brintos

brintos / linux-shallow public Read only

0
0
Text · 164.3 KiB · ed97984 Raw
5940 lines · c
1// SPDX-License-Identifier: MIT2/*3 * Copyright © 2014 Intel Corporation4 */5 6#include <linux/circ_buf.h>7 8#include "gem/i915_gem_context.h"9#include "gem/i915_gem_lmem.h"10#include "gt/gen8_engine_cs.h"11#include "gt/intel_breadcrumbs.h"12#include "gt/intel_context.h"13#include "gt/intel_engine_heartbeat.h"14#include "gt/intel_engine_pm.h"15#include "gt/intel_engine_regs.h"16#include "gt/intel_gpu_commands.h"17#include "gt/intel_gt.h"18#include "gt/intel_gt_clock_utils.h"19#include "gt/intel_gt_irq.h"20#include "gt/intel_gt_pm.h"21#include "gt/intel_gt_regs.h"22#include "gt/intel_gt_requests.h"23#include "gt/intel_lrc.h"24#include "gt/intel_lrc_reg.h"25#include "gt/intel_mocs.h"26#include "gt/intel_ring.h"27 28#include "intel_guc_ads.h"29#include "intel_guc_capture.h"30#include "intel_guc_print.h"31#include "intel_guc_submission.h"32 33#include "i915_drv.h"34#include "i915_reg.h"35#include "i915_irq.h"36#include "i915_trace.h"37 38/**39 * DOC: GuC-based command submission40 *41 * The Scratch registers:42 * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes43 * a value to the action register (SOFT_SCRATCH_0) along with any data. It then44 * triggers an interrupt on the GuC via another register write (0xC4C8).45 * Firmware writes a success/fail code back to the action register after46 * processes the request. The kernel driver polls waiting for this update and47 * then proceeds.48 *49 * Command Transport buffers (CTBs):50 * Covered in detail in other sections but CTBs (Host to GuC - H2G, GuC to Host51 * - G2H) are a message interface between the i915 and GuC.52 *53 * Context registration:54 * Before a context can be submitted it must be registered with the GuC via a55 * H2G. A unique guc_id is associated with each context. The context is either56 * registered at request creation time (normal operation) or at submission time57 * (abnormal operation, e.g. after a reset).58 *59 * Context submission:60 * The i915 updates the LRC tail value in memory. The i915 must enable the61 * scheduling of the context within the GuC for the GuC to actually consider it.62 * Therefore, the first time a disabled context is submitted we use a schedule63 * enable H2G, while follow up submissions are done via the context submit H2G,64 * which informs the GuC that a previously enabled context has new work65 * available.66 *67 * Context unpin:68 * To unpin a context a H2G is used to disable scheduling. When the69 * corresponding G2H returns indicating the scheduling disable operation has70 * completed it is safe to unpin the context. While a disable is in flight it71 * isn't safe to resubmit the context so a fence is used to stall all future72 * requests of that context until the G2H is returned. Because this interaction73 * with the GuC takes a non-zero amount of time we delay the disabling of74 * scheduling after the pin count goes to zero by a configurable period of time75 * (see SCHED_DISABLE_DELAY_MS). The thought is this gives the user a window of76 * time to resubmit something on the context before doing this costly operation.77 * This delay is only done if the context isn't closed and the guc_id usage is78 * less than a threshold (see NUM_SCHED_DISABLE_GUC_IDS_THRESHOLD).79 *80 * Context deregistration:81 * Before a context can be destroyed or if we steal its guc_id we must82 * deregister the context with the GuC via H2G. If stealing the guc_id it isn't83 * safe to submit anything to this guc_id until the deregister completes so a84 * fence is used to stall all requests associated with this guc_id until the85 * corresponding G2H returns indicating the guc_id has been deregistered.86 *87 * submission_state.guc_ids:88 * Unique number associated with private GuC context data passed in during89 * context registration / submission / deregistration. 64k available. Simple ida90 * is used for allocation.91 *92 * Stealing guc_ids:93 * If no guc_ids are available they can be stolen from another context at94 * request creation time if that context is unpinned. If a guc_id can't be found95 * we punt this problem to the user as we believe this is near impossible to hit96 * during normal use cases.97 *98 * Locking:99 * In the GuC submission code we have 3 basic spin locks which protect100 * everything. Details about each below.101 *102 * sched_engine->lock103 * This is the submission lock for all contexts that share an i915 schedule104 * engine (sched_engine), thus only one of the contexts which share a105 * sched_engine can be submitting at a time. Currently only one sched_engine is106 * used for all of GuC submission but that could change in the future.107 *108 * guc->submission_state.lock109 * Global lock for GuC submission state. Protects guc_ids and destroyed contexts110 * list.111 *112 * ce->guc_state.lock113 * Protects everything under ce->guc_state. Ensures that a context is in the114 * correct state before issuing a H2G. e.g. We don't issue a schedule disable115 * on a disabled context (bad idea), we don't issue a schedule enable when a116 * schedule disable is in flight, etc... Also protects list of inflight requests117 * on the context and the priority management state. Lock is individual to each118 * context.119 *120 * Lock ordering rules:121 * sched_engine->lock -> ce->guc_state.lock122 * guc->submission_state.lock -> ce->guc_state.lock123 *124 * Reset races:125 * When a full GT reset is triggered it is assumed that some G2H responses to126 * H2Gs can be lost as the GuC is also reset. Losing these G2H can prove to be127 * fatal as we do certain operations upon receiving a G2H (e.g. destroy128 * contexts, release guc_ids, etc...). When this occurs we can scrub the129 * context state and cleanup appropriately, however this is quite racey.130 * To avoid races, the reset code must disable submission before scrubbing for131 * the missing G2H, while the submission code must check for submission being132 * disabled and skip sending H2Gs and updating context states when it is. Both133 * sides must also make sure to hold the relevant locks.134 */135 136/* GuC Virtual Engine */137struct guc_virtual_engine {138	struct intel_engine_cs base;139	struct intel_context context;140};141 142static struct intel_context *143guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count,144		   unsigned long flags);145 146static struct intel_context *147guc_create_parallel(struct intel_engine_cs **engines,148		    unsigned int num_siblings,149		    unsigned int width);150 151#define GUC_REQUEST_SIZE 64 /* bytes */152 153/*154 * We reserve 1/16 of the guc_ids for multi-lrc as these need to be contiguous155 * per the GuC submission interface. A different allocation algorithm is used156 * (bitmap vs. ida) between multi-lrc and single-lrc hence the reason to157 * partition the guc_id space. We believe the number of multi-lrc contexts in158 * use should be low and 1/16 should be sufficient. Minimum of 32 guc_ids for159 * multi-lrc.160 */161#define NUMBER_MULTI_LRC_GUC_ID(guc)	\162	((guc)->submission_state.num_guc_ids / 16)163 164/*165 * Below is a set of functions which control the GuC scheduling state which166 * require a lock.167 */168#define SCHED_STATE_WAIT_FOR_DEREGISTER_TO_REGISTER	BIT(0)169#define SCHED_STATE_DESTROYED				BIT(1)170#define SCHED_STATE_PENDING_DISABLE			BIT(2)171#define SCHED_STATE_BANNED				BIT(3)172#define SCHED_STATE_ENABLED				BIT(4)173#define SCHED_STATE_PENDING_ENABLE			BIT(5)174#define SCHED_STATE_REGISTERED				BIT(6)175#define SCHED_STATE_POLICY_REQUIRED			BIT(7)176#define SCHED_STATE_CLOSED				BIT(8)177#define SCHED_STATE_BLOCKED_SHIFT			9178#define SCHED_STATE_BLOCKED		BIT(SCHED_STATE_BLOCKED_SHIFT)179#define SCHED_STATE_BLOCKED_MASK	(0xfff << SCHED_STATE_BLOCKED_SHIFT)180 181static inline void init_sched_state(struct intel_context *ce)182{183	lockdep_assert_held(&ce->guc_state.lock);184	ce->guc_state.sched_state &= SCHED_STATE_BLOCKED_MASK;185}186 187/*188 * Kernel contexts can have SCHED_STATE_REGISTERED after suspend.189 * A context close can race with the submission path, so SCHED_STATE_CLOSED190 * can be set immediately before we try to register.191 */192#define SCHED_STATE_VALID_INIT \193	(SCHED_STATE_BLOCKED_MASK | \194	 SCHED_STATE_CLOSED | \195	 SCHED_STATE_REGISTERED)196 197__maybe_unused198static bool sched_state_is_init(struct intel_context *ce)199{200	return !(ce->guc_state.sched_state & ~SCHED_STATE_VALID_INIT);201}202 203static inline bool204context_wait_for_deregister_to_register(struct intel_context *ce)205{206	return ce->guc_state.sched_state &207		SCHED_STATE_WAIT_FOR_DEREGISTER_TO_REGISTER;208}209 210static inline void211set_context_wait_for_deregister_to_register(struct intel_context *ce)212{213	lockdep_assert_held(&ce->guc_state.lock);214	ce->guc_state.sched_state |=215		SCHED_STATE_WAIT_FOR_DEREGISTER_TO_REGISTER;216}217 218static inline void219clr_context_wait_for_deregister_to_register(struct intel_context *ce)220{221	lockdep_assert_held(&ce->guc_state.lock);222	ce->guc_state.sched_state &=223		~SCHED_STATE_WAIT_FOR_DEREGISTER_TO_REGISTER;224}225 226static inline bool227context_destroyed(struct intel_context *ce)228{229	return ce->guc_state.sched_state & SCHED_STATE_DESTROYED;230}231 232static inline void233set_context_destroyed(struct intel_context *ce)234{235	lockdep_assert_held(&ce->guc_state.lock);236	ce->guc_state.sched_state |= SCHED_STATE_DESTROYED;237}238 239static inline void240clr_context_destroyed(struct intel_context *ce)241{242	lockdep_assert_held(&ce->guc_state.lock);243	ce->guc_state.sched_state &= ~SCHED_STATE_DESTROYED;244}245 246static inline bool context_pending_disable(struct intel_context *ce)247{248	return ce->guc_state.sched_state & SCHED_STATE_PENDING_DISABLE;249}250 251static inline void set_context_pending_disable(struct intel_context *ce)252{253	lockdep_assert_held(&ce->guc_state.lock);254	ce->guc_state.sched_state |= SCHED_STATE_PENDING_DISABLE;255}256 257static inline void clr_context_pending_disable(struct intel_context *ce)258{259	lockdep_assert_held(&ce->guc_state.lock);260	ce->guc_state.sched_state &= ~SCHED_STATE_PENDING_DISABLE;261}262 263static inline bool context_banned(struct intel_context *ce)264{265	return ce->guc_state.sched_state & SCHED_STATE_BANNED;266}267 268static inline void set_context_banned(struct intel_context *ce)269{270	lockdep_assert_held(&ce->guc_state.lock);271	ce->guc_state.sched_state |= SCHED_STATE_BANNED;272}273 274static inline void clr_context_banned(struct intel_context *ce)275{276	lockdep_assert_held(&ce->guc_state.lock);277	ce->guc_state.sched_state &= ~SCHED_STATE_BANNED;278}279 280static inline bool context_enabled(struct intel_context *ce)281{282	return ce->guc_state.sched_state & SCHED_STATE_ENABLED;283}284 285static inline void set_context_enabled(struct intel_context *ce)286{287	lockdep_assert_held(&ce->guc_state.lock);288	ce->guc_state.sched_state |= SCHED_STATE_ENABLED;289}290 291static inline void clr_context_enabled(struct intel_context *ce)292{293	lockdep_assert_held(&ce->guc_state.lock);294	ce->guc_state.sched_state &= ~SCHED_STATE_ENABLED;295}296 297static inline bool context_pending_enable(struct intel_context *ce)298{299	return ce->guc_state.sched_state & SCHED_STATE_PENDING_ENABLE;300}301 302static inline void set_context_pending_enable(struct intel_context *ce)303{304	lockdep_assert_held(&ce->guc_state.lock);305	ce->guc_state.sched_state |= SCHED_STATE_PENDING_ENABLE;306}307 308static inline void clr_context_pending_enable(struct intel_context *ce)309{310	lockdep_assert_held(&ce->guc_state.lock);311	ce->guc_state.sched_state &= ~SCHED_STATE_PENDING_ENABLE;312}313 314static inline bool context_registered(struct intel_context *ce)315{316	return ce->guc_state.sched_state & SCHED_STATE_REGISTERED;317}318 319static inline void set_context_registered(struct intel_context *ce)320{321	lockdep_assert_held(&ce->guc_state.lock);322	ce->guc_state.sched_state |= SCHED_STATE_REGISTERED;323}324 325static inline void clr_context_registered(struct intel_context *ce)326{327	lockdep_assert_held(&ce->guc_state.lock);328	ce->guc_state.sched_state &= ~SCHED_STATE_REGISTERED;329}330 331static inline bool context_policy_required(struct intel_context *ce)332{333	return ce->guc_state.sched_state & SCHED_STATE_POLICY_REQUIRED;334}335 336static inline void set_context_policy_required(struct intel_context *ce)337{338	lockdep_assert_held(&ce->guc_state.lock);339	ce->guc_state.sched_state |= SCHED_STATE_POLICY_REQUIRED;340}341 342static inline void clr_context_policy_required(struct intel_context *ce)343{344	lockdep_assert_held(&ce->guc_state.lock);345	ce->guc_state.sched_state &= ~SCHED_STATE_POLICY_REQUIRED;346}347 348static inline bool context_close_done(struct intel_context *ce)349{350	return ce->guc_state.sched_state & SCHED_STATE_CLOSED;351}352 353static inline void set_context_close_done(struct intel_context *ce)354{355	lockdep_assert_held(&ce->guc_state.lock);356	ce->guc_state.sched_state |= SCHED_STATE_CLOSED;357}358 359static inline u32 context_blocked(struct intel_context *ce)360{361	return (ce->guc_state.sched_state & SCHED_STATE_BLOCKED_MASK) >>362		SCHED_STATE_BLOCKED_SHIFT;363}364 365static inline void incr_context_blocked(struct intel_context *ce)366{367	lockdep_assert_held(&ce->guc_state.lock);368 369	ce->guc_state.sched_state += SCHED_STATE_BLOCKED;370 371	GEM_BUG_ON(!context_blocked(ce));	/* Overflow check */372}373 374static inline void decr_context_blocked(struct intel_context *ce)375{376	lockdep_assert_held(&ce->guc_state.lock);377 378	GEM_BUG_ON(!context_blocked(ce));	/* Underflow check */379 380	ce->guc_state.sched_state -= SCHED_STATE_BLOCKED;381}382 383static struct intel_context *384request_to_scheduling_context(struct i915_request *rq)385{386	return intel_context_to_parent(rq->context);387}388 389static inline bool context_guc_id_invalid(struct intel_context *ce)390{391	return ce->guc_id.id == GUC_INVALID_CONTEXT_ID;392}393 394static inline void set_context_guc_id_invalid(struct intel_context *ce)395{396	ce->guc_id.id = GUC_INVALID_CONTEXT_ID;397}398 399static inline struct intel_guc *ce_to_guc(struct intel_context *ce)400{401	return gt_to_guc(ce->engine->gt);402}403 404static inline struct i915_priolist *to_priolist(struct rb_node *rb)405{406	return rb_entry(rb, struct i915_priolist, node);407}408 409/*410 * When using multi-lrc submission a scratch memory area is reserved in the411 * parent's context state for the process descriptor, work queue, and handshake412 * between the parent + children contexts to insert safe preemption points413 * between each of the BBs. Currently the scratch area is sized to a page.414 *415 * The layout of this scratch area is below:416 * 0						guc_process_desc417 * + sizeof(struct guc_process_desc)		child go418 * + CACHELINE_BYTES				child join[0]419 * ...420 * + CACHELINE_BYTES				child join[n - 1]421 * ...						unused422 * PARENT_SCRATCH_SIZE / 2			work queue start423 * ...						work queue424 * PARENT_SCRATCH_SIZE - 1			work queue end425 */426#define WQ_SIZE			(PARENT_SCRATCH_SIZE / 2)427#define WQ_OFFSET		(PARENT_SCRATCH_SIZE - WQ_SIZE)428 429struct sync_semaphore {430	u32 semaphore;431	u8 unused[CACHELINE_BYTES - sizeof(u32)];432};433 434struct parent_scratch {435	union guc_descs {436		struct guc_sched_wq_desc wq_desc;437		struct guc_process_desc_v69 pdesc;438	} descs;439 440	struct sync_semaphore go;441	struct sync_semaphore join[MAX_ENGINE_INSTANCE + 1];442 443	u8 unused[WQ_OFFSET - sizeof(union guc_descs) -444		sizeof(struct sync_semaphore) * (MAX_ENGINE_INSTANCE + 2)];445 446	u32 wq[WQ_SIZE / sizeof(u32)];447};448 449static u32 __get_parent_scratch_offset(struct intel_context *ce)450{451	GEM_BUG_ON(!ce->parallel.guc.parent_page);452 453	return ce->parallel.guc.parent_page * PAGE_SIZE;454}455 456static u32 __get_wq_offset(struct intel_context *ce)457{458	BUILD_BUG_ON(offsetof(struct parent_scratch, wq) != WQ_OFFSET);459 460	return __get_parent_scratch_offset(ce) + WQ_OFFSET;461}462 463static struct parent_scratch *464__get_parent_scratch(struct intel_context *ce)465{466	BUILD_BUG_ON(sizeof(struct parent_scratch) != PARENT_SCRATCH_SIZE);467	BUILD_BUG_ON(sizeof(struct sync_semaphore) != CACHELINE_BYTES);468 469	/*470	 * Need to subtract LRC_STATE_OFFSET here as the471	 * parallel.guc.parent_page is the offset into ce->state while472	 * ce->lrc_reg_reg is ce->state + LRC_STATE_OFFSET.473	 */474	return (struct parent_scratch *)475		(ce->lrc_reg_state +476		 ((__get_parent_scratch_offset(ce) -477		   LRC_STATE_OFFSET) / sizeof(u32)));478}479 480static struct guc_process_desc_v69 *481__get_process_desc_v69(struct intel_context *ce)482{483	struct parent_scratch *ps = __get_parent_scratch(ce);484 485	return &ps->descs.pdesc;486}487 488static struct guc_sched_wq_desc *489__get_wq_desc_v70(struct intel_context *ce)490{491	struct parent_scratch *ps = __get_parent_scratch(ce);492 493	return &ps->descs.wq_desc;494}495 496static u32 *get_wq_pointer(struct intel_context *ce, u32 wqi_size)497{498	/*499	 * Check for space in work queue. Caching a value of head pointer in500	 * intel_context structure in order reduce the number accesses to shared501	 * GPU memory which may be across a PCIe bus.502	 */503#define AVAILABLE_SPACE	\504	CIRC_SPACE(ce->parallel.guc.wqi_tail, ce->parallel.guc.wqi_head, WQ_SIZE)505	if (wqi_size > AVAILABLE_SPACE) {506		ce->parallel.guc.wqi_head = READ_ONCE(*ce->parallel.guc.wq_head);507 508		if (wqi_size > AVAILABLE_SPACE)509			return NULL;510	}511#undef AVAILABLE_SPACE512 513	return &__get_parent_scratch(ce)->wq[ce->parallel.guc.wqi_tail / sizeof(u32)];514}515 516static inline struct intel_context *__get_context(struct intel_guc *guc, u32 id)517{518	struct intel_context *ce = xa_load(&guc->context_lookup, id);519 520	GEM_BUG_ON(id >= GUC_MAX_CONTEXT_ID);521 522	return ce;523}524 525static struct guc_lrc_desc_v69 *__get_lrc_desc_v69(struct intel_guc *guc, u32 index)526{527	struct guc_lrc_desc_v69 *base = guc->lrc_desc_pool_vaddr_v69;528 529	if (!base)530		return NULL;531 532	GEM_BUG_ON(index >= GUC_MAX_CONTEXT_ID);533 534	return &base[index];535}536 537static int guc_lrc_desc_pool_create_v69(struct intel_guc *guc)538{539	u32 size;540	int ret;541 542	size = PAGE_ALIGN(sizeof(struct guc_lrc_desc_v69) *543			  GUC_MAX_CONTEXT_ID);544	ret = intel_guc_allocate_and_map_vma(guc, size, &guc->lrc_desc_pool_v69,545					     (void **)&guc->lrc_desc_pool_vaddr_v69);546	if (ret)547		return ret;548 549	return 0;550}551 552static void guc_lrc_desc_pool_destroy_v69(struct intel_guc *guc)553{554	if (!guc->lrc_desc_pool_vaddr_v69)555		return;556 557	guc->lrc_desc_pool_vaddr_v69 = NULL;558	i915_vma_unpin_and_release(&guc->lrc_desc_pool_v69, I915_VMA_RELEASE_MAP);559}560 561static inline bool guc_submission_initialized(struct intel_guc *guc)562{563	return guc->submission_initialized;564}565 566static inline void _reset_lrc_desc_v69(struct intel_guc *guc, u32 id)567{568	struct guc_lrc_desc_v69 *desc = __get_lrc_desc_v69(guc, id);569 570	if (desc)571		memset(desc, 0, sizeof(*desc));572}573 574static inline bool ctx_id_mapped(struct intel_guc *guc, u32 id)575{576	return __get_context(guc, id);577}578 579static inline void set_ctx_id_mapping(struct intel_guc *guc, u32 id,580				      struct intel_context *ce)581{582	unsigned long flags;583 584	/*585	 * xarray API doesn't have xa_save_irqsave wrapper, so calling the586	 * lower level functions directly.587	 */588	xa_lock_irqsave(&guc->context_lookup, flags);589	__xa_store(&guc->context_lookup, id, ce, GFP_ATOMIC);590	xa_unlock_irqrestore(&guc->context_lookup, flags);591}592 593static inline void clr_ctx_id_mapping(struct intel_guc *guc, u32 id)594{595	unsigned long flags;596 597	if (unlikely(!guc_submission_initialized(guc)))598		return;599 600	_reset_lrc_desc_v69(guc, id);601 602	/*603	 * xarray API doesn't have xa_erase_irqsave wrapper, so calling604	 * the lower level functions directly.605	 */606	xa_lock_irqsave(&guc->context_lookup, flags);607	__xa_erase(&guc->context_lookup, id);608	xa_unlock_irqrestore(&guc->context_lookup, flags);609}610 611static void decr_outstanding_submission_g2h(struct intel_guc *guc)612{613	if (atomic_dec_and_test(&guc->outstanding_submission_g2h))614		wake_up_all(&guc->ct.wq);615}616 617static int guc_submission_send_busy_loop(struct intel_guc *guc,618					 const u32 *action,619					 u32 len,620					 u32 g2h_len_dw,621					 bool loop)622{623	int ret;624 625	/*626	 * We always loop when a send requires a reply (i.e. g2h_len_dw > 0),627	 * so we don't handle the case where we don't get a reply because we628	 * aborted the send due to the channel being busy.629	 */630	GEM_BUG_ON(g2h_len_dw && !loop);631 632	if (g2h_len_dw)633		atomic_inc(&guc->outstanding_submission_g2h);634 635	ret = intel_guc_send_busy_loop(guc, action, len, g2h_len_dw, loop);636	if (ret)637		atomic_dec(&guc->outstanding_submission_g2h);638 639	return ret;640}641 642int intel_guc_wait_for_pending_msg(struct intel_guc *guc,643				   atomic_t *wait_var,644				   bool interruptible,645				   long timeout)646{647	const int state = interruptible ?648		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;649	DEFINE_WAIT(wait);650 651	might_sleep();652	GEM_BUG_ON(timeout < 0);653 654	if (!atomic_read(wait_var))655		return 0;656 657	if (!timeout)658		return -ETIME;659 660	for (;;) {661		prepare_to_wait(&guc->ct.wq, &wait, state);662 663		if (!atomic_read(wait_var))664			break;665 666		if (signal_pending_state(state, current)) {667			timeout = -EINTR;668			break;669		}670 671		if (!timeout) {672			timeout = -ETIME;673			break;674		}675 676		timeout = io_schedule_timeout(timeout);677	}678	finish_wait(&guc->ct.wq, &wait);679 680	return (timeout < 0) ? timeout : 0;681}682 683int intel_guc_wait_for_idle(struct intel_guc *guc, long timeout)684{685	if (!intel_uc_uses_guc_submission(&guc_to_gt(guc)->uc))686		return 0;687 688	return intel_guc_wait_for_pending_msg(guc,689					      &guc->outstanding_submission_g2h,690					      true, timeout);691}692 693static int guc_context_policy_init_v70(struct intel_context *ce, bool loop);694static int try_context_registration(struct intel_context *ce, bool loop);695 696static int __guc_add_request(struct intel_guc *guc, struct i915_request *rq)697{698	int err = 0;699	struct intel_context *ce = request_to_scheduling_context(rq);700	u32 action[3];701	int len = 0;702	u32 g2h_len_dw = 0;703	bool enabled;704 705	lockdep_assert_held(&rq->engine->sched_engine->lock);706 707	/*708	 * Corner case where requests were sitting in the priority list or a709	 * request resubmitted after the context was banned.710	 */711	if (unlikely(!intel_context_is_schedulable(ce))) {712		i915_request_put(i915_request_mark_eio(rq));713		intel_engine_signal_breadcrumbs(ce->engine);714		return 0;715	}716 717	GEM_BUG_ON(!atomic_read(&ce->guc_id.ref));718	GEM_BUG_ON(context_guc_id_invalid(ce));719 720	if (context_policy_required(ce)) {721		err = guc_context_policy_init_v70(ce, false);722		if (err)723			return err;724	}725 726	spin_lock(&ce->guc_state.lock);727 728	/*729	 * The request / context will be run on the hardware when scheduling730	 * gets enabled in the unblock. For multi-lrc we still submit the731	 * context to move the LRC tails.732	 */733	if (unlikely(context_blocked(ce) && !intel_context_is_parent(ce)))734		