit_in_vram(xe, tile, &ct->bo);
312 		if (ret)
313 			return ret;
314 	}
315 
316 	devm_remove_action(xe->drm.dev, guc_action_disable_ct, ct);
317 	return devm_add_action_or_reset(xe->drm.dev, guc_action_disable_ct, ct);
318 }
319 
320 #define desc_read(xe_, guc_ctb__, field_)			\
321 	xe_map_rd_field(xe_, &guc_ctb__->desc, 0,		\
322 			struct guc_ct_buffer_desc, field_)
323 
324 #define desc_write(xe_, guc_ctb__, field_, val_)		\
325 	xe_map_wr_field(xe_, &guc_ctb__->desc, 0,		\
326 			struct guc_ct_buffer_desc, field_, val_)
327 
guc_ct_ctb_h2g_init(struct xe_device * xe,struct guc_ctb * h2g,struct iosys_map * map)328 static void guc_ct_ctb_h2g_init(struct xe_device *xe, struct guc_ctb *h2g,
329 				struct iosys_map *map)
330 {
331 	h2g->info.size = CTB_H2G_BUFFER_SIZE / sizeof(u32);
332 	h2g->info.resv_space = 0;
333 	h2g->info.tail = 0;
334 	h2g->info.head = 0;
335 	h2g->info.space = CIRC_SPACE(h2g->info.tail, h2g->info.head,
336 				     h2g->info.size) -
337 			  h2g->info.resv_space;
338 	h2g->info.broken = false;
339 
340 	h2g->desc = *map;
341 	xe_map_memset(xe, &h2g->desc, 0, 0, sizeof(struct guc_ct_buffer_desc));
342 
343 	h2g->cmds = IOSYS_MAP_INIT_OFFSET(map, CTB_H2G_BUFFER_OFFSET);
344 }
345 
guc_ct_ctb_g2h_init(struct xe_device * xe,struct guc_ctb * g2h,struct iosys_map * map)346 static void guc_ct_ctb_g2h_init(struct xe_device *xe, struct guc_ctb *g2h,
347 				struct iosys_map *map)
348 {
349 	g2h->info.size = CTB_G2H_BUFFER_SIZE / sizeof(u32);
350 	g2h->info.resv_space = G2H_ROOM_BUFFER_SIZE / sizeof(u32);
351 	g2h->info.head = 0;
352 	g2h->info.tail = 0;
353 	g2h->info.space = CIRC_SPACE(g2h->info.tail, g2h->info.head,
354 				     g2h->info.size) -
355 			  g2h->info.resv_space;
356 	g2h->info.broken = false;
357 
358 	g2h->desc = IOSYS_MAP_INIT_OFFSET(map, CTB_DESC_SIZE);
359 	xe_map_memset(xe, &g2h->desc, 0, 0, sizeof(struct guc_ct_buffer_desc));
360 
361 	g2h->cmds = IOSYS_MAP_INIT_OFFSET(map, CTB_H2G_BUFFER_OFFSET +
362 					    CTB_H2G_BUFFER_SIZE);
363 }
364 
guc_ct_ctb_h2g_register(struct xe_guc_ct * ct)365 static int guc_ct_ctb_h2g_register(struct xe_guc_ct *ct)
366 {
367 	struct xe_guc *guc = ct_to_guc(ct);
368 	u32 desc_addr, ctb_addr, size;
369 	int err;
370 
371 	desc_addr = xe_bo_ggtt_addr(ct->bo);
372 	ctb_addr = xe_bo_ggtt_addr(ct->bo) + CTB_H2G_BUFFER_OFFSET;
373 	size = ct->ctbs.h2g.info.size * sizeof(u32);
374 
375 	err = xe_guc_self_cfg64(guc,
376 				GUC_KLV_SELF_CFG_H2G_CTB_DESCRIPTOR_ADDR_KEY,
377 				desc_addr);
378 	if (err)
379 		return err;
380 
381 	err = xe_guc_self_cfg64(guc,
382 				GUC_KLV_SELF_CFG_H2G_CTB_ADDR_KEY,
383 				ctb_addr);
384 	if (err)
385 		return err;
386 
387 	return xe_guc_self_cfg32(guc,
388 				 GUC_KLV_SELF_CFG_H2G_CTB_SIZE_KEY,
389 				 size);
390 }
391 
guc_ct_ctb_g2h_register(struct xe_guc_ct * ct)392 static int guc_ct_ctb_g2h_register(struct xe_guc_ct *ct)
393 {
394 	struct xe_guc *guc = ct_to_guc(ct);
395 	u32 desc_addr, ctb_addr, size;
396 	int err;
397 
398 	desc_addr = xe_bo_ggtt_addr(ct->bo) + CTB_DESC_SIZE;
399 	ctb_addr = xe_bo_ggtt_addr(ct->bo) + CTB_H2G_BUFFER_OFFSET +
400 		CTB_H2G_BUFFER_SIZE;
401 	size = ct->ctbs.g2h.info.size * sizeof(u32);
402 
403 	err = xe_guc_self_cfg64(guc,
404 				GUC_KLV_SELF_CFG_G2H_CTB_DESCRIPTOR_ADDR_KEY,
405 				desc_addr);
406 	if (err)
407 		return err;
408 
409 	err = xe_guc_self_cfg64(guc,
410 				GUC_KLV_SELF_CFG_G2H_CTB_ADDR_KEY,
411 				ctb_addr);
412 	if (err)
413 		return err;
414 
415 	return xe_guc_self_cfg32(guc,
416 				 GUC_KLV_SELF_CFG_G2H_CTB_SIZE_KEY,
417 				 size);
418 }
419 
guc_ct_control_toggle(struct xe_guc_ct * ct,bool enable)420 static int guc_ct_control_toggle(struct xe_guc_ct *ct, bool enable)
421 {
422 	u32 request[HOST2GUC_CONTROL_CTB_REQUEST_MSG_LEN] = {
423 		FIELD_PREP(GUC_HXG_MSG_0_ORIGIN, GUC_HXG_ORIGIN_HOST) |
424 		FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_REQUEST) |
425 		FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION,
426 			   GUC_ACTION_HOST2GUC_CONTROL_CTB),
427 		FIELD_PREP(HOST2GUC_CONTROL_CTB_REQUEST_MSG_1_CONTROL,
428 			   enable ? GUC_CTB_CONTROL_ENABLE :
429 			   GUC_CTB_CONTROL_DISABLE),
430 	};
431 	int ret = xe_guc_mmio_send(ct_to_guc(ct), request, ARRAY_SIZE(request));
432 
433 	return ret > 0 ? -EPROTO : ret;
434 }
435 
guc_ct_change_state(struct xe_guc_ct * ct,enum xe_guc_ct_state state)436 static void guc_ct_change_state(struct xe_guc_ct *ct,
437 				enum xe_guc_ct_state state)
438 {
439 	struct xe_gt *gt = ct_to_gt(ct);
440 	struct g2h_fence *g2h_fence;
441 	unsigned long idx;
442 
443 	mutex_lock(&ct->lock);		/* Serialise dequeue_one_g2h() */
444 	spin_lock_irq(&ct->fast_lock);	/* Serialise CT fast-path */
445 
446 	xe_gt_assert(ct_to_gt(ct), ct->g2h_outstanding == 0 ||
447 		     state == XE_GUC_CT_STATE_STOPPED);
448 
449 	if (ct->g2h_outstanding)
450 		xe_pm_runtime_put(ct_to_xe(ct));
451 	ct->g2h_outstanding = 0;
452 	ct->state = state;
453 
454 	xe_gt_dbg(gt, "GuC CT communication channel %s\n",
455 		  state == XE_GUC_CT_STATE_STOPPED ? "stopped" :
456 		  str_enabled_disabled(state == XE_GUC_CT_STATE_ENABLED));
457 
458 	spin_unlock_irq(&ct->fast_lock);
459 
460 	/* cancel all in-flight send-recv requests */
461 	xa_for_each(&ct->fence_lookup, idx, g2h_fence)
462 		g2h_fence_cancel(g2h_fence);
463 
464 	/* make sure guc_ct_send_recv() will see g2h_fence changes */
465 	smp_mb();
466 	wake_up_all(&ct->g2h_fence_wq);
467 
468 	/*
469 	 * Lockdep doesn't like this under the fast lock and he destroy only
470 	 * needs to be serialized with the send path which ct lock provides.
471 	 */
472 	xa_destroy(&ct->fence_lookup);
473 
474 	mutex_unlock(&ct->lock);
475 }
476 
ct_needs_safe_mode(struct xe_guc_ct * ct)477 static bool ct_needs_safe_mode(struct xe_guc_ct *ct)
478 {
479 	return !pci_dev_msi_enabled(to_pci_dev(ct_to_xe(ct)->drm.dev));
480 }
481 
ct_restart_safe_mode_worker(struct xe_guc_ct * ct)482 static bool ct_restart_safe_mode_worker(struct xe_guc_ct *ct)
483 {
484 	if (!ct_needs_safe_mode(ct))
485 		return false;
486 
487 	queue_delayed_work(ct->g2h_wq, &ct->safe_mode_worker, HZ / 10);
488 	return true;
489 }
490 
safe_mode_worker_func(struct work_struct * w)491 static void safe_mode_worker_func(struct work_struct *w)
492 {
493 	struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, safe_mode_worker.work);
494 
495 	receive_g2h(ct);
496 
497 	if (!ct_restart_safe_mode_worker(ct))
498 		xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode canceled\n");
499 }
500 
ct_enter_safe_mode(struct xe_guc_ct * ct)501 static void ct_enter_safe_mode(struct xe_guc_ct *ct)
502 {
503 	if (ct_restart_safe_mode_worker(ct))
504 		xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode enabled\n");
505 }
506 
ct_exit_safe_mode(struct xe_guc_ct * ct)507 static void ct_exit_safe_mode(struct xe_guc_ct *ct)
508 {
509 	if (cancel_delayed_work_sync(&ct->safe_mode_worker))
510 		xe_gt_dbg(ct_to_gt(ct), "GuC CT safe-mode disabled\n");
511 }
512 
__xe_guc_ct_start(struct xe_guc_ct * ct,bool needs_register)513 static int __xe_guc_ct_start(struct xe_guc_ct *ct, bool needs_register)
514 {
515 	struct xe_device *xe = ct_to_xe(ct);
516 	struct xe_gt *gt = ct_to_gt(ct);
517 	int err;
518 
519 	xe_gt_assert(gt, !xe_guc_ct_enabled(ct));
520 
521 	if (needs_register) {
522 		xe_map_memset(xe, &ct->bo->vmap, 0, 0, xe_bo_size(ct->bo));
523 		guc_ct_ctb_h2g_init(xe, &ct->ctbs.h2g, &ct->bo->vmap);
524 		guc_ct_ctb_g2h_init(xe, &ct->ctbs.g2h, &ct->bo->vmap);
525 
526 		err = guc_ct_ctb_h2g_register(ct);
527 		if (err)
528 			goto err_out;
529 
530 		err = guc_ct_ctb_g2h_register(ct);
531 		if (err)
532 			goto err_out;
533 
534 		err = guc_ct_control_toggle(ct, true);
535 		if (err)
536 			goto err_out;
537 	} else {
538 		ct->ctbs.h2g.info.broken = false;
539 		ct->ctbs.g2h.info.broken = false;
540 		/* Skip everything in H2G buffer */
541 		xe_map_memset(xe, &ct->bo->vmap, CTB_H2G_BUFFER_OFFSET, 0,
542 			      CTB_H2G_BUFFER_SIZE);
543 	}
544 
545 	guc_ct_change_state(ct, XE_GUC_CT_STATE_ENABLED);
546 
547 	smp_mb();
548 	wake_up_all(&ct->wq);
549 
550 	if (ct_needs_safe_mode(ct))
551 		ct_enter_safe_mode(ct);
552 
553 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
554 	/*
555 	 * The CT has now been reset so the dumper can be re-armed
556 	 * after any existing dead state has been dumped.
