Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 7161EEB64D8 for ; Fri, 16 Jun 2023 15:26:13 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 3101B10E089; Fri, 16 Jun 2023 15:26:13 +0000 (UTC) Received: from mga04.intel.com (mga04.intel.com [192.55.52.120]) by gabe.freedesktop.org (Postfix) with ESMTPS id 699A210E089 for ; Fri, 16 Jun 2023 15:26:11 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1686929171; x=1718465171; h=message-id:date:mime-version:subject:to:references:from: in-reply-to:content-transfer-encoding; bh=ikK8CYz6n26tB1tsob0IjhiUxZZoz7sVr2upsC26Pak=; b=eU6pSA/+xXqW6jepfXM7AV8ifaRMX/+4wRUYxFRGq5YLboQSOrTvcKMQ eAixfusjVSimC9djDX6hCMvSiF+BhDrOUC2O0RECgX6Q5x6MObaXVvSPk S43QwKZ+am3rDUtRXLVNw/Su+1wILZwTS8aSZBfgqZ5Sla2fUXqos2xho eHKUpPAB7VcKlktSGKHFpiV3r3+T4QzXRYaSQR9a9BMX0drKHYAT7nWoE kldz12P+Kpj90q0DXAIw/3YZTIMEfqSMypWkEnyJ97JlqTb1+gcT7jHOU aCh+L4yIzFxunSmBMvVWasO++Pc2cxKL6zIDiIMM0uPKptJP4uZq/3uM6 g==; X-IronPort-AV: E=McAfee;i="6600,9927,10742"; a="358110002" X-IronPort-AV: E=Sophos;i="6.00,247,1681196400"; d="scan'208";a="358110002" Received: from orsmga002.jf.intel.com ([10.7.209.21]) by fmsmga104.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 16 Jun 2023 08:26:10 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=McAfee;i="6600,9927,10742"; a="712916234" X-IronPort-AV: E=Sophos;i="6.00,247,1681196400"; d="scan'208";a="712916234" Received: from nurha65x-mobl2.gar.corp.intel.com (HELO [10.249.254.148]) ([10.249.254.148]) by orsmga002-auth.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 16 Jun 2023 08:26:08 -0700 Message-ID: <394c84ff-0035-732a-268d-5c7577e2e1db@linux.intel.com> Date: Fri, 16 Jun 2023 17:26:06 +0200 MIME-Version: 1.0 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101 Thunderbird/102.11.1 Content-Language: en-US To: Matthew Brost , intel-xe@lists.freedesktop.org References: <20230615173014.258703-1-matthew.brost@intel.com> <20230615173014.258703-6-matthew.brost@intel.com> From: =?UTF-8?Q?Thomas_Hellstr=c3=b6m?= In-Reply-To: <20230615173014.258703-6-matthew.brost@intel.com> Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 8bit Subject: Re: [Intel-xe] [PATCH v4 5/8] drm/xe: Long running job update X-BeenThere: intel-xe@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Intel Xe graphics driver List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: intel-xe-bounces@lists.freedesktop.org Sender: "Intel-xe" On 6/15/23 19:30, Matthew Brost wrote: > For long running (LR) jobs with the DRM scheduler we must return NULL in > run_job which results in signaling the job's finished fence immediately. > This prevents LR jobs from creating infinite dma-fences. > > Signaling job's finished fence immediately breaks flow controlling ring > with the DRM scheduler. To work around this, the ring is flow controlled > and written in the exec IOCTL. Signaling job's finished fence > immediately also breaks the TDR which is used in reset / cleanup entity > paths so write a new path for LR entities. > > v2: Better commit, white space, remove rmb(), better comment next to > emit_job() > v3 (Thomas): Change LR reference counting, fix working in commit Ideally someone with more familiarity with the GuC code should take a second look on this, small nit below, but regardless Reviewed-by: Thomas Hellström > Signed-off-by: Matthew Brost > --- > drivers/gpu/drm/xe/xe_engine.c | 32 ++++++++ > drivers/gpu/drm/xe/xe_engine.h | 4 + > drivers/gpu/drm/xe/xe_exec.c | 8 ++ > drivers/gpu/drm/xe/xe_guc_engine_types.h | 2 + > drivers/gpu/drm/xe/xe_guc_submit.c | 96 ++++++++++++++++++++---- > drivers/gpu/drm/xe/xe_trace.h | 5 ++ > 6 files changed, 134 insertions(+), 13 deletions(-) > > diff --git a/drivers/gpu/drm/xe/xe_engine.c b/drivers/gpu/drm/xe/xe_engine.c > index b3036c4a8ec3..eb5f66f07da8 100644 > --- a/drivers/gpu/drm/xe/xe_engine.c > +++ b/drivers/gpu/drm/xe/xe_engine.c > @@ -18,6 +18,7 @@ > #include "xe_macros.h" > #include "xe_migrate.h" > #include "xe_pm.h" > +#include "xe_ring_ops_types.h" > #include "xe_trace.h" > #include "xe_vm.h" > > @@ -679,6 +680,37 @@ static void engine_kill_compute(struct xe_engine *e) > up_write(&e->vm->lock); > } > > +/** > + * xe_engine_is_lr() - Whether an engine is long-running > + * @e: The engine > + * > + * Return: True if the engine is long-running, false otherwise. > + */ > +bool xe_engine_is_lr(struct xe_engine *e) > +{ > + return e->vm && xe_vm_no_dma_fences(e->vm) && > + !(e->flags & ENGINE_FLAG_VM); > +} > + > +static s32 xe_engine_num_job_inflight(struct xe_engine *e) > +{ > + return e->lrc->fence_ctx.next_seqno - xe_lrc_seqno(e->lrc) - 1; > +} > + > +/** > + * xe_engine_ring_full() - Whether an engine's ring is full > + * @e: The engine > + * > + * Return: True if the engine's ring is full, false otherwise. > + */ > +bool xe_engine_ring_full(struct xe_engine *e) > +{ > + struct xe_lrc *lrc = e->lrc; > + s32 max_job = lrc->ring.size / MAX_JOB_SIZE_BYTES; > + > + return xe_engine_num_job_inflight(e) >= max_job; > +} > + > /** > * xe_engine_is_idle() - Whether an engine is idle. > * @engine: The engine > diff --git a/drivers/gpu/drm/xe/xe_engine.h b/drivers/gpu/drm/xe/xe_engine.h > index a49cf2ab405e..2e60f6d90226 100644 > --- a/drivers/gpu/drm/xe/xe_engine.h > +++ b/drivers/gpu/drm/xe/xe_engine.h > @@ -42,6 +42,10 @@ static inline bool xe_engine_is_parallel(struct xe_engine *engine) > return engine->width > 1; > } > > +bool xe_engine_is_lr(struct xe_engine *e); > + > +bool xe_engine_ring_full(struct xe_engine *e); > + > bool xe_engine_is_idle(struct xe_engine *engine); > > void xe_engine_kill(struct xe_engine *e); > diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c > index e44076ee2e11..9e9523939de8 100644 > --- a/drivers/gpu/drm/xe/xe_exec.c > +++ b/drivers/gpu/drm/xe/xe_exec.c > @@ -14,6 +14,7 @@ > #include "xe_device.h" > #include "xe_engine.h" > #include "xe_macros.h" > +#include "xe_ring_ops_types.h" > #include "xe_sched_job.h" > #include "xe_sync.h" > #include "xe_vm.h" > @@ -300,6 +301,11 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file) > goto err_engine_end; > } > > + if (xe_engine_is_lr(engine) && xe_engine_ring_full(engine)) { > + err = -EWOULDBLOCK; > + goto err_engine_end; > + } > + > job = xe_sched_job_create(engine, xe_engine_is_parallel(engine) ? > addresses : &args->address); > if (IS_ERR(job)) { > @@ -386,6 +392,8 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file) > xe_sync_entry_signal(&syncs[i], job, > &job->drm.s_fence->finished); > > + if (xe_engine_is_lr(engine)) > + engine->ring_ops->emit_job(job); Would it make sense to teach xe_sched_job_xxx() about the LR jobs, so that the above is done in xe_sched_job_push()? > xe_sched_job_push(job); > xe_vm_reactivate_rebind(vm); > > diff --git a/drivers/gpu/drm/xe/xe_guc_engine_types.h b/drivers/gpu/drm/xe/xe_guc_engine_types.h > index cbfb13026ec1..5d83132034a6 100644 > --- a/drivers/gpu/drm/xe/xe_guc_engine_types.h > +++ b/drivers/gpu/drm/xe/xe_guc_engine_types.h > @@ -31,6 +31,8 @@ struct xe_guc_engine { > */ > #define MAX_STATIC_MSG_TYPE 3 > struct drm_sched_msg static_msgs[MAX_STATIC_MSG_TYPE]; > + /** @lr_tdr: long running TDR worker */ > + struct work_struct lr_tdr; > /** @fini_async: do final fini async from this worker */ > struct work_struct fini_async; > /** @resume_time: time of last resume */ > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c > index a7a0d9f806bd..86c445903560 100644 > --- a/drivers/gpu/drm/xe/xe_guc_submit.c > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c > @@ -482,6 +482,14 @@ static void register_engine(struct xe_engine *e) > parallel_write(xe, map, wq_desc.wq_status, WQ_STATUS_ACTIVE); > } > > + /* > + * We must keep a reference for LR