goto out;735 736	enabled = context_enabled(ce) || context_blocked(ce);737 738	if (!enabled) {739		action[len++] = INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET;740		action[len++] = ce->guc_id.id;741		action[len++] = GUC_CONTEXT_ENABLE;742		set_context_pending_enable(ce);743		intel_context_get(ce);744		g2h_len_dw = G2H_LEN_DW_SCHED_CONTEXT_MODE_SET;745	} else {746		action[len++] = INTEL_GUC_ACTION_SCHED_CONTEXT;747		action[len++] = ce->guc_id.id;748	}749 750	err = intel_guc_send_nb(guc, action, len, g2h_len_dw);751	if (!enabled && !err) {752		trace_intel_context_sched_enable(ce);753		atomic_inc(&guc->outstanding_submission_g2h);754		set_context_enabled(ce);755 756		/*757		 * Without multi-lrc KMD does the submission step (moving the758		 * lrc tail) so enabling scheduling is sufficient to submit the759		 * context. This isn't the case in multi-lrc submission as the760		 * GuC needs to move the tails, hence the need for another H2G761		 * to submit a multi-lrc context after enabling scheduling.762		 */763		if (intel_context_is_parent(ce)) {764			action[0] = INTEL_GUC_ACTION_SCHED_CONTEXT;765			err = intel_guc_send_nb(guc, action, len - 1, 0);766		}767	} else if (!enabled) {768		clr_context_pending_enable(ce);769		intel_context_put(ce);770	}771	if (likely(!err))772		trace_i915_request_guc_submit(rq);773 774out:775	spin_unlock(&ce->guc_state.lock);776	return err;777}778 779static int guc_add_request(struct intel_guc *guc, struct i915_request *rq)780{781	int ret = __guc_add_request(guc, rq);782 783	if (unlikely(ret == -EBUSY)) {784		guc->stalled_request = rq;785		guc->submission_stall_reason = STALL_ADD_REQUEST;786	}787 788	return ret;789}790 791static inline void guc_set_lrc_tail(struct i915_request *rq)792{793	rq->context->lrc_reg_state[CTX_RING_TAIL] =794		intel_ring_set_tail(rq->ring, rq->tail);795}796 797static inline int rq_prio(const struct i915_request *rq)798{799	return rq->sched.attr.priority;800}801 802static bool is_multi_lrc_rq(struct i915_request *rq)803{804	return intel_context_is_parallel(rq->context);805}806 807static bool can_merge_rq(struct i915_request *rq,808			 struct i915_request *last)809{810	return request_to_scheduling_context(rq) ==811		request_to_scheduling_context(last);812}813 814static u32 wq_space_until_wrap(struct intel_context *ce)815{816	return (WQ_SIZE - ce->parallel.guc.wqi_tail);817}818 819static void write_wqi(struct intel_context *ce, u32 wqi_size)820{821	BUILD_BUG_ON(!is_power_of_2(WQ_SIZE));822 823	/*824	 * Ensure WQI are visible before updating tail825	 */826	intel_guc_write_barrier(ce_to_guc(ce));827 828	ce->parallel.guc.wqi_tail = (ce->parallel.guc.wqi_tail + wqi_size) &829		(WQ_SIZE - 1);830	WRITE_ONCE(*ce->parallel.guc.wq_tail, ce->parallel.guc.wqi_tail);831}832 833static int guc_wq_noop_append(struct intel_context *ce)834{835	u32 *wqi = get_wq_pointer(ce, wq_space_until_wrap(ce));836	u32 len_dw = wq_space_until_wrap(ce) / sizeof(u32) - 1;837 838	if (!wqi)839		return -EBUSY;840 841	GEM_BUG_ON(!FIELD_FIT(WQ_LEN_MASK, len_dw));842 843	*wqi = FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |844		FIELD_PREP(WQ_LEN_MASK, len_dw);845	ce->parallel.guc.wqi_tail = 0;846 847	return 0;848}849 850static int __guc_wq_item_append(struct i915_request *rq)851{852	struct intel_context *ce = request_to_scheduling_context(rq);853	struct intel_context *child;854	unsigned int wqi_size = (ce->parallel.number_children + 4) *855		sizeof(u32);856	u32 *wqi;857	u32 len_dw = (wqi_size / sizeof(u32)) - 1;858	int ret;859 860	/* Ensure context is in correct state updating work queue */861	GEM_BUG_ON(!atomic_read(&ce->guc_id.ref));862	GEM_BUG_ON(context_guc_id_invalid(ce));863	GEM_BUG_ON(context_wait_for_deregister_to_register(ce));864	GEM_BUG_ON(!ctx_id_mapped(ce_to_guc(ce), ce->guc_id.id));865 866	/* Insert NOOP if this work queue item will wrap the tail pointer. */867	if (wqi_size > wq_space_until_wrap(ce)) {868		ret = guc_wq_noop_append(ce);869		if (ret)870			return ret;871	}872 873	wqi = get_wq_pointer(ce, wqi_size);874	if (!wqi)875		return -EBUSY;876 877	GEM_BUG_ON(!FIELD_FIT(WQ_LEN_MASK, len_dw));878 879	*wqi++ = FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_MULTI_LRC) |880		FIELD_PREP(WQ_LEN_MASK, len_dw);881	*wqi++ = ce->lrc.lrca;882	*wqi++ = FIELD_PREP(WQ_GUC_ID_MASK, ce->guc_id.id) |883	       FIELD_PREP(WQ_RING_TAIL_MASK, ce->ring->tail / sizeof(u64));884	*wqi++ = 0;	/* fence_id */885	for_each_child(ce, child)886		*wqi++ = child->ring->tail / sizeof(u64);887 888	write_wqi(ce, wqi_size);889 890	return 0;891}892 893static int guc_wq_item_append(struct intel_guc *guc,894			      struct i915_request *rq)895{896	struct intel_context *ce = request_to_scheduling_context(rq);897	int ret;898 899	if (unlikely(!intel_context_is_schedulable(ce)))900		return 0;901 902	ret = __guc_wq_item_append(rq);903	if (unlikely(ret == -EBUSY)) {904		guc->stalled_request = rq;905		guc->submission_stall_reason = STALL_MOVE_LRC_TAIL;906	}907 908	return ret;909}910 911static bool multi_lrc_submit(struct i915_request *rq)912{913	struct intel_context *ce = request_to_scheduling_context(rq);914 915	intel_ring_set_tail(rq->ring, rq->tail);916 917	/*918	 * We expect the front end (execbuf IOCTL) to set this flag on the last919	 * request generated from a multi-BB submission. This indicates to the920	 * backend (GuC interface) that we should submit this context thus921	 * submitting all the requests generated in parallel.922	 */923	return test_bit(I915_FENCE_FLAG_SUBMIT_PARALLEL, &rq->fence.flags) ||924	       !intel_context_is_schedulable(ce);925}926 927static int guc_dequeue_one_context(struct intel_guc *guc)928{929	struct i915_sched_engine * const sched_engine = guc->sched_engine;930	struct i915_request *last = NULL;931	bool submit = false;932	struct rb_node *rb;933	int ret;934 935	lockdep_assert_held(&sched_engine->lock);936 937	if (guc->stalled_request) {938		submit = true;939		last = guc->stalled_request;940 941		switch (guc->submission_stall_reason) {942		case STALL_REGISTER_CONTEXT:943			goto register_context;944		case STALL_MOVE_LRC_TAIL:945			goto move_lrc_tail;946		case STALL_ADD_REQUEST:947			goto add_request;948		default:949			MISSING_CASE(guc->submission_stall_reason);950		}951	}952 953	while ((rb = rb_first_cached(&sched_engine->queue))) {954		struct i915_priolist *p = to_priolist(rb);955		struct i915_request *rq, *rn;956 957		priolist_for_each_request_consume(rq, rn, p) {958			if (last && !can_merge_rq(rq, last))959				goto register_context;960 961			list_del_init(&rq->sched.link);962 963			__i915_request_submit(rq);964 965			trace_i915_request_in(rq, 0);966			last = rq;967 968			if (is_multi_lrc_rq(rq)) {969				/*970				 * We need to coalesce all multi-lrc requests in971				 * a relationship into a single H2G. We are972				 * guaranteed that all of these requests will be973				 * submitted sequentially.974				 */975				if (multi_lrc_submit(rq)) {976					submit = true;977					goto register_context;978				}979			} else {980				submit = true;981			}982		}983 984		rb_erase_cached(&p->node, &sched_engine->queue);985		i915_priolist_free(p);986	}987 988register_context:989	if (submit) {990		struct intel_context *ce = request_to_scheduling_context(last);991 992		if (unlikely(!ctx_id_mapped(guc, ce->guc_id.id) &&993			     intel_context_is_schedulable(ce))) {994			ret = try_context_registration(ce, false);995			if (unlikely(ret == -EPIPE)) {996				goto deadlk;997			} else if (ret == -EBUSY) {998				guc->stalled_request = last;999				guc->submission_stall_reason =1000					STALL_REGISTER_CONTEXT;1001				goto schedule_tasklet;1002			} else if (ret != 0) {1003				GEM_WARN_ON(ret);	/* Unexpected */1004				goto deadlk;1005			}1006		}1007 1008move_lrc_tail:1009		if (is_multi_lrc_rq(last)) {1010			ret = guc_wq_item_append(guc, last);1011			if (ret == -EBUSY) {1012				goto schedule_tasklet;1013			} else if (ret != 0) {1014				GEM_WARN_ON(ret);	/* Unexpected */1015				goto deadlk;1016			}1017		} else {1018			guc_set_lrc_tail(last);1019		}1020 1021add_request:1022		ret = guc_add_request(guc, last);1023		if (unlikely(ret == -EPIPE)) {1024			goto deadlk;1025		} else if (ret == -EBUSY) {1026			goto schedule_tasklet;1027		} else if (ret != 0) {1028			GEM_WARN_ON(ret);	/* Unexpected */1029			goto deadlk;1030		}1031	}1032 1033	guc->stalled_request = NULL;1034	guc->submission_stall_reason = STALL_NONE;1035	return submit;1036 1037deadlk:1038	sched_engine->tasklet.callback = NULL;1039	tasklet_disable_nosync(&sched_engine->tasklet);1040	return false;1041 1042schedule_tasklet:1043	tasklet_schedule(&sched_engine->tasklet);1044	return false;1045}1046 1047static void guc_submission_tasklet(struct tasklet_struct *t)1048{1049	struct i915_sched_engine *sched_engine =1050		from_tasklet(sched_engine, t, tasklet);1051	unsigned long flags;1052	bool loop;1053 1054	spin_lock_irqsave(&sched_engine->lock, flags);1055 1056	do {1057		loop = guc_dequeue_one_context(sched_engine->private_data);1058	} while (loop);1059 1060	i915_sched_engine_reset_on_empty(sched_engine);1061 1062	spin_unlock_irqrestore(&sched_engine->lock, flags);1063}1064 1065static void cs_irq_handler(struct intel_engine_cs *engine, u16 iir)1066{1067	if (iir & GT_RENDER_USER_INTERRUPT)1068		intel_engine_signal_breadcrumbs(engine);1069}1070 1071static void __guc_context_destroy(struct intel_context *ce);1072static void release_guc_id(struct intel_guc *guc, struct intel_context *ce);1073static void guc_signal_context_fence(struct intel_context *ce);1074static void guc_cancel_context_requests(struct intel_context *ce);1075static void guc_blocked_fence_complete(struct intel_context *ce);1076 1077static void scrub_guc_desc_for_outstanding_g2h(struct intel_guc *guc)1078{1079	struct intel_context *ce;1080	unsigned long index, flags;1081	bool pending_disable, pending_enable, deregister, destroyed, banned;1082 1083	xa_lock_irqsave(&guc->context_lookup, flags);1084	xa_for_each(&guc->context_lookup, index, ce) {1085		/*1086		 * Corner case where the ref count on the object is zero but and1087		 * deregister G2H was lost. In this case we don't touch the ref1088		 * count and finish the destroy of the context.1089		 */1090		bool do_put = kref_get_unless_zero(&ce->ref);1091 1092		xa_unlock(&guc->context_lookup);1093 1094		if (test_bit(CONTEXT_GUC_INIT, &ce->flags) &&1095		    (cancel_delayed_work(&ce->guc_state.sched_disable_delay_work))) {1096			/* successful cancel so jump straight to close it */1097			intel_context_sched_disable_unpin(ce);1098		}1099 1100		spin_lock(&ce->guc_state.lock);1101 1102		/*1103		 * Once we are at this point submission_disabled() is guaranteed1104		 * to be visible to all callers who set the below flags (see above1105		 * flush and flushes in reset_prepare). If submission_disabled()1106		 * is set, the caller shouldn't set these flags.1107		 */1108 1109		destroyed = context_destroyed(ce);1110		pending_enable = context_pending_enable(ce);1111		pending_disable = context_pending_disable(ce);1112		deregister = context_wait_for_deregister_to_register(ce);1113		banned = context_banned(ce);1114		init_sched_state(ce);1115 1116		spin_unlock(&ce->guc_state.lock);1117 1118		if (pending_enable || destroyed || deregister) {1119			decr_outstanding_submission_g2h(guc);1120			if (deregister)1121				guc_signal_context_fence(ce);1122			if (destroyed) {1123				intel_gt_pm_put_async_untracked(guc_to_gt(guc));1124				release_guc_id(guc, ce);1125				__guc_context_destroy(ce);1126			}1127			if (pending_enable || deregister)1128				intel_context_put(ce);1129		}1130 1131		/* Not mutualy exclusive with above if statement. */1132		if (pending_disable) {1133			guc_signal_context_fence(ce);1134			if (banned) {1135				guc_cancel_context_requests(ce);1136				intel_engine_signal_breadcrumbs(ce->engine);1137			}1138			intel_context_sched_disable_unpin(ce);1139			decr_outstanding_submission_g2h(guc);1140 1141			spin_lock(&ce->guc_state.lock);1142			guc_blocked_fence_complete(ce);1143			spin_unlock(&ce->guc_state.lock);1144 1145			intel_context_put(ce);1146		}1147 1148		if (do_put)1149			intel_context_put(ce);1150		xa_lock(&guc->context_lookup);1151	}1152	xa_unlock_irqrestore(&guc->context_lookup, flags);1153}1154 1155/*1156 * GuC stores busyness stats for each engine at context in/out boundaries. A1157 * context 'in' logs execution start time, 'out' adds in -> out delta to total.1158 * i915/kmd accesses 'start', 'total' and 'context id' from memory shared with1159 * GuC.1160 *1161 * __i915_pmu_event_read samples engine busyness. When sampling, if context id1162 * is valid (!= ~0) and start is non-zero, the engine is considered to be1163 * active. For an active engine total busyness = total + (now - start), where1164 * 'now' is the time at which the busyness is sampled. For inactive engine,1165 * total busyness = total.1166 *1167 * All times are captured from GUCPMTIMESTAMP reg and are in gt clock domain.1168 *1169 * The start and total values provided by GuC are 32 bits and wrap around in a1170 * few minutes. Since perf pmu provides busyness as 64 bit monotonically1171 * increasing ns values, there is a need for this implementation to account for1172 * overflows and extend the GuC provided values to 64 bits before returning1173 * busyness to the user. In order to do that, a worker runs periodically at1174 * frequency = 1/8th the time it takes for the timestamp to wrap (i.e. once in1175 * 27 seconds for a gt clock frequency of 19.2 MHz).1176 */1177 1178#define WRAP_TIME_CLKS U32_MAX1179#define POLL_TIME_CLKS (WRAP_TIME_CLKS >> 3)1180 1181static void1182__extend_last_switch(struct intel_guc *guc, u64 *prev_start, u32 new_start)1183{1184	u32 gt_stamp_hi = upper_32_bits(guc->timestamp.gt_stamp);1185	u32 gt_stamp_last = lower_32_bits(guc->timestamp.gt_stamp);1186 1187	if (new_start == lower_32_bits(*prev_start))1188		return;1189 1190	/*1191	 * When gt is unparked, we update the gt timestamp and start the ping1192	 * worker that updates the gt_stamp every POLL_TIME_CLKS. As long as gt1193	 * is unparked, all switched in contexts will have a start time that is1194	 * within +/- POLL_TIME_CLKS of the most recent gt_stamp.1195	 *1196	 * If neither gt_stamp nor new_start has rolled over, then the1197	 * gt_stamp_hi does not need to be adjusted, however if one of them has1198	 * rolled over, we need to adjust gt_stamp_hi accordingly.1199	 *1200	 * The below conditions address the cases of new_start rollover and1201	 * gt_stamp_last rollover respectively.1202	 */1203	if (new_start < gt_stamp_last &&1204	    (new_start - gt_stamp_last) <= POLL_TIME_CLKS)1205		gt_stamp_hi++;1206 1207	if (new_start > gt_stamp_last &&1208	    (gt_stamp_last - new_start) <= POLL_TIME_CLKS && gt_stamp_hi)1209		gt_stamp_hi--;1210 1211	*prev_start = ((u64)gt_stamp_hi << 32) | new_start;1212}1213 1214#define record_read(map_, field_) \1215	iosys_map_rd_field(map_, 0, struct guc_engine_usage_record, field_)1216 1217/*1218 * GuC updates shared memory and KMD reads it. Since this is not synchronized,1219 * we run into a race where the value read is inconsistent. Sometimes the1220 * inconsistency is in reading the upper MSB bytes of the last_in value when1221 * this race occurs. 2 types of cases are seen - upper 8 bits are zero and upper1222 * 24 bits are zero. Since these are non-zero values, it is non-trivial to1223 * determine validity of these values. Instead we read the values multiple times1224 * until they are consistent. In test runs, 3 attempts results in consistent1225 * values. The upper bound is set to 6 attempts and may need to be tuned as per1226 * any new occurences.1227 */1228static void __get_engine_usage_record(struct intel_engine_cs *engine,1229				      u32 *last_in, u32 *id, u32 *total)1230{1231	struct iosys_map rec_map = intel_guc_engine_usage_record_map(engine);1232	int i = 0;1233 1234	do {1235		*last_in = record_read(&rec_map, last_switch_in_stamp);1236		*id = record_read(&rec_map, current_context_index);1237		*total = record_read(&rec_map, total_runtime);1238 1239		if (record_read(&rec_map, last_switch_in_stamp) == *last_in &&1240		    record_read(&rec_map, current_context_index) == *id &&1241		    record_read(&rec_map, total_runtime) == *total)1242			break;1243	} while (++i < 6);1244}1245 1246static void guc_update_engine_gt_clks(struct intel_engine_cs *engine)1247{1248	struct intel_engine_guc_stats *stats = &engine->stats.guc;1249	struct intel_guc *guc = gt_to_guc(engine->gt);1250	u32 last_switch, ctx_id, total;1251 1252	lockdep_assert_held(&guc->timestamp.lock);1253 1254	__get_engine_usage_record(engine, &last_switch, &ctx_id, &total);1255 1256	stats->running = ctx_id != ~0U && last_switch;1257	if (stats->running)1258		__extend_last_switch(guc, &stats->start_gt_clk, last_switch);1259 1260	/*1261	 * Instead of adjusting the total for overflow, just add the1262	 * difference from previous sample stats->total_gt_clks1263	 */1264	if (total && total != ~0U) {1265		stats->total_gt_clks += (u32)(total - stats->prev_total);1266		stats->prev_total = total;1267	}1268}1269 1270static u32 gpm_timestamp_shift(struct intel_gt *gt)1271{1272	intel_wakeref_t wakeref;1273	u32 reg, shift;1274 1275	with_intel_runtime_pm(gt->uncore->rpm, wakeref)1276		reg = intel_uncore_read(gt->uncore, RPM_CONFIG0);1277 1278	shift = (reg & GEN10_RPM_CONFIG0_CTC_SHIFT_PARAMETER_MASK) >>1279		GEN10_RPM_CONFIG0_CTC_SHIFT_PARAMETER_SHIFT;1280 1281	return 3 - shift;1282}1283 1284static void guc_update_pm_timestamp(struct intel_guc *guc, ktime_t *now)1285{1286	struct intel_gt *gt = guc_to_gt(guc);1287	u32 gt_stamp_lo, gt_stamp_hi;1288	u64 gpm_ts;1289 1290	lockdep_assert_held(&guc->timestamp.lock);1291 1292	gt_stamp_hi = upper_32_bits(guc->timestamp.gt_stamp);1293	gpm_ts = intel_uncore_read64_2x32(gt->uncore, MISC_STATUS0,1294					  MISC_STATUS1) >> guc->timestamp.shift;1295	gt_stamp_lo = lower_32_bits(gpm_ts);1296	*now = ktime_get();1297 1298	if (gt_stamp_lo < lower_32_bits(guc->timestamp.gt_stamp))1299		gt_stamp_hi++;1300 1301	guc->timestamp.gt_stamp = ((u64)gt_stamp_hi << 32) | gt_stamp_lo;1302}1303 1304/*1305 * Unlike the execlist mode of submission total and active times are in terms of1306 * gt clocks. The *now parameter is retained to return the cpu time at which the1307 * busyness was sampled.1308 */1309static ktime_t guc_engine_busyness(struct intel_engine_cs *engine, ktime_t *now)1310{1311	struct intel_engine_guc_stats stats_saved, *stats = &engine->stats.guc;1312	struct i915_gpu_error *gpu_error = &engine->i915->gpu_error;1313	struct intel_gt *gt = engine->gt;1314	struct intel_guc *guc = gt_to_guc(gt);1315	u64 total, gt_stamp_saved;1316	unsigned long flags;1317	u32 reset_count;1318	bool in_reset;1319	intel_wakeref_t wakeref;1320 1321	spin_lock_irqsave(&guc->timestamp.lock, flags);1322 1323	/*1324	 * If a reset happened, we risk reading partially updated engine1325	 * busyness from GuC, so we just use the driver stored copy of busyness.1326	 * Synchronize with gt reset using reset_count and the1327	 * I915_RESET_BACKOFF flag. Note that reset flow updates the reset_count1328	 * after I915_RESET_BACKOFF flag, so ensure that the reset_count is1329	 * usable by checking the flag afterwards.1330	 */1331	reset_count = i915_reset_count(gpu_error);1332	in_reset = test_bit(I915_RESET_BACKOFF, &gt->reset.flags);1333 1334	*now = ktime_get();1335 1336	/*1337	 * The active busyness depends on start_gt_clk and gt_stamp.1338	 * gt_stamp is updated by i915 only when gt is awake and the1339	 * start_gt_clk is derived from GuC state. To get a consistent1340	 * view of activity, we query the GuC state only if gt is awake.1341	 */1342	wakeref = in_reset ? 0 : intel_gt_pm_get_if_awake(gt);1343	if (wakeref) {1344		stats_saved = *stats;1345		gt_stamp_saved = guc->timestamp.gt_stamp;1346		/*1347		 * Update gt_clks, then gt timestamp to simplify the 'gt_stamp -1348		 * start_gt_clk' calculation below for active engines.1349		 */1350		guc_update_engine_gt_clks(engine);1351		guc_update_pm_timestamp(guc, now);1352		intel_gt_pm_put_async(gt, wakeref);1353		if (i915_reset_count(gpu_error) != reset_count) {1354			*stats = stats_saved;1355			guc->timestamp.gt_stamp = gt_stamp_saved;1356		}1357	}1358 1359	total = intel_gt_clock_interval_to_ns(gt, stats->total_gt_clks);1360	if (stats->running) {1361		u64 clk = guc->timestamp.gt_stamp - stats->start_gt_clk;1362 1363		total += intel_gt_clock_interval_to_ns(gt, clk);1364	}1365 1366	spin_unlock_irqrestore(&guc->timestamp.lock, flags);1367 1368	return ns_to_ktime(total);1369}1370 1371static void guc_enable_busyness_worker(struct intel_guc *guc)1372{1373	mod_delayed_work(system_highpri_wq, &guc->timestamp.work, guc->timestamp.ping_delay);1374}1375 1376static void guc_cancel_busyness_worker(struct intel_guc *guc)1377{1378	/*1379	 * There are many different call stacks that can get here. Some of them1380	 * hold the reset mutex. The busyness worker also attempts to acquire the1381	 * reset mutex. Synchronously flushing a worker thread requires acquiring1382	 * the worker mutex. Lockdep sees this as a conflict. It thinks that the1383	 * flush can deadlock because it holds the worker mutex while waiting for1384	 * the reset mutex, but another thread is holding the reset mutex and might1385	 * attempt to use other worker functions.1386	 *1387	 * In practice, this scenario does not exist because the busyness worker1388	 * does not block waiting for the reset mutex. It does a try-lock on it and1389	 * immediately exits if the lock is already held. Unfortunately, the mutex1390	 * in question (I915_RESET_BACKOFF) is an i915 implementation which has lockdep1391	 * annotation but not to the extent of explaining the 'might lock' is also a1392	 * 'does not need to lock'. So one option would be to add more complex lockdep1393	 * annotations to ignore the issue (if at all possible). A simpler option is to1394	 * just not flush synchronously when a rest in progress. Given that the worker1395	 * will just early exit and re-schedule itself anyway, there is no advantage1396	 * to running it immediately.1397	 *1398	 * If a reset is not in progress, then the synchronous flush may be required.1399	 * As noted many call stacks lead here, some during suspend and driver unload1400	 * which do require a synchronous flush to make sure the worker is stopped1401	 * before memory is freed.1402	 *1403	 * Trying to pass a 'need_sync' or 'in_reset' flag all the way down through1404	 * every possible call stack is unfeasible. It would be too intrusive to many1405	 * areas that really don't care about the GuC backend. However, there is the1406	 * I915_RESET_BACKOFF flag and the gt->reset.mutex can be tested for is_locked.1407	 * So just use those. Note that testing both is required due to the hideously1408	 * complex nature of the i915 driver's reset code paths.1409	 *1410	 * And note that in the case of a reset occurring during driver unload1411	 * (wedged_on_fini), skipping the cancel in reset_prepare/reset_fini (when the1412	 * reset flag/mutex are set) is fine because there is another explicit cancel in1413	 * intel_guc_submission_fini (when the reset flag/mutex are not).1414	 */1415	if (mutex_is_locked(&guc_to_gt(guc)->reset.mutex) ||1416	    test_bit(I915_RESET_BACKOFF, &guc_to_gt(guc)->reset.flags))1417		cancel_delayed_work(&guc->timestamp.work);1418	