557 	 */
558 	spin_lock_irq(&ct->dead.lock);
559 	if (ct->dead.reason) {
560 		ct->dead.reason |= (1 << CT_DEAD_STATE_REARM);
561 		queue_work(system_unbound_wq, &ct->dead.worker);
562 	}
563 	spin_unlock_irq(&ct->dead.lock);
564 #endif
565 
566 	return 0;
567 
568 err_out:
569 	xe_gt_err(gt, "Failed to enable GuC CT (%pe)\n", ERR_PTR(err));
570 	CT_DEAD(ct, NULL, SETUP);
571 
572 	return err;
573 }
574 
575 /**
576  * xe_guc_ct_restart() - Restart GuC CT
577  * @ct: the &xe_guc_ct
578  *
579  * Restart GuC CT to an empty state without issuing a CT register MMIO command.
580  *
581  * Return: 0 on success, or a negative errno on failure.
582  */
xe_guc_ct_restart(struct xe_guc_ct * ct)583 int xe_guc_ct_restart(struct xe_guc_ct *ct)
584 {
585 	return __xe_guc_ct_start(ct, false);
586 }
587 
588 /**
589  * xe_guc_ct_enable() - Enable GuC CT
590  * @ct: the &xe_guc_ct
591  *
592  * Enable GuC CT to an empty state and issue a CT register MMIO command.
593  *
594  * Return: 0 on success, or a negative errno on failure.
595  */
xe_guc_ct_enable(struct xe_guc_ct * ct)596 int xe_guc_ct_enable(struct xe_guc_ct *ct)
597 {
598 	return __xe_guc_ct_start(ct, true);
599 }
600 
stop_g2h_handler(struct xe_guc_ct * ct)601 static void stop_g2h_handler(struct xe_guc_ct *ct)
602 {
603 	cancel_work_sync(&ct->g2h_worker);
604 }
605 
606 /**
607  * xe_guc_ct_disable - Set GuC to disabled state
608  * @ct: the &xe_guc_ct
609  *
610  * Set GuC CT to disabled state and stop g2h handler. No outstanding g2h expected
611  * in this transition.
612  */
xe_guc_ct_disable(struct xe_guc_ct * ct)613 void xe_guc_ct_disable(struct xe_guc_ct *ct)
614 {
615 	guc_ct_change_state(ct, XE_GUC_CT_STATE_DISABLED);
616 	ct_exit_safe_mode(ct);
617 	stop_g2h_handler(ct);
618 }
619 
620 /**
621  * xe_guc_ct_flush_and_stop - Flush and stop all processing of G2H / H2G
622  * @ct: the &xe_guc_ct
623  */
xe_guc_ct_flush_and_stop(struct xe_guc_ct * ct)624 void xe_guc_ct_flush_and_stop(struct xe_guc_ct *ct)
625 {
626 	receive_g2h(ct);
627 	xe_guc_ct_stop(ct);
628 }
629 
630 /**
631  * xe_guc_ct_stop - Set GuC to stopped state
632  * @ct: the &xe_guc_ct
633  *
634  * Set GuC CT to stopped state, stop g2h handler, and clear any outstanding g2h
635  */
xe_guc_ct_stop(struct xe_guc_ct * ct)636 void xe_guc_ct_stop(struct xe_guc_ct *ct)
637 {
638 	if (!xe_guc_ct_initialized(ct))
639 		return;
640 
641 	guc_ct_change_state(ct, XE_GUC_CT_STATE_STOPPED);
642 	stop_g2h_handler(ct);
643 }
644 
h2g_has_room(struct xe_guc_ct * ct,u32 cmd_len)645 static bool h2g_has_room(struct xe_guc_ct *ct, u32 cmd_len)
646 {
647 	struct guc_ctb *h2g = &ct->ctbs.h2g;
648 
649 	lockdep_assert_held(&ct->lock);
650 
651 	if (cmd_len > h2g->info.space) {
652 		h2g->info.head = desc_read(ct_to_xe(ct), h2g, head);
653 
654 		if (h2g->info.head > h2g->info.size) {
655 			struct xe_device *xe = ct_to_xe(ct);
656 			u32 desc_status = desc_read(xe, h2g, status);
657 
658 			desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW);
659 
660 			xe_gt_err(ct_to_gt(ct), "CT: invalid head offset %u >= %u)\n",
661 				  h2g->info.head, h2g->info.size);
662 			CT_DEAD(ct, h2g, H2G_HAS_ROOM);
663 			return false;
664 		}
665 
666 		h2g->info.space = CIRC_SPACE(h2g->info.tail, h2g->info.head,
667 					     h2g->info.size) -
668 				  h2g->info.resv_space;
669 		if (cmd_len > h2g->info.space)
670 			return false;
671 	}
672 
673 	return true;
674 }
675 
g2h_has_room(struct xe_guc_ct * ct,u32 g2h_len)676 static bool g2h_has_room(struct xe_guc_ct *ct, u32 g2h_len)
677 {
678 	if (!g2h_len)
679 		return true;
680 
681 	lockdep_assert_held(&ct->fast_lock);
682 
683 	return ct->ctbs.g2h.info.space > g2h_len;
684 }
685 
has_room(struct xe_guc_ct * ct,u32 cmd_len,u32 g2h_len)686 static int has_room(struct xe_guc_ct *ct, u32 cmd_len, u32 g2h_len)
687 {
688 	lockdep_assert_held(&ct->lock);
689 
690 	if (!g2h_has_room(ct, g2h_len) || !h2g_has_room(ct, cmd_len))
691 		return -EBUSY;
692 
693 	return 0;
694 }
695 
h2g_reserve_space(struct xe_guc_ct * ct,u32 cmd_len)696 static void h2g_reserve_space(struct xe_guc_ct *ct, u32 cmd_len)
697 {
698 	lockdep_assert_held(&ct->lock);
699 	ct->ctbs.h2g.info.space -= cmd_len;
700 }
701 
__g2h_reserve_space(struct xe_guc_ct * ct,u32 g2h_len,u32 num_g2h)702 static void __g2h_reserve_space(struct xe_guc_ct *ct, u32 g2h_len, u32 num_g2h)
703 {
704 	xe_gt_assert(ct_to_gt(ct), g2h_len <= ct->ctbs.g2h.info.space);
705 	xe_gt_assert(ct_to_gt(ct), (!g2h_len && !num_g2h) ||
706 		     (g2h_len && num_g2h));
707 
708 	if (g2h_len) {
709 		lockdep_assert_held(&ct->fast_lock);
710 
711 		if (!ct->g2h_outstanding)
712 			xe_pm_runtime_get_noresume(ct_to_xe(ct));
713 
714 		ct->ctbs.g2h.info.space -= g2h_len;
715 		ct->g2h_outstanding += num_g2h;
716 	}
717 }
718 
__g2h_release_space(struct xe_guc_ct * ct,u32 g2h_len)719 static void __g2h_release_space(struct xe_guc_ct *ct, u32 g2h_len)
720 {
721 	bool bad = false;
722 
723 	lockdep_assert_held(&ct->fast_lock);
724 
725 	bad = ct->ctbs.g2h.info.space + g2h_len >
726 		     ct->ctbs.g2h.info.size - ct->ctbs.g2h.info.resv_space;
727 	bad |= !ct->g2h_outstanding;
728 
729 	if (bad) {
730 		xe_gt_err(ct_to_gt(ct), "Invalid G2H release: %d + %d vs %d - %d -> %d vs %d, outstanding = %d!\n",
731 			  ct->ctbs.g2h.info.space, g2h_len,
732 			  ct->ctbs.g2h.info.size, ct->ctbs.g2h.info.resv_space,
733 			  ct->ctbs.g2h.info.space + g2h_len,
734 			  ct->ctbs.g2h.info.size - ct->ctbs.g2h.info.resv_space,
735 			  ct->g2h_outstanding);
736 		CT_DEAD(ct, &ct->ctbs.g2h, G2H_RELEASE);
737 		return;
738 	}
739 
740 	ct->ctbs.g2h.info.space += g2h_len;
741 	if (!--ct->g2h_outstanding)
742 		xe_pm_runtime_put(ct_to_xe(ct));
743 }
744 
g2h_release_space(struct xe_guc_ct * ct,u32 g2h_len)745 static void g2h_release_space(struct xe_guc_ct *ct, u32 g2h_len)
746 {
747 	spin_lock_irq(&ct->fast_lock);
748 	__g2h_release_space(ct, g2h_len);
749 	spin_unlock_irq(&ct->fast_lock);
750 }
751 
752 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
fast_req_track(struct xe_guc_ct * ct,u16 fence,u16 action)753 static void fast_req_track(struct xe_guc_ct *ct, u16 fence, u16 action)
754 {
755 	unsigned int slot = fence % ARRAY_SIZE(ct->fast_req);
756 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC)
757 	unsigned long entries[SZ_32];
758 	unsigned int n;
759 
760 	n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
761 
762 	/* May be called under spinlock, so avoid sleeping */
763 	ct->fast_req[slot].stack = stack_depot_save(entries, n, GFP_NOWAIT);
764 #endif
765 	ct->fast_req[slot].fence = fence;
766 	ct->fast_req[slot].action = action;
767 }
768 #else
fast_req_track(struct xe_guc_ct * ct,u16 fence,u16 action)769 static void fast_req_track(struct xe_guc_ct *ct, u16 fence, u16 action)
770 {
771 }
772 #endif
773 
774 /*
775  * The CT protocol accepts a 16 bits fence. This field is fully owned by the
776  * driver, the GuC will just copy it to the reply message. Since we need to
777  * be able to distinguish between replies to REQUEST and FAST_REQUEST messages,
778  * we use one bit of the seqno as an indicator for that and a rolling counter
779  * for the remaining 15 bits.