else1419		cancel_delayed_work_sync(&guc->timestamp.work);1420}1421 1422static void __reset_guc_busyness_stats(struct intel_guc *guc)1423{1424	struct intel_gt *gt = guc_to_gt(guc);1425	struct intel_engine_cs *engine;1426	enum intel_engine_id id;1427	unsigned long flags;1428	ktime_t unused;1429 1430	spin_lock_irqsave(&guc->timestamp.lock, flags);1431 1432	guc_update_pm_timestamp(guc, &unused);1433	for_each_engine(engine, gt, id) {1434		guc_update_engine_gt_clks(engine);1435		engine->stats.guc.prev_total = 0;1436	}1437 1438	spin_unlock_irqrestore(&guc->timestamp.lock, flags);1439}1440 1441static void __update_guc_busyness_stats(struct intel_guc *guc)1442{1443	struct intel_gt *gt = guc_to_gt(guc);1444	struct intel_engine_cs *engine;1445	enum intel_engine_id id;1446	unsigned long flags;1447	ktime_t unused;1448 1449	guc->timestamp.last_stat_jiffies = jiffies;1450 1451	spin_lock_irqsave(&guc->timestamp.lock, flags);1452 1453	guc_update_pm_timestamp(guc, &unused);1454	for_each_engine(engine, gt, id)1455		guc_update_engine_gt_clks(engine);1456 1457	spin_unlock_irqrestore(&guc->timestamp.lock, flags);1458}1459 1460static void __guc_context_update_stats(struct intel_context *ce)1461{1462	struct intel_guc *guc = ce_to_guc(ce);1463	unsigned long flags;1464 1465	spin_lock_irqsave(&guc->timestamp.lock, flags);1466	lrc_update_runtime(ce);1467	spin_unlock_irqrestore(&guc->timestamp.lock, flags);1468}1469 1470static void guc_context_update_stats(struct intel_context *ce)1471{1472	if (!intel_context_pin_if_active(ce))1473		return;1474 1475	__guc_context_update_stats(ce);1476	intel_context_unpin(ce);1477}1478 1479static void guc_timestamp_ping(struct work_struct *wrk)1480{1481	struct intel_guc *guc = container_of(wrk, typeof(*guc),1482					     timestamp.work.work);1483	struct intel_uc *uc = container_of(guc, typeof(*uc), guc);1484	struct intel_gt *gt = guc_to_gt(guc);1485	struct intel_context *ce;1486	intel_wakeref_t wakeref;1487	unsigned long index;1488	int srcu, ret;1489 1490	/*1491	 * Ideally the busyness worker should take a gt pm wakeref because the1492	 * worker only needs to be active while gt is awake. However, the1493	 * gt_park path cancels the worker synchronously and this complicates1494	 * the flow if the worker is also running at the same time. The cancel1495	 * waits for the worker and when the worker releases the wakeref, that1496	 * would call gt_park and would lead to a deadlock.1497	 *1498	 * The resolution is to take the global pm wakeref if runtime pm is1499	 * already active. If not, we don't need to update the busyness stats as1500	 * the stats would already be updated when the gt was parked.1501	 *1502	 * Note:1503	 * - We do not requeue the worker if we cannot take a reference to runtime1504	 *   pm since intel_guc_busyness_unpark would requeue the worker in the1505	 *   resume path.1506	 *1507	 * - If the gt was parked longer than time taken for GT timestamp to roll1508	 *   over, we ignore those rollovers since we don't care about tracking1509	 *   the exact GT time. We only care about roll overs when the gt is1510	 *   active and running workloads.1511	 *1512	 * - There is a window of time between gt_park and runtime suspend,1513	 *   where the worker may run. This is acceptable since the worker will1514	 *   not find any new data to update busyness.1515	 */1516	wakeref = intel_runtime_pm_get_if_active(&gt->i915->runtime_pm);1517	if (!wakeref)1518		return;1519 1520	/*1521	 * Synchronize with gt reset to make sure the worker does not1522	 * corrupt the engine/guc stats. NB: can't actually block waiting1523	 * for a reset to complete as the reset requires flushing out1524	 * this worker thread if started. So waiting would deadlock.1525	 */1526	ret = intel_gt_reset_trylock(gt, &srcu);1527	if (ret)1528		goto err_trylock;1529 1530	__update_guc_busyness_stats(guc);1531 1532	/* adjust context stats for overflow */1533	xa_for_each(&guc->context_lookup, index, ce)1534		guc_context_update_stats(ce);1535 1536	intel_gt_reset_unlock(gt, srcu);1537 1538	guc_enable_busyness_worker(guc);1539 1540err_trylock:1541	intel_runtime_pm_put(&gt->i915->runtime_pm, wakeref);1542}1543 1544static int guc_action_enable_usage_stats(struct intel_guc *guc)1545{1546	u32 offset = intel_guc_engine_usage_offset(guc);1547	u32 action[] = {1548		INTEL_GUC_ACTION_SET_ENG_UTIL_BUFF,1549		offset,1550		0,1551	};1552 1553	return intel_guc_send(guc, action, ARRAY_SIZE(action));1554}1555 1556static int guc_init_engine_stats(struct intel_guc *guc)1557{1558	struct intel_gt *gt = guc_to_gt(guc);1559	intel_wakeref_t wakeref;1560	int ret;1561 1562	with_intel_runtime_pm(&gt->i915->runtime_pm, wakeref)1563		ret = guc_action_enable_usage_stats(guc);1564 1565	if (ret)1566		guc_err(guc, "Failed to enable usage stats: %pe\n", ERR_PTR(ret));1567	else1568		guc_enable_busyness_worker(guc);1569 1570	return ret;1571}1572 1573static void guc_fini_engine_stats(struct intel_guc *guc)1574{1575	guc_cancel_busyness_worker(guc);1576}1577 1578void intel_guc_busyness_park(struct intel_gt *gt)1579{1580	struct intel_guc *guc = gt_to_guc(gt);1581 1582	if (!guc_submission_initialized(guc))1583		return;1584 1585	/*1586	 * There is a race with suspend flow where the worker runs after suspend1587	 * and causes an unclaimed register access warning. Cancel the worker1588	 * synchronously here.1589	 */1590	guc_cancel_busyness_worker(guc);1591 1592	/*1593	 * Before parking, we should sample engine busyness stats if we need to.1594	 * We can skip it if we are less than half a ping from the last time we1595	 * sampled the busyness stats.1596	 */1597	if (guc->timestamp.last_stat_jiffies &&1598	    !time_after(jiffies, guc->timestamp.last_stat_jiffies +1599			(guc->timestamp.ping_delay / 2)))1600		return;1601 1602	__update_guc_busyness_stats(guc);1603}1604 1605void intel_guc_busyness_unpark(struct intel_gt *gt)1606{1607	struct intel_guc *guc = gt_to_guc(gt);1608	unsigned long flags;1609	ktime_t unused;1610 1611	if (!guc_submission_initialized(guc))1612		return;1613 1614	spin_lock_irqsave(&guc->timestamp.lock, flags);1615	guc_update_pm_timestamp(guc, &unused);1616	spin_unlock_irqrestore(&guc->timestamp.lock, flags);1617	guc_enable_busyness_worker(guc);1618}1619 1620static inline bool1621submission_disabled(struct intel_guc *guc)1622{1623	struct i915_sched_engine * const sched_engine = guc->sched_engine;1624 1625	return unlikely(!sched_engine ||1626			!__tasklet_is_enabled(&sched_engine->tasklet) ||1627			intel_gt_is_wedged(guc_to_gt(guc)));1628}1629 1630static void disable_submission(struct intel_guc *guc)1631{1632	struct i915_sched_engine * const sched_engine = guc->sched_engine;1633 1634	if (__tasklet_is_enabled(&sched_engine->tasklet)) {1635		GEM_BUG_ON(!guc->ct.enabled);1636		__tasklet_disable_sync_once(&sched_engine->tasklet);1637		sched_engine->tasklet.callback = NULL;1638	}1639}1640 1641static void enable_submission(struct intel_guc *guc)1642{1643	struct i915_sched_engine * const sched_engine = guc->sched_engine;1644	unsigned long flags;1645 1646	spin_lock_irqsave(&guc->sched_engine->lock, flags);1647	sched_engine->tasklet.callback = guc_submission_tasklet;1648	wmb();	/* Make sure callback visible */1649	if (!__tasklet_is_enabled(&sched_engine->tasklet) &&1650	    __tasklet_enable(&sched_engine->tasklet)) {1651		GEM_BUG_ON(!guc->ct.enabled);1652 1653		/* And kick in case we missed a new request submission. */1654		tasklet_hi_schedule(&sched_engine->tasklet);1655	}1656	spin_unlock_irqrestore(&guc->sched_engine->lock, flags);1657}1658 1659static void guc_flush_submissions(struct intel_guc *guc)1660{1661	struct i915_sched_engine * const sched_engine = guc->sched_engine;1662	unsigned long flags;1663 1664	spin_lock_irqsave(&sched_engine->lock, flags);1665	spin_unlock_irqrestore(&sched_engine->lock, flags);1666}1667 1668void intel_guc_submission_flush_work(struct intel_guc *guc)1669{1670	flush_work(&guc->submission_state.destroyed_worker);1671}1672 1673static void guc_flush_destroyed_contexts(struct intel_guc *guc);1674 1675void intel_guc_submission_reset_prepare(struct intel_guc *guc)1676{1677	if (unlikely(!guc_submission_initialized(guc))) {1678		/* Reset called during driver load? GuC not yet initialised! */1679		return;1680	}1681 1682	intel_gt_park_heartbeats(guc_to_gt(guc));1683	disable_submission(guc);1684	guc->interrupts.disable(guc);1685	__reset_guc_busyness_stats(guc);1686 1687	/* Flush IRQ handler */1688	spin_lock_irq(guc_to_gt(guc)->irq_lock);1689	spin_unlock_irq(guc_to_gt(guc)->irq_lock);1690 1691	guc_flush_submissions(guc);1692	guc_flush_destroyed_contexts(guc);1693	flush_work(&guc->ct.requests.worker);1694 1695	scrub_guc_desc_for_outstanding_g2h(guc);1696}1697 1698static struct intel_engine_cs *1699guc_virtual_get_sibling(struct intel_engine_cs *ve, unsigned int sibling)1700{1701	struct intel_engine_cs *engine;1702	intel_engine_mask_t tmp, mask = ve->mask;1703	unsigned int num_siblings = 0;1704 1705	for_each_engine_masked(engine, ve->gt, mask, tmp)1706		if (num_siblings++ == sibling)1707			return engine;1708 1709	return NULL;1710}1711 1712static inline struct intel_engine_cs *1713__context_to_physical_engine(struct intel_context *ce)1714{1715	struct intel_engine_cs *engine = ce->engine;1716 1717	if (intel_engine_is_virtual(engine))1718		engine = guc_virtual_get_sibling(engine, 0);1719 1720	return engine;1721}1722 1723static void guc_reset_state(struct intel_context *ce, u32 head, bool scrub)1724{1725	struct intel_engine_cs *engine = __context_to_physical_engine(ce);1726 1727	if (!intel_context_is_schedulable(ce))1728		return;1729 1730	GEM_BUG_ON(!intel_context_is_pinned(ce));1731 1732	/*1733	 * We want a simple context + ring to execute the breadcrumb update.1734	 * We cannot rely on the context being intact across the GPU hang,1735	 * so clear it and rebuild just what we need for the breadcrumb.1736	 * All pending requests for this context will be zapped, and any1737	 * future request will be after userspace has had the opportunity1738	 * to recreate its own state.1739	 */1740	if (scrub)1741		lrc_init_regs(ce, engine, true);1742 1743	/* Rerun the request; its payload has been neutered (if guilty). */1744	lrc_update_regs(ce, engine, head);1745}1746 1747static void guc_engine_reset_prepare(struct intel_engine_cs *engine)1748{1749	/*1750	 * Wa_22011802037: In addition to stopping the cs, we need1751	 * to wait for any pending mi force wakeups1752	 */1753	if (intel_engine_reset_needs_wa_22011802037(engine->gt)) {1754		intel_engine_stop_cs(engine);1755		intel_engine_wait_for_pending_mi_fw(engine);1756	}1757}1758 1759static void guc_reset_nop(struct intel_engine_cs *engine)1760{1761}1762 1763static void guc_rewind_nop(struct intel_engine_cs *engine, bool stalled)1764{1765}1766 1767static void1768__unwind_incomplete_requests(struct intel_context *ce)1769{1770	struct i915_request *rq, *rn;1771	struct list_head *pl;1772	int prio = I915_PRIORITY_INVALID;1773	struct i915_sched_engine * const sched_engine =1774		ce->engine->sched_engine;1775	unsigned long flags;1776 1777	spin_lock_irqsave(&sched_engine->lock, flags);1778	spin_lock(&ce->guc_state.lock);1779	list_for_each_entry_safe_reverse(rq, rn,1780					 &ce->guc_state.requests,1781					 sched.link) {1782		if (i915_request_completed(rq))1783			continue;1784 1785		list_del_init(&rq->sched.link);1786		__i915_request_unsubmit(rq);1787 1788		/* Push the request back into the queue for later resubmission. */1789		GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);1790		if (rq_prio(rq) != prio) {1791			prio = rq_prio(rq);1792			pl = i915_sched_lookup_priolist(sched_engine, prio);1793		}1794		GEM_BUG_ON(i915_sched_engine_is_empty(sched_engine));1795 1796		list_add(&rq->sched.link, pl);1797		set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);1798	}1799	spin_unlock(&ce->guc_state.lock);1800	spin_unlock_irqrestore(&sched_engine->lock, flags);1801}1802 1803static void __guc_reset_context(struct intel_context *ce, intel_engine_mask_t stalled)1804{1805	bool guilty;1806	struct i915_request *rq;1807	unsigned long flags;1808	u32 head;1809	int i, number_children = ce->parallel.number_children;1810	struct intel_context *parent = ce;1811 1812	GEM_BUG_ON(intel_context_is_child(ce));1813 1814	intel_context_get(ce);1815 1816	/*1817	 * GuC will implicitly mark the context as non-schedulable when it sends1818	 * the reset notification. Make sure our state reflects this change. The1819	 * context will be marked enabled on resubmission.1820	 */1821	spin_lock_irqsave(&ce->guc_state.lock, flags);1822	clr_context_enabled(ce);1823	spin_unlock_irqrestore(&ce->guc_state.lock, flags);1824 1825	/*1826	 * For each context in the relationship find the hanging request1827	 * resetting each context / request as needed1828	 */1829	for (i = 0; i < number_children + 1; ++i) {1830		if (!intel_context_is_pinned(ce))1831			goto next_context;1832 1833		guilty = false;1834		rq = intel_context_get_active_request(ce);1835		if (!rq) {1836			head = ce->ring->tail;1837			goto out_replay;1838		}1839 1840		if (i915_request_started(rq))1841			guilty = stalled & ce->engine->mask;1842 1843		GEM_BUG_ON(i915_active_is_idle(&ce->active));1844		head = intel_ring_wrap(ce->ring, rq->head);1845 1846		__i915_request_reset(rq, guilty);1847		i915_request_put(rq);1848out_replay:1849		guc_reset_state(ce, head, guilty);1850next_context:1851		if (i != number_children)1852			ce = list_next_entry(ce, parallel.child_link);1853	}1854 1855	__unwind_incomplete_requests(parent);1856	intel_context_put(parent);1857}1858 1859void wake_up_all_tlb_invalidate(struct intel_guc *guc)1860{1861	struct intel_guc_tlb_wait *wait;1862	unsigned long i;1863 1864	if (!intel_guc_tlb_invalidation_is_available(guc))1865		return;1866 1867	xa_lock_irq(&guc->tlb_lookup);1868	xa_for_each(&guc->tlb_lookup, i, wait)1869		wake_up(&wait->wq);1870	xa_unlock_irq(&guc->tlb_lookup);1871}1872 1873void intel_guc_submission_reset(struct intel_guc *guc, intel_engine_mask_t stalled)1874{1875	struct intel_context *ce;1876	unsigned long index;1877	unsigned long flags;1878 1879	if (unlikely(!guc_submission_initialized(guc))) {1880		/* Reset called during driver load? GuC not yet initialised! */1881		return;1882	}1883 1884	xa_lock_irqsave(&guc->context_lookup, flags);1885	xa_for_each(&guc->context_lookup, index, ce) {1886		if (!kref_get_unless_zero(&ce->ref))1887			continue;1888 1889		xa_unlock(&guc->context_lookup);1890 1891		if (intel_context_is_pinned(ce) &&1892		    !intel_context_is_child(ce))1893			__guc_reset_context(ce, stalled);1894 1895		intel_context_put(ce);1896 1897		xa_lock(&guc->context_lookup);1898	}1899	xa_unlock_irqrestore(&guc->context_lookup, flags);1900 1901	/* GuC is blown away, drop all references to contexts */1902	xa_destroy(&guc->context_lookup);1903}1904 1905static void guc_cancel_context_requests(struct intel_context *ce)1906{1907	struct i915_sched_engine *sched_engine = ce_to_guc(ce)->sched_engine;1908	struct i915_request *rq;1909	unsigned long flags;1910 1911	/* Mark all executing requests as skipped. */1912	spin_lock_irqsave(&sched_engine->lock, flags);1913	spin_lock(&ce->guc_state.lock);1914	list_for_each_entry(rq, &ce->guc_state.requests, sched.link)1915		i915_request_put(i915_request_mark_eio(rq));1916	spin_unlock(&ce->guc_state.lock);1917	spin_unlock_irqrestore(&sched_engine->lock, flags);1918}1919 1920static void1921guc_cancel_sched_engine_requests(struct i915_sched_engine *sched_engine)1922{1923	struct i915_request *rq, *rn;1924	struct rb_node *rb;1925	unsigned long flags;1926 1927	/* Can be called during boot if GuC fails to load */1928	if (!sched_engine)1929		return;1930 1931	/*1932	 * Before we call engine->cancel_requests(), we should have exclusive1933	 * access to the submission state. This is arranged for us by the1934	 * caller disabling the interrupt generation, the tasklet and other1935	 * threads that may then access the same state, giving us a free hand1936	 * to reset state. However, we still need to let lockdep be aware that1937	 * we know this state may be accessed in hardirq context, so we1938	 * disable the irq around this manipulation and we want to keep1939	 * the spinlock focused on its duties and not accidentally conflate1940	 * coverage to the submission's irq state. (Similarly, although we1941	 * shouldn't need to disable irq around the manipulation of the1942	 * submission's irq state, we also wish to remind ourselves that1943	 * it is irq state.)1944	 */1945	spin_lock_irqsave(&sched_engine->lock, flags);1946 1947	/* Flush the queued requests to the timeline list (for retiring). */1948	while ((rb = rb_first_cached(&sched_engine->queue))) {1949		struct i915_priolist *p = to_priolist(rb);1950 1951		priolist_for_each_request_consume(rq, rn, p) {1952			list_del_init(&rq->sched.link);1953 1954			__i915_request_submit(rq);1955 1956			i915_request_put(i915_request_mark_eio(rq));1957		}1958 1959		rb_erase_cached(&p->node, &sched_engine->queue);1960		i915_priolist_free(p);1961	}1962 1963	/* Remaining _unready_ requests will be nop'ed when submitted */1964 1965	sched_engine->queue_priority_hint = INT_MIN;1966	sched_engine->queue = RB_ROOT_CACHED;1967 1968	spin_unlock_irqrestore(&sched_engine->lock, flags);1969}1970 1971void intel_guc_submission_cancel_requests(struct intel_guc *guc)1972{1973	struct intel_context *ce;1974	unsigned long index;1975	unsigned long flags;1976 1977	xa_lock_irqsave(&guc->context_lookup, flags);1978	xa_for_each(&guc->context_lookup, index, ce) {1979		if (!kref_get_unless_zero(&ce->ref))1980			continue;1981 1982		xa_unlock(&guc->context_lookup);1983 1984		if (intel_context_is_pinned(ce) &&1985		    !intel_context_is_child(ce))1986			guc_cancel_context_requests(ce);1987 1988		intel_context_put(ce);1989 1990		xa_lock(&guc->context_lookup);1991	}1992	xa_unlock_irqrestore(&guc->context_lookup, flags);1993 1994	guc_cancel_sched_engine_requests(guc->sched_engine);1995 1996	/* GuC is blown away, drop all references to contexts */1997	xa_destroy(&guc->context_lookup);1998 1999	/*2000	 * Wedged GT won't respond to any TLB invalidation request. Simply2001	 * release all the blocked waiters.2002	 */2003	wake_up_all_tlb_invalidate(guc);2004}2005 2006void intel_guc_submission_reset_finish(struct intel_guc *guc)2007{2008	/* Reset called during driver load or during wedge? */2009	if (unlikely(!guc_submission_initialized(guc) ||2010		     !intel_guc_is_fw_running(guc) ||2011		     intel_gt_is_wedged(guc_to_gt(guc)))) {2012		return;2013	}2014 2015	/*2016	 * Technically possible for either of these values to be non-zero here,2017	 * but very unlikely + harmless. Regardless let's add an error so we can2018	 * see in CI if this happens frequently / a precursor to taking down the2019	 * machine.2020	 */2021	if (atomic_read(&guc->outstanding_submission_g2h))2022		guc_err(guc, "Unexpected outstanding GuC to Host in reset finish\n");2023	atomic_set(&guc->outstanding_submission_g2h, 0);2024 2025	intel_guc_global_policies_update(guc);2026	enable_submission(guc);2027	intel_gt_unpark_heartbeats(guc_to_gt(guc));2028 2029	/*2030	 * The full GT reset will have cleared the TLB caches and flushed the2031	 * G2H message queue; we can release all the blocked waiters.2032	 */2033	wake_up_all_tlb_invalidate(guc);2034}2035 2036static void destroyed_worker_func(struct work_struct *w);2037static void reset_fail_worker_func(struct work_struct *w);2038 2039bool intel_guc_tlb_invalidation_is_available(struct intel_guc *guc)2040{2041	return HAS_GUC_TLB_INVALIDATION(guc_to_gt(guc)->i915) &&2042		intel_guc_is_ready(guc);2043}2044 2045static int init_tlb_lookup(struct intel_guc *guc)2046{2047	struct intel_guc_tlb_wait *wait;2048	int err;2049 2050	if (!HAS_GUC_TLB_INVALIDATION(guc_to_gt(guc)->i915))2051		return 0;2052 2053	xa_init_flags(&guc->tlb_lookup, XA_FLAGS_ALLOC);2054 2055	wait = kzalloc(sizeof(*wait), GFP_KERNEL);2056	if (!wait)2057		return -ENOMEM;2058 2059	init_waitqueue_head(&wait->wq);2060 2061	/* Preallocate a shared id for use under memory pressure. */2062	err = xa_alloc_cyclic_irq(&guc->tlb_lookup, &guc->serial_slot, wait,2063				  xa_limit_32b, &guc->next_seqno, GFP_KERNEL);2064	if (err < 0) {2065		kfree(wait);2066		return err;2067	}2068 2069	return 0;2070}2071 2072static void fini_tlb_lookup(struct intel_guc *guc)2073{2074	struct intel_guc_tlb_wait *wait;2075 2076	if (!HAS_GUC_TLB_INVALIDATION(guc_to_gt(guc)->i915))2077		return;2078 2079	wait = xa_load(&guc->tlb_lookup, guc->serial_slot);2080	if (wait && wait->busy)2081		guc_err(guc, "Unexpected busy item in tlb_lookup on fini\n");2082	kfree(wait);2083 2084	xa_destroy(&guc->tlb_lookup);2085}2086 2087/*2088 * Set up the memory resources to be shared with the GuC (via the GGTT)2089 * at firmware loading time.2090 */2091int intel_guc_submission_init(struct intel_guc *guc)2092{2093	struct intel_gt *gt = guc_to_gt(guc);2094	int ret;2095 2096	if (guc->submission_initialized)2097		return 0;2098 2099	if (GUC_SUBMIT_VER(guc) < MAKE_GUC_VER(1, 0, 0)) {2100		ret = guc_lrc_desc_pool_create_v69(guc);2101		if (ret)2102			return ret;2103	}2104 2105	ret = init_tlb_lookup(guc);2106	if (ret)2107		goto destroy_pool;2108 2109	guc->submission_state.guc_ids_bitmap =2110		bitmap_zalloc(NUMBER_MULTI_LRC_GUC_ID(guc), GFP_KERNEL);2111	if (!guc->submission_state.guc_ids_bitmap) {2112		ret = -ENOMEM;2113		goto destroy_tlb;2114	}2115 2116	guc->timestamp.ping_delay = (POLL_TIME_CLKS / gt->clock_frequency + 1) * HZ;2117	guc->timestamp.shift = gpm_timestamp_shift(gt);2118	guc->submission_initialized = true;2119 2120	return 0;2121 2122destroy_tlb:2123	fini_tlb_lookup(guc);2124destroy_pool:2125	guc_lrc_desc_pool_destroy_v69(guc);2126	return ret;2127}2128 2129void intel_guc_submission_fini(struct intel_guc *guc)2130{2131	if (!guc->submission_initialized)2132		return;2133 2134	guc_fini_engine_stats(guc);2135	guc_flush_destroyed_contexts(guc);2136	guc_lrc_desc_pool_destroy_v69(guc);2137	i915_sched_engine_put(guc->sched_engine);2138	bitmap_free(guc->submission_state.guc_ids_bitmap);2139	fini_tlb_lookup(guc);2140	guc->submission_initialized = false;2141}2142 2143static inline void queue_request(struct i915_sched_engine *sched_engine,2144				 struct i915_request *rq,2145				 int prio)2146{2147	GEM_BUG_ON(!list_empty(&rq->sched.link));2148	list_add_tail(&rq->sched.link,2149		      i915_sched_lookup_priolist(sched_engine, prio));2150	set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);2151	tasklet_hi_schedule(&sched_engine->tasklet);2152}2153 2154static int guc_bypass_tasklet_submit(struct intel_guc *guc,2155				     struct i915_request *rq)2156{2157	int ret = 0;2158 2159	__i915_request_submit(rq);2160 2161	trace_i915_request_in(rq, 0);2162 2163	if (is_multi_lrc_rq(rq)) {2164		if (multi_lrc_submit(rq)) {2165			ret = guc_wq_item_append(guc, rq);2166			if (!ret)2167				ret = guc_add_request(guc, rq);2168		}2169	} else {2170		guc_set_lrc_tail(rq);2171		ret = guc_add_request(guc, rq);2172	}2173 2174	if (unlikely(ret == -EPIPE))2175		disable_submission(guc);2176 2177	return ret;2178}2179 2180static bool need_tasklet(struct intel_guc *guc, struct i915_request *rq)2181{2182	struct i915_sched_engine *sched_engine = rq->engine->sched_engine;2183	struct intel_context *ce = request_to_scheduling_context(rq);2184 2185	return submission_disabled(guc) || guc->stalled_request ||2186		!i915_sched_engine_is_empty(sched_engine) ||2187		