780  */
781 #define CT_SEQNO_MASK GENMASK(14, 0)
782 #define CT_SEQNO_UNTRACKED BIT(15)
next_ct_seqno(struct xe_guc_ct * ct,bool is_g2h_fence)783 static u16 next_ct_seqno(struct xe_guc_ct *ct, bool is_g2h_fence)
784 {
785 	u32 seqno = ct->fence_seqno++ & CT_SEQNO_MASK;
786 
787 	if (!is_g2h_fence)
788 		seqno |= CT_SEQNO_UNTRACKED;
789 
790 	return seqno;
791 }
792 
793 #define MAKE_ACTION(type, __action)				\
794 ({								\
795 	FIELD_PREP(GUC_HXG_MSG_0_TYPE, type) |			\
796 	FIELD_PREP(GUC_HXG_EVENT_MSG_0_ACTION |			\
797 		   GUC_HXG_EVENT_MSG_0_DATA0, __action);	\
798 })
799 
vf_action_can_safely_fail(struct xe_device * xe,u32 action)800 static bool vf_action_can_safely_fail(struct xe_device *xe, u32 action)
801 {
802 	/*
803 	 * When resuming a VF, we can't reliably track whether context
804 	 * registration has completed in the GuC state machine. It is harmless
805 	 * to resend the request, as it will fail silently if GUC_HXG_TYPE_EVENT
806 	 * is used. Additionally, if there is an H2G protocol issue on a VF,
807 	 * subsequent H2G messages sent as GUC_HXG_TYPE_FAST_REQUEST will likely
808 	 * fail.
809 	 */
810 	return IS_SRIOV_VF(xe) && xe_sriov_vf_migration_supported(xe) &&
811 		(action == XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC ||
812 		 action == XE_GUC_ACTION_REGISTER_CONTEXT);
813 }
814 
815 #define H2G_CT_HEADERS (GUC_CTB_HDR_LEN + 1) /* one DW CTB header and one DW HxG header */
816 
h2g_write(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 ct_fence_value,bool want_response)817 static int h2g_write(struct xe_guc_ct *ct, const u32 *action, u32 len,
818 		     u32 ct_fence_value, bool want_response)
819 {
820 	struct xe_device *xe = ct_to_xe(ct);
821 	struct xe_gt *gt = ct_to_gt(ct);
822 	struct guc_ctb *h2g = &ct->ctbs.h2g;
823 	u32 cmd[H2G_CT_HEADERS];
824 	u32 tail = h2g->info.tail;
825 	u32 full_len;
826 	struct iosys_map map = IOSYS_MAP_INIT_OFFSET(&h2g->cmds,
827 							 tail * sizeof(u32));
828 	u32 desc_status;
829 
830 	full_len = len + GUC_CTB_HDR_LEN;
831 
832 	lockdep_assert_held(&ct->lock);
833 	xe_gt_assert(gt, full_len <= GUC_CTB_MSG_MAX_LEN);
834 
835 	desc_status = desc_read(xe, h2g, status);
836 	if (desc_status) {
837 		xe_gt_err(gt, "CT write: non-zero status: %u\n", desc_status);
838 		goto corrupted;
839 	}
840 
841 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
842 		u32 desc_tail = desc_read(xe, h2g, tail);
843 		u32 desc_head = desc_read(xe, h2g, head);
844 
845 		if (tail != desc_tail) {
846 			desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_MISMATCH);
847 			xe_gt_err(gt, "CT write: tail was modified %u != %u\n", desc_tail, tail);
848 			goto corrupted;
849 		}
850 
851 		if (tail > h2g->info.size) {
852 			desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW);
853 			xe_gt_err(gt, "CT write: tail out of range: %u vs %u\n",
854 				  tail, h2g->info.size);
855 			goto corrupted;
856 		}
857 
858 		if (desc_head >= h2g->info.size) {
859 			desc_write(xe, h2g, status, desc_status | GUC_CTB_STATUS_OVERFLOW);
860 			xe_gt_err(gt, "CT write: invalid head offset %u >= %u)\n",
861 				  desc_head, h2g->info.size);
862 			goto corrupted;
863 		}
864 	}
865 
866 	/* Command will wrap, zero fill (NOPs), return and check credits again */
867 	if (tail + full_len > h2g->info.size) {
868 		xe_map_memset(xe, &map, 0, 0,
869 			      (h2g->info.size - tail) * sizeof(u32));
870 		h2g_reserve_space(ct, (h2g->info.size - tail));
871 		h2g->info.tail = 0;
872 		desc_write(xe, h2g, tail, h2g->info.tail);
873 
874 		return -EAGAIN;
875 	}
876 
877 	/*
878 	 * dw0: CT header (including fence)
879 	 * dw1: HXG header (including action code)
880 	 * dw2+: action data
881 	 */
882 	cmd[0] = FIELD_PREP(GUC_CTB_MSG_0_FORMAT, GUC_CTB_FORMAT_HXG) |
883 		FIELD_PREP(GUC_CTB_MSG_0_NUM_DWORDS, len) |
884 		FIELD_PREP(GUC_CTB_MSG_0_FENCE, ct_fence_value);
885 	if (want_response) {
886 		cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_REQUEST, action[0]);
887 	} else if (vf_action_can_safely_fail(xe, action[0])) {
888 		cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_EVENT, action[0]);
889 	} else {
890 		fast_req_track(ct, ct_fence_value,
891 			       FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, action[0]));
892 
893 		cmd[1] = MAKE_ACTION(GUC_HXG_TYPE_FAST_REQUEST, action[0]);
894 	}
895 
896 	/* H2G header in cmd[1] replaces action[0] so: */
897 	--len;
898 	++action;
899 
900 	/* Write H2G ensuring visible before descriptor update */
901 	xe_map_memcpy_to(xe, &map, 0, cmd, H2G_CT_HEADERS * sizeof(u32));
902 	xe_map_memcpy_to(xe, &map, H2G_CT_HEADERS * sizeof(u32), action, len * sizeof(u32));
903 	xe_device_wmb(xe);
904 
905 	/* Update local copies */
906 	h2g->info.tail = (tail + full_len) % h2g->info.size;
907 	h2g_reserve_space(ct, full_len);
908 
909 	/* Update descriptor */
910 	desc_write(xe, h2g, tail, h2g->info.tail);
911 
912 	trace_xe_guc_ctb_h2g(xe, gt->info.id, *(action - 1), full_len,
913 			     desc_read(xe, h2g, head), h2g->info.tail);
914 
915 	return 0;
916 
917 corrupted:
918 	CT_DEAD(ct, &ct->ctbs.h2g, H2G_WRITE);
919 	return -EPIPE;
920 }
921 
__guc_ct_send_locked(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 g2h_len,u32 num_g2h,struct g2h_fence * g2h_fence)922 static int __guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action,
923 				u32 len, u32 g2h_len, u32 num_g2h,
924 				struct g2h_fence *g2h_fence)
925 {
926 	struct xe_gt *gt = ct_to_gt(ct);
927 	u16 seqno;
928 	int ret;
929 
930 	xe_gt_assert(gt, xe_guc_ct_initialized(ct));
931 	xe_gt_assert(gt, !g2h_len || !g2h_fence);
932 	xe_gt_assert(gt, !num_g2h || !g2h_fence);
933 	xe_gt_assert(gt, !g2h_len || num_g2h);
934 	xe_gt_assert(gt, g2h_len || !num_g2h);
935 	lockdep_assert_held(&ct->lock);
936 
937 	if (unlikely(ct->ctbs.h2g.info.broken)) {
938 		ret = -EPIPE;
939 		goto out;
940 	}
941 
942 	if (ct->state == XE_GUC_CT_STATE_DISABLED) {
943 		ret = -ENODEV;
944 		goto out;
945 	}
946 
947 	if (ct->state == XE_GUC_CT_STATE_STOPPED || xe_gt_recovery_pending(gt)) {
948 		ret = -ECANCELED;
949 		goto out;
950 	}
951 
952 	xe_gt_assert(gt, xe_guc_ct_enabled(ct));
953 
954 	if (g2h_fence) {
955 		g2h_len = GUC_CTB_HXG_MSG_MAX_LEN;
956 		num_g2h = 1;
957 
958 		if (g2h_fence_needs_alloc(g2h_fence)) {
959 			g2h_fence->seqno = next_ct_seqno(ct, true);
960 			ret = xa_err(xa_store(&ct->fence_lookup,
961 					      g2h_fence->seqno, g2h_fence,
962 					      GFP_ATOMIC));
963 			if (ret)
964 				goto out;
965 		}
966 
967 		seqno = g2h_fence->seqno;
968 	} else {
969 		seqno = next_ct_seqno(ct, false);
970 	}
971 
972 	if (g2h_len)
973 		spin_lock_irq(&ct->fast_lock);
974 retry:
975 	ret = has_room(ct, len + GUC_CTB_HDR_LEN, g2h_len);
976 	if (unlikely(ret))
977 		goto out_unlock;
978 
979 	ret = h2g_write(ct, action, len, seqno, !!g2h_fence);
980 	if (unlikely(ret)) {
981 		if (ret == -EAGAIN)
982 			goto retry;
983 		goto out_unlock;
984 	}
985 
986 	__g2h_reserve_space(ct, g2h_len, num_g2h);
987 	xe_guc_notify(ct_to_guc(ct));
988 out_unlock:
989 	if (g2h_len)
990 		spin_unlock_irq(&ct->fast_lock);
991 out:
992 	return ret;
993 }
994 
kick_reset(struct xe_guc_ct * ct)995 static void kick_reset(struct xe_guc_ct *ct)
996 {
997 	xe_gt_reset_async(ct_to_gt(ct));
998 }
999 
1000 static int dequeue_one_g2h(struct xe_guc_ct *ct);
1001 
1002 /*
1003  * wait before retry of sending h2g message
1004  * Return: true if ready for retry, false if the wait timeouted
1005  */
guc_ct_send_wait_for_retry(struct xe_guc_ct * ct,u32 len,u32 g2h_len,struct g2h_fence * g2h_fence,unsigned int * sleep_period_ms)1006 static bool guc_ct_send_wait_for_retry(struct xe_guc_ct *ct, u32 len,
1007 				       u32 g2h_len, struct g2h_fence *g2h_fence,
1008 				       unsigned int *sleep_period_ms)
1009 {
1010 	struct xe_device *xe = ct_to_xe(ct);
1011 
1012 	/*
1013 	 * We wait to try to restore credits for about 1 second before bailing.
1014 	 * In the case of H2G credits we have no choice but just to wait for the
1015 	 * GuC to consume H2Gs in the channel so we use a wait / sleep loop. In
1016 	 * the case of G2H we process any G2H in the channel, hopefully freeing
1017 	 * credits as we consume the G2H messages.
1018 	 */
1019 	if (!h2g_has_room(ct, len + GUC_CTB_HDR_LEN)) {
1020 		struct guc_ctb *h2g = &ct->ctbs.h2g;
1021 
1022 		if (*sleep_period_ms == 1024)
1023 			return false;
1024 
1025 		trace_xe_guc_ct_h2g_flow_control(xe, h2g->info.head, h2g->info.tail,
1026 						 h2g->info.size,
1027 						 h2g->info.space,
1028 						 len + GUC_CTB_HDR_LEN);
1029 		msleep(*sleep_period_ms);
1030 		*sleep_period_ms <<= 1;
1031 	} else {
1032 		struct xe_device *xe = ct_to_xe(ct);
1033 		struct guc_ctb *g2h = &ct->ctbs.g2h;
1034 		int ret;
1035 
1036 		trace_xe_guc_ct_g2h_flow_control(xe, g2h->info.head,
1037 						 desc_read(xe, g2h, tail),
1038 						 g2h->info.size,
1039 						 g2h->info.space,
1040 						 g2h_fence ?
1041 						 GUC_CTB_HXG_MSG_MAX_LEN :
1042 						 g2h_len);
1043 
1044 #define g2h_avail(ct)	\
1045 	(desc_read(ct_to_xe(ct), (&ct->ctbs.g2h), tail) != ct->ctbs.g2h.info.head)
1046 		if (!wait_event_timeout(ct->wq, !ct->g2h_outstanding ||
1047 					g2h_avail(ct), HZ))
1048 			return false;
1049 #undef g2h_avail
1050 
1051 		ret = dequeue_one_g2h(ct);
1052 		if (ret < 0) {
1053 			if (ret != -ECANCELED)
1054 				xe_gt_err(ct_to_gt(ct), "CTB receive failed (%pe)",
1055 					  ERR_PTR(ret));
1056 			return false;
1057 		}
1058 	}
1059 	return true;
1060 }
1061 
guc_ct_send_locked(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 g2h_len,u32 num_g2h,struct g2h_fence * g2h_fence)1062 static int guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action, u32 len,
1063 			      u32 g2h_len, u32 num_g2h,
1064 			      struct g2h_fence *g2h_fence)
1065 {
1066 	struct xe_gt *gt = ct_to_gt(ct);
1067 	unsigned int sleep_period_ms = 1;
1068 	int ret;
1069 
1070 	xe_gt_assert(gt, !g2h_len || !g2h_fence);
1071 	lockdep_assert_held(&ct->lock);
1072 	xe_device_assert_mem_access(ct_to_xe(ct));
1073 
1074 try_again:
1075 	ret = __guc_ct_send_locked(ct, action, len, g2h_len, num_g2h,
1076 				   g2h_fence);
1077 
1078 	if (unlikely(ret == -EBUSY)) {
1079 		if (!guc_ct_send_wait_for_retry(ct, len, g2h_len, g2h_fence,
1080 						&sleep_period_ms))
1081 			goto broken;
1082 		goto try_again;
1083 	}
1084 
1085 	return ret;
1086 
1087 broken:
1088 	xe_gt_err(gt, "No forward process on H2G, reset required\n");
1089 	CT_DEAD(ct, &ct->ctbs.h2g, DEADLOCK);
1090 
1091 	return -EDEADLK;
1092 }
1093 
guc_ct_send(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 g2h_len,u32 num_g2h,struct g2h_fence * g2h_fence)1094 static int guc_ct_send(struct xe_guc_ct *ct, const u32 *action, u32 len,
1095 		       u32 g2h_len, u32 num_g2h, struct g2h_fence *g2h_fence)
1096 {
1097 	int ret;
1098 
1099 	xe_gt_assert(ct_to_gt(ct), !g2h_len || !g2h_fence);
1100 
1101 	mutex_lock(&ct->lock);
1102 	ret = guc_ct_send_locked(ct, action, len, g2h_len, num_g2h, g2h_fence);
1103 	mutex_unlock(&ct->lock);
1104 
1105 	return ret;
1106 }
1107 
xe_guc_ct_send(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 g2h_len,u32 num_g2h)1108 int xe_guc_ct_send(struct xe_guc_ct *ct, const u32 *action, u32 len,
1109 		   u32 g2h_len, u32 num_g2h)
1110 {
1111 	int ret;
1112 
1113 	ret = guc_ct_send(ct, action, len, g2h_len, num_g2h, NULL);
1114 	if (ret == -EDEADLK)
1115 		kick_reset(ct);
1116 
1117 	return ret;
1118 }
1119 
xe_guc_ct_send_locked(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 g2h_len,u32 num_g2h)1120 int xe_guc_ct_send_locked(struct xe_guc_ct *ct, const u32 *action, u32 len,
1121 			  u32 g2h_len, u32 num_g2h)
1122 {
1123 	int ret;
1124 
1125 	ret = guc_ct_send_locked(ct, action, len, g2h_len, num_g2h, NULL);
1126 	if (ret == -EDEADLK)
1127 		kick_reset(ct);
1128 
1129 	return ret;
1130 }
1131 
xe_guc_ct_send_g2h_handler(struct xe_guc_ct * ct,const u32 * action,u32 len)1132 int xe_guc_ct_send_g2h_handler(struct xe_guc_ct *ct, const u32 *action, u32 len)
1133 {
1134 	int ret;
1135 
1136 	lockdep_assert_held(&ct->lock);
1137 
1138 	ret = guc_ct_send_locked(ct, action, len, 0, 0, NULL);
1139 	if (ret == -EDEADLK)
1140 		kick_reset(ct);
1141 
1142 	return ret;
1143 }
1144 
1145 /*
1146  * Check if a GT reset is in progress or will occur and if GT reset brought the
1147  * CT back up. Randomly picking 5 seconds for an upper limit to do a GT a reset.
1148  */
retry_failure(struct xe_guc_ct * ct,int ret)1149 static bool retry_failure(struct xe_guc_ct *ct, int ret)
1150 {
1151 	if (!(ret == -EDEADLK || ret == -EPIPE || ret == -ENODEV))
1152 		return false;
1153 
1154 #define ct_alive(ct)	\
1155 	(xe_guc_ct_enabled(ct) && !ct->ctbs.h2g.info.broken && \
1156 	 !ct->ctbs.g2h.info.broken)
1157 	if (!wait_event_interruptible_timeout(ct->wq, ct_alive(ct), HZ * 5))
1158 		return false;
1159 #undef ct_alive
1160 
1161 	return true;
1162 }
1163 
1164 #define GUC_SEND_RETRY_LIMIT	50
1165 #define GUC_SEND_RETRY_MSLEEP	5
1166 
guc_ct_send_recv(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 * response_buffer,bool no_fail)1167 static int guc_ct_send_recv(struct xe_guc_ct *ct, const u32 *action, u32 len,
1168 			    u32 *response_buffer, bool no_fail)
1169 {
1170 	struct xe_gt *gt = ct_to_gt(ct);
1171 	struct g2h_fence g2h_fence;
1172 	unsigned int retries = 0;
1173 	int ret = 0;
1174 
1175 	/*
1176 	 * We use a fence to implement blocking sends / receiving response data.
1177 	 * The seqno of the fence is sent in the H2G, returned in the G2H, and
1178 	 * an xarray is used as storage media with the seqno being to key.
1179 	 * Fields in the fence hold success, failure, retry status and the
1180 	 * response data. Safe to allocate on the stack as the xarray is the
1181 	 * only reference and it cannot be present after this function exits.
1182 	 */
1183 retry:
1184 	g2h_fence_init(&g2h_fence, response_buffer);
1185 retry_same_fence:
1186 	ret = guc_ct_send(ct, action, len, 0, 0, &g2h_fence);
1187 	if (unlikely(ret == -ENOMEM)) {
1188 		/* Retry allocation /w GFP_KERNEL */
1189 		ret = xa_err(xa_store(&ct->fence_lookup, g2h_fence.seqno,
1190 				      &g2h_fence, GFP_KERNEL));
1191 		if (ret)
1192 			return ret;
1193 
1194 		goto retry_same_fence;
1195 	} else if (unlikely(ret)) {
1196 		if (ret == -EDEADLK)
1197 			kick_reset(ct);
1198 
1199 		if (no_fail && retry_failure(ct, ret))
1200 			goto retry_same_fence;
1201 
1202 		if (!g2h_fence_needs_alloc(&g2h_fence))
1203 			xa_erase(&ct->fence_lookup, g2h_fence.seqno);
1204 
1205 		return ret;
1206 	}
1207 
1208 	/* READ_ONCEs pairs with WRITE_ONCEs in parse_g2h_response
1209 	 * and g2h_fence_cancel.
1210 	 */
1211 	ret = wait_event_timeout(ct->g2h_fence_wq, READ_ONCE(g2h_fence.done), HZ);
1212 	if (!ret) {
1213 		LNL_FLUSH_WORK(&ct->g2h_worker);
1214 		if (READ_ONCE(g2h_fence.done)) {
1215 			xe_gt_warn(gt, "G2H fence %u, action %04x, done\n",
1216 				   g2h_fence.seqno, action[0]);
1217 			ret = 1;
1218 		}
1219 	}
1220 
1221 	/*
1222 	 * Ensure we serialize with completion side to prevent UAF with fence going out of scope on
1223 	 * the stack, since we have no clue if it will fire after the timeout before we can erase
1224 	 * from the xa. Also we have some dependent loads and stores below for which we need the
1225 	 * correct ordering, and we lack the needed barriers.
1226 	 */
1227 	mutex_lock(&ct->lock);
1228 	if (!ret) {
1229 		xe_gt_err(gt, "Timed out wait for G2H, fence %u, action %04x, done %s",
1230 			  g2h_fence.seqno, action[0], str_yes_no(g2h_fence.done));
1231 		xa_erase(&ct->fence_lookup, g2h_fence.seqno);
1232 		mutex_unlock(&ct->lock);
1233 		return -ETIME;
1234 	}
1235 
1236 	if (g2h_fence.retry) {
1237 		xe_gt_dbg(gt, "H2G action %#x retrying: reason %#x\n",
1238 			  action[0], g2h_fence.reason);
1239 		mutex_unlock(&ct->lock);
1240 		if (++retries > GUC_SEND_RETRY_LIMIT) {
1241 			xe_gt_err(gt, "H2G action %#x reached retry limit=%u, aborting\n",
1242 				  action[0], GUC_SEND_RETRY_LIMIT);
1243 			return -ELOOP;
1244 		}
1245 		msleep(GUC_SEND_RETRY_MSLEEP * retries);
1246 		goto retry;
1247 	}
1248 	if (g2h_fence.fail) {
1249 		if (g2h_fence.cancel) {
1250 			xe_gt_dbg(gt, "H2G request %#x canceled!\n", action[0]);
1251 			ret = -ECANCELED;
1252 			goto unlock;
1253 		}
1254 		xe_gt_err(gt, "H2G request %#x failed: error %#x hint %#x\n",
1255 			  action[0], g2h_fence.error, g2h_fence.hint);
1256 		ret = -EIO;
1257 	}
1258 
1259 	if (ret > 0)
1260 		ret = response_buffer ? g2h_fence.response_len : g2h_fence.response_data;
1261 
1262 unlock:
1263 	mutex_unlock(&ct->lock);
1264 
1265 	return ret;
1266 }
1267 
1268 /**
1269  * xe_guc_ct_send_recv - Send and receive HXG to the GuC
1270  * @ct: the &xe_guc_ct
1271  * @action: the dword array with `HXG Request`_ message (can't be NULL)
1272  * @len: length of the `HXG Request`_ message (in dwords, can't be 0)
1273  * @response_buffer: placeholder for the `HXG Response`_ message (can be NULL)
1274  *
1275  * Send a `HXG Request`_ message to the GuC over CT communication channel and
1276  * blocks until GuC replies with a `HXG Response`_ message.