!ctx_id_mapped(guc, ce->guc_id.id);2188}2189 2190static void guc_submit_request(struct i915_request *rq)2191{2192	struct i915_sched_engine *sched_engine = rq->engine->sched_engine;2193	struct intel_guc *guc = gt_to_guc(rq->engine->gt);2194	unsigned long flags;2195 2196	/* Will be called from irq-context when using foreign fences. */2197	spin_lock_irqsave(&sched_engine->lock, flags);2198 2199	if (need_tasklet(guc, rq))2200		queue_request(sched_engine, rq, rq_prio(rq));2201	else if (guc_bypass_tasklet_submit(guc, rq) == -EBUSY)2202		tasklet_hi_schedule(&sched_engine->tasklet);2203 2204	spin_unlock_irqrestore(&sched_engine->lock, flags);2205}2206 2207static int new_guc_id(struct intel_guc *guc, struct intel_context *ce)2208{2209	int ret;2210 2211	GEM_BUG_ON(intel_context_is_child(ce));2212 2213	if (intel_context_is_parent(ce))2214		ret = bitmap_find_free_region(guc->submission_state.guc_ids_bitmap,2215					      NUMBER_MULTI_LRC_GUC_ID(guc),2216					      order_base_2(ce->parallel.number_children2217							   + 1));2218	else2219		ret = ida_alloc_range(&guc->submission_state.guc_ids,2220				      NUMBER_MULTI_LRC_GUC_ID(guc),2221				      guc->submission_state.num_guc_ids - 1,2222				      GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);2223	if (unlikely(ret < 0))2224		return ret;2225 2226	if (!intel_context_is_parent(ce))2227		++guc->submission_state.guc_ids_in_use;2228 2229	ce->guc_id.id = ret;2230	return 0;2231}2232 2233static void __release_guc_id(struct intel_guc *guc, struct intel_context *ce)2234{2235	GEM_BUG_ON(intel_context_is_child(ce));2236 2237	if (!context_guc_id_invalid(ce)) {2238		if (intel_context_is_parent(ce)) {2239			bitmap_release_region(guc->submission_state.guc_ids_bitmap,2240					      ce->guc_id.id,2241					      order_base_2(ce->parallel.number_children2242							   + 1));2243		} else {2244			--guc->submission_state.guc_ids_in_use;2245			ida_free(&guc->submission_state.guc_ids,2246				 ce->guc_id.id);2247		}2248		clr_ctx_id_mapping(guc, ce->guc_id.id);2249		set_context_guc_id_invalid(ce);2250	}2251	if (!list_empty(&ce->guc_id.link))2252		list_del_init(&ce->guc_id.link);2253}2254 2255static void release_guc_id(struct intel_guc *guc, struct intel_context *ce)2256{2257	unsigned long flags;2258 2259	spin_lock_irqsave(&guc->submission_state.lock, flags);2260	__release_guc_id(guc, ce);2261	spin_unlock_irqrestore(&guc->submission_state.lock, flags);2262}2263 2264static int steal_guc_id(struct intel_guc *guc, struct intel_context *ce)2265{2266	struct intel_context *cn;2267 2268	lockdep_assert_held(&guc->submission_state.lock);2269	GEM_BUG_ON(intel_context_is_child(ce));2270	GEM_BUG_ON(intel_context_is_parent(ce));2271 2272	if (!list_empty(&guc->submission_state.guc_id_list)) {2273		cn = list_first_entry(&guc->submission_state.guc_id_list,2274				      struct intel_context,2275				      guc_id.link);2276 2277		GEM_BUG_ON(atomic_read(&cn->guc_id.ref));2278		GEM_BUG_ON(context_guc_id_invalid(cn));2279		GEM_BUG_ON(intel_context_is_child(cn));2280		GEM_BUG_ON(intel_context_is_parent(cn));2281 2282		list_del_init(&cn->guc_id.link);2283		ce->guc_id.id = cn->guc_id.id;2284 2285		spin_lock(&cn->guc_state.lock);2286		clr_context_registered(cn);2287		spin_unlock(&cn->guc_state.lock);2288 2289		set_context_guc_id_invalid(cn);2290 2291#ifdef CONFIG_DRM_I915_SELFTEST2292		guc->number_guc_id_stolen++;2293#endif2294 2295		return 0;2296	} else {2297		return -EAGAIN;2298	}2299}2300 2301static int assign_guc_id(struct intel_guc *guc, struct intel_context *ce)2302{2303	int ret;2304 2305	lockdep_assert_held(&guc->submission_state.lock);2306	GEM_BUG_ON(intel_context_is_child(ce));2307 2308	ret = new_guc_id(guc, ce);2309	if (unlikely(ret < 0)) {2310		if (intel_context_is_parent(ce))2311			return -ENOSPC;2312 2313		ret = steal_guc_id(guc, ce);2314		if (ret < 0)2315			return ret;2316	}2317 2318	if (intel_context_is_parent(ce)) {2319		struct intel_context *child;2320		int i = 1;2321 2322		for_each_child(ce, child)2323			child->guc_id.id = ce->guc_id.id + i++;2324	}2325 2326	return 0;2327}2328 2329#define PIN_GUC_ID_TRIES	42330static int pin_guc_id(struct intel_guc *guc, struct intel_context *ce)2331{2332	int ret = 0;2333	unsigned long flags, tries = PIN_GUC_ID_TRIES;2334 2335	GEM_BUG_ON(atomic_read(&ce->guc_id.ref));2336 2337try_again:2338	spin_lock_irqsave(&guc->submission_state.lock, flags);2339 2340	might_lock(&ce->guc_state.lock);2341 2342	if (context_guc_id_invalid(ce)) {2343		ret = assign_guc_id(guc, ce);2344		if (ret)2345			goto out_unlock;2346		ret = 1;	/* Indidcates newly assigned guc_id */2347	}2348	if (!list_empty(&ce->guc_id.link))2349		list_del_init(&ce->guc_id.link);2350	atomic_inc(&ce->guc_id.ref);2351 2352out_unlock:2353	spin_unlock_irqrestore(&guc->submission_state.lock, flags);2354 2355	/*2356	 * -EAGAIN indicates no guc_id are available, let's retire any2357	 * outstanding requests to see if that frees up a guc_id. If the first2358	 * retire didn't help, insert a sleep with the timeslice duration before2359	 * attempting to retire more requests. Double the sleep period each2360	 * subsequent pass before finally giving up. The sleep period has max of2361	 * 100ms and minimum of 1ms.2362	 */2363	if (ret == -EAGAIN && --tries) {2364		if (PIN_GUC_ID_TRIES - tries > 1) {2365			unsigned int timeslice_shifted =2366				ce->engine->props.timeslice_duration_ms <<2367				(PIN_GUC_ID_TRIES - tries - 2);2368			unsigned int max = min_t(unsigned int, 100,2369						 timeslice_shifted);2370 2371			msleep(max_t(unsigned int, max, 1));2372		}2373		intel_gt_retire_requests(guc_to_gt(guc));2374		goto try_again;2375	}2376 2377	return ret;2378}2379 2380static void unpin_guc_id(struct intel_guc *guc, struct intel_context *ce)2381{2382	unsigned long flags;2383 2384	GEM_BUG_ON(atomic_read(&ce->guc_id.ref) < 0);2385	GEM_BUG_ON(intel_context_is_child(ce));2386 2387	if (unlikely(context_guc_id_invalid(ce) ||2388		     intel_context_is_parent(ce)))2389		return;2390 2391	spin_lock_irqsave(&guc->submission_state.lock, flags);2392	if (!context_guc_id_invalid(ce) && list_empty(&ce->guc_id.link) &&2393	    !atomic_read(&ce->guc_id.ref))2394		list_add_tail(&ce->guc_id.link,2395			      &guc->submission_state.guc_id_list);2396	spin_unlock_irqrestore(&guc->submission_state.lock, flags);2397}2398 2399static int __guc_action_register_multi_lrc_v69(struct intel_guc *guc,2400					       struct intel_context *ce,2401					       u32 guc_id,2402					       u32 offset,2403					       bool loop)2404{2405	struct intel_context *child;2406	u32 action[4 + MAX_ENGINE_INSTANCE];2407	int len = 0;2408 2409	GEM_BUG_ON(ce->parallel.number_children > MAX_ENGINE_INSTANCE);2410 2411	action[len++] = INTEL_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC;2412	action[len++] = guc_id;2413	action[len++] = ce->parallel.number_children + 1;2414	action[len++] = offset;2415	for_each_child(ce, child) {2416		offset += sizeof(struct guc_lrc_desc_v69);2417		action[len++] = offset;2418	}2419 2420	return guc_submission_send_busy_loop(guc, action, len, 0, loop);2421}2422 2423static int __guc_action_register_multi_lrc_v70(struct intel_guc *guc,2424					       struct intel_context *ce,2425					       struct guc_ctxt_registration_info *info,2426					       bool loop)2427{2428	struct intel_context *child;2429	u32 action[13 + (MAX_ENGINE_INSTANCE * 2)];2430	int len = 0;2431	u32 next_id;2432 2433	GEM_BUG_ON(ce->parallel.number_children > MAX_ENGINE_INSTANCE);2434 2435	action[len++] = INTEL_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC;2436	action[len++] = info->flags;2437	action[len++] = info->context_idx;2438	action[len++] = info->engine_class;2439	action[len++] = info->engine_submit_mask;2440	action[len++] = info->wq_desc_lo;2441	action[len++] = info->wq_desc_hi;2442	action[len++] = info->wq_base_lo;2443	action[len++] = info->wq_base_hi;2444	action[len++] = info->wq_size;2445	action[len++] = ce->parallel.number_children + 1;2446	action[len++] = info->hwlrca_lo;2447	action[len++] = info->hwlrca_hi;2448 2449	next_id = info->context_idx + 1;2450	for_each_child(ce, child) {2451		GEM_BUG_ON(next_id++ != child->guc_id.id);2452 2453		/*2454		 * NB: GuC interface supports 64 bit LRCA even though i915/HW2455		 * only supports 32 bit currently.2456		 */2457		action[len++] = lower_32_bits(child->lrc.lrca);2458		action[len++] = upper_32_bits(child->lrc.lrca);2459	}2460 2461	GEM_BUG_ON(len > ARRAY_SIZE(action));2462 2463	return guc_submission_send_busy_loop(guc, action, len, 0, loop);2464}2465 2466static int __guc_action_register_context_v69(struct intel_guc *guc,2467					     u32 guc_id,2468					     u32 offset,2469					     bool loop)2470{2471	u32 action[] = {2472		INTEL_GUC_ACTION_REGISTER_CONTEXT,2473		guc_id,2474		offset,2475	};2476 2477	return guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),2478					     0, loop);2479}2480 2481static int __guc_action_register_context_v70(struct intel_guc *guc,2482					     struct guc_ctxt_registration_info *info,2483					     bool loop)2484{2485	u32 action[] = {2486		INTEL_GUC_ACTION_REGISTER_CONTEXT,2487		info->flags,2488		info->context_idx,2489		info->engine_class,2490		info->engine_submit_mask,2491		info->wq_desc_lo,2492		info->wq_desc_hi,2493		info->wq_base_lo,2494		info->wq_base_hi,2495		info->wq_size,2496		info->hwlrca_lo,2497		info->hwlrca_hi,2498	};2499 2500	return guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),2501					     0, loop);2502}2503 2504static void prepare_context_registration_info_v69(struct intel_context *ce);2505static void prepare_context_registration_info_v70(struct intel_context *ce,2506						  struct guc_ctxt_registration_info *info);2507 2508static int2509register_context_v69(struct intel_guc *guc, struct intel_context *ce, bool loop)2510{2511	u32 offset = intel_guc_ggtt_offset(guc, guc->lrc_desc_pool_v69) +2512		ce->guc_id.id * sizeof(struct guc_lrc_desc_v69);2513 2514	prepare_context_registration_info_v69(ce);2515 2516	if (intel_context_is_parent(ce))2517		return __guc_action_register_multi_lrc_v69(guc, ce, ce->guc_id.id,2518							   offset, loop);2519	else2520		return __guc_action_register_context_v69(guc, ce->guc_id.id,2521							 offset, loop);2522}2523 2524static int2525register_context_v70(struct intel_guc *guc, struct intel_context *ce, bool loop)2526{2527	struct guc_ctxt_registration_info info;2528 2529	prepare_context_registration_info_v70(ce, &info);2530 2531	if (intel_context_is_parent(ce))2532		return __guc_action_register_multi_lrc_v70(guc, ce, &info, loop);2533	else2534		return __guc_action_register_context_v70(guc, &info, loop);2535}2536 2537static int register_context(struct intel_context *ce, bool loop)2538{2539	struct intel_guc *guc = ce_to_guc(ce);2540	int ret;2541 2542	GEM_BUG_ON(intel_context_is_child(ce));2543	trace_intel_context_register(ce);2544 2545	if (GUC_SUBMIT_VER(guc) >= MAKE_GUC_VER(1, 0, 0))2546		ret = register_context_v70(guc, ce, loop);2547	else2548		ret = register_context_v69(guc, ce, loop);2549 2550	if (likely(!ret)) {2551		unsigned long flags;2552 2553		spin_lock_irqsave(&ce->guc_state.lock, flags);2554		set_context_registered(ce);2555		spin_unlock_irqrestore(&ce->guc_state.lock, flags);2556 2557		if (GUC_SUBMIT_VER(guc) >= MAKE_GUC_VER(1, 0, 0))2558			guc_context_policy_init_v70(ce, loop);2559	}2560 2561	return ret;2562}2563 2564static int __guc_action_deregister_context(struct intel_guc *guc,2565					   u32 guc_id)2566{2567	u32 action[] = {2568		INTEL_GUC_ACTION_DEREGISTER_CONTEXT,2569		guc_id,2570	};2571 2572	return guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),2573					     G2H_LEN_DW_DEREGISTER_CONTEXT,2574					     true);2575}2576 2577static int deregister_context(struct intel_context *ce, u32 guc_id)2578{2579	struct intel_guc *guc = ce_to_guc(ce);2580 2581	GEM_BUG_ON(intel_context_is_child(ce));2582	trace_intel_context_deregister(ce);2583 2584	return __guc_action_deregister_context(guc, guc_id);2585}2586 2587static inline void clear_children_join_go_memory(struct intel_context *ce)2588{2589	struct parent_scratch *ps = __get_parent_scratch(ce);2590	int i;2591 2592	ps->go.semaphore = 0;2593	for (i = 0; i < ce->parallel.number_children + 1; ++i)2594		ps->join[i].semaphore = 0;2595}2596 2597static inline u32 get_children_go_value(struct intel_context *ce)2598{2599	return __get_parent_scratch(ce)->go.semaphore;2600}2601 2602static inline u32 get_children_join_value(struct intel_context *ce,2603					  u8 child_index)2604{2605	return __get_parent_scratch(ce)->join[child_index].semaphore;2606}2607 2608struct context_policy {2609	u32 count;2610	struct guc_update_context_policy h2g;2611};2612 2613static u32 __guc_context_policy_action_size(struct context_policy *policy)2614{2615	size_t bytes = sizeof(policy->h2g.header) +2616		       (sizeof(policy->h2g.klv[0]) * policy->count);2617 2618	return bytes / sizeof(u32);2619}2620 2621static void __guc_context_policy_start_klv(struct context_policy *policy, u16 guc_id)2622{2623	policy->h2g.header.action = INTEL_GUC_ACTION_HOST2GUC_UPDATE_CONTEXT_POLICIES;2624	policy->h2g.header.ctx_id = guc_id;2625	policy->count = 0;2626}2627 2628#define MAKE_CONTEXT_POLICY_ADD(func, id) \2629static void __guc_context_policy_add_##func(struct context_policy *policy, u32 data) \2630{ \2631	GEM_BUG_ON(policy->count >= GUC_CONTEXT_POLICIES_KLV_NUM_IDS); \2632	policy->h2g.klv[policy->count].kl = \2633		FIELD_PREP(GUC_KLV_0_KEY, GUC_CONTEXT_POLICIES_KLV_ID_##id) | \2634		FIELD_PREP(GUC_KLV_0_LEN, 1); \2635	policy->h2g.klv[policy->count].value = data; \2636	policy->count++; \2637}2638 2639MAKE_CONTEXT_POLICY_ADD(execution_quantum, EXECUTION_QUANTUM)2640MAKE_CONTEXT_POLICY_ADD(preemption_timeout, PREEMPTION_TIMEOUT)2641MAKE_CONTEXT_POLICY_ADD(priority, SCHEDULING_PRIORITY)2642MAKE_CONTEXT_POLICY_ADD(preempt_to_idle, PREEMPT_TO_IDLE_ON_QUANTUM_EXPIRY)2643MAKE_CONTEXT_POLICY_ADD(slpc_ctx_freq_req, SLPM_GT_FREQUENCY)2644 2645#undef MAKE_CONTEXT_POLICY_ADD2646 2647static int __guc_context_set_context_policies(struct intel_guc *guc,2648					      struct context_policy *policy,2649					      bool loop)2650{2651	return guc_submission_send_busy_loop(guc, (u32 *)&policy->h2g,2652					__guc_context_policy_action_size(policy),2653					0, loop);2654}2655 2656static int guc_context_policy_init_v70(struct intel_context *ce, bool loop)2657{2658	struct intel_engine_cs *engine = ce->engine;2659	struct intel_guc *guc = gt_to_guc(engine->gt);2660	struct context_policy policy;2661	u32 execution_quantum;2662	u32 preemption_timeout;2663	u32 slpc_ctx_freq_req = 0;2664	unsigned long flags;2665	int ret;2666 2667	/* NB: For both of these, zero means disabled. */2668	GEM_BUG_ON(overflows_type(engine->props.timeslice_duration_ms * 1000,2669				  execution_quantum));2670	GEM_BUG_ON(overflows_type(engine->props.preempt_timeout_ms * 1000,2671				  preemption_timeout));2672	execution_quantum = engine->props.timeslice_duration_ms * 1000;2673	preemption_timeout = engine->props.preempt_timeout_ms * 1000;2674 2675	if (ce->flags & BIT(CONTEXT_LOW_LATENCY))2676		slpc_ctx_freq_req |= SLPC_CTX_FREQ_REQ_IS_COMPUTE;2677 2678	__guc_context_policy_start_klv(&policy, ce->guc_id.id);2679 2680	__guc_context_policy_add_priority(&policy, ce->guc_state.prio);2681	__guc_context_policy_add_execution_quantum(&policy, execution_quantum);2682	__guc_context_policy_add_preemption_timeout(&policy, preemption_timeout);2683	__guc_context_policy_add_slpc_ctx_freq_req(&policy, slpc_ctx_freq_req);2684 2685	if (engine->flags & I915_ENGINE_WANT_FORCED_PREEMPTION)2686		__guc_context_policy_add_preempt_to_idle(&policy, 1);2687 2688	ret = __guc_context_set_context_policies(guc, &policy, loop);2689 2690	spin_lock_irqsave(&ce->guc_state.lock, flags);2691	if (ret != 0)2692		set_context_policy_required(ce);2693	else2694		clr_context_policy_required(ce);2695	spin_unlock_irqrestore(&ce->guc_state.lock, flags);2696 2697	return ret;2698}2699 2700static void guc_context_policy_init_v69(struct intel_engine_cs *engine,2701					struct guc_lrc_desc_v69 *desc)2702{2703	desc->policy_flags = 0;2704 2705	if (engine->flags & I915_ENGINE_WANT_FORCED_PREEMPTION)2706		desc->policy_flags |= CONTEXT_POLICY_FLAG_PREEMPT_TO_IDLE_V69;2707 2708	/* NB: For both of these, zero means disabled. */2709	GEM_BUG_ON(overflows_type(engine->props.timeslice_duration_ms * 1000,2710				  desc->execution_quantum));2711	GEM_BUG_ON(overflows_type(engine->props.preempt_timeout_ms * 1000,2712				  desc->preemption_timeout));2713	desc->execution_quantum = engine->props.timeslice_duration_ms * 1000;2714	desc->preemption_timeout = engine->props.preempt_timeout_ms * 1000;2715}2716 2717static u32 map_guc_prio_to_lrc_desc_prio(u8 prio)2718{2719	/*2720	 * this matches the mapping we do in map_i915_prio_to_guc_prio()2721	 * (e.g. prio < I915_PRIORITY_NORMAL maps to GUC_CLIENT_PRIORITY_NORMAL)2722	 */2723	switch (prio) {2724	default:2725		MISSING_CASE(prio);2726		fallthrough;2727	case GUC_CLIENT_PRIORITY_KMD_NORMAL:2728		return GEN12_CTX_PRIORITY_NORMAL;2729	case GUC_CLIENT_PRIORITY_NORMAL:2730		return GEN12_CTX_PRIORITY_LOW;2731	case GUC_CLIENT_PRIORITY_HIGH:2732	case GUC_CLIENT_PRIORITY_KMD_HIGH:2733		return GEN12_CTX_PRIORITY_HIGH;2734	}2735}2736 2737static void prepare_context_registration_info_v69(struct intel_context *ce)2738{2739	struct intel_engine_cs *engine = ce->engine;2740	struct intel_guc *guc = gt_to_guc(engine->gt);2741	u32 ctx_id = ce->guc_id.id;2742	struct guc_lrc_desc_v69 *desc;2743	struct intel_context *child;2744 2745	GEM_BUG_ON(!engine->mask);2746 2747	/*2748	 * Ensure LRC + CT vmas are is same region as write barrier is done2749	 * based on CT vma region.2750	 */2751	GEM_BUG_ON(i915_gem_object_is_lmem(guc->ct.vma->obj) !=2752		   i915_gem_object_is_lmem(ce->ring->vma->obj));2753 2754	desc = __get_lrc_desc_v69(guc, ctx_id);2755	GEM_BUG_ON(!desc);2756	desc->engine_class = engine_class_to_guc_class(engine->class);2757	desc->engine_submit_mask = engine->logical_mask;2758	desc->hw_context_desc = ce->lrc.lrca;2759	desc->priority = ce->guc_state.prio;2760	desc->context_flags = CONTEXT_REGISTRATION_FLAG_KMD;2761	guc_context_policy_init_v69(engine, desc);2762 2763	/*2764	 * If context is a parent, we need to register a process descriptor2765	 * describing a work queue and register all child contexts.2766	 */2767	if (intel_context_is_parent(ce)) {2768		struct guc_process_desc_v69 *pdesc;2769 2770		ce->parallel.guc.wqi_tail = 0;2771		ce->parallel.guc.wqi_head = 0;2772 2773		desc->process_desc = i915_ggtt_offset(ce->state) +2774			__get_parent_scratch_offset(ce);2775		desc->wq_addr = i915_ggtt_offset(ce->state) +2776			__get_wq_offset(ce);2777		desc->wq_size = WQ_SIZE;2778 2779		pdesc = __get_process_desc_v69(ce);2780		memset(pdesc, 0, sizeof(*(pdesc)));2781		pdesc->stage_id = ce->guc_id.id;2782		pdesc->wq_base_addr = desc->wq_addr;2783		pdesc->wq_size_bytes = desc->wq_size;2784		pdesc->wq_status = WQ_STATUS_ACTIVE;2785 2786		ce->parallel.guc.wq_head = &pdesc->head;2787		ce->parallel.guc.wq_tail = &pdesc->tail;2788		ce->parallel.guc.wq_status = &pdesc->wq_status;2789 2790		for_each_child(ce, child) {2791			desc = __get_lrc_desc_v69(guc, child->guc_id.id);2792 2793			desc->engine_class =2794				engine_class_to_guc_class(engine->class);2795			desc->hw_context_desc = child->lrc.lrca;2796			desc->priority = ce->guc_state.prio;2797			desc->context_flags = CONTEXT_REGISTRATION_FLAG_KMD;2798			guc_context_policy_init_v69(engine, desc);2799		}2800 2801		clear_children_join_go_memory(ce);2802	}2803}2804 2805static void prepare_context_registration_info_v70(struct intel_context *ce,2806						  struct guc_ctxt_registration_info *info)2807{2808	struct intel_engine_cs *engine = ce->engine;2809	struct intel_guc *guc = gt_to_guc(engine->gt);2810	u32 ctx_id = ce->guc_id.id;2811 2812	GEM_BUG_ON(!engine->mask);2813 2814	/*2815	 * Ensure LRC + CT vmas are is same region as write barrier is done2816	 * based on CT vma region.2817	 */2818	GEM_BUG_ON(i915_gem_object_is_lmem(guc->ct.vma->obj) !=2819		   i915_gem_object_is_lmem(ce->ring->vma->obj));2820 2821	memset(info, 0, sizeof(*info));2822	info->context_idx = ctx_id;2823	info->engine_class = engine_class_to_guc_class(engine->class);2824	info->engine_submit_mask = engine->logical_mask;2825	/*2826	 * NB: GuC interface supports 64 bit LRCA even though i915/HW2827	 * only supports 32 bit currently.2828	 */2829	info->hwlrca_lo = lower_32_bits(ce->lrc.lrca);2830	info->hwlrca_hi = upper_32_bits(ce->lrc.lrca);2831	if (engine->flags & I915_ENGINE_HAS_EU_PRIORITY)2832		info->hwlrca_lo |= map_guc_prio_to_lrc_desc_prio(ce->guc_state.prio);2833	info->flags = CONTEXT_REGISTRATION_FLAG_KMD;2834 2835	/*2836	 * If context is a parent, we need to register a process descriptor2837	 * describing a work queue and register all child contexts.2838	 */2839	if (intel_context_is_parent(ce)) {2840		struct guc_sched_wq_desc *wq_desc;2841		u64 wq_desc_offset, wq_base_offset;2842 2843		ce->parallel.guc.wqi_tail = 0;2844		ce->parallel.guc.wqi_head = 0;2845 2846		wq_desc_offset = (u64)i915_ggtt_offset(ce->state) +2847				 __get_parent_scratch_offset(ce);2848		wq_base_offset = (u64)i915_ggtt_offset(ce->state) +2849				 __get_wq_offset(ce);2850		info->wq_desc_lo = lower_32_bits(wq_desc_offset);2851		info->wq_desc_hi = upper_32_bits(wq_desc_offset);2852		info->wq_base_lo = lower_32_bits(wq_base_offset);2853		info->wq_base_hi = upper_32_bits(wq_base_offset);2854		info->wq_size = WQ_SIZE;2855 2856		wq_desc = __get_wq_desc_v70(ce);2857		memset(wq_desc, 0, sizeof(*wq_desc));2858		wq_desc->wq_status = WQ_STATUS_ACTIVE;2859 2860		ce->parallel.guc.wq_head = &wq_desc->head;2861		ce->parallel.guc.wq_tail = &wq_desc->tail;2862		ce->parallel.guc.wq_status = &wq_desc->wq_status;2863 2864		clear_children_join_go_memory(ce);2865	}2866}2867 2868static int try_context_registration(struct intel_context *ce, bool loop)2869{2870	struct intel_engine_cs *engine = ce->engine;2871	struct intel_runtime_pm *runtime_pm = engine->uncore->rpm;2872	struct intel_guc *guc = gt_to_guc(engine->gt);2873	intel_wakeref_t wakeref;2874	u32 ctx_id = ce->guc_id.id;2875	bool context_registered;2876	int ret = 0;2877 2878	GEM_BUG_ON(!sched_state_is_init(ce));2879 2880	context_registered = ctx_id_mapped(guc, ctx_id);2881 2882	clr_ctx_id_mapping(guc, ctx_id);2883	set_ctx_id_mapping(guc, ctx_id, ce);2884 2885	/*2886	 * The context_lookup xarray is used to determine if the hardware2887	 * context is currently registered. There are two cases in which it2888	 * could be registered either the guc_id has been stolen from another2889	 * context or the lrc descriptor address of this context has changed. In2890	 * either case the context needs to be deregistered with the GuC before2891	 * registering this context.2892	 */2893	if (context_registered) {2894		bool disabled;2895		unsigned long flags;2896 2897		trace_intel_context_steal_guc_id(ce);2898		GEM_BUG_ON(!loop);2899 2900		/* Seal race with Reset */2901		spin_lock_irqsave(&ce->guc_state.lock, flags);2902		disabled = submission_disabled(guc);2903		if (likely(!disabled)) {2904			set_context_wait_for_deregister_to_register(ce);2905			intel_context_get(ce);2906		}2907		spin_unlock_irqrestore(&ce->guc_state.lock, flags);2908		if (unlikely(disabled)) {2909			clr_ctx_id_mapping(guc, ctx_id);2910			return 0;	/* Will get registered later */2911		}2912 2913		/*2914		 * If stealing the guc_id, this ce has the same guc_id as the2915		 * context whose guc_id was stolen.2916		 */2917		with_intel_runtime_pm(runtime_pm, wakeref)2918			ret = deregister_context(ce, ce->guc_id.id);2919		if (unlikely(ret == -ENODEV))2920			ret = 0;	/* Will get registered later */2921	} else {2922		with_intel_runtime_pm(runtime_pm, wakeref)2923			ret = register_context(ce, loop);2924		if (unlikely(ret == -EBUSY)) {2925			clr_ctx_id_mapping(guc, ctx_id);2926		} else if (unlikely(ret == -ENODEV)) {2927			clr_ctx_id_mapping(guc, ctx_id);2928			ret = 0;	/* Will get registered later */2929		}2930	}2931 2932	return ret;2933}2934 2935static int __guc_context_pre_pin(struct intel_context *ce,2936				 struct intel_engine_cs *engine,2937				 struct i915_gem_ww_ctx *ww,2938				 void **vaddr)2939{2940	return lrc_pre_pin(ce, engine, ww, vaddr);2941}2942 2943static int __guc_context_pin(struct intel_context *ce,2944			     struct intel_engine_cs *engine,2945			     void *vaddr)2946{2947	if (i915_ggtt_offset(ce->state) !