1277  *
1278  * For non-blocking communication with GuC use xe_guc_ct_send().
1279  *
1280  * Note: The size of &response_buffer must be at least GUC_CTB_MAX_DWORDS_.
1281  *
1282  * Return: response length (in dwords) if &response_buffer was not NULL, or
1283  *         DATA0 from `HXG Response`_ if &response_buffer was NULL, or
1284  *         a negative error code on failure.
1285  */
xe_guc_ct_send_recv(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 * response_buffer)1286 int xe_guc_ct_send_recv(struct xe_guc_ct *ct, const u32 *action, u32 len,
1287 			u32 *response_buffer)
1288 {
1289 	KUNIT_STATIC_STUB_REDIRECT(xe_guc_ct_send_recv, ct, action, len, response_buffer);
1290 	return guc_ct_send_recv(ct, action, len, response_buffer, false);
1291 }
1292 ALLOW_ERROR_INJECTION(xe_guc_ct_send_recv, ERRNO);
1293 
xe_guc_ct_send_recv_no_fail(struct xe_guc_ct * ct,const u32 * action,u32 len,u32 * response_buffer)1294 int xe_guc_ct_send_recv_no_fail(struct xe_guc_ct *ct, const u32 *action,
1295 				u32 len, u32 *response_buffer)
1296 {
1297 	return guc_ct_send_recv(ct, action, len, response_buffer, true);
1298 }
1299 
msg_to_hxg(u32 * msg)1300 static u32 *msg_to_hxg(u32 *msg)
1301 {
1302 	return msg + GUC_CTB_MSG_MIN_LEN;
1303 }
1304 
msg_len_to_hxg_len(u32 len)1305 static u32 msg_len_to_hxg_len(u32 len)
1306 {
1307 	return len - GUC_CTB_MSG_MIN_LEN;
1308 }
1309 
parse_g2h_event(struct xe_guc_ct * ct,u32 * msg,u32 len)1310 static int parse_g2h_event(struct xe_guc_ct *ct, u32 *msg, u32 len)
1311 {
1312 	u32 *hxg = msg_to_hxg(msg);
1313 	u32 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1314 
1315 	lockdep_assert_held(&ct->lock);
1316 
1317 	switch (action) {
1318 	case XE_GUC_ACTION_SCHED_CONTEXT_MODE_DONE:
1319 	case XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE:
1320 	case XE_GUC_ACTION_SCHED_ENGINE_MODE_DONE:
1321 	case XE_GUC_ACTION_TLB_INVALIDATION_DONE:
1322 		g2h_release_space(ct, len);
1323 	}
1324 
1325 	return 0;
1326 }
1327 
guc_crash_process_msg(struct xe_guc_ct * ct,u32 action)1328 static int guc_crash_process_msg(struct xe_guc_ct *ct, u32 action)
1329 {
1330 	struct xe_gt *gt = ct_to_gt(ct);
1331 
1332 	if (action == XE_GUC_ACTION_NOTIFY_CRASH_DUMP_POSTED)
1333 		xe_gt_err(gt, "GuC Crash dump notification\n");
1334 	else if (action == XE_GUC_ACTION_NOTIFY_EXCEPTION)
1335 		xe_gt_err(gt, "GuC Exception notification\n");
1336 	else
1337 		xe_gt_err(gt, "Unknown GuC crash notification: 0x%04X\n", action);
1338 
1339 	CT_DEAD(ct, NULL, CRASH);
1340 
1341 	kick_reset(ct);
1342 
1343 	return 0;
1344 }
1345 
1346 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
fast_req_report(struct xe_guc_ct * ct,u16 fence)1347 static void fast_req_report(struct xe_guc_ct *ct, u16 fence)
1348 {
1349 	u16 fence_min = U16_MAX, fence_max = 0;
1350 	struct xe_gt *gt = ct_to_gt(ct);
1351 	bool found = false;
1352 	unsigned int n;
1353 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC)
1354 	char *buf;
1355 #endif
1356 
1357 	lockdep_assert_held(&ct->lock);
1358 
1359 	for (n = 0; n < ARRAY_SIZE(ct->fast_req); n++) {
1360 		if (ct->fast_req[n].fence < fence_min)
1361 			fence_min = ct->fast_req[n].fence;
1362 		if (ct->fast_req[n].fence > fence_max)
1363 			fence_max = ct->fast_req[n].fence;
1364 
1365 		if (ct->fast_req[n].fence != fence)
1366 			continue;
1367 		found = true;
1368 
1369 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC)
1370 		buf = kmalloc(SZ_4K, GFP_NOWAIT);
1371 		if (buf && stack_depot_snprint(ct->fast_req[n].stack, buf, SZ_4K, 0))
1372 			xe_gt_err(gt, "Fence 0x%x was used by action %#04x sent at:\n%s",
1373 				  fence, ct->fast_req[n].action, buf);
1374 		else
1375 			xe_gt_err(gt, "Fence 0x%x was used by action %#04x [failed to retrieve stack]\n",
1376 				  fence, ct->fast_req[n].action);
1377 		kfree(buf);
1378 #else
1379 		xe_gt_err(gt, "Fence 0x%x was used by action %#04x\n",
1380 			  fence, ct->fast_req[n].action);
1381 #endif
1382 		break;
1383 	}
1384 
1385 	if (!found)
1386 		xe_gt_warn(gt, "Fence 0x%x not found - tracking buffer wrapped? [range = 0x%x -> 0x%x, next = 0x%X]\n",
1387 			   fence, fence_min, fence_max, ct->fence_seqno);
1388 }
1389 #else
fast_req_report(struct xe_guc_ct * ct,u16 fence)1390 static void fast_req_report(struct xe_guc_ct *ct, u16 fence)
1391 {
1392 }
1393 #endif
1394 
parse_g2h_response(struct xe_guc_ct * ct,u32 * msg,u32 len)1395 static int parse_g2h_response(struct xe_guc_ct *ct, u32 *msg, u32 len)
1396 {
1397 	struct xe_gt *gt =  ct_to_gt(ct);
1398 	u32 *hxg = msg_to_hxg(msg);
1399 	u32 hxg_len = msg_len_to_hxg_len(len);
1400 	u32 fence = FIELD_GET(GUC_CTB_MSG_0_FENCE, msg[0]);
1401 	u32 type = FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]);
1402 	struct g2h_fence *g2h_fence;
1403 
1404 	lockdep_assert_held(&ct->lock);
1405 
1406 	/*
1407 	 * Fences for FAST_REQUEST messages are not tracked in ct->fence_lookup.
1408 	 * Those messages should never fail, so if we do get an error back it
1409 	 * means we're likely doing an illegal operation and the GuC is
1410 	 * rejecting it. We have no way to inform the code that submitted the
1411 	 * H2G that the message was rejected, so we need to escalate the
1412 	 * failure to trigger a reset.
1413 	 */
1414 	if (fence & CT_SEQNO_UNTRACKED) {
1415 		if (type == GUC_HXG_TYPE_RESPONSE_FAILURE)
1416 			xe_gt_err(gt, "FAST_REQ H2G fence 0x%x failed! e=0x%x, h=%u\n",
1417 				  fence,
1418 				  FIELD_GET(GUC_HXG_FAILURE_MSG_0_ERROR, hxg[0]),
1419 				  FIELD_GET(GUC_HXG_FAILURE_MSG_0_HINT, hxg[0]));
1420 		else
1421 			xe_gt_err(gt, "unexpected response %u for FAST_REQ H2G fence 0x%x!\n",
1422 				  type, fence);
1423 
1424 		fast_req_report(ct, fence);
1425 
1426 		/* FIXME: W/A race in the GuC, will get in firmware soon */
1427 		if (xe_gt_recovery_pending(gt))
1428 			return 0;
1429 
1430 		CT_DEAD(ct, NULL, PARSE_G2H_RESPONSE);
1431 
1432 		return -EPROTO;
1433 	}
1434 
1435 	g2h_fence = xa_erase(&ct->fence_lookup, fence);
1436 	if (unlikely(!g2h_fence)) {
1437 		/* Don't tear down channel, as send could've timed out */
1438 		/* CT_DEAD(ct, NULL, PARSE_G2H_UNKNOWN); */
1439 		xe_gt_warn(gt, "G2H fence (%u) not found!\n", fence);
1440 		g2h_release_space(ct, GUC_CTB_HXG_MSG_MAX_LEN);
1441 		return 0;
1442 	}
1443 
1444 	xe_gt_assert(gt, fence == g2h_fence->seqno);
1445 
1446 	if (type == GUC_HXG_TYPE_RESPONSE_FAILURE) {
1447 		g2h_fence->fail = true;
1448 		g2h_fence->error = FIELD_GET(GUC_HXG_FAILURE_MSG_0_ERROR, hxg[0]);
1449 		g2h_fence->hint = FIELD_GET(GUC_HXG_FAILURE_MSG_0_HINT, hxg[0]);
1450 	} else if (type == GUC_HXG_TYPE_NO_RESPONSE_RETRY) {
1451 		g2h_fence->retry = true;
1452 		g2h_fence->reason = FIELD_GET(GUC_HXG_RETRY_MSG_0_REASON, hxg[0]);
1453 	} else if (g2h_fence->response_buffer) {
1454 		g2h_fence->response_len = hxg_len;
1455 		memcpy(g2h_fence->response_buffer, hxg, hxg_len * sizeof(u32));
1456 	} else {
1457 		g2h_fence->response_data = FIELD_GET(GUC_HXG_RESPONSE_MSG_0_DATA0, hxg[0]);
1458 	}
1459 
1460 	g2h_release_space(ct, GUC_CTB_HXG_MSG_MAX_LEN);
1461 
1462 	/* WRITE_ONCE pairs with READ_ONCEs in guc_ct_send_recv. */
1463 	WRITE_ONCE(g2h_fence->done, true);
1464 	smp_mb();
1465 
1466 	wake_up_all(&ct->g2h_fence_wq);
1467 
1468 	return 0;
1469 }
1470 
parse_g2h_msg(struct xe_guc_ct * ct,u32 * msg,u32 len)1471 static int parse_g2h_msg(struct xe_guc_ct *ct, u32 *msg, u32 len)
1472 {
1473 	struct xe_gt *gt = ct_to_gt(ct);
1474 	u32 *hxg = msg_to_hxg(msg);
1475 	u32 origin, type;
1476 	int ret;
1477 
1478 	lockdep_assert_held(&ct->lock);
1479 
1480 	origin = FIELD_GET(GUC_HXG_MSG_0_ORIGIN, hxg[0]);
1481 	if (unlikely(origin != GUC_HXG_ORIGIN_GUC)) {
1482 		xe_gt_err(gt, "G2H channel broken on read, origin=%u, reset required\n",
1483 			  origin);
1484 		CT_DEAD(ct, &ct->ctbs.g2h, PARSE_G2H_ORIGIN);
1485 
1486 		return -EPROTO;
1487 	}
1488 
1489 	type = FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]);
1490 	switch (type) {
1491 	case GUC_HXG_TYPE_EVENT:
1492 		ret = parse_g2h_event(ct, msg, len);
1493 		break;
1494 	case GUC_HXG_TYPE_RESPONSE_SUCCESS:
1495 	case GUC_HXG_TYPE_RESPONSE_FAILURE:
1496 	case GUC_HXG_TYPE_NO_RESPONSE_RETRY:
1497 		ret = parse_g2h_response(ct, msg, len);
1498 		break;
1499 	default:
1500 		xe_gt_err(gt, "G2H channel broken on read, type=%u, reset required\n",
1501 			  type);
1502 		CT_DEAD(ct, &ct->ctbs.g2h, PARSE_G2H_TYPE);
1503 
1504 		ret = -EOPNOTSUPP;
1505 	}
1506 
1507 	return ret;
1508 }
1509 
process_g2h_msg(struct xe_guc_ct * ct,u32 * msg,u32 len)1510 static int process_g2h_msg(struct xe_guc_ct *ct, u32 *msg, u32 len)
1511 {
1512 	struct xe_guc *guc = ct_to_guc(ct);
1513 	struct xe_gt *gt = ct_to_gt(ct);
1514 	u32 hxg_len = msg_len_to_hxg_len(len);
1515 	u32 *hxg = msg_to_hxg(msg);
1516 	u32 action, adj_len;
1517 	u32 *payload;
1518 	int ret = 0;
1519 
1520 	if (FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_EVENT)
1521 		return 0;
1522 
1523 	action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1524 	payload = hxg + GUC_HXG_EVENT_MSG_MIN_LEN;
1525 	adj_len = hxg_len - GUC_HXG_EVENT_MSG_MIN_LEN;
1526 
1527 	switch (action) {
1528 	case XE_GUC_ACTION_SCHED_CONTEXT_MODE_DONE:
1529 		ret = xe_guc_sched_done_handler(guc, payload, adj_len);
1530 		break;
1531 	case XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE:
1532 		ret = xe_guc_deregister_done_handler(guc, payload, adj_len);
1533 		break;
1534 	case XE_GUC_ACTION_CONTEXT_RESET_NOTIFICATION:
1535 		ret = xe_guc_exec_queue_reset_handler(guc, payload, adj_len);
1536 		break;
1537 	case XE_GUC_ACTION_ENGINE_FAILURE_NOTIFICATION:
1538 		ret = xe_guc_exec_queue_reset_failure_handler(guc, payload,
1539 							      adj_len);
1540 		break;
1541 	case XE_GUC_ACTION_SCHED_ENGINE_MODE_DONE:
1542 		/* Selftest only at the moment */
1543 		break;
1544 	case XE_GUC_ACTION_STATE_CAPTURE_NOTIFICATION:
1545 		ret = xe_guc_error_capture_handler(guc, payload, adj_len);
1546 		break;
1547 	case XE_GUC_ACTION_NOTIFY_FLUSH_LOG_BUFFER_TO_FILE:
1548 		/* FIXME: Handle this */
1549 		break;
1550 	case XE_GUC_ACTION_NOTIFY_MEMORY_CAT_ERROR:
1551 		ret = xe_guc_exec_queue_memory_cat_error_handler(guc, payload,
1552 								 adj_len);
1553 		break;
1554 	case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC:
1555 		ret = xe_guc_pagefault_handler(guc, payload, adj_len);
1556 		break;
1557 	case XE_GUC_ACTION_TLB_INVALIDATION_DONE:
1558 		ret = xe_guc_tlb_inval_done_handler(guc, payload, adj_len);
1559 		break;
1560 	case XE_GUC_ACTION_GUC2PF_RELAY_FROM_VF:
1561 		ret = xe_guc_relay_process_guc2pf(&guc->relay, hxg, hxg_len);
1562 		break;
1563 	case XE_GUC_ACTION_GUC2VF_RELAY_FROM_PF:
1564 		ret = xe_guc_relay_process_guc2vf(&guc->relay, hxg, hxg_len);
1565 		break;
1566 	case GUC_ACTION_GUC2PF_VF_STATE_NOTIFY:
1567 		ret = xe_gt_sriov_pf_control_process_guc2pf(gt, hxg, hxg_len);
1568 		break;
1569 	case GUC_ACTION_GUC2PF_ADVERSE_EVENT:
1570 		ret = xe_gt_sriov_pf_monitor_process_guc2pf(gt, hxg, hxg_len);
1571 		break;
1572 	case XE_GUC_ACTION_NOTIFY_CRASH_DUMP_POSTED:
1573 	case XE_GUC_ACTION_NOTIFY_EXCEPTION:
1574 		ret = guc_crash_process_msg(ct, action);
1575 		break;
1576 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST)
1577 	case XE_GUC_ACTION_TEST_G2G_RECV:
1578 		ret = xe_guc_g2g_test_notification(guc, payload, adj_len);
1579 		break;
1580 #endif
1581 	default:
1582 		xe_gt_err(gt, "unexpected G2H action 0x%04x\n", action);
1583 	}
1584 
1585 	if (ret) {
1586 		xe_gt_err(gt, "G2H action %#04x failed (%pe) len %u msg %*ph\n",
1587 			  action, ERR_PTR(ret), hxg_len, (int)sizeof(u32) * hxg_len, hxg);
1588 		CT_DEAD(ct, NULL, PROCESS_FAILED);
1589 	}
1590 
1591 	return 0;
1592 }
1593 
g2h_read(struct xe_guc_ct * ct,u32 * msg,bool fast_path)1594 static int g2h_read(struct xe_guc_ct *ct, u32 *msg, bool fast_path)
1595 {
1596 	struct xe_device *xe = ct_to_xe(ct);
1597 	struct xe_gt *gt = ct_to_gt(ct);
1598 	struct guc_ctb *g2h = &ct->ctbs.g2h;
1599 	u32 tail, head, len, desc_status;
1600 	s32 avail;
1601 	u32 action;
1602 	u32 *hxg;
1603 
1604 	xe_gt_assert(gt, xe_guc_ct_initialized(ct));
1605 	lockdep_assert_held(&ct->fast_lock);
1606 
1607 	if (ct->state == XE_GUC_CT_STATE_DISABLED)
1608 		return -ENODEV;
1609 
1610 	if (ct->state == XE_GUC_CT_STATE_STOPPED)
1611 		return -ECANCELED;
1612 
1613 	if (g2h->info.broken)
1614 		return -EPIPE;
1615 
1616 	xe_gt_assert(gt, xe_guc_ct_enabled(ct));
1617 
1618 	desc_status = desc_read(xe, g2h, status);
1619 	if (desc_status) {
1620 		if (desc_status & GUC_CTB_STATUS_DISABLED) {
1621 			/*
1622 			 * Potentially valid if a CLIENT_RESET request resulted in
1623 			 * contexts/engines being reset. But should never happen as
1624 			 * no contexts should be active when CLIENT_RESET is sent.
1625 			 */
1626 			xe_gt_err(gt, "CT read: unexpected G2H after GuC has stopped!\n");
1627 			desc_status &= ~GUC_CTB_STATUS_DISABLED;
1628 		}
1629 
1630 		if (desc_status) {
1631 			xe_gt_err(gt, "CT read: non-zero status: %u\n", desc_status);
1632 			goto corrupted;
1633 		}
1634 	}
1635 
1636 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
1637 		u32 desc_tail = desc_read(xe, g2h, tail);
1638 		/*
1639 		u32 desc_head = desc_read(xe, g2h, head);
1640 
1641 		 * info.head and desc_head are updated back-to-back at the end of
1642 		 * this function and nowhere else. Hence, they cannot be different
1643 		 * unless two g2h_read calls are running concurrently. Which is not
1644 		 * possible because it is guarded by ct->fast_lock. And yet, some
1645 		 * discrete platforms are regularly hitting this error :(.
1646 		 *
1647 		 * desc_head rolling backwards shouldn't cause any noticeable
1648 		 * problems - just a delay in GuC being allowed to proceed past that
1649 		 * point in the queue. So for now, just disable the error until it
1650 		 * can be root caused.