=2948	    (ce->lrc.lrca & CTX_GTT_ADDRESS_MASK))2949		set_bit(CONTEXT_LRCA_DIRTY, &ce->flags);2950 2951	/*2952	 * GuC context gets pinned in guc_request_alloc. See that function for2953	 * explaination of why.2954	 */2955 2956	return lrc_pin(ce, engine, vaddr);2957}2958 2959static int guc_context_pre_pin(struct intel_context *ce,2960			       struct i915_gem_ww_ctx *ww,2961			       void **vaddr)2962{2963	return __guc_context_pre_pin(ce, ce->engine, ww, vaddr);2964}2965 2966static int guc_context_pin(struct intel_context *ce, void *vaddr)2967{2968	int ret = __guc_context_pin(ce, ce->engine, vaddr);2969 2970	if (likely(!ret && !intel_context_is_barrier(ce)))2971		intel_engine_pm_get(ce->engine);2972 2973	return ret;2974}2975 2976static void guc_context_unpin(struct intel_context *ce)2977{2978	struct intel_guc *guc = ce_to_guc(ce);2979 2980	__guc_context_update_stats(ce);2981	unpin_guc_id(guc, ce);2982	lrc_unpin(ce);2983 2984	if (likely(!intel_context_is_barrier(ce)))2985		intel_engine_pm_put_async(ce->engine);2986}2987 2988static void guc_context_post_unpin(struct intel_context *ce)2989{2990	lrc_post_unpin(ce);2991}2992 2993static void __guc_context_sched_enable(struct intel_guc *guc,2994				       struct intel_context *ce)2995{2996	u32 action[] = {2997		INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET,2998		ce->guc_id.id,2999		GUC_CONTEXT_ENABLE3000	};3001 3002	trace_intel_context_sched_enable(ce);3003 3004	guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),3005				      G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, true);3006}3007 3008static void __guc_context_sched_disable(struct intel_guc *guc,3009					struct intel_context *ce,3010					u16 guc_id)3011{3012	u32 action[] = {3013		INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET,3014		guc_id,	/* ce->guc_id.id not stable */3015		GUC_CONTEXT_DISABLE3016	};3017 3018	GEM_BUG_ON(guc_id == GUC_INVALID_CONTEXT_ID);3019 3020	GEM_BUG_ON(intel_context_is_child(ce));3021	trace_intel_context_sched_disable(ce);3022 3023	guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),3024				      G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, true);3025}3026 3027static void guc_blocked_fence_complete(struct intel_context *ce)3028{3029	lockdep_assert_held(&ce->guc_state.lock);3030 3031	if (!i915_sw_fence_done(&ce->guc_state.blocked))3032		i915_sw_fence_complete(&ce->guc_state.blocked);3033}3034 3035static void guc_blocked_fence_reinit(struct intel_context *ce)3036{3037	lockdep_assert_held(&ce->guc_state.lock);3038	GEM_BUG_ON(!i915_sw_fence_done(&ce->guc_state.blocked));3039 3040	/*3041	 * This fence is always complete unless a pending schedule disable is3042	 * outstanding. We arm the fence here and complete it when we receive3043	 * the pending schedule disable complete message.3044	 */3045	i915_sw_fence_fini(&ce->guc_state.blocked);3046	i915_sw_fence_reinit(&ce->guc_state.blocked);3047	i915_sw_fence_await(&ce->guc_state.blocked);3048	i915_sw_fence_commit(&ce->guc_state.blocked);3049}3050 3051static u16 prep_context_pending_disable(struct intel_context *ce)3052{3053	lockdep_assert_held(&ce->guc_state.lock);3054 3055	set_context_pending_disable(ce);3056	clr_context_enabled(ce);3057	guc_blocked_fence_reinit(ce);3058	intel_context_get(ce);3059 3060	return ce->guc_id.id;3061}3062 3063static struct i915_sw_fence *guc_context_block(struct intel_context *ce)3064{3065	struct intel_guc *guc = ce_to_guc(ce);3066	unsigned long flags;3067	struct intel_runtime_pm *runtime_pm = ce->engine->uncore->rpm;3068	intel_wakeref_t wakeref;3069	u16 guc_id;3070	bool enabled;3071 3072	GEM_BUG_ON(intel_context_is_child(ce));3073 3074	spin_lock_irqsave(&ce->guc_state.lock, flags);3075 3076	incr_context_blocked(ce);3077 3078	enabled = context_enabled(ce);3079	if (unlikely(!enabled || submission_disabled(guc))) {3080		if (enabled)3081			clr_context_enabled(ce);3082		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3083		return &ce->guc_state.blocked;3084	}3085 3086	/*3087	 * We add +2 here as the schedule disable complete CTB handler calls3088	 * intel_context_sched_disable_unpin (-2 to pin_count).3089	 */3090	atomic_add(2, &ce->pin_count);3091 3092	guc_id = prep_context_pending_disable(ce);3093 3094	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3095 3096	with_intel_runtime_pm(runtime_pm, wakeref)3097		__guc_context_sched_disable(guc, ce, guc_id);3098 3099	return &ce->guc_state.blocked;3100}3101 3102#define SCHED_STATE_MULTI_BLOCKED_MASK \3103	(SCHED_STATE_BLOCKED_MASK & ~SCHED_STATE_BLOCKED)3104#define SCHED_STATE_NO_UNBLOCK \3105	(SCHED_STATE_MULTI_BLOCKED_MASK | \3106	 SCHED_STATE_PENDING_DISABLE | \3107	 SCHED_STATE_BANNED)3108 3109static bool context_cant_unblock(struct intel_context *ce)3110{3111	lockdep_assert_held(&ce->guc_state.lock);3112 3113	return (ce->guc_state.sched_state & SCHED_STATE_NO_UNBLOCK) ||3114		context_guc_id_invalid(ce) ||3115		!ctx_id_mapped(ce_to_guc(ce), ce->guc_id.id) ||3116		!intel_context_is_pinned(ce);3117}3118 3119static void guc_context_unblock(struct intel_context *ce)3120{3121	struct intel_guc *guc = ce_to_guc(ce);3122	unsigned long flags;3123	struct intel_runtime_pm *runtime_pm = ce->engine->uncore->rpm;3124	intel_wakeref_t wakeref;3125	bool enable;3126 3127	GEM_BUG_ON(context_enabled(ce));3128	GEM_BUG_ON(intel_context_is_child(ce));3129 3130	spin_lock_irqsave(&ce->guc_state.lock, flags);3131 3132	if (unlikely(submission_disabled(guc) ||3133		     context_cant_unblock(ce))) {3134		enable = false;3135	} else {3136		enable = true;3137		set_context_pending_enable(ce);3138		set_context_enabled(ce);3139		intel_context_get(ce);3140	}3141 3142	decr_context_blocked(ce);3143 3144	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3145 3146	if (enable) {3147		with_intel_runtime_pm(runtime_pm, wakeref)3148			__guc_context_sched_enable(guc, ce);3149	}3150}3151 3152static void guc_context_cancel_request(struct intel_context *ce,3153				       struct i915_request *rq)3154{3155	struct intel_context *block_context =3156		request_to_scheduling_context(rq);3157 3158	if (i915_sw_fence_signaled(&rq->submit)) {3159		struct i915_sw_fence *fence;3160 3161		intel_context_get(ce);3162		fence = guc_context_block(block_context);3163		i915_sw_fence_wait(fence);3164		if (!i915_request_completed(rq)) {3165			__i915_request_skip(rq);3166			guc_reset_state(ce, intel_ring_wrap(ce->ring, rq->head),3167					true);3168		}3169 3170		guc_context_unblock(block_context);3171		intel_context_put(ce);3172	}3173}3174 3175static void __guc_context_set_preemption_timeout(struct intel_guc *guc,3176						 u16 guc_id,3177						 u32 preemption_timeout)3178{3179	if (GUC_SUBMIT_VER(guc) >= MAKE_GUC_VER(1, 0, 0)) {3180		struct context_policy policy;3181 3182		__guc_context_policy_start_klv(&policy, guc_id);3183		__guc_context_policy_add_preemption_timeout(&policy, preemption_timeout);3184		__guc_context_set_context_policies(guc, &policy, true);3185	} else {3186		u32 action[] = {3187			INTEL_GUC_ACTION_V69_SET_CONTEXT_PREEMPTION_TIMEOUT,3188			guc_id,3189			preemption_timeout3190		};3191 3192		intel_guc_send_busy_loop(guc, action, ARRAY_SIZE(action), 0, true);3193	}3194}3195 3196static void3197guc_context_revoke(struct intel_context *ce, struct i915_request *rq,3198		   unsigned int preempt_timeout_ms)3199{3200	struct intel_guc *guc = ce_to_guc(ce);3201	struct intel_runtime_pm *runtime_pm =3202		&ce->engine->gt->i915->runtime_pm;3203	intel_wakeref_t wakeref;3204	unsigned long flags;3205 3206	GEM_BUG_ON(intel_context_is_child(ce));3207 3208	guc_flush_submissions(guc);3209 3210	spin_lock_irqsave(&ce->guc_state.lock, flags);3211	set_context_banned(ce);3212 3213	if (submission_disabled(guc) ||3214	    (!context_enabled(ce) && !context_pending_disable(ce))) {3215		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3216 3217		guc_cancel_context_requests(ce);3218		intel_engine_signal_breadcrumbs(ce->engine);3219	} else if (!context_pending_disable(ce)) {3220		u16 guc_id;3221 3222		/*3223		 * We add +2 here as the schedule disable complete CTB handler3224		 * calls intel_context_sched_disable_unpin (-2 to pin_count).3225		 */3226		atomic_add(2, &ce->pin_count);3227 3228		guc_id = prep_context_pending_disable(ce);3229		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3230 3231		/*3232		 * In addition to disabling scheduling, set the preemption3233		 * timeout to the minimum value (1 us) so the banned context3234		 * gets kicked off the HW ASAP.3235		 */3236		with_intel_runtime_pm(runtime_pm, wakeref) {3237			__guc_context_set_preemption_timeout(guc, guc_id,3238							     preempt_timeout_ms);3239			__guc_context_sched_disable(guc, ce, guc_id);3240		}3241	} else {3242		if (!context_guc_id_invalid(ce))3243			with_intel_runtime_pm(runtime_pm, wakeref)3244				__guc_context_set_preemption_timeout(guc,3245								     ce->guc_id.id,3246								     preempt_timeout_ms);3247		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3248	}3249}3250 3251static void do_sched_disable(struct intel_guc *guc, struct intel_context *ce,3252			     unsigned long flags)3253	__releases(ce->guc_state.lock)3254{3255	struct intel_runtime_pm *runtime_pm = &ce->engine->gt->i915->runtime_pm;3256	intel_wakeref_t wakeref;3257	u16 guc_id;3258 3259	lockdep_assert_held(&ce->guc_state.lock);3260	guc_id = prep_context_pending_disable(ce);3261 3262	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3263 3264	with_intel_runtime_pm(runtime_pm, wakeref)3265		__guc_context_sched_disable(guc, ce, guc_id);3266}3267 3268static bool bypass_sched_disable(struct intel_guc *guc,3269				 struct intel_context *ce)3270{3271	lockdep_assert_held(&ce->guc_state.lock);3272	GEM_BUG_ON(intel_context_is_child(ce));3273 3274	if (submission_disabled(guc) || context_guc_id_invalid(ce) ||3275	    !ctx_id_mapped(guc, ce->guc_id.id)) {3276		clr_context_enabled(ce);3277		return true;3278	}3279 3280	return !context_enabled(ce);3281}3282 3283static void __delay_sched_disable(struct work_struct *wrk)3284{3285	struct intel_context *ce =3286		container_of(wrk, typeof(*ce), guc_state.sched_disable_delay_work.work);3287	struct intel_guc *guc = ce_to_guc(ce);3288	unsigned long flags;3289 3290	spin_lock_irqsave(&ce->guc_state.lock, flags);3291 3292	if (bypass_sched_disable(guc, ce)) {3293		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3294		intel_context_sched_disable_unpin(ce);3295	} else {3296		do_sched_disable(guc, ce, flags);3297	}3298}3299 3300static bool guc_id_pressure(struct intel_guc *guc, struct intel_context *ce)3301{3302	/*3303	 * parent contexts are perma-pinned, if we are unpinning do schedule3304	 * disable immediately.3305	 */3306	if (intel_context_is_parent(ce))3307		return true;3308 3309	/*3310	 * If we are beyond the threshold for avail guc_ids, do schedule disable immediately.3311	 */3312	return guc->submission_state.guc_ids_in_use >3313		guc->submission_state.sched_disable_gucid_threshold;3314}3315 3316static void guc_context_sched_disable(struct intel_context *ce)3317{3318	struct intel_guc *guc = ce_to_guc(ce);3319	u64 delay = guc->submission_state.sched_disable_delay_ms;3320	unsigned long flags;3321 3322	spin_lock_irqsave(&ce->guc_state.lock, flags);3323 3324	if (bypass_sched_disable(guc, ce)) {3325		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3326		intel_context_sched_disable_unpin(ce);3327	} else if (!intel_context_is_closed(ce) && !guc_id_pressure(guc, ce) &&3328		   delay) {3329		spin_unlock_irqrestore(&ce->guc_state.lock, flags);3330		mod_delayed_work(system_unbound_wq,3331				 &ce->guc_state.sched_disable_delay_work,3332				 msecs_to_jiffies(delay));3333	} else {3334		do_sched_disable(guc, ce, flags);3335	}3336}3337 3338static void guc_context_close(struct intel_context *ce)3339{3340	unsigned long flags;3341 3342	if (test_bit(CONTEXT_GUC_INIT, &ce->flags) &&3343	    cancel_delayed_work(&ce->guc_state.sched_disable_delay_work))3344		__delay_sched_disable(&ce->guc_state.sched_disable_delay_work.work);3345 3346	spin_lock_irqsave(&ce->guc_state.lock, flags);3347	set_context_close_done(ce);3348	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3349}3350 3351static inline int guc_lrc_desc_unpin(struct intel_context *ce)3352{3353	struct intel_guc *guc = ce_to_guc(ce);3354	struct intel_gt *gt = guc_to_gt(guc);3355	unsigned long flags;3356	bool disabled;3357	int ret;3358 3359	GEM_BUG_ON(!intel_gt_pm_is_awake(gt));3360	GEM_BUG_ON(!ctx_id_mapped(guc, ce->guc_id.id));3361	GEM_BUG_ON(ce != __get_context(guc, ce->guc_id.id));3362	GEM_BUG_ON(context_enabled(ce));3363 3364	/* Seal race with Reset */3365	spin_lock_irqsave(&ce->guc_state.lock, flags);3366	disabled = submission_disabled(guc);3367	if (likely(!disabled)) {3368		/*3369		 * Take a gt-pm ref and change context state to be destroyed.3370		 * NOTE: a G2H IRQ that comes after will put this gt-pm ref back3371		 */3372		__intel_gt_pm_get(gt);3373		set_context_destroyed(ce);3374		clr_context_registered(ce);3375	}3376	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3377 3378	if (unlikely(disabled)) {3379		release_guc_id(guc, ce);3380		__guc_context_destroy(ce);3381		return 0;3382	}3383 3384	/*3385	 * GuC is active, lets destroy this context, but at this point we can still be racing3386	 * with suspend, so we undo everything if the H2G fails in deregister_context so3387	 * that GuC reset will find this context during clean up.3388	 */3389	ret = deregister_context(ce, ce->guc_id.id);3390	if (ret) {3391		spin_lock(&ce->guc_state.lock);3392		set_context_registered(ce);3393		clr_context_destroyed(ce);3394		spin_unlock(&ce->guc_state.lock);3395		/*3396		 * As gt-pm is awake at function entry, intel_wakeref_put_async merely decrements3397		 * the wakeref immediately but per function spec usage call this after unlock.3398		 */3399		intel_wakeref_put_async(&gt->wakeref);3400	}3401 3402	return ret;3403}3404 3405static void __guc_context_destroy(struct intel_context *ce)3406{3407	GEM_BUG_ON(ce->guc_state.prio_count[GUC_CLIENT_PRIORITY_KMD_HIGH] ||3408		   ce->guc_state.prio_count[GUC_CLIENT_PRIORITY_HIGH] ||3409		   ce->guc_state.prio_count[GUC_CLIENT_PRIORITY_KMD_NORMAL] ||3410		   ce->guc_state.prio_count[GUC_CLIENT_PRIORITY_NORMAL]);3411 3412	lrc_fini(ce);3413	intel_context_fini(ce);3414 3415	if (intel_engine_is_virtual(ce->engine)) {3416		struct guc_virtual_engine *ve =3417			container_of(ce, typeof(*ve), context);3418 3419		if (ve->base.breadcrumbs)3420			intel_breadcrumbs_put(ve->base.breadcrumbs);3421 3422		kfree(ve);3423	} else {3424		intel_context_free(ce);3425	}3426}3427 3428static void guc_flush_destroyed_contexts(struct intel_guc *guc)3429{3430	struct intel_context *ce;3431	unsigned long flags;3432 3433	GEM_BUG_ON(!submission_disabled(guc) &&3434		   guc_submission_initialized(guc));3435 3436	while (!list_empty(&guc->submission_state.destroyed_contexts)) {3437		spin_lock_irqsave(&guc->submission_state.lock, flags);3438		ce = list_first_entry_or_null(&guc->submission_state.destroyed_contexts,3439					      struct intel_context,3440					      destroyed_link);3441		if (ce)3442			list_del_init(&ce->destroyed_link);3443		spin_unlock_irqrestore(&guc->submission_state.lock, flags);3444 3445		if (!ce)3446			break;3447 3448		release_guc_id(guc, ce);3449		__guc_context_destroy(ce);3450	}3451}3452 3453static void deregister_destroyed_contexts(struct intel_guc *guc)3454{3455	struct intel_context *ce;3456	unsigned long flags;3457 3458	while (!list_empty(&guc->submission_state.destroyed_contexts)) {3459		spin_lock_irqsave(&guc->submission_state.lock, flags);3460		ce = list_first_entry_or_null(&guc->submission_state.destroyed_contexts,3461					      struct intel_context,3462					      destroyed_link);3463		if (ce)3464			list_del_init(&ce->destroyed_link);3465		spin_unlock_irqrestore(&guc->submission_state.lock, flags);3466 3467		if (!ce)3468			break;3469 3470		if (guc_lrc_desc_unpin(ce)) {3471			/*3472			 * This means GuC's CT link severed mid-way which could happen3473			 * in suspend-resume corner cases. In this case, put the3474			 * context back into the destroyed_contexts list which will3475			 * get picked up on the next context deregistration event or3476			 * purged in a GuC sanitization event (reset/unload/wedged/...).3477			 */3478			spin_lock_irqsave(&guc->submission_state.lock, flags);3479			list_add_tail(&ce->destroyed_link,3480				      &guc->submission_state.destroyed_contexts);3481			spin_unlock_irqrestore(&guc->submission_state.lock, flags);3482			/* Bail now since the list might never be emptied if h2gs fail */3483			break;3484		}3485 3486	}3487}3488 3489static void destroyed_worker_func(struct work_struct *w)3490{3491	struct intel_guc *guc = container_of(w, struct intel_guc,3492					     submission_state.destroyed_worker);3493	struct intel_gt *gt = guc_to_gt(guc);3494	intel_wakeref_t wakeref;3495 3496	/*3497	 * In rare cases we can get here via async context-free fence-signals that3498	 * come very late in suspend flow or very early in resume flows. In these3499	 * cases, GuC won't be ready but just skipping it here is fine as these3500	 * pending-destroy-contexts get destroyed totally at GuC reset time at the3501	 * end of suspend.. OR.. this worker can be picked up later on the next3502	 * context destruction trigger after resume-completes3503	 */3504	if (!intel_guc_is_ready(guc))3505		return;3506 3507	with_intel_gt_pm(gt, wakeref)3508		deregister_destroyed_contexts(guc);3509}3510 3511static void guc_context_destroy(struct kref *kref)3512{3513	struct intel_context *ce = container_of(kref, typeof(*ce), ref);3514	struct intel_guc *guc = ce_to_guc(ce);3515	unsigned long flags;3516	bool destroy;3517 3518	/*3519	 * If the guc_id is invalid this context has been stolen and we can free3520	 * it immediately. Also can be freed immediately if the context is not3521	 * registered with the GuC or the GuC is in the middle of a reset.3522	 */3523	spin_lock_irqsave(&guc->submission_state.lock, flags);3524	destroy = submission_disabled(guc) || context_guc_id_invalid(ce) ||3525		!ctx_id_mapped(guc, ce->guc_id.id);3526	if (likely(!destroy)) {3527		if (!list_empty(&ce->guc_id.link))3528			list_del_init(&ce->guc_id.link);3529		list_add_tail(&ce->destroyed_link,3530			      &guc->submission_state.destroyed_contexts);3531	} else {3532		__release_guc_id(guc, ce);3533	}3534	spin_unlock_irqrestore(&guc->submission_state.lock, flags);3535	if (unlikely(destroy)) {3536		__guc_context_destroy(ce);3537		return;3538	}3539 3540	/*3541	 * We use a worker to issue the H2G to deregister the context as we can3542	 * take the GT PM for the first time which isn't allowed from an atomic3543	 * context.3544	 */3545	queue_work(system_unbound_wq, &guc->submission_state.destroyed_worker);3546}3547 3548static int guc_context_alloc(struct intel_context *ce)3549{3550	return lrc_alloc(ce, ce->engine);3551}3552 3553static void __guc_context_set_prio(struct intel_guc *guc,3554				   struct intel_context *ce)3555{3556	if (GUC_SUBMIT_VER(guc) >= MAKE_GUC_VER(1, 0, 0)) {3557		struct context_policy policy;3558 3559		__guc_context_policy_start_klv(&policy, ce->guc_id.id);3560		__guc_context_policy_add_priority(&policy, ce->guc_state.prio);3561		__guc_context_set_context_policies(guc, &policy, true);3562	} else {3563		u32 action[] = {3564			INTEL_GUC_ACTION_V69_SET_CONTEXT_PRIORITY,3565			ce->guc_id.id,3566			ce->guc_state.prio,3567		};3568 3569		guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action), 0, true);3570	}3571}3572 3573static void guc_context_set_prio(struct intel_guc *guc,3574				 struct intel_context *ce,3575				 u8 prio)3576{3577	GEM_BUG_ON(prio < GUC_CLIENT_PRIORITY_KMD_HIGH ||3578		   prio > GUC_CLIENT_PRIORITY_NORMAL);3579	lockdep_assert_held(&ce->guc_state.lock);3580 3581	if (ce->guc_state.prio == prio || submission_disabled(guc) ||3582	    !context_registered(ce)) {3583		ce->guc_state.prio = prio;3584		return;3585	}3586 3587	ce->guc_state.prio = prio;3588	__guc_context_set_prio(guc, ce);3589 3590	trace_intel_context_set_prio(ce);3591}3592 3593static inline u8 map_i915_prio_to_guc_prio(int prio)3594{3595	if (prio == I915_PRIORITY_NORMAL)3596		return GUC_CLIENT_PRIORITY_KMD_NORMAL;3597	else if (prio < I915_PRIORITY_NORMAL)3598		return GUC_CLIENT_PRIORITY_NORMAL;3599	else if (prio < I915_PRIORITY_DISPLAY)3600		return GUC_CLIENT_PRIORITY_HIGH;3601	else3602		return GUC_CLIENT_PRIORITY_KMD_HIGH;3603}3604 3605static inline void add_context_inflight_prio(struct intel_context *ce,3606					     u8 guc_prio)3607{3608	lockdep_assert_held(&ce->guc_state.lock);3609	GEM_BUG_ON(guc_prio >= ARRAY_SIZE(ce->guc_state.prio_count));3610 3611	++ce->guc_state.prio_count[guc_prio];3612 3613	/* Overflow protection */3614	GEM_WARN_ON(!ce->guc_state.prio_count[guc_prio]);3615}3616 3617static inline void sub_context_inflight_prio(struct intel_context *ce,3618					     u8 guc_prio)3619{3620	lockdep_assert_held(&ce->guc_state.lock);3621	GEM_BUG_ON(guc_prio >= ARRAY_SIZE(ce->guc_state.prio_count));3622 3623	/* Underflow protection */3624	GEM_WARN_ON(!ce->guc_state.prio_count[guc_prio]);3625 3626	--ce->guc_state.prio_count[guc_prio];3627}3628 3629static inline void update_context_prio(struct intel_context *ce)3630{3631	struct intel_guc *guc = &ce->engine->gt->uc.guc;3632	int i;3633 3634	BUILD_BUG_ON(GUC_CLIENT_PRIORITY_KMD_HIGH != 0);3635	BUILD_BUG_ON(GUC_CLIENT_PRIORITY_KMD_HIGH > GUC_CLIENT_PRIORITY_NORMAL);3636 3637	lockdep_assert_held(&ce->guc_state.lock);3638 3639	for (i = 0; i < ARRAY_SIZE(ce->guc_state.prio_count); ++i) {3640		if (ce->guc_state.prio_count[i]) {3641			guc_context_set_prio(guc, ce, i);3642			break;3643		}3644	}3645}3646 3647static inline bool new_guc_prio_higher(u8 old_guc_prio, u8 new_guc_prio)3648{3649	/* Lower value is higher priority */3650	return new_guc_prio < old_guc_prio;3651}3652 3653static void add_to_context(struct i915_request *rq)3654{3655	struct intel_context *ce = request_to_scheduling_context(rq);3656	u8 new_guc_prio = map_i915_prio_to_guc_prio(rq_prio(rq));3657 3658	GEM_BUG_ON(intel_context_is_child(ce));3659	GEM_BUG_ON(rq->guc_prio == GUC_PRIO_FINI);3660 3661	spin_lock(&ce->guc_state.lock);3662	list_move_tail(&rq->sched.link, &ce->guc_state.requests);3663 3664	if (rq->guc_prio == GUC_PRIO_INIT) {3665		rq->guc_prio = new_guc_prio;3666		add_context_inflight_prio(ce, rq->guc_prio);3667	} else if (new_guc_prio_higher(rq->guc_prio, new_guc_prio)) {3668		sub_context_inflight_prio(ce, rq->guc_prio);3669		rq->guc_prio = new_guc_prio;3670		add_context_inflight_prio(ce, rq->guc_prio);3671	}3672	update_context_prio(ce);3673 3674	spin_unlock(&ce->guc_state.lock);3675}3676 3677static void guc_prio_fini(struct i915_request *rq, struct intel_context *ce)3678{3679	lockdep_assert_held(&ce->guc_state.lock);3680 3681	if (rq->guc_prio != GUC_PRIO_INIT &&3682	    rq->guc_prio != GUC_PRIO_FINI) {3683		sub_context_inflight_prio(ce, rq->guc_prio);3684		update_context_prio(ce);3685	}3686	rq->guc_prio = GUC_PRIO_FINI;3687}3688 3689static void remove_from_context(struct i915_request *rq)3690{3691	struct intel_context *ce = request_to_scheduling_context(rq);3692 3693	GEM_BUG_ON(intel_context_is_child(ce));3694 3695	spin_lock_irq(&ce->guc_state.lock);3696 3697	list_del_init(&rq->sched.link);3698	clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);3699 3700	/* Prevent further __await_execution() registering a cb, then flush */3701	set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);3702 3703	guc_prio_fini(rq, ce);3704 3705	spin_unlock_irq(&ce->guc_state.lock);3706 3707	atomic_dec(&ce->guc_id.ref);3708	i915_request_notify_execute_cb_imm(rq);3709}3710 3711static const struct intel_context_ops guc_context_ops = {3712	.flags = COPS_RUNTIME_CYCLES,3713	.alloc = guc_context_alloc,3714 3715	.close = guc_context_close,3716 3717	.pre_pin = guc_context_pre_pin,3718	.pin = guc_context_pin,3719	.unpin = guc_context_unpin,3720	.post_unpin = guc_context_post_unpin,3721 3722	.revoke = guc_context_revoke,3723 3724	.cancel_request = guc_context_cancel_request,3725 3726	.enter = intel_context_enter_engine,3727	.exit = intel_context_exit_engine,3728 3729	.sched_disable = guc_context_sched_disable,3730 3731	.update_stats = guc_context_update_stats,3732 3733	.reset = lrc_reset,3734	.destroy = guc_context_destroy,3735 3736	.create_virtual = guc_create_virtual,3737	.create_parallel = guc_create_parallel,3738};3739 3740static void submit_work_cb(struct irq_work *wrk)3741{3742	struct i915_request *rq = container_of(wrk, typeof(*rq), submit_work);3743 3744	might_lock(&rq->engine->sched_engine->lock);3745	i915_sw_fence_complete(&rq->submit);3746}3747 3748static void __guc_signal_context_fence(struct intel_context *ce)3749{3750	struct i915_request *rq, *rn;3751 3752	lockdep_assert_held(&ce->guc_state.lock);3753 3754	if (!list_empty(&ce->guc_state.fences))3755		trace_intel_context_fence_release(ce);3756 3757	/*3758	 * Use an IRQ to ensure locking order of sched_engine->lock ->3759	 * ce->guc_state.lock is preserved.3760	 */3761	list_for_each_entry_safe(rq, rn, &ce->guc_state.fences,3762				 guc_fence_link) {3763		list_del(&rq->guc_fence_link);3764		irq_work_queue(&rq->submit_work);3765	}3766 3767	INIT_LIST_HEAD(&ce->guc_state.fences);3768}3769 3770static void guc_signal_context_fence(struct intel_context *ce)3771{3772	unsigned long flags;3773 3774	GEM_BUG_ON(intel_context_is_child(ce));3775 3776	spin_lock_irqsave(&ce->guc_state.lock, flags);3777	clr_context_wait_for_deregister_to_register(ce);3778	__guc_signal_context_fence(ce);3779	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3780}3781 3782static bool context_needs_register(struct intel_context *ce, bool new_guc_id)3783{3784	return (new_guc_id || test_bit(CONTEXT_LRCA_DIRTY, &ce->flags) ||3785		!ctx_id_mapped(ce_to_guc(ce), ce->guc_id.id)) &&3786		!submission_disabled(ce_to_guc(ce));3787}3788 3789static void guc_context_init(struct intel_context *ce)3790{3791	const struct i915_gem_context *ctx;3792	int prio = I915_CONTEXT_DEFAULT_PRIORITY;3793 3794	rcu_read_lock();3795	ctx = rcu_dereference(ce->gem_context);3796	if (ctx)3797		prio = ctx->sched.priority;3798	rcu_read_unlock();3799 3800	ce->guc_state.prio = map_i915_prio_to_guc_prio(prio);3801 3802	INIT_DELAYED_WORK(&ce->guc_state.sched_disable_delay_work,3803			  __delay_sched_disable);3804 3805	set_bit(CONTEXT_GUC_INIT, &ce->flags);3806}3807 3808static int guc_request_alloc(struct i915_request *rq)3809{3810	struct intel_context *ce = request_to_scheduling_context(rq);3811	struct intel_guc *guc = ce_to_guc(ce);3812	unsigned long flags;3813	int ret;3814 3815	GEM_BUG_ON(!intel_context_is_pinned(rq->context));3816 3817	/*3818	 * Flush enough space to reduce the likelihood of waiting after3819	 * we start building the request - in which case we will just3820	 * have to repeat work.3821	 */3822	rq->reserved_space += GUC_REQUEST_SIZE;3823 3824	/*3825	 * Note that after this point, we have committed to using3826	 * this request as it is being used to both track the3827	 * state of engine initialisation and liveness of the3828	 * golden renderstate above. Think twice before you try3829	 * to cancel/unwind this request now.3830	 */3831 3832	/* Unconditionally invalidate GPU caches and TLBs. */3833	ret = rq->engine->emit_flush(rq, EMIT_INVALIDATE);3834	if (ret)3835		return ret;3836 3837	rq->reserved_space -= GUC_REQUEST_SIZE;3838 3839	if (unlikely(!test_bit(CONTEXT_GUC_INIT, &ce->flags)))3840		guc_context_init(ce);3841 3842	/*3843	 * If the context gets closed while the execbuf is ongoing, the context3844	 * close code will race with the below code to cancel the delayed work.3845	 * If the context close wins the race and cancels the work, it will3846	 * immediately call the sched disable (see guc_context_close), so there3847	 * is a chance we can get past this check while the sched_disable code3848	 * is being executed. To make sure that code completes before we check3849	 * the status further down, we wait for the close process to complete.3850	 * Else, this code path could send a request down thinking that the3851	 * context is still in a schedule-enable mode while the GuC ends up3852	 * dropping the request completely because the disable did go from the3853	 * context_close path right to GuC just prior. In the event the CT is3854	 * full, we could potentially need to wait up to 1.5 seconds.3855	 */3856	if (cancel_delayed_work_sync(&ce->guc_state.sched_disable_delay_work))3857		intel_context_sched_disable_unpin(ce);3858	else if (intel_context_is_closed(ce))3859		if (wait_for(context_close_done(ce), 1500))3860			guc_warn(guc, "timed out waiting on context sched close before realloc\n");3861	/*3862	 * Call pin_guc_id here rather than in the pinning step as with3863	 * dma_resv, contexts can be repeatedly pinned / unpinned trashing the3864	 * guc_id and creating horrible race conditions. This is especially bad3865	 * when guc_id are being stolen due to over subscription. By the time3866	 * this function is reached, it is guaranteed that the guc_id will be3867	 * persistent until the generated request is retired. Thus, sealing these3868	 * race conditions. It is still safe to fail here if guc_id are3869	 * exhausted and return -EAGAIN to the user indicating that they can try3870	 * again in the future.3871	 *3872	 * There is no need for a lock here as the timeline mutex ensures at3873	 * most one context can be executing this code path at once. The3874	 * guc_id_ref is incremented once for every request in flight and3875	 * decremented on each retire. When it is zero, a lock around the3876	 * increment (in pin_guc_id) is needed to seal a race with unpin_guc_id.3877	 */3878	if (atomic_add_unless(&ce->guc_id.ref, 1, 0))3879		goto out;3880 3881	ret = pin_guc_id(guc, ce);	/* returns 1 if new guc_id assigned */3882	if (unlikely(ret < 0))3883		return ret;3884	if (context_needs_register(ce, !!ret)) {3885		ret = try_context_registration(ce, true);3886		if (unlikely(ret)) {	/* unwind */3887			if (ret == -EPIPE) {3888				disable_submission(guc);3889				goto out;	/* GPU will be reset */3890			}3891			atomic_dec(&ce->guc_id.ref);3892			unpin_guc_id(guc, ce);3893			return ret;3894		}3895	}3896 3897	clear_bit(CONTEXT_LRCA_DIRTY, &ce->flags);3898 3899out:3900	/*3901	 * We block all requests on this context if a G2H is pending for a3902	 * schedule disable or context deregistration as the GuC will fail a3903	 * schedule enable or context registration if either G2H is pending3904	 * respectfully. Once a G2H returns, the fence is released that is3905	 * blocking these requests (see guc_signal_context_fence).3906	 */3907	spin_lock_irqsave(&ce->guc_state.lock, flags);3908	if (context_wait_for_deregister_to_register(ce) ||3909	    context_pending_disable(ce)) {3910		init_irq_work(&rq->submit_work, submit_work_cb);3911		i915_sw_fence_await(&rq->submit);3912 3913		list_add_tail(&rq->guc_fence_link, &ce->guc_state.fences);3914	}3915	spin_unlock_irqrestore(&ce->guc_state.lock, flags);3916 3917	return 0;3918}3919 3920static int guc_virtual_context_pre_pin(struct intel_context *ce,3921				       struct i915_gem_ww_ctx *ww,3922				       void **vaddr)3923{3924	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);3925 3926	return __guc_context_pre_pin(ce, engine, ww, vaddr);3927}3928 3929static int guc_virtual_context_pin(struct intel_context *ce, void *vaddr)3930{3931	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);3932	int ret = __guc_context_pin(ce, engine, vaddr);3933	intel_engine_mask_t tmp, mask = ce->engine->mask;3934 3935	if (likely(!ret))3936		for_each_engine_masked(engine, ce->engine->gt, mask, tmp)3937			intel_engine_pm_get(engine);3938 3939	return ret;3940}3941 3942static void guc_virtual_context_unpin(struct intel_context *ce)3943{3944	intel_engine_mask_t tmp, mask = ce->engine->mask;3945	struct intel_engine_cs *engine;3946	struct intel_guc *guc = ce_to_guc(ce);3947 3948	GEM_BUG_ON(context_enabled(ce));3949	GEM_BUG_ON(intel_context_is_barrier(ce));3950 3951	unpin_guc_id(guc, ce);3952	lrc_unpin(ce);3953 3954	for_each_engine_masked(engine, ce->engine->gt, mask, tmp)3955		intel_engine_pm_put_async(engine);3956}3957 3958static void guc_virtual_context_enter(struct intel_context *ce)3959{3960	intel_engine_mask_t tmp, mask = ce->engine->mask;3961	struct intel_engine_cs *engine;3962 3963	for_each_engine_masked(engine, ce->engine->gt, mask, tmp)3964		intel_engine_pm_get(engine);3965 3966	intel_timeline_enter(ce->timeline);3967}3968 3969static void guc_virtual_context_exit(struct intel_context *ce)3970{3971	intel_engine_mask_t tmp, mask = ce->engine->mask;3972	struct intel_engine_cs *engine;3973 3974	for_each_engine_masked(engine, ce->engine->gt, mask, tmp)3975		intel_engine_pm_put(engine);3976 3977	intel_timeline_exit(ce->timeline);3978}3979 3980static int guc_virtual_context_alloc(struct intel_context *ce)3981{3982	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);3983 3984	return lrc_alloc(ce, engine);3985}3986 3987static const struct intel_context_ops virtual_guc_context_ops = {3988	.flags = COPS_RUNTIME_CYCLES,3989	.alloc = guc_virtual_context_alloc,3990 3991	.close = guc_context_close,3992 3993	.pre_pin = guc_virtual_context_pre_pin,3994	.pin = guc_virtual_context_pin,3995	.unpin = guc_virtual_context_unpin,3996	.post_unpin = guc_context_post_unpin,3997 3998	.revoke = guc_context_revoke,3999 4000	.cancel_request = guc_context_cancel_request,4001 4002	.enter = guc_virtual_context_enter,4003	.exit = guc_virtual_context_exit,4004 4005	.sched_disable = guc_context_sched_disable,4006	.update_stats = guc_context_update_stats,4007 4008	.destroy = guc_context_destroy,4009 4010	.get_sibling = guc_virtual_get_sibling,4011};4012 4013static int guc_parent_context_pin(struct intel_context *ce, void *vaddr)4014{4015	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);4016	struct intel_guc *guc = ce_to_guc(ce);4017	int ret;4018 4019	GEM_BUG_ON(!intel_context_is_parent(ce));4020	GEM_BUG_ON(!intel_engine_is_virtual(ce->engine));4021 4022	ret = pin_guc_id(guc, ce);4023	if (unlikely(ret < 0))4024		return ret;4025 4026	return __guc_context_pin(ce, engine, vaddr);4027}4028 4029static int guc_child_context_pin(struct intel_context *ce, void *vaddr)4030{4031	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);4032 4033	GEM_BUG_ON(!intel_context_is_child(ce));4034	GEM_BUG_ON(!intel_engine_is_virtual(ce->engine));4035 4036	__intel_context_pin(ce->parallel.parent);4037	return __guc_context_pin(ce, engine, vaddr);4038}4039 4040static void guc_parent_context_unpin(struct intel_context *ce)4041{4042	struct intel_guc *guc = ce_to_guc(ce);4043 4044	GEM_BUG_ON(context_enabled(ce));4045	GEM_BUG_ON(intel_context_is_barrier(ce));4046	GEM_BUG_ON(!intel_context_is_parent(ce));4047	GEM_BUG_ON(!intel_engine_is_virtual(ce->engine));4048 4049	unpin_guc_id(guc, ce);4050	lrc_unpin(ce);4051}4052 4053static void guc_child_context_unpin(struct intel_context *ce)4054{4055	GEM_BUG_ON(context_enabled(ce));4056	GEM_BUG_ON(intel_context_is_barrier(ce));4057	GEM_BUG_ON(!intel_context_is_child(ce));4058	GEM_BUG_ON(!intel_engine_is_virtual(ce->engine));4059 4060	lrc_unpin(ce);4061}4062 4063static void guc_child_context_post_unpin(struct intel_context *ce)4064{4065	GEM_BUG_ON(!intel_context_is_child(ce));4066	GEM_BUG_ON(!intel_context_is_pinned(ce->parallel.parent));4067	GEM_BUG_ON(!intel_engine_is_virtual(ce->engine));4068 4069	lrc_post_unpin(ce);4070	intel_context_unpin(ce->parallel.parent);4071}4072 4073static void guc_child_context_destroy(struct kref *kref)4074{4075	struct intel_context *ce = container_of(kref, typeof(*ce), ref);4076 4077	__guc_context_destroy(ce);4078}4079 4080static const struct intel_context_ops virtual_parent_context_ops = {4081	.alloc = guc_virtual_context_alloc,4082 4083	.close = guc_context_close,4084 4085	.pre_pin = guc_context_pre_pin,4086	.pin = guc_parent_context_pin,4087	.unpin = guc_parent_context_unpin,4088	.post_unpin = guc_context_post_unpin,4089 4090	.revoke = guc_context_revoke,4091 4092	.cancel_request = guc_context_cancel_request,4093 4094	.enter = guc_virtual_context_enter,4095	.exit = guc_virtual_context_exit,4096 4097	.sched_disable = guc_context_sched_disable,4098 4099	.destroy = guc_context_destroy,4100 4101	.get_sibling = guc_virtual_get_sibling,4102};4103 4104static const struct intel_context_ops virtual_child_context_ops = {4105	.alloc = guc_virtual_context_alloc,4106 4107	.pre_pin = guc_context_pre_pin,4108	.pin = guc_child_context_pin,4109	.unpin = guc_child_context_unpin,4110	.post_unpin = guc_child_context_post_unpin,4111 4112	.cancel_request = guc_context_cancel_request,4113 4114	.enter = guc_virtual_context_enter,4115	.exit = guc_virtual_context_exit,4116 4117	.destroy = guc_child_context_destroy,4118 4119	.get_sibling = guc_virtual_get_sibling,4120};4121 4122/*4123 * The below override of the breadcrumbs is enabled when the user configures a4124 * context for parallel submission (multi-lrc, parent-child).4125 *4126 * The overridden breadcrumbs implements an algorithm which allows the GuC to4127 * safely preempt all the hw contexts configured for parallel submission4128 * between each BB. The contract between the i915 and GuC is if the parent4129 * context can be preempted, all the children can be preempted, and the GuC will4130 * always try to preempt the parent before the children. A handshake between the4131 * parent / children breadcrumbs ensures the i915 holds up its end of the deal4132 * creating a window to preempt between each set of BBs.4133 */4134static int emit_bb_start_parent_no_preempt_mid_batch(struct i915_request *rq,4135						     u64 offset, u32 len,4136						     const unsigned int flags);4137static int emit_bb_start_child_no_preempt_mid_batch(struct i915_request *rq,4138						    u64 offset, u32 len,4139						    const unsigned int flags);4140static u32 *4141emit_fini_breadcrumb_parent_no_preempt_mid_batch(struct i915_request *rq,4142						 u32 *cs);4143static u32 *4144emit_fini_breadcrumb_child_no_preempt_mid_batch(struct i915_request *rq,4145						u32 *cs);4146 4147static struct intel_context *4148guc_create_parallel(struct intel_engine_cs **engines,4149		    unsigned int num_siblings,4150		    unsigned int width)4151{4152	struct intel_engine_cs **siblings = NULL;4153	struct intel_context *parent = NULL, *ce, *err;4154	int i, j;4155 4156	siblings = kmalloc_array(num_siblings,4157				 sizeof(*siblings),4158				 GFP_KERNEL);4159	if (!siblings)4160		return ERR_PTR(-ENOMEM);4161 4162	for (i = 0; i < width; ++i) {4163		for (j = 0; j < num_siblings; ++j)4164			siblings[j] = engines[i * num_siblings + j];4165 4166		ce = intel_engine_create_virtual(siblings, num_siblings,4167						 FORCE_VIRTUAL);4168		if (IS_ERR(ce)) {4169			err = ERR_CAST(ce);4170			goto unwind;4171		}4172 4173		if (i == 0) {4174			parent = ce;4175			parent->ops = &virtual_parent_context_ops;4176		} else {4177			ce->ops = &virtual_child_context_ops;4178			intel_context_bind_parent_child(parent, ce);4179		}4180	}4181 4182	parent->parallel.fence_context = dma_fence_context_alloc(1);4183 4184	parent->engine->emit_bb_start =4185		emit_bb_start_parent_no_preempt_mid_batch;4186	parent->engine->emit_fini_breadcrumb =4187		emit_fini_breadcrumb_parent_no_preempt_mid_batch;4188	parent->engine->emit_fini_breadcrumb_dw =4189		12 + 4 * parent->parallel.number_children;4190	for_each_child(parent, ce) {4191		ce->engine->emit_bb_start =4192			emit_bb_start_child_no_preempt_mid_batch;4193		ce->engine->emit_fini_breadcrumb =4194			emit_fini_breadcrumb_child_no_preempt_mid_batch;4195		ce->engine->emit_fini_breadcrumb_dw = 16;4196	}4197 4198	kfree(siblings);4199	return parent;4200 4201unwind:4202	if (parent)4203		intel_context_put(parent);4204	kfree(siblings);4205	return err;4206}4207 4208static bool4209guc_irq_enable_breadcrumbs(struct intel_breadcrumbs *b)4210{4211	struct intel_engine_cs *sibling;4212	intel_engine_mask_t tmp, mask = b->engine_mask;4213	bool result = false;4214 4215	for_each_engine_masked(sibling, b->irq_engine->gt, mask, tmp)4216		result |= intel_engine_irq_enable(sibling);4217 4218	return result;4219}4220 4221static void4222guc_irq_disable_breadcrumbs(struct intel_breadcrumbs *b)4223{4224	struct intel_engine_cs *sibling;4225	intel_engine_mask_t tmp, mask = b->engine_mask;4226 4227	for_each_engine_masked(sibling, b->irq_engine->gt, mask, tmp)4228		intel_engine_irq_disable(sibling);4229}4230 4231static void guc_init_breadcrumbs(struct intel_engine_cs *engine)4232{4233	int i;4234 4235	/*4236	 * In GuC submission mode we do not know which physical engine a request4237	 * will be scheduled on, this creates a problem because the breadcrumb4238	 * interrupt is per physical engine. To work around this we attach4239	 * requests and direct all breadcrumb interrupts to the first instance4240	 * of an engine per class. In addition all breadcrumb interrupts are4241	 * enabled / disabled across an engine class in unison.4242	 */4243	for (i = 0; i < MAX_ENGINE_INSTANCE; ++i) {4244		struct intel_engine_cs *sibling =4245			engine->gt->engine_class[engine->class][i];4246 4247		if (sibling) {4248			if (engine->breadcrumbs != sibling->breadcrumbs) {4249				intel_breadcrumbs_put(engine->breadcrumbs);4250				engine->breadcrumbs =4251					intel_breadcrumbs_get(sibling->breadcrumbs);4252			}4253			break;4254		}4255	}4256 4257	if (engine->breadcrumbs) {4258		engine->breadcrumbs->engine_mask |= engine->mask;4259		engine->breadcrumbs->irq_enable = guc_irq_enable_breadcrumbs;4260		engine->breadcrumbs->irq_disable = guc_irq_disable_breadcrumbs;4261	}4262}4263 4264static void guc_bump_inflight_request_prio(struct i915_request *rq,4265					   int prio)4266{4267	struct intel_context *ce = request_to_scheduling_context(rq);4268	u8 new_guc_prio = map_i915_prio_to_guc_prio(prio);4269 4270	/* Short circuit function */4271	if (prio < I915_PRIORITY_NORMAL)4272		return;4273 4274	spin_lock(&ce->guc_state.lock);4275 4276	if (rq->guc_prio == GUC_PRIO_FINI)4277		goto exit;4278 4279	if (!new_guc_prio_higher(rq->guc_prio, new_guc_prio))4280		goto exit;4281 4282	if (rq->guc_prio != GUC_PRIO_INIT)4283		sub_context_inflight_prio(ce, rq->guc_prio);4284 4285	rq->guc_prio = new_guc_prio;4286	add_context_inflight_prio(ce, rq->guc_prio);4287	update_context_prio(ce);4288 4289exit:4290	spin_unlock(&ce->guc_state.lock);4291}4292 4293static void guc_retire_inflight_request_prio(struct i915_request *rq)4294{4295	struct intel_context *ce = request_to_scheduling_context(rq);4296 4297	spin_lock(&ce->guc_state.lock);4298	guc_prio_fini(rq, ce);4299	spin_unlock(&ce->guc_state.lock);4300}4301 4302static void sanitize_hwsp(struct intel_engine_cs *engine)4303{4304	struct intel_timeline *tl;4305 4306	list_for_each_entry(tl, &engine->status_page.timelines, engine_link)4307		intel_timeline_reset_seqno(tl);4308}4309 4310static void guc_sanitize(struct intel_engine_cs *engine)4311{4312	/*4313	 * Poison residual state on resume, in case the suspend didn't!4314	 *4315	 * We have to assume that across suspend/resume (or other loss4316	 * of control) that the contents of our pinned buffers has been4317	 * lost, replaced by garbage. Since this doesn't always happen,4318	 * let's poison such state so that we more quickly spot when4319	 * we falsely assume it has been preserved.4320	 */4321	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))4322		memset(engine->status_page.addr, POISON_INUSE, PAGE_SIZE);4323 4324	/*4325	 * The kernel_context HWSP is stored in the status_page. As above,4326	 * that may be lost on resume/initialisation, and so we need to4327	 * reset the value in the HWSP.4328	 */4329	sanitize_hwsp(engine);4330 4331	/* And scrub the dirty cachelines for the HWSP */4332	drm_clflush_virt_range(engine->status_page.addr, PAGE_SIZE);4333 4334	intel_engine_reset_pinned_contexts(engine);4335}4336 4337static void setup_hwsp(struct intel_engine_cs *engine)4338{4339	intel_engine_set_hwsp_writemask(engine, ~0u); /* HWSTAM */4340 4341	ENGINE_WRITE_FW(engine,4342			RING_HWS_PGA,4343			i915_ggtt_offset(engine->status_page.vma));4344}4345 4346static void start_engine(struct intel_engine_cs *engine)4347{4348	ENGINE_WRITE_FW(engine,4349			RING_MODE_GEN7,4350			_MASKED_BIT_ENABLE(GEN11_GFX_DISABLE_LEGACY_MODE));4351 4352	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));4353	ENGINE_POSTING_READ(engine, RING_MI_MODE);4354}4355 4356static int guc_resume(struct intel_engine_cs *engine)4357{4358	assert_forcewakes_active(engine->uncore, FORCEWAKE_ALL);4359 4360	intel_mocs_init_engine(engine);4361 4362	intel_breadcrumbs_reset(engine->breadcrumbs);4363 4364	setup_hwsp(engine);4365	start_engine(engine);4366 4367	if (engine->flags & I915_ENGINE_FIRST_RENDER_COMPUTE)4368		xehp_enable_ccs_engines(engine);4369 4370	return 0;4371}4372 4373static bool guc_sched_engine_disabled(struct i915_sched_engine *sched_engine)4374{4375	return !sched_engine->tasklet.callback;4376}4377 4378static void guc_set_default_submission(struct intel_engine_cs *engine)4379{4380	engine->submit_request = guc_submit_request;4381}4382 4383static inline int guc_kernel_context_pin(struct intel_guc *guc,4384					 struct intel_context *ce)4385{4386	int ret;4387 4388	/*4389	 * Note: we purposefully do not check the returns below because4390	 * the registration can only fail if a reset is just starting.4391	 * This is called at the end of reset so presumably another reset4392	 * isn't happening and even it did this code would be run again.4393	 */4394 4395	if (context_guc_id_invalid(ce)) {4396		ret = pin_guc_id(guc, ce);4397 4398		if (ret < 0)4399			return ret;4400	}4401 4402	if (!test_bit(CONTEXT_GUC_INIT, &ce->flags))4403		guc_context_init(ce);4404 4405	ret = try_context_registration(ce, true);4406	