1651 		 *
1652 		if (g2h->info.head != desc_head) {
1653 			desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_MISMATCH);
1654 			xe_gt_err(gt, "CT read: head was modified %u != %u\n",
1655 				  desc_head, g2h->info.head);
1656 			goto corrupted;
1657 		}
1658 		 */
1659 
1660 		if (g2h->info.head > g2h->info.size) {
1661 			desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_OVERFLOW);
1662 			xe_gt_err(gt, "CT read: head out of range: %u vs %u\n",
1663 				  g2h->info.head, g2h->info.size);
1664 			goto corrupted;
1665 		}
1666 
1667 		if (desc_tail >= g2h->info.size) {
1668 			desc_write(xe, g2h, status, desc_status | GUC_CTB_STATUS_OVERFLOW);
1669 			xe_gt_err(gt, "CT read: invalid tail offset %u >= %u)\n",
1670 				  desc_tail, g2h->info.size);
1671 			goto corrupted;
1672 		}
1673 	}
1674 
1675 	/* Calculate DW available to read */
1676 	tail = desc_read(xe, g2h, tail);
1677 	avail = tail - g2h->info.head;
1678 	if (unlikely(avail == 0))
1679 		return 0;
1680 
1681 	if (avail < 0)
1682 		avail += g2h->info.size;
1683 
1684 	/* Read header */
1685 	xe_map_memcpy_from(xe, msg, &g2h->cmds, sizeof(u32) * g2h->info.head,
1686 			   sizeof(u32));
1687 	len = FIELD_GET(GUC_CTB_MSG_0_NUM_DWORDS, msg[0]) + GUC_CTB_MSG_MIN_LEN;
1688 	if (len > avail) {
1689 		xe_gt_err(gt, "G2H channel broken on read, avail=%d, len=%d, reset required\n",
1690 			  avail, len);
1691 		goto corrupted;
1692 	}
1693 
1694 	head = (g2h->info.head + 1) % g2h->info.size;
1695 	avail = len - 1;
1696 
1697 	/* Read G2H message */
1698 	if (avail + head > g2h->info.size) {
1699 		u32 avail_til_wrap = g2h->info.size - head;
1700 
1701 		xe_map_memcpy_from(xe, msg + 1,
1702 				   &g2h->cmds, sizeof(u32) * head,
1703 				   avail_til_wrap * sizeof(u32));
1704 		xe_map_memcpy_from(xe, msg + 1 + avail_til_wrap,
1705 				   &g2h->cmds, 0,
1706 				   (avail - avail_til_wrap) * sizeof(u32));
1707 	} else {
1708 		xe_map_memcpy_from(xe, msg + 1,
1709 				   &g2h->cmds, sizeof(u32) * head,
1710 				   avail * sizeof(u32));
1711 	}
1712 
1713 	hxg = msg_to_hxg(msg);
1714 	action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1715 
1716 	if (fast_path) {
1717 		if (FIELD_GET(GUC_HXG_MSG_0_TYPE, hxg[0]) != GUC_HXG_TYPE_EVENT)
1718 			return 0;
1719 
1720 		switch (action) {
1721 		case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC:
1722 		case XE_GUC_ACTION_TLB_INVALIDATION_DONE:
1723 			break;	/* Process these in fast-path */
1724 		default:
1725 			return 0;
1726 		}
1727 	}
1728 
1729 	/* Update local / descriptor header */
1730 	g2h->info.head = (head + avail) % g2h->info.size;
1731 	desc_write(xe, g2h, head, g2h->info.head);
1732 
1733 	trace_xe_guc_ctb_g2h(xe, ct_to_gt(ct)->info.id,
1734 			     action, len, g2h->info.head, tail);
1735 
1736 	return len;
1737 
1738 corrupted:
1739 	CT_DEAD(ct, &ct->ctbs.g2h, G2H_READ);
1740 	return -EPROTO;
1741 }
1742 
g2h_fast_path(struct xe_guc_ct * ct,u32 * msg,u32 len)1743 static void g2h_fast_path(struct xe_guc_ct *ct, u32 *msg, u32 len)
1744 {
1745 	struct xe_gt *gt = ct_to_gt(ct);
1746 	struct xe_guc *guc = ct_to_guc(ct);
1747 	u32 hxg_len = msg_len_to_hxg_len(len);
1748 	u32 *hxg = msg_to_hxg(msg);
1749 	u32 action = FIELD_GET(GUC_HXG_EVENT_MSG_0_ACTION, hxg[0]);
1750 	u32 *payload = hxg + GUC_HXG_MSG_MIN_LEN;
1751 	u32 adj_len = hxg_len - GUC_HXG_MSG_MIN_LEN;
1752 	int ret = 0;
1753 
1754 	switch (action) {
1755 	case XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC:
1756 		ret = xe_guc_pagefault_handler(guc, payload, adj_len);
1757 		break;
1758 	case XE_GUC_ACTION_TLB_INVALIDATION_DONE:
1759 		__g2h_release_space(ct, len);
1760 		ret = xe_guc_tlb_inval_done_handler(guc, payload, adj_len);
1761 		break;
1762 	default:
1763 		xe_gt_warn(gt, "NOT_POSSIBLE");
1764 	}
1765 
1766 	if (ret) {
1767 		xe_gt_err(gt, "G2H action 0x%04x failed (%pe)\n",
1768 			  action, ERR_PTR(ret));
1769 		CT_DEAD(ct, NULL, FAST_G2H);
1770 	}
1771 }
1772 
1773 /**
1774  * xe_guc_ct_fast_path - process critical G2H in the IRQ handler
1775  * @ct: GuC CT object
1776  *
1777  * Anything related to page faults is critical for performance, process these
1778  * critical G2H in the IRQ. This is safe as these handlers either just wake up
1779  * waiters or queue another worker.
1780  */
xe_guc_ct_fast_path(struct xe_guc_ct * ct)1781 void xe_guc_ct_fast_path(struct xe_guc_ct *ct)
1782 {
1783 	struct xe_device *xe = ct_to_xe(ct);
1784 	bool ongoing;
1785 	int len;
1786 
1787 	ongoing = xe_pm_runtime_get_if_active(ct_to_xe(ct));
1788 	if (!ongoing && xe_pm_read_callback_task(ct_to_xe(ct)) == NULL)
1789 		return;
1790 
1791 	spin_lock(&ct->fast_lock);
1792 	do {
1793 		len = g2h_read(ct, ct->fast_msg, true);
1794 		if (len > 0)
1795 			g2h_fast_path(ct, ct->fast_msg, len);
1796 	} while (len > 0);
1797 	spin_unlock(&ct->fast_lock);
1798 
1799 	if (ongoing)
1800 		xe_pm_runtime_put(xe);
1801 }
1802 
1803 /* Returns less than zero on error, 0 on done, 1 on more available */
dequeue_one_g2h(struct xe_guc_ct * ct)1804 static int dequeue_one_g2h(struct xe_guc_ct *ct)
1805 {
1806 	int len;
1807 	int ret;
1808 
1809 	lockdep_assert_held(&ct->lock);
1810 
1811 	spin_lock_irq(&ct->fast_lock);
1812 	len = g2h_read(ct, ct->msg, false);
1813 	spin_unlock_irq(&ct->fast_lock);
1814 	if (len <= 0)
1815 		return len;
1816 
1817 	ret = parse_g2h_msg(ct, ct->msg, len);
1818 	if (unlikely(ret < 0))
1819 		return ret;
1820 
1821 	ret = process_g2h_msg(ct, ct->msg, len);
1822 	if (unlikely(ret < 0))
1823 		return ret;
1824 
1825 	return 1;
1826 }
1827 
receive_g2h(struct xe_guc_ct * ct)1828 static void receive_g2h(struct xe_guc_ct *ct)
1829 {
1830 	bool ongoing;
1831 	int ret;
1832 
1833 	/*
1834 	 * Normal users must always hold mem_access.ref around CT calls. However
1835 	 * during the runtime pm callbacks we rely on CT to talk to the GuC, but
1836 	 * at this stage we can't rely on mem_access.ref and even the
1837 	 * callback_task will be different than current.  For such cases we just
1838 	 * need to ensure we always process the responses from any blocking
1839 	 * ct_send requests or where we otherwise expect some response when
1840 	 * initiated from those callbacks (which will need to wait for the below
1841 	 * dequeue_one_g2h()).  The dequeue_one_g2h() will gracefully fail if
1842 	 * the device has suspended to the point that the CT communication has
1843 	 * been disabled.
1844 	 *
1845 	 * If we are inside the runtime pm callback, we can be the only task
1846 	 * still issuing CT requests (since that requires having the
1847 	 * mem_access.ref).  It seems like it might in theory be possible to
1848 	 * receive unsolicited events from the GuC just as we are
1849 	 * suspending-resuming, but those will currently anyway be lost when
1850 	 * eventually exiting from suspend, hence no need to wake up the device
1851 	 * here. If we ever need something stronger than get_if_ongoing() then
1852 	 * we need to be careful with blocking the pm callbacks from getting CT
1853 	 * responses, if the worker here is blocked on those callbacks
1854 	 * completing, creating a deadlock.
1855 	 */
1856 	ongoing = xe_pm_runtime_get_if_active(ct_to_xe(ct));
1857 	if (!ongoing && xe_pm_read_callback_task(ct_to_xe(ct)) == NULL)
1858 		return;
1859 
1860 	do {
1861 		mutex_lock(&ct->lock);
1862 		ret = dequeue_one_g2h(ct);
1863 		mutex_unlock(&ct->lock);
1864 
1865 		if (unlikely(ret == -EPROTO || ret == -EOPNOTSUPP)) {
1866 			xe_gt_err(ct_to_gt(ct), "CT dequeue failed: %d", ret);
1867 			CT_DEAD(ct, NULL, G2H_RECV);
1868 			kick_reset(ct);
1869 		}
1870 	} while (ret == 1);
1871 
1872 	if (ongoing)
1873 		xe_pm_runtime_put(ct_to_xe(ct));
1874 }
1875 
g2h_worker_func(struct work_struct * w)1876 static void g2h_worker_func(struct work_struct *w)
1877 {
1878 	struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, g2h_worker);
1879 
1880 	receive_g2h(ct);
1881 }
1882 
guc_ct_snapshot_alloc(struct xe_guc_ct * ct,bool atomic,bool want_ctb)1883 static struct xe_guc_ct_snapshot *guc_ct_snapshot_alloc(struct xe_guc_ct *ct, bool atomic,
1884 							bool want_ctb)
1885 {
1886 	struct xe_guc_ct_snapshot *snapshot;
1887 
1888 	snapshot = kzalloc(sizeof(*snapshot), atomic ? GFP_ATOMIC : GFP_KERNEL);
1889 	if (!snapshot)
1890 		return NULL;
1891 
1892 	if (ct->bo && want_ctb) {
1893 		snapshot->ctb_size = xe_bo_size(ct->bo);
1894 		snapshot->ctb = kmalloc(snapshot->ctb_size, atomic ? GFP_ATOMIC : GFP_KERNEL);
1895 	}
1896 
1897 	return snapshot;
1898 }
1899 
guc_ctb_snapshot_capture(struct xe_device * xe,struct guc_ctb * ctb,struct guc_ctb_snapshot * snapshot)1900 static void guc_ctb_snapshot_capture(struct xe_device *xe, struct guc_ctb *ctb,
1901 				     struct guc_ctb_snapshot *snapshot)
1902 {
1903 	xe_map_memcpy_from(xe, &snapshot->desc, &ctb->desc, 0,
1904 			   sizeof(struct guc_ct_buffer_desc));
1905 	memcpy(&snapshot->info, &ctb->info, sizeof(struct guc_ctb_info));
1906 }
1907 
guc_ctb_snapshot_print(struct guc_ctb_snapshot * snapshot,struct drm_printer * p)1908 static void guc_ctb_snapshot_print(struct guc_ctb_snapshot *snapshot,
1909 				   struct drm_printer *p)
1910 {
1911 	drm_printf(p, "\tsize: %d\n", snapshot->info.size);
1912 	drm_printf(p, "\tresv_space: %d\n", snapshot->info.resv_space);
1913 	drm_printf(p, "\thead: %d\n", snapshot->info.head);
1914 	drm_printf(p, "\ttail: %d\n", snapshot->info.tail);
1915 	drm_printf(p, "\tspace: %d\n", snapshot->info.space);
1916 	drm_printf(p, "\tbroken: %d\n", snapshot->info.broken);
1917 	drm_printf(p, "\thead (memory): %d\n", snapshot->desc.head);
1918 	drm_printf(p, "\ttail (memory): %d\n", snapshot->desc.tail);
1919 	drm_printf(p, "\tstatus (memory): 0x%x\n", snapshot->desc.status);
1920 }
1921 
guc_ct_snapshot_capture(struct xe_guc_ct * ct,bool atomic,bool want_ctb)1922 static struct xe_guc_ct_snapshot *guc_ct_snapshot_capture(struct xe_guc_ct *ct, bool atomic,
1923 							  bool want_ctb)
1924 {
1925 	struct xe_device *xe = ct_to_xe(ct);
1926 	struct xe_guc_ct_snapshot *snapshot;
1927 
1928 	snapshot = guc_ct_snapshot_alloc(ct, atomic, want_ctb);
1929 	if (!snapshot) {
1930 		xe_gt_err(ct_to_gt(ct), "Skipping CTB snapshot entirely.\n");
1931 		return NULL;
1932 	}
1933 
1934 	if (xe_guc_ct_enabled(ct) || ct->state == XE_GUC_CT_STATE_STOPPED) {
1935 		snapshot->ct_enabled = true;
1936 		snapshot->g2h_outstanding = READ_ONCE(ct->g2h_outstanding);
1937 		guc_ctb_snapshot_capture(xe, &ct->ctbs.h2g, &snapshot->h2g);
1938 		guc_ctb_snapshot_capture(xe, &ct->ctbs.g2h, &snapshot->g2h);
1939 	}
1940 
1941 	if (ct->bo && snapshot->ctb)
1942 		xe_map_memcpy_from(xe, snapshot->ctb, &ct->bo->vmap, 0, snapshot->ctb_size);
1943 
1944 	return snapshot;
1945 }
1946 
1947 /**
1948  * xe_guc_ct_snapshot_capture - Take a quick snapshot of the CT state.