if (ret)4407		unpin_guc_id(guc, ce);4408 4409	return ret;4410}4411 4412static inline int guc_init_submission(struct intel_guc *guc)4413{4414	struct intel_gt *gt = guc_to_gt(guc);4415	struct intel_engine_cs *engine;4416	enum intel_engine_id id;4417 4418	/* make sure all descriptors are clean... */4419	xa_destroy(&guc->context_lookup);4420 4421	/*4422	 * A reset might have occurred while we had a pending stalled request,4423	 * so make sure we clean that up.4424	 */4425	guc->stalled_request = NULL;4426	guc->submission_stall_reason = STALL_NONE;4427 4428	/*4429	 * Some contexts might have been pinned before we enabled GuC4430	 * submission, so we need to add them to the GuC bookeeping.4431	 * Also, after a reset the of the GuC we want to make sure that the4432	 * information shared with GuC is properly reset. The kernel LRCs are4433	 * not attached to the gem_context, so they need to be added separately.4434	 */4435	for_each_engine(engine, gt, id) {4436		struct intel_context *ce;4437 4438		list_for_each_entry(ce, &engine->pinned_contexts_list,4439				    pinned_contexts_link) {4440			int ret = guc_kernel_context_pin(guc, ce);4441 4442			if (ret) {4443				/* No point in trying to clean up as i915 will wedge on failure */4444				return ret;4445			}4446		}4447	}4448 4449	return 0;4450}4451 4452static void guc_release(struct intel_engine_cs *engine)4453{4454	engine->sanitize = NULL; /* no longer in control, nothing to sanitize */4455 4456	intel_engine_cleanup_common(engine);4457	lrc_fini_wa_ctx(engine);4458}4459 4460static void virtual_guc_bump_serial(struct intel_engine_cs *engine)4461{4462	struct intel_engine_cs *e;4463	intel_engine_mask_t tmp, mask = engine->mask;4464 4465	for_each_engine_masked(e, engine->gt, mask, tmp)4466		e->serial++;4467}4468 4469static void guc_default_vfuncs(struct intel_engine_cs *engine)4470{4471	/* Default vfuncs which can be overridden by each engine. */4472 4473	engine->resume = guc_resume;4474 4475	engine->cops = &guc_context_ops;4476	engine->request_alloc = guc_request_alloc;4477	engine->add_active_request = add_to_context;4478	engine->remove_active_request = remove_from_context;4479 4480	engine->sched_engine->schedule = i915_schedule;4481 4482	engine->reset.prepare = guc_engine_reset_prepare;4483	engine->reset.rewind = guc_rewind_nop;4484	engine->reset.cancel = guc_reset_nop;4485	engine->reset.finish = guc_reset_nop;4486 4487	engine->emit_flush = gen8_emit_flush_xcs;4488	engine->emit_init_breadcrumb = gen8_emit_init_breadcrumb;4489	engine->emit_fini_breadcrumb = gen8_emit_fini_breadcrumb_xcs;4490	if (GRAPHICS_VER(engine->i915) >= 12) {4491		engine->emit_fini_breadcrumb = gen12_emit_fini_breadcrumb_xcs;4492		engine->emit_flush = gen12_emit_flush_xcs;4493	}4494	engine->set_default_submission = guc_set_default_submission;4495	engine->busyness = guc_engine_busyness;4496 4497	engine->flags |= I915_ENGINE_SUPPORTS_STATS;4498	engine->flags |= I915_ENGINE_HAS_PREEMPTION;4499	engine->flags |= I915_ENGINE_HAS_TIMESLICES;4500 4501	/* Wa_14014475959:dg2 */4502	if (engine->class == COMPUTE_CLASS)4503		if (IS_GFX_GT_IP_STEP(engine->gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||4504		    IS_DG2(engine->i915))4505			engine->flags |= I915_ENGINE_USES_WA_HOLD_SWITCHOUT;4506 4507	/* Wa_16019325821 */4508	/* Wa_14019159160 */4509	if ((engine->class == COMPUTE_CLASS || engine->class == RENDER_CLASS) &&4510	    IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 70), IP_VER(12, 74)))4511		engine->flags |= I915_ENGINE_USES_WA_HOLD_SWITCHOUT;4512 4513	/*4514	 * TODO: GuC supports timeslicing and semaphores as well, but they're4515	 * handled by the firmware so some minor tweaks are required before4516	 * enabling.4517	 *4518	 * engine->flags |= I915_ENGINE_HAS_SEMAPHORES;4519	 */4520 4521	engine->emit_bb_start = gen8_emit_bb_start;4522	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))4523		engine->emit_bb_start = xehp_emit_bb_start;4524}4525 4526static void rcs_submission_override(struct intel_engine_cs *engine)4527{4528	switch (GRAPHICS_VER(engine->i915)) {4529	case 12:4530		engine->emit_flush = gen12_emit_flush_rcs;4531		engine->emit_fini_breadcrumb = gen12_emit_fini_breadcrumb_rcs;4532		break;4533	case 11:4534		engine->emit_flush = gen11_emit_flush_rcs;4535		engine->emit_fini_breadcrumb = gen11_emit_fini_breadcrumb_rcs;4536		break;4537	default:4538		engine->emit_flush = gen8_emit_flush_rcs;4539		engine->emit_fini_breadcrumb = gen8_emit_fini_breadcrumb_rcs;4540		break;4541	}4542}4543 4544static inline void guc_default_irqs(struct intel_engine_cs *engine)4545{4546	engine->irq_keep_mask = GT_RENDER_USER_INTERRUPT;4547	intel_engine_set_irq_handler(engine, cs_irq_handler);4548}4549 4550static void guc_sched_engine_destroy(struct kref *kref)4551{4552	struct i915_sched_engine *sched_engine =4553		container_of(kref, typeof(*sched_engine), ref);4554	struct intel_guc *guc = sched_engine->private_data;4555 4556	guc->sched_engine = NULL;4557	tasklet_kill(&sched_engine->tasklet); /* flush the callback */4558	kfree(sched_engine);4559}4560 4561int intel_guc_submission_setup(struct intel_engine_cs *engine)4562{4563	struct drm_i915_private *i915 = engine->i915;4564	struct intel_guc *guc = gt_to_guc(engine->gt);4565 4566	/*4567	 * The setup relies on several assumptions (e.g. irqs always enabled)4568	 * that are only valid on gen11+4569	 */4570	GEM_BUG_ON(GRAPHICS_VER(i915) < 11);4571 4572	if (!guc->sched_engine) {4573		guc->sched_engine = i915_sched_engine_create(ENGINE_VIRTUAL);4574		if (!guc->sched_engine)4575			return -ENOMEM;4576 4577		guc->sched_engine->schedule = i915_schedule;4578		guc->sched_engine->disabled = guc_sched_engine_disabled;4579		guc->sched_engine->private_data = guc;4580		guc->sched_engine->destroy = guc_sched_engine_destroy;4581		guc->sched_engine->bump_inflight_request_prio =4582			guc_bump_inflight_request_prio;4583		guc->sched_engine->retire_inflight_request_prio =4584			guc_retire_inflight_request_prio;4585		tasklet_setup(&guc->sched_engine->tasklet,4586			      guc_submission_tasklet);4587	}4588	i915_sched_engine_put(engine->sched_engine);4589	engine->sched_engine = i915_sched_engine_get(guc->sched_engine);4590 4591	guc_default_vfuncs(engine);4592	guc_default_irqs(engine);4593	guc_init_breadcrumbs(engine);4594 4595	if (engine->flags & I915_ENGINE_HAS_RCS_REG_STATE)4596		rcs_submission_override(engine);4597 4598	lrc_init_wa_ctx(engine);4599 4600	/* Finally, take ownership and responsibility for cleanup! */4601	engine->sanitize = guc_sanitize;4602	engine->release = guc_release;4603 4604	return 0;4605}4606 4607struct scheduling_policy {4608	/* internal data */4609	u32 max_words, num_words;4610	u32 count;4611	/* API data */4612	struct guc_update_scheduling_policy h2g;4613};4614 4615static u32 __guc_scheduling_policy_action_size(struct scheduling_policy *policy)4616{4617	u32 *start = (void *)&policy->h2g;4618	u32 *end = policy->h2g.data + policy->num_words;4619	size_t delta = end - start;4620 4621	return delta;4622}4623 4624static struct scheduling_policy *__guc_scheduling_policy_start_klv(struct scheduling_policy *policy)4625{4626	policy->h2g.header.action = INTEL_GUC_ACTION_UPDATE_SCHEDULING_POLICIES_KLV;4627	policy->max_words = ARRAY_SIZE(policy->h2g.data);4628	policy->num_words = 0;4629	policy->count = 0;4630 4631	return policy;4632}4633 4634static void __guc_scheduling_policy_add_klv(struct scheduling_policy *policy,4635					    u32 action, u32 *data, u32 len)4636{4637	u32 *klv_ptr = policy->h2g.data + policy->num_words;4638 4639	GEM_BUG_ON((policy->num_words + 1 + len) > policy->max_words);4640	*(klv_ptr++) = FIELD_PREP(GUC_KLV_0_KEY, action) |4641		       FIELD_PREP(GUC_KLV_0_LEN, len);4642	memcpy(klv_ptr, data, sizeof(u32) * len);4643	policy->num_words += 1 + len;4644	policy->count++;4645}4646 4647static int __guc_action_set_scheduling_policies(struct intel_guc *guc,4648						struct scheduling_policy *policy)4649{4650	int ret;4651 4652	ret = intel_guc_send(guc, (u32 *)&policy->h2g,4653			     __guc_scheduling_policy_action_size(policy));4654	if (ret < 0) {4655		guc_probe_error(guc, "Failed to configure global scheduling policies: %pe!\n",4656				ERR_PTR(ret));4657		return ret;4658	}4659 4660	if (ret != policy->count) {4661		guc_warn(guc, "global scheduler policy processed %d of %d KLVs!",4662			 ret, policy->count);4663		if (ret > policy->count)4664			return -EPROTO;4665	}4666 4667	return 0;4668}4669 4670static int guc_init_global_schedule_policy(struct intel_guc *guc)4671{4672	struct scheduling_policy policy;4673	struct intel_gt *gt = guc_to_gt(guc);4674	intel_wakeref_t wakeref;4675	int ret;4676 4677	if (GUC_SUBMIT_VER(guc) < MAKE_GUC_VER(1, 1, 0))4678		return 0;4679 4680	__guc_scheduling_policy_start_klv(&policy);4681 4682	with_intel_runtime_pm(&gt->i915->runtime_pm, wakeref) {4683		u32 yield[] = {4684			GLOBAL_SCHEDULE_POLICY_RC_YIELD_DURATION,4685			GLOBAL_SCHEDULE_POLICY_RC_YIELD_RATIO,4686		};4687 4688		__guc_scheduling_policy_add_klv(&policy,4689						GUC_SCHEDULING_POLICIES_KLV_ID_RENDER_COMPUTE_YIELD,4690						yield, ARRAY_SIZE(yield));4691 4692		ret = __guc_action_set_scheduling_policies(guc, &policy);4693	}4694 4695	return ret;4696}4697 4698static void guc_route_semaphores(struct intel_guc *guc, bool to_guc)4699{4700	struct intel_gt *gt = guc_to_gt(guc);4701	u32 val;4702 4703	if (GRAPHICS_VER(gt->i915) < 12)4704		return;4705 4706	if (to_guc)4707		val = GUC_SEM_INTR_ROUTE_TO_GUC | GUC_SEM_INTR_ENABLE_ALL;4708	else4709		val = 0;4710 4711	intel_uncore_write(gt->uncore, GEN12_GUC_SEM_INTR_ENABLES, val);4712}4713 4714int intel_guc_submission_enable(struct intel_guc *guc)4715{4716	int ret;4717 4718	/* Semaphore interrupt enable and route to GuC */4719	guc_route_semaphores(guc, true);4720 4721	ret = guc_init_submission(guc);4722	if (ret)4723		goto fail_sem;4724 4725	ret = guc_init_engine_stats(guc);4726	if (ret)4727		goto fail_sem;4728 4729	ret = guc_init_global_schedule_policy(guc);4730	if (ret)4731		goto fail_stats;4732 4733	return 0;4734 4735fail_stats:4736	guc_fini_engine_stats(guc);4737fail_sem:4738	guc_route_semaphores(guc, false);4739	return ret;4740}4741 4742/* Note: By the time we're here, GuC may have already been reset */4743void intel_guc_submission_disable(struct intel_guc *guc)4744{4745	guc_cancel_busyness_worker(guc);4746 4747	/* Semaphore interrupt disable and route to host */4748	guc_route_semaphores(guc, false);4749}4750 4751static bool __guc_submission_supported(struct intel_guc *guc)4752{4753	/* GuC submission is unavailable for pre-Gen11 */4754	return intel_guc_is_supported(guc) &&4755	       GRAPHICS_VER(guc_to_i915(guc)) >= 11;4756}4757 4758static bool __guc_submission_selected(struct intel_guc *guc)4759{4760	struct drm_i915_private *i915 = guc_to_i915(guc);4761 4762	if (!intel_guc_submission_is_supported(guc))4763		return false;4764 4765	return i915->params.enable_guc & ENABLE_GUC_SUBMISSION;4766}4767 4768int intel_guc_sched_disable_gucid_threshold_max(struct intel_guc *guc)4769{4770	return guc->submission_state.num_guc_ids - NUMBER_MULTI_LRC_GUC_ID(guc);4771}4772 4773/*4774 * This default value of 33 milisecs (+1 milisec round up) ensures 30fps or higher4775 * workloads are able to enjoy the latency reduction when delaying the schedule-disable4776 * operation. This matches the 30fps game-render + encode (real world) workload this4777 * knob was tested against.4778 */4779#define SCHED_DISABLE_DELAY_MS	344780 4781/*4782 * A threshold of 75% is a reasonable starting point considering that real world apps4783 * generally don't get anywhere near this.4784 */4785#define NUM_SCHED_DISABLE_GUCIDS_DEFAULT_THRESHOLD(__guc) \4786	(((intel_guc_sched_disable_gucid_threshold_max(guc)) * 3) / 4)4787 4788void intel_guc_submission_init_early(struct intel_guc *guc)4789{4790	xa_init_flags(&guc->context_lookup, XA_FLAGS_LOCK_IRQ);4791 4792	spin_lock_init(&guc->submission_state.lock);4793	INIT_LIST_HEAD(&guc->submission_state.guc_id_list);4794	ida_init(&guc->submission_state.guc_ids);4795	INIT_LIST_HEAD(&guc->submission_state.destroyed_contexts);4796	INIT_WORK(&guc->submission_state.destroyed_worker,4797		  destroyed_worker_func);4798	INIT_WORK(&guc->submission_state.reset_fail_worker,4799		  reset_fail_worker_func);4800 4801	spin_lock_init(&guc->timestamp.lock);4802	INIT_DELAYED_WORK(&guc->timestamp.work, guc_timestamp_ping);4803 4804	guc->submission_state.sched_disable_delay_ms = SCHED_DISABLE_DELAY_MS;4805	guc->submission_state.num_guc_ids = GUC_MAX_CONTEXT_ID;4806	guc->submission_state.sched_disable_gucid_threshold =4807		NUM_SCHED_DISABLE_GUCIDS_DEFAULT_THRESHOLD(guc);4808	guc->submission_supported = __guc_submission_supported(guc);4809	guc->submission_selected = __guc_submission_selected(guc);4810}4811 4812static inline struct intel_context *4813g2h_context_lookup(struct intel_guc *guc, u32 ctx_id)4814{4815	struct intel_context *ce;4816 4817	if (unlikely(ctx_id >= GUC_MAX_CONTEXT_ID)) {4818		guc_err(guc, "Invalid ctx_id %u\n", ctx_id);4819		return NULL;4820	}4821 4822	ce = __get_context(guc, ctx_id);4823	if (unlikely(!ce)) {4824		guc_err(guc, "Context is NULL, ctx_id %u\n", ctx_id);4825		return NULL;4826	}4827 4828	if (unlikely(intel_context_is_child(ce))) {4829		guc_err(guc, "Context is child, ctx_id %u\n", ctx_id);4830		return NULL;4831	}4832 4833	return ce;4834}4835 4836static void wait_wake_outstanding_tlb_g2h(struct intel_guc *guc, u32 seqno)4837{4838	struct intel_guc_tlb_wait *wait;4839	unsigned long flags;4840 4841	xa_lock_irqsave(&guc->tlb_lookup, flags);4842	wait = xa_load(&guc->tlb_lookup, seqno);4843 4844	if (wait)4845		wake_up(&wait->wq);4846	else4847		guc_dbg(guc,4848			"Stale TLB invalidation response with seqno %d\n", seqno);4849 4850	xa_unlock_irqrestore(&guc->tlb_lookup, flags);4851}4852 4853int intel_guc_tlb_invalidation_done(struct intel_guc *guc,4854				    const u32 *payload, u32 len)4855{4856	if (len < 1)4857		return -EPROTO;4858 4859	wait_wake_outstanding_tlb_g2h(guc, payload[0]);4860	return 0;4861}4862 4863static long must_wait_woken(struct wait_queue_entry *wq_entry, long timeout)4864{4865	/*4866	 * This is equivalent to wait_woken() with the exception that4867	 * we do not wake up early if the kthread task has been completed.4868	 * As we are called from page reclaim in any task context,4869	 * we may be invoked from stopped kthreads, but we *must*4870	 * complete the wait from the HW.4871	 */4872	do {4873		set_current_state(TASK_UNINTERRUPTIBLE);4874		if (wq_entry->flags & WQ_FLAG_WOKEN)4875			break;4876 4877		timeout = schedule_timeout(timeout);4878	} while (timeout);4879 4880	/* See wait_woken() and woken_wake_function() */4881	__set_current_state(TASK_RUNNING);4882	smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN);4883 4884	return timeout;4885}4886 4887static bool intel_gt_is_enabled(const struct intel_gt *gt)4888{4889	/* Check if GT is wedged or suspended */4890	if (intel_gt_is_wedged(gt) || !intel_irqs_enabled(gt->i915))4891		return false;4892	return true;4893}4894 4895static int guc_send_invalidate_tlb(struct intel_guc *guc,4896				   enum intel_guc_tlb_invalidation_type type)4897{4898	struct intel_guc_tlb_wait _wq, *wq = &_wq;4899	struct intel_gt *gt = guc_to_gt(guc);4900	DEFINE_WAIT_FUNC(wait, woken_wake_function);4901	int err;4902	u32 seqno;4903	u32 action[] = {4904		INTEL_GUC_ACTION_TLB_INVALIDATION,4905		0,4906		REG_FIELD_PREP(INTEL_GUC_TLB_INVAL_TYPE_MASK, type) |4907			REG_FIELD_PREP(INTEL_GUC_TLB_INVAL_MODE_MASK,4908				       INTEL_GUC_TLB_INVAL_MODE_HEAVY) |4909			INTEL_GUC_TLB_INVAL_FLUSH_CACHE,4910	};4911	u32 size = ARRAY_SIZE(action);4912 4913	/*4914	 * Early guard against GT enablement.  TLB invalidation should not be4915	 * attempted if the GT is disabled due to suspend/wedge.4916	 */4917	if (!intel_gt_is_enabled(gt))4918		return -EINVAL;4919 4920	init_waitqueue_head(&_wq.wq);4921 4922	if (xa_alloc_cyclic_irq(&guc->tlb_lookup, &seqno, wq,4923				xa_limit_32b, &guc->next_seqno,4924				GFP_ATOMIC | __GFP_NOWARN) < 0) {4925		/* Under severe memory pressure? Serialise TLB allocations */4926		xa_lock_irq(&guc->tlb_lookup);4927		wq = xa_load(&guc->tlb_lookup, guc->serial_slot);4928		wait_event_lock_irq(wq->wq,4929				    !READ_ONCE(wq->busy),4930				    guc->tlb_lookup.xa_lock);4931		/*4932		 * Update wq->busy under lock to ensure only one waiter can4933		 * issue the TLB invalidation command using the serial slot at a4934		 * time. The condition is set to true before releasing the lock4935		 * so that other caller continue to wait until woken up again.4936		 */4937		wq->busy = true;4938		xa_unlock_irq(&guc->tlb_lookup);4939 4940		seqno = guc->serial_slot;4941	}4942 4943	action[1] = seqno;4944 4945	add_wait_queue(&wq->wq, &wait);4946 4947	/* This is a critical reclaim path and thus we must loop here. */4948	err = intel_guc_send_busy_loop(guc, action, size, G2H_LEN_DW_INVALIDATE_TLB, true);4949	if (err)4950		goto out;4951 4952	/*4953	 * Late guard against GT enablement.  It is not an error for the TLB4954	 * invalidation to time out if the GT is disabled during the process4955	 * due to suspend/wedge.  In fact, the TLB invalidation is cancelled4956	 * in this case.4957	 */4958	if (!must_wait_woken(&wait, intel_guc_ct_max_queue_time_jiffies()) &&4959	    intel_gt_is_enabled(gt)) {4960		guc_err(guc,4961			"TLB invalidation response timed out for seqno %u\n", seqno);4962		err = -ETIME;4963	}4964out:4965	remove_wait_queue(&wq->wq, &wait);4966	if (seqno != guc->serial_slot)4967		xa_erase_irq(&guc->tlb_lookup, seqno);4968 4969	return err;4970}4971 4972/* Send a H2G command to invalidate the TLBs at engine level and beyond. */4973int intel_guc_invalidate_tlb_engines(struct intel_guc *guc)4974{4975	return guc_send_invalidate_tlb(guc, INTEL_GUC_TLB_INVAL_ENGINES);4976}4977 4978/* Send a H2G command to invalidate the GuC's internal TLB. */4979int intel_guc_invalidate_tlb_guc(struct intel_guc *guc)4980{4981	return guc_send_invalidate_tlb(guc, INTEL_GUC_TLB_INVAL_GUC);4982}4983 4984int intel_guc_deregister_done_process_msg(struct intel_guc *guc,4985					  const u32 *msg,4986					  u32 len)4987{4988	struct intel_context *ce;4989	u32 ctx_id;4990 4991	if (unlikely(len < 1)) {4992		guc_err(guc, "Invalid length %u\n", len);4993		return -EPROTO;4994	}4995	ctx_id = msg[0];4996 4997	ce = g2h_context_lookup(guc, ctx_id);4998	if (unlikely(!ce))4999		return -EPROTO;5000 5001	trace_intel_context_deregister_done(ce);5002 5003#ifdef CONFIG_DRM_I915_SELFTEST5004	if (unlikely(ce->drop_deregister)) {5005		ce->drop_deregister = false;5006		return 0;5007	}5008#endif5009 5010	if (context_wait_for_deregister_to_register(ce)) {5011		struct intel_runtime_pm *runtime_pm =5012			&ce->engine->gt->i915->runtime_pm;5013		intel_wakeref_t wakeref;5014 5015		/*5016		 * Previous owner of this guc_id has been deregistered, now safe5017		 * register this context.5018		 */5019		with_intel_runtime_pm(runtime_pm, wakeref)5020			register_context(ce, true);5021		guc_signal_context_fence(ce);5022		intel_context_put(ce);5023	} else if (context_destroyed(ce)) {5024		/* Context has been destroyed */5025		intel_gt_pm_put_async_untracked(guc_to_gt(guc));5026		release_guc_id(guc, ce);5027		__guc_context_destroy(ce);5028	}5029 5030	decr_outstanding_submission_g2h(guc);5031 5032	return 0;5033}5034 5035int intel_guc_sched_done_process_msg(struct intel_guc *guc,5036				     const u32 *msg,5037				     u32 len)5038{5039	struct intel_context *ce;5040	unsigned long flags;5041	u32 ctx_id;5042 5043	if (unlikely(len < 2)) {5044		guc_err(guc, "Invalid length %u\n", len);5045		return -EPROTO;5046	}5047	ctx_id = msg[0];5048 5049	ce = g2h_context_lookup(guc, ctx_id);5050	if (unlikely(!ce))5051		return -EPROTO;5052 5053	if (unlikely(context_destroyed(ce) ||5054		     (!context_pending_enable(ce) &&5055		     !context_pending_disable(ce)))) {5056		guc_err(guc, "Bad context sched_state 0x%x, ctx_id %u\n",5057			ce->guc_state.sched_state, ctx_id);5058		return -EPROTO;5059	}5060 5061	trace_intel_context_sched_done(ce);5062 5063	if (context_pending_enable(ce)) {5064#ifdef CONFIG_DRM_I915_SELFTEST5065		if (unlikely(ce->drop_schedule_enable)) {5066			ce->drop_schedule_enable = false;5067			return 0;5068		}5069#endif5070 5071		spin_lock_irqsave(&ce->guc_state.lock, flags);5072		clr_context_pending_enable(ce);5073		spin_unlock_irqrestore(&ce->guc_state.lock, flags);5074	} else if (context_pending_disable(ce)) {5075		bool banned;5076 5077#ifdef CONFIG_DRM_I915_SELFTEST5078		if (unlikely(ce->drop_schedule_disable)) {5079			ce->drop_schedule_disable = false;5080			return 0;5081		}5082#endif5083 5084		/*5085		 * Unpin must be done before __guc_signal_context_fence,5086		 * otherwise a race exists between the requests getting5087		 * submitted + retired before this unpin completes resulting in5088		 * the pin_count going to zero and the context still being5089		 * enabled.5090		 */5091		intel_context_sched_disable_unpin(ce);5092 5093		spin_lock_irqsave(&ce->guc_state.lock, flags);5094		banned = context_banned(ce);5095		clr_context_banned(ce);5096		clr_context_pending_disable(ce);5097		__guc_signal_context_fence(ce);5098		guc_blocked_fence_complete(ce);5099		spin_unlock_irqrestore(&ce->guc_state.lock, flags);5100 5101		if (banned) {5102			guc_cancel_context_requests(ce);5103			intel_engine_signal_breadcrumbs(ce->engine);5104		}5105	}5106 5107	decr_outstanding_submission_g2h(guc);5108	intel_context_put(ce);5109 5110	return 0;5111}5112 5113static void capture_error_state(struct intel_guc *guc,5114				struct intel_context *ce)5115{5116	struct intel_gt *gt = guc_to_gt(guc);5117	struct drm_i915_private *i915 = gt->i915;5118	intel_wakeref_t wakeref;5119	intel_engine_mask_t engine_mask;5120 5121	if (intel_engine_is_virtual(ce->engine)) {5122		struct intel_engine_cs *e;5123		intel_engine_mask_t tmp, virtual_mask = ce->engine->mask;5124 5125		engine_mask = 0;5126		for_each_engine_masked(e, ce->engine->gt, virtual_mask, tmp) {5127			bool match = intel_guc_capture_is_matching_engine(gt, ce, e);5128 5129			if (match) {5130				intel_engine_set_hung_context(e, ce);5131				engine_mask |= e->mask;5132				i915_increase_reset_engine_count(&i915->gpu_error,5133								 e);5134			}5135		}5136 5137		if (!engine_mask) {5138			guc_warn(guc, "No matching physical engine capture for virtual engine context 0x%04X / %s",5139				 ce->guc_id.id, ce->engine->name);5140			engine_mask = ~0U;5141		}5142	} else {5143		intel_engine_set_hung_context(ce->engine, ce);5144		engine_mask = ce->engine->mask;5145		i915_increase_reset_engine_count(&i915->gpu_error, ce->engine);5146	}5147 5148	with_intel_runtime_pm(&i915->runtime_pm, wakeref)5149		i915_capture_error_state(gt, engine_mask, CORE_DUMP_FLAG_IS_GUC_CAPTURE);5150}5151 5152static void guc_context_replay(struct intel_context *ce)5153{5154	struct i915_sched_engine *sched_engine = ce->engine->sched_engine;5155 5156	__guc_reset_context(ce, ce->engine->mask);5157	tasklet_hi_schedule(&sched_engine->tasklet);5158}5159 5160static void guc_handle_context_reset(struct intel_guc *guc,5161				     struct intel_context *ce)5162{5163	bool capture = intel_context_is_schedulable(ce);5164 5165	trace_intel_context_reset(ce);5166 5167	guc_dbg(guc, "%s context reset notification: 0x%04X on %s, exiting = %s, banned = %s\n",5168		capture ? "Got" : "Ignoring",5169		ce->guc_id.id, ce->engine->name,5170		str_yes_no(intel_context_is_exiting(ce)),5171		