1949  * @ct: GuC CT object.
1950  *
1951  * This can be printed out in a later stage like during dev_coredump
1952  * analysis. This is safe to be called during atomic context.
1953  *
1954  * Returns: a GuC CT snapshot object that must be freed by the caller
1955  * by using `xe_guc_ct_snapshot_free`.
1956  */
xe_guc_ct_snapshot_capture(struct xe_guc_ct * ct)1957 struct xe_guc_ct_snapshot *xe_guc_ct_snapshot_capture(struct xe_guc_ct *ct)
1958 {
1959 	return guc_ct_snapshot_capture(ct, true, true);
1960 }
1961 
1962 /**
1963  * xe_guc_ct_snapshot_print - Print out a given GuC CT snapshot.
1964  * @snapshot: GuC CT snapshot object.
1965  * @p: drm_printer where it will be printed out.
1966  *
1967  * This function prints out a given GuC CT snapshot object.
1968  */
xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot * snapshot,struct drm_printer * p)1969 void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot,
1970 			      struct drm_printer *p)
1971 {
1972 	if (!snapshot)
1973 		return;
1974 
1975 	if (snapshot->ct_enabled) {
1976 		drm_puts(p, "H2G CTB (all sizes in DW):\n");
1977 		guc_ctb_snapshot_print(&snapshot->h2g, p);
1978 
1979 		drm_puts(p, "G2H CTB (all sizes in DW):\n");
1980 		guc_ctb_snapshot_print(&snapshot->g2h, p);
1981 		drm_printf(p, "\tg2h outstanding: %d\n",
1982 			   snapshot->g2h_outstanding);
1983 
1984 		if (snapshot->ctb) {
1985 			drm_printf(p, "[CTB].length: 0x%zx\n", snapshot->ctb_size);
1986 			xe_print_blob_ascii85(p, "[CTB].data", '\n',
1987 					      snapshot->ctb, 0, snapshot->ctb_size);
1988 		}
1989 	} else {
1990 		drm_puts(p, "CT disabled\n");
1991 	}
1992 }
1993 
1994 /**
1995  * xe_guc_ct_snapshot_free - Free all allocated objects for a given snapshot.
1996  * @snapshot: GuC CT snapshot object.
1997  *
1998  * This function free all the memory that needed to be allocated at capture
1999  * time.
2000  */
xe_guc_ct_snapshot_free(struct xe_guc_ct_snapshot * snapshot)2001 void xe_guc_ct_snapshot_free(struct xe_guc_ct_snapshot *snapshot)
2002 {
2003 	if (!snapshot)
2004 		return;
2005 
2006 	kfree(snapshot->ctb);
2007 	kfree(snapshot);
2008 }
2009 
2010 /**
2011  * xe_guc_ct_print - GuC CT Print.
2012  * @ct: GuC CT.
2013  * @p: drm_printer where it will be printed out.
2014  * @want_ctb: Should the full CTB content be dumped (vs just the headers)
2015  *
2016  * This function will quickly capture a snapshot of the CT state
2017  * and immediately print it out.
2018  */
xe_guc_ct_print(struct xe_guc_ct * ct,struct drm_printer * p,bool want_ctb)2019 void xe_guc_ct_print(struct xe_guc_ct *ct, struct drm_printer *p, bool want_ctb)
2020 {
2021 	struct xe_guc_ct_snapshot *snapshot;
2022 
2023 	snapshot = guc_ct_snapshot_capture(ct, false, want_ctb);
2024 	xe_guc_ct_snapshot_print(snapshot, p);
2025 	xe_guc_ct_snapshot_free(snapshot);
2026 }
2027 
2028 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
2029 
2030 #ifdef CONFIG_FUNCTION_ERROR_INJECTION
2031 /*
2032  * This is a helper function which assists the driver in identifying if a fault
2033  * injection test is currently active, allowing it to reduce unnecessary debug
2034  * output. Typically, the function returns zero, but the fault injection
2035  * framework can alter this to return an error. Since faults are injected
2036  * through this function, it's important to ensure the compiler doesn't optimize
2037  * it into an inline function. To avoid such optimization, the 'noinline'
2038  * attribute is applied. Compiler optimizes the static function defined in the
2039  * header file as an inline function.
2040  */
xe_is_injection_active(void)2041 noinline int xe_is_injection_active(void) { return 0; }
2042 ALLOW_ERROR_INJECTION(xe_is_injection_active, ERRNO);
2043 #else
xe_is_injection_active(void)2044 int xe_is_injection_active(void) { return 0; }
2045 #endif
2046 
ct_dead_capture(struct xe_guc_ct * ct,struct guc_ctb * ctb,u32 reason_code)2047 static void ct_dead_capture(struct xe_guc_ct *ct, struct guc_ctb *ctb, u32 reason_code)
2048 {
2049 	struct xe_guc_log_snapshot *snapshot_log;
2050 	struct xe_guc_ct_snapshot *snapshot_ct;
2051 	struct xe_guc *guc = ct_to_guc(ct);
2052 	unsigned long flags;
2053 	bool have_capture;
2054 
2055 	if (ctb)
2056 		ctb->info.broken = true;
2057 	/*
2058 	 * Huge dump is getting generated when injecting error for guc CT/MMIO
2059 	 * functions. So, let us suppress the dump when fault is injected.
2060 	 */
2061 	if (xe_is_injection_active())
2062 		return;
2063 
2064 	/* Ignore further errors after the first dump until a reset */
2065 	if (ct->dead.reported)
2066 		return;
2067 
2068 	spin_lock_irqsave(&ct->dead.lock, flags);
2069 
2070 	/* And only capture one dump at a time */
2071 	have_capture = ct->dead.reason & (1 << CT_DEAD_STATE_CAPTURE);
2072 	ct->dead.reason |= (1 << reason_code) |
2073 			   (1 << CT_DEAD_STATE_CAPTURE);
2074 
2075 	spin_unlock_irqrestore(&ct->dead.lock, flags);
2076 
2077 	if (have_capture)
2078 		return;
2079 
2080 	snapshot_log = xe_guc_log_snapshot_capture(&guc->log, true);
2081 	snapshot_ct = xe_guc_ct_snapshot_capture((ct));
2082 
2083 	spin_lock_irqsave(&ct->dead.lock, flags);
2084 
2085 	if (ct->dead.snapshot_log || ct->dead.snapshot_ct) {
2086 		xe_gt_err(ct_to_gt(ct), "Got unexpected dead CT capture!\n");
2087 		xe_guc_log_snapshot_free(snapshot_log);
2088 		xe_guc_ct_snapshot_free(snapshot_ct);
2089 	} else {
2090 		ct->dead.snapshot_log = snapshot_log;
2091 		ct->dead.snapshot_ct = snapshot_ct;
2092 	}
2093 
2094 	spin_unlock_irqrestore(&ct->dead.lock, flags);
2095 
2096 	queue_work(system_unbound_wq, &(ct)->dead.worker);
2097 }
2098 
ct_dead_print(struct xe_dead_ct * dead)2099 static void ct_dead_print(struct xe_dead_ct *dead)
2100 {
2101 	struct xe_guc_ct *ct = container_of(dead, struct xe_guc_ct, dead);
2102 	struct xe_device *xe = ct_to_xe(ct);
2103 	struct xe_gt *gt = ct_to_gt(ct);
2104 	static int g_count;
2105 	struct drm_printer ip = xe_gt_info_printer(gt);
2106 	struct drm_printer lp = drm_line_printer(&ip, "Capture", ++g_count);
2107 
2108 	if (!dead->reason) {
2109 		xe_gt_err(gt, "CTB is dead for no reason!?\n");
2110 		return;
2111 	}
2112 
2113 	/* Can't generate a genuine core dump at this point, so just do the good bits */
2114 	drm_puts(&lp, "**** Xe Device Coredump ****\n");
2115 	drm_printf(&lp, "Reason: CTB is dead - 0x%X\n", dead->reason);
2116 	xe_device_snapshot_print(xe, &lp);
2117 
2118 	drm_printf(&lp, "**** GT #%d ****\n", gt->info.id);
2119 	drm_printf(&lp, "\tTile: %d\n", gt->tile->id);
2120 
2121 	drm_puts(&lp, "**** GuC Log ****\n");
2122 	xe_guc_log_snapshot_print(dead->snapshot_log, &lp);
2123 
2124 	drm_puts(&lp, "**** GuC CT ****\n");
2125 	xe_guc_ct_snapshot_print(dead->snapshot_ct, &lp);
2126 
2127 	drm_puts(&lp, "Done.\n");
2128 }
2129 
ct_dead_worker_func(struct work_struct * w)2130 static void ct_dead_worker_func(struct work_struct *w)
2131 {
2132 	struct xe_guc_ct *ct = container_of(w, struct xe_guc_ct, dead.worker);
2133 
2134 	if (!ct->dead.reported) {
2135 		ct->dead.reported = true;
2136 		ct_dead_print(&ct->dead);
2137 	}
2138 
2139 	spin_lock_irq(&ct->dead.lock);
2140 
2141 	xe_guc_log_snapshot_free(ct->dead.snapshot_log);
2142 	ct->dead.snapshot_log = NULL;
2143 	xe_guc_ct_snapshot_free(ct->dead.snapshot_ct);
2144 	ct->dead.snapshot_ct = NULL;
2145 
2146 	if (ct->dead.reason & (1 << CT_DEAD_STATE_REARM)) {
2147 		/* A reset has occurred so re-arm the error reporting */
2148 		ct->dead.reason = 0;
2149 		ct->dead.reported = false;
2150 	}
2151 
2152 	spin_unlock_irq(&ct->dead.lock);
2153 }
2154 #endif
2155