str_yes_no(intel_context_is_banned(ce)));5172 5173	if (capture) {5174		capture_error_state(guc, ce);5175		guc_context_replay(ce);5176	}5177}5178 5179int intel_guc_context_reset_process_msg(struct intel_guc *guc,5180					const u32 *msg, u32 len)5181{5182	struct intel_context *ce;5183	unsigned long flags;5184	int ctx_id;5185 5186	if (unlikely(len != 1)) {5187		guc_err(guc, "Invalid length %u", len);5188		return -EPROTO;5189	}5190 5191	ctx_id = msg[0];5192 5193	/*5194	 * The context lookup uses the xarray but lookups only require an RCU lock5195	 * not the full spinlock. So take the lock explicitly and keep it until the5196	 * context has been reference count locked to ensure it can't be destroyed5197	 * asynchronously until the reset is done.5198	 */5199	xa_lock_irqsave(&guc->context_lookup, flags);5200	ce = g2h_context_lookup(guc, ctx_id);5201	if (ce)5202		intel_context_get(ce);5203	xa_unlock_irqrestore(&guc->context_lookup, flags);5204 5205	if (unlikely(!ce))5206		return -EPROTO;5207 5208	guc_handle_context_reset(guc, ce);5209	intel_context_put(ce);5210 5211	return 0;5212}5213 5214int intel_guc_error_capture_process_msg(struct intel_guc *guc,5215					const u32 *msg, u32 len)5216{5217	u32 status;5218 5219	if (unlikely(len != 1)) {5220		guc_dbg(guc, "Invalid length %u", len);5221		return -EPROTO;5222	}5223 5224	status = msg[0] & INTEL_GUC_STATE_CAPTURE_EVENT_STATUS_MASK;5225	if (status == INTEL_GUC_STATE_CAPTURE_EVENT_STATUS_NOSPACE)5226		guc_warn(guc, "No space for error capture");5227 5228	intel_guc_capture_process(guc);5229 5230	return 0;5231}5232 5233struct intel_engine_cs *5234intel_guc_lookup_engine(struct intel_guc *guc, u8 guc_class, u8 instance)5235{5236	struct intel_gt *gt = guc_to_gt(guc);5237	u8 engine_class = guc_class_to_engine_class(guc_class);5238 5239	/* Class index is checked in class converter */5240	GEM_BUG_ON(instance > MAX_ENGINE_INSTANCE);5241 5242	return gt->engine_class[engine_class][instance];5243}5244 5245static void reset_fail_worker_func(struct work_struct *w)5246{5247	struct intel_guc *guc = container_of(w, struct intel_guc,5248					     submission_state.reset_fail_worker);5249	struct intel_gt *gt = guc_to_gt(guc);5250	intel_engine_mask_t reset_fail_mask;5251	unsigned long flags;5252 5253	spin_lock_irqsave(&guc->submission_state.lock, flags);5254	reset_fail_mask = guc->submission_state.reset_fail_mask;5255	guc->submission_state.reset_fail_mask = 0;5256	spin_unlock_irqrestore(&guc->submission_state.lock, flags);5257 5258	if (likely(reset_fail_mask)) {5259		struct intel_engine_cs *engine;5260		enum intel_engine_id id;5261 5262		/*5263		 * GuC is toast at this point - it dead loops after sending the failed5264		 * reset notification. So need to manually determine the guilty context.5265		 * Note that it should be reliable to do this here because the GuC is5266		 * toast and will not be scheduling behind the KMD's back.5267		 */5268		for_each_engine_masked(engine, gt, reset_fail_mask, id)5269			intel_guc_find_hung_context(engine);5270 5271		intel_gt_handle_error(gt, reset_fail_mask,5272				      I915_ERROR_CAPTURE,5273				      "GuC failed to reset engine mask=0x%x",5274				      reset_fail_mask);5275	}5276}5277 5278int intel_guc_engine_failure_process_msg(struct intel_guc *guc,5279					 const u32 *msg, u32 len)5280{5281	struct intel_engine_cs *engine;5282	u8 guc_class, instance;5283	u32 reason;5284	unsigned long flags;5285 5286	if (unlikely(len != 3)) {5287		guc_err(guc, "Invalid length %u", len);5288		return -EPROTO;5289	}5290 5291	guc_class = msg[0];5292	instance = msg[1];5293	reason = msg[2];5294 5295	engine = intel_guc_lookup_engine(guc, guc_class, instance);5296	if (unlikely(!engine)) {5297		guc_err(guc, "Invalid engine %d:%d", guc_class, instance);5298		return -EPROTO;5299	}5300 5301	/*5302	 * This is an unexpected failure of a hardware feature. So, log a real5303	 * error message not just the informational that comes with the reset.5304	 */5305	guc_err(guc, "Engine reset failed on %d:%d (%s) because 0x%08X",5306		guc_class, instance, engine->name, reason);5307 5308	spin_lock_irqsave(&guc->submission_state.lock, flags);5309	guc->submission_state.reset_fail_mask |= engine->mask;5310	spin_unlock_irqrestore(&guc->submission_state.lock, flags);5311 5312	/*5313	 * A GT reset flushes this worker queue (G2H handler) so we must use5314	 * another worker to trigger a GT reset.5315	 */5316	queue_work(system_unbound_wq, &guc->submission_state.reset_fail_worker);5317 5318	return 0;5319}5320 5321void intel_guc_find_hung_context(struct intel_engine_cs *engine)5322{5323	struct intel_guc *guc = gt_to_guc(engine->gt);5324	struct intel_context *ce;5325	struct i915_request *rq;5326	unsigned long index;5327	unsigned long flags;5328 5329	/* Reset called during driver load? GuC not yet initialised! */5330	if (unlikely(!guc_submission_initialized(guc)))5331		return;5332 5333	xa_lock_irqsave(&guc->context_lookup, flags);5334	xa_for_each(&guc->context_lookup, index, ce) {5335		bool found;5336 5337		if (!kref_get_unless_zero(&ce->ref))5338			continue;5339 5340		xa_unlock(&guc->context_lookup);5341 5342		if (!intel_context_is_pinned(ce))5343			goto next;5344 5345		if (intel_engine_is_virtual(ce->engine)) {5346			if (!(ce->engine->mask & engine->mask))5347				goto next;5348		} else {5349			if (ce->engine != engine)5350				goto next;5351		}5352 5353		found = false;5354		spin_lock(&ce->guc_state.lock);5355		list_for_each_entry(rq, &ce->guc_state.requests, sched.link) {5356			if (i915_test_request_state(rq) != I915_REQUEST_ACTIVE)5357				continue;5358 5359			found = true;5360			break;5361		}5362		spin_unlock(&ce->guc_state.lock);5363 5364		if (found) {5365			intel_engine_set_hung_context(engine, ce);5366 5367			/* Can only cope with one hang at a time... */5368			intel_context_put(ce);5369			xa_lock(&guc->context_lookup);5370			goto done;5371		}5372 5373next:5374		intel_context_put(ce);5375		xa_lock(&guc->context_lookup);5376	}5377done:5378	xa_unlock_irqrestore(&guc->context_lookup, flags);5379}5380 5381void intel_guc_dump_active_requests(struct intel_engine_cs *engine,5382				    struct i915_request *hung_rq,5383				    struct drm_printer *m)5384{5385	struct intel_guc *guc = gt_to_guc(engine->gt);5386	struct intel_context *ce;5387	unsigned long index;5388	unsigned long flags;5389 5390	/* Reset called during driver load? GuC not yet initialised! */5391	if (unlikely(!guc_submission_initialized(guc)))5392		return;5393 5394	xa_lock_irqsave(&guc->context_lookup, flags);5395	xa_for_each(&guc->context_lookup, index, ce) {5396		if (!kref_get_unless_zero(&ce->ref))5397			continue;5398 5399		xa_unlock(&guc->context_lookup);5400 5401		if (!intel_context_is_pinned(ce))5402			goto next;5403 5404		if (intel_engine_is_virtual(ce->engine)) {5405			if (!(ce->engine->mask & engine->mask))5406				goto next;5407		} else {5408			if (ce->engine != engine)5409				goto next;5410		}5411 5412		spin_lock(&ce->guc_state.lock);5413		intel_engine_dump_active_requests(&ce->guc_state.requests,5414						  hung_rq, m);5415		spin_unlock(&ce->guc_state.lock);5416 5417next:5418		intel_context_put(ce);5419		xa_lock(&guc->context_lookup);5420	}5421	xa_unlock_irqrestore(&guc->context_lookup, flags);5422}5423 5424void intel_guc_submission_print_info(struct intel_guc *guc,5425				     struct drm_printer *p)5426{5427	struct i915_sched_engine *sched_engine = guc->sched_engine;5428	struct rb_node *rb;5429	unsigned long flags;5430 5431	if (!sched_engine)5432		return;5433 5434	drm_printf(p, "GuC Submission API Version: %d.%d.%d\n",5435		   guc->submission_version.major, guc->submission_version.minor,5436		   guc->submission_version.patch);5437	drm_printf(p, "GuC Number Outstanding Submission G2H: %u\n",5438		   atomic_read(&guc->outstanding_submission_g2h));5439	drm_printf(p, "GuC tasklet count: %u\n",5440		   atomic_read(&sched_engine->tasklet.count));5441 5442	spin_lock_irqsave(&sched_engine->lock, flags);5443	drm_printf(p, "Requests in GuC submit tasklet:\n");5444	for (rb = rb_first_cached(&sched_engine->queue); rb; rb = rb_next(rb)) {5445		struct i915_priolist *pl = to_priolist(rb);5446		struct i915_request *rq;5447 5448		priolist_for_each_request(rq, pl)5449			drm_printf(p, "guc_id=%u, seqno=%llu\n",5450				   rq->context->guc_id.id,5451				   rq->fence.seqno);5452	}5453	spin_unlock_irqrestore(&sched_engine->lock, flags);5454	drm_printf(p, "\n");5455}5456 5457static inline void guc_log_context_priority(struct drm_printer *p,5458					    struct intel_context *ce)5459{5460	int i;5461 5462	drm_printf(p, "\t\tPriority: %d\n", ce->guc_state.prio);5463	drm_printf(p, "\t\tNumber Requests (lower index == higher priority)\n");5464	for (i = GUC_CLIENT_PRIORITY_KMD_HIGH;5465	     i < GUC_CLIENT_PRIORITY_NUM; ++i) {5466		drm_printf(p, "\t\tNumber requests in priority band[%d]: %d\n",5467			   i, ce->guc_state.prio_count[i]);5468	}5469	drm_printf(p, "\n");5470}5471 5472static inline void guc_log_context(struct drm_printer *p,5473				   struct intel_context *ce)5474{5475	drm_printf(p, "GuC lrc descriptor %u:\n", ce->guc_id.id);5476	drm_printf(p, "\tHW Context Desc: 0x%08x\n", ce->lrc.lrca);5477	drm_printf(p, "\t\tLRC Head: Internal %u, Memory %u\n",5478		   ce->ring->head,5479		   ce->lrc_reg_state[CTX_RING_HEAD]);5480	drm_printf(p, "\t\tLRC Tail: Internal %u, Memory %u\n",5481		   ce->ring->tail,5482		   ce->lrc_reg_state[CTX_RING_TAIL]);5483	drm_printf(p, "\t\tContext Pin Count: %u\n",5484		   atomic_read(&ce->pin_count));5485	drm_printf(p, "\t\tGuC ID Ref Count: %u\n",5486		   atomic_read(&ce->guc_id.ref));5487	drm_printf(p, "\t\tSchedule State: 0x%x\n",5488		   ce->guc_state.sched_state);5489}5490 5491void intel_guc_submission_print_context_info(struct intel_guc *guc,5492					     struct drm_printer *p)5493{5494	struct intel_context *ce;5495	unsigned long index;5496	unsigned long flags;5497 5498	xa_lock_irqsave(&guc->context_lookup, flags);5499	xa_for_each(&guc->context_lookup, index, ce) {5500		GEM_BUG_ON(intel_context_is_child(ce));5501 5502		guc_log_context(p, ce);5503		guc_log_context_priority(p, ce);5504 5505		if (intel_context_is_parent(ce)) {5506			struct intel_context *child;5507 5508			drm_printf(p, "\t\tNumber children: %u\n",5509				   ce->parallel.number_children);5510 5511			if (ce->parallel.guc.wq_status) {5512				drm_printf(p, "\t\tWQI Head: %u\n",5513					   READ_ONCE(*ce->parallel.guc.wq_head));5514				drm_printf(p, "\t\tWQI Tail: %u\n",5515					   READ_ONCE(*ce->parallel.guc.wq_tail));5516				drm_printf(p, "\t\tWQI Status: %u\n",5517					   READ_ONCE(*ce->parallel.guc.wq_status));5518			}5519 5520			if (ce->engine->emit_bb_start ==5521			    emit_bb_start_parent_no_preempt_mid_batch) {5522				u8 i;5523 5524				drm_printf(p, "\t\tChildren Go: %u\n",5525					   get_children_go_value(ce));5526				for (i = 0; i < ce->parallel.number_children; ++i)5527					drm_printf(p, "\t\tChildren Join: %u\n",5528						   get_children_join_value(ce, i));5529			}5530 5531			for_each_child(ce, child)5532				guc_log_context(p, child);5533		}5534	}5535	xa_unlock_irqrestore(&guc->context_lookup, flags);5536}5537 5538static inline u32 get_children_go_addr(struct intel_context *ce)5539{5540	GEM_BUG_ON(!intel_context_is_parent(ce));5541 5542	return i915_ggtt_offset(ce->state) +5543		__get_parent_scratch_offset(ce) +5544		offsetof(struct parent_scratch, go.semaphore);5545}5546 5547static inline u32 get_children_join_addr(struct intel_context *ce,5548					 u8 child_index)5549{5550	GEM_BUG_ON(!intel_context_is_parent(ce));5551 5552	return i915_ggtt_offset(ce->state) +5553		__get_parent_scratch_offset(ce) +5554		offsetof(struct parent_scratch, join[child_index].semaphore);5555}5556 5557#define PARENT_GO_BB			15558#define PARENT_GO_FINI_BREADCRUMB	05559#define CHILD_GO_BB			15560#define CHILD_GO_FINI_BREADCRUMB	05561static int emit_bb_start_parent_no_preempt_mid_batch(struct i915_request *rq,5562						     u64 offset, u32 len,5563						     const unsigned int flags)5564{5565	struct intel_context *ce = rq->context;5566	u32 *cs;5567	u8 i;5568 5569	GEM_BUG_ON(!intel_context_is_parent(ce));5570 5571	cs = intel_ring_begin(rq, 10 + 4 * ce->parallel.number_children);5572	if (IS_ERR(cs))5573		return PTR_ERR(cs);5574 5575	/* Wait on children */5576	for (i = 0; i < ce->parallel.number_children; ++i) {5577		*cs++ = (MI_SEMAPHORE_WAIT |5578			 MI_SEMAPHORE_GLOBAL_GTT |5579			 MI_SEMAPHORE_POLL |5580			 MI_SEMAPHORE_SAD_EQ_SDD);5581		*cs++ = PARENT_GO_BB;5582		*cs++ = get_children_join_addr(ce, i);5583		*cs++ = 0;5584	}5585 5586	/* Turn off preemption */5587	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;5588	*cs++ = MI_NOOP;5589 5590	/* Tell children go */5591	cs = gen8_emit_ggtt_write(cs,5592				  CHILD_GO_BB,5593				  get_children_go_addr(ce),5594				  0);5595 5596	/* Jump to batch */5597	*cs++ = MI_BATCH_BUFFER_START_GEN8 |5598		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));5599	*cs++ = lower_32_bits(offset);5600	*cs++ = upper_32_bits(offset);5601	*cs++ = MI_NOOP;5602 5603	intel_ring_advance(rq, cs);5604 5605	return 0;5606}5607 5608static int emit_bb_start_child_no_preempt_mid_batch(struct i915_request *rq,5609						    u64 offset, u32 len,5610						    const unsigned int flags)5611{5612	struct intel_context *ce = rq->context;5613	struct intel_context *parent = intel_context_to_parent(ce);5614	u32 *cs;5615 5616	GEM_BUG_ON(!intel_context_is_child(ce));5617 5618	cs = intel_ring_begin(rq, 12);5619	if (IS_ERR(cs))5620		return PTR_ERR(cs);5621 5622	/* Signal parent */5623	cs = gen8_emit_ggtt_write(cs,5624				  PARENT_GO_BB,5625				  get_children_join_addr(parent,5626							 ce->parallel.child_index),5627				  0);5628 5629	/* Wait on parent for go */5630	*cs++ = (MI_SEMAPHORE_WAIT |5631		 MI_SEMAPHORE_GLOBAL_GTT |5632		 MI_SEMAPHORE_POLL |5633		 MI_SEMAPHORE_SAD_EQ_SDD);5634	*cs++ = CHILD_GO_BB;5635	*cs++ = get_children_go_addr(parent);5636	*cs++ = 0;5637 5638	/* Turn off preemption */5639	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;5640 5641	/* Jump to batch */5642	*cs++ = MI_BATCH_BUFFER_START_GEN8 |5643		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));5644	*cs++ = lower_32_bits(offset);5645	*cs++ = upper_32_bits(offset);5646 5647	intel_ring_advance(rq, cs);5648 5649	return 0;5650}5651 5652static u32 *5653__emit_fini_breadcrumb_parent_no_preempt_mid_batch(struct i915_request *rq,5654						   u32 *cs)5655{5656	struct intel_context *ce = rq->context;5657	u8 i;5658 5659	GEM_BUG_ON(!intel_context_is_parent(ce));5660 5661	/* Wait on children */5662	for (i = 0; i < ce->parallel.number_children; ++i) {5663		*cs++ = (MI_SEMAPHORE_WAIT |5664			 MI_SEMAPHORE_GLOBAL_GTT |5665			 MI_SEMAPHORE_POLL |5666			 MI_SEMAPHORE_SAD_EQ_SDD);5667		*cs++ = PARENT_GO_FINI_BREADCRUMB;5668		*cs++ = get_children_join_addr(ce, i);5669		*cs++ = 0;5670	}5671 5672	/* Turn on preemption */5673	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;5674	*cs++ = MI_NOOP;5675 5676	/* Tell children go */5677	cs = gen8_emit_ggtt_write(cs,5678				  CHILD_GO_FINI_BREADCRUMB,5679				  get_children_go_addr(ce),5680				  0);5681 5682	return cs;5683}5684 5685/*5686 * If this true, a submission of multi-lrc requests had an error and the5687 * requests need to be skipped. The front end (execuf IOCTL) should've called5688 * i915_request_skip which squashes the BB but we still need to emit the fini5689 * breadrcrumbs seqno write. At this point we don't know how many of the5690 * requests in the multi-lrc submission were generated so we can't do the5691 * handshake between the parent and children (e.g. if 4 requests should be5692 * generated but 2nd hit an error only 1 would be seen by the GuC backend).5693 * Simply skip the handshake, but still emit the breadcrumbd seqno, if an error5694 * has occurred on any of the requests in submission / relationship.5695 */5696static inline bool skip_handshake(struct i915_request *rq)5697{5698	return test_bit(I915_FENCE_FLAG_SKIP_PARALLEL, &rq->fence.flags);5699}5700 5701#define NON_SKIP_LEN	65702static u32 *5703emit_fini_breadcrumb_parent_no_preempt_mid_batch(struct i915_request *rq,5704						 u32 *cs)5705{5706	struct intel_context *ce = rq->context;5707	__maybe_unused u32 *before_fini_breadcrumb_user_interrupt_cs;5708	__maybe_unused u32 *start_fini_breadcrumb_cs = cs;5709 5710	GEM_BUG_ON(!intel_context_is_parent(ce));5711 5712	if (unlikely(skip_handshake(rq))) {5713		/*5714		 * NOP everything in __emit_fini_breadcrumb_parent_no_preempt_mid_batch,5715		 * the NON_SKIP_LEN comes from the length of the emits below.5716		 */5717		memset(cs, 0, sizeof(u32) *5718		       (ce->engine->emit_fini_breadcrumb_dw - NON_SKIP_LEN));5719		cs += ce->engine->emit_fini_breadcrumb_dw - NON_SKIP_LEN;5720	} else {5721		cs = __emit_fini_breadcrumb_parent_no_preempt_mid_batch(rq, cs);5722	}5723 5724	/* Emit fini breadcrumb */5725	before_fini_breadcrumb_user_interrupt_cs = cs;5726	cs = gen8_emit_ggtt_write(cs,5727				  rq->fence.seqno,5728				  i915_request_active_timeline(rq)->hwsp_offset,5729				  0);5730 5731	/* User interrupt */5732	*cs++ = MI_USER_INTERRUPT;5733	*cs++ = MI_NOOP;5734 5735	/* Ensure our math for skip + emit is correct */5736	GEM_BUG_ON(before_fini_breadcrumb_user_interrupt_cs + NON_SKIP_LEN !=5737		   cs);5738	GEM_BUG_ON(start_fini_breadcrumb_cs +5739		   ce->engine->emit_fini_breadcrumb_dw != cs);5740 5741	rq->tail = intel_ring_offset(rq, cs);5742 5743	return cs;5744}5745 5746static u32 *5747__emit_fini_breadcrumb_child_no_preempt_mid_batch(struct i915_request *rq,5748						  u32 *cs)5749{5750	struct intel_context *ce = rq->context;5751	struct intel_context *parent = intel_context_to_parent(ce);5752 5753	GEM_BUG_ON(!intel_context_is_child(ce));5754 5755	/* Turn on preemption */5756	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;5757	*cs++ = MI_NOOP;5758 5759	/* Signal parent */5760	cs = gen8_emit_ggtt_write(cs,5761				  PARENT_GO_FINI_BREADCRUMB,5762				  get_children_join_addr(parent,5763							 ce->parallel.child_index),5764				  0);5765 5766	/* Wait parent on for go */5767	*cs++ = (MI_SEMAPHORE_WAIT |5768		 MI_SEMAPHORE_GLOBAL_GTT |5769		 MI_SEMAPHORE_POLL |5770		 MI_SEMAPHORE_SAD_EQ_SDD);5771	*cs++ = CHILD_GO_FINI_BREADCRUMB;5772	*cs++ = get_children_go_addr(parent);5773	*cs++ = 0;5774 5775	return cs;5776}5777 5778static u32 *5779emit_fini_breadcrumb_child_no_preempt_mid_batch(struct i915_request *rq,5780						u32 *cs)5781{5782	struct intel_context *ce = rq->context;5783	__maybe_unused u32 *before_fini_breadcrumb_user_interrupt_cs;5784	__maybe_unused u32 *start_fini_breadcrumb_cs = cs;5785 5786	GEM_BUG_ON(!intel_context_is_child(ce));5787 5788	if (unlikely(skip_handshake(rq))) {5789		/*5790		 * NOP everything in __emit_fini_breadcrumb_child_no_preempt_mid_batch,5791		 * the NON_SKIP_LEN comes from the length of the emits below.5792		 */5793		memset(cs, 0, sizeof(u32) *5794		       (ce->engine->emit_fini_breadcrumb_dw - NON_SKIP_LEN));5795		cs += ce->engine->emit_fini_breadcrumb_dw - NON_SKIP_LEN;5796	} else {5797		cs = __emit_fini_breadcrumb_child_no_preempt_mid_batch(rq, cs);5798	}5799 5800	/* Emit fini breadcrumb */5801	before_fini_breadcrumb_user_interrupt_cs = cs;5802	cs = gen8_emit_ggtt_write(cs,5803				  rq->fence.seqno,5804				  i915_request_active_timeline(rq)->hwsp_offset,5805				  0);5806 5807	/* User interrupt */5808	*cs++ = MI_USER_INTERRUPT;5809	*cs++ = MI_NOOP;5810 5811	/* Ensure our math for skip + emit is correct */5812	GEM_BUG_ON(before_fini_breadcrumb_user_interrupt_cs + NON_SKIP_LEN !=5813		   cs);5814	GEM_BUG_ON(start_fini_breadcrumb_cs +5815		   ce->engine->emit_fini_breadcrumb_dw != cs);5816 5817	rq->tail = intel_ring_offset(rq, cs);5818 5819	return cs;5820}5821 5822#undef NON_SKIP_LEN5823 5824static struct intel_context *5825guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count,5826		   unsigned long flags)5827{5828	struct guc_virtual_engine *ve;5829	struct intel_guc *guc;5830	unsigned int n;5831	int err;5832 5833	ve = kzalloc(sizeof(*ve), GFP_KERNEL);5834	if (!ve)5835		return ERR_PTR(-ENOMEM);5836 5837	guc = gt_to_guc(siblings[0]->gt);5838 5839	ve->base.i915 = siblings[0]->i915;5840	ve->base.gt = siblings[0]->gt;5841	ve->base.uncore = siblings[0]->uncore;5842	ve->base.id = -1;5843 5844	ve->base.uabi_class = I915_ENGINE_CLASS_INVALID;5845	ve->base.instance = I915_ENGINE_CLASS_INVALID_VIRTUAL;5846	ve->base.uabi_instance = I915_ENGINE_CLASS_INVALID_VIRTUAL;5847	ve->base.saturated = ALL_ENGINES;5848 5849	snprintf(ve->base.name, sizeof(ve->base.name), "virtual");5850 5851	ve->base.sched_engine = i915_sched_engine_get(guc->sched_engine);5852 5853	ve->base.cops = &virtual_guc_context_ops;5854	ve->base.request_alloc = guc_request_alloc;5855	ve->base.bump_serial = virtual_guc_bump_serial;5856 5857	ve->base.submit_request = guc_submit_request;5858 5859	ve->base.flags = I915_ENGINE_IS_VIRTUAL;5860 5861	BUILD_BUG_ON(ilog2(VIRTUAL_ENGINES) < I915_NUM_ENGINES);5862	ve->base.mask = VIRTUAL_ENGINES;5863 5864	intel_context_init(&ve->context, &ve->base);5865 5866	for (n = 0; n < count; n++) {5867		struct intel_engine_cs *sibling = siblings[n];5868 5869		GEM_BUG_ON(!is_power_of_2(sibling->mask));5870		if (sibling->mask & ve->base.mask) {5871			guc_dbg(guc, "duplicate %s entry in load balancer\n",5872				sibling->name);5873			err = -EINVAL;5874			goto err_put;5875		}5876 5877		ve->base.mask |= sibling->mask;5878		ve->base.logical_mask |= sibling->logical_mask;5879 5880		if (n != 0 && ve->base.class != sibling->class) {5881			guc_dbg(guc, "invalid mixing of engine class, sibling %d, already %d\n",5882				sibling->class, ve->base.class);5883			err = -EINVAL;5884			goto err_put;5885		} else if (n == 0) {5886			ve->base.class = sibling->class;5887			ve->base.uabi_class = sibling->uabi_class;5888			snprintf(ve->base.name, sizeof(ve->base.name),5889				 "v%dx%d", ve->base.class, count);5890			ve->base.context_size = sibling->context_size;5891 5892			ve->base.add_active_request =5893				sibling->add_active_request;5894			ve->base.remove_active_request =5895				sibling->remove_active_request;5896			ve->base.emit_bb_start = sibling->emit_bb_start;5897			ve->base.emit_flush = sibling->emit_flush;5898			ve->base.emit_init_breadcrumb =5899				sibling->emit_init_breadcrumb;5900			ve->base.emit_fini_breadcrumb =5901				sibling->emit_fini_breadcrumb;5902			ve->base.emit_fini_breadcrumb_dw =5903				sibling->emit_fini_breadcrumb_dw;5904			ve->base.breadcrumbs =5905				intel_breadcrumbs_get(sibling->breadcrumbs);5906 5907			ve->base.flags |= sibling->flags;5908 5909			ve->base.props.timeslice_duration_ms =5910				sibling->props.timeslice_duration_ms;5911			ve->base.props.preempt_timeout_ms =5912				sibling->props.preempt_timeout_ms;5913		}5914	}5915 5916	return &ve->context;5917 5918err_put:5919	intel_context_put(&ve->context);5920	return ERR_PTR(err);5921}5922 5923bool intel_guc_virtual_engine_has_heartbeat(const struct intel_engine_cs *ve)5924{5925	struct intel_engine_cs *engine;5926	intel_engine_mask_t tmp, mask = ve->mask;5927 5928	for_each_engine_masked(engine, ve->gt, mask, tmp)5929		if (READ_ONCE(engine->props.heartbeat_interval_ms))5930			return true;5931 5932	return false;5933}5934 5935#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)5936#include "selftest_guc.c"5937#include "selftest_guc_multi_lrc.c"5938#include "selftest_guc_hangcheck.c"5939#endif5940