engines if engine is registered with > + * the GuC as jobs signal immediately and can't destroy an engine if the > + * GuC has a reference to it. > + */ > + if (xe_engine_is_lr(e)) > + xe_engine_get(e); > + > set_engine_registered(e); > trace_xe_engine_register(e); > if (xe_engine_is_parallel(e)) > @@ -644,6 +652,7 @@ guc_engine_run_job(struct drm_sched_job *drm_job) > { > struct xe_sched_job *job = to_xe_sched_job(drm_job); > struct xe_engine *e = job->engine; > + bool lr = xe_engine_is_lr(e); > > XE_BUG_ON((engine_destroyed(e) || engine_pending_disable(e)) && > !engine_banned(e) && !engine_suspended(e)); > @@ -653,14 +662,19 @@ guc_engine_run_job(struct drm_sched_job *drm_job) > if (!engine_killed_or_banned(e) && !xe_sched_job_is_error(job)) { > if (!engine_registered(e)) > register_engine(e); > - e->ring_ops->emit_job(job); > + if (!lr) /* LR jobs are emitted in the exec IOCTL */ > + e->ring_ops->emit_job(job); > submit_engine(e); > } > > - if (test_and_set_bit(JOB_FLAG_SUBMIT, &job->fence->flags)) > + if (lr) { > + xe_sched_job_set_error(job, -EOPNOTSUPP); > + return NULL; > + } else if (test_and_set_bit(JOB_FLAG_SUBMIT, &job->fence->flags)) { > return job->fence; > - else > + } else { > return dma_fence_get(job->fence); > + } > } > > static void guc_engine_free_job(struct drm_sched_job *drm_job) > @@ -764,6 +778,55 @@ static void simple_error_capture(struct xe_engine *e) > } > #endif > > +static void xe_guc_engine_trigger_cleanup(struct xe_engine *e) > +{ > + struct xe_guc *guc = engine_to_guc(e); > + > + if (xe_engine_is_lr(e)) > + queue_work(guc_to_gt(guc)->ordered_wq, &e->guc->lr_tdr); > + else > + drm_sched_set_timeout(&e->guc->sched, MIN_SCHED_TIMEOUT); > +} > + > +static void xe_guc_engine_lr_cleanup(struct work_struct *w) > +{ > + struct xe_guc_engine *ge = > + container_of(w, struct xe_guc_engine, lr_tdr); > + struct xe_engine *e = ge->engine; > + struct drm_gpu_scheduler *sched = &ge->sched; > + > + XE_WARN_ON(!xe_engine_is_lr(e)); > + trace_xe_engine_lr_cleanup(e); > + > + /* Kill the run_job / process_msg entry points */ > + drm_sched_run_wq_stop(sched); > + > + /* Engine state now stable, disable scheduling / deregister if needed */ > + if (engine_registered(e)) { > + struct xe_guc *guc = engine_to_guc(e); > + int ret; > + > + set_engine_banned(e); > + disable_scheduling_deregister(guc, e); > + > + /* > + * Must wait for scheduling to be disabled before signalling > + * any fences, if GT broken the GT reset code should signal us. > + */ > + ret = wait_event_timeout(guc->ct.wq, > + !engine_pending_disable(e) || > + guc_read_stopped(guc), HZ * 5); > + if (!ret) { > + XE_WARN_ON("Schedule disable failed to respond"); > + drm_sched_run_wq_start(sched); > + xe_gt_reset_async(e->gt); > + return; > + } > + } > + > + drm_sched_run_wq_start(sched); > +} > + > static enum drm_gpu_sched_stat > guc_engine_timedout_job(struct drm_sched_job *drm_job) > { > @@ -815,7 +878,7 @@ guc_engine_timedout_job(struct drm_sched_job *drm_job) > err = -EIO; > set_engine_banned(e); > xe_engine_get(e); > - disable_scheduling_deregister(engine_to_guc(e), e); > + disable_scheduling_deregister(guc, e); > > /* > * Must wait for scheduling to be disabled before signalling > @@ -848,7 +911,7 @@ guc_engine_timedout_job(struct drm_sched_job *drm_job) > */ > list_add(&drm_job->list, &sched->pending_list); > drm_sched_run_wq_start(sched); > - drm_sched_set_timeout(&e->guc->sched, MIN_SCHED_TIMEOUT); > + xe_guc_engine_trigger_cleanup(e); > > /* Mark all outstanding jobs as bad, thus completing them */ > spin_lock(&sched->job_list_lock); > @@ -872,6 +935,8 @@ static void __guc_engine_fini_async(struct work_struct *w) > > trace_xe_engine_destroy(e); > > + if (xe_engine_is_lr(e)) > + cancel_work_sync(&ge->lr_tdr); > if (e->flags & ENGINE_FLAG_PERSISTENT) > xe_device_remove_persistent_engines(gt_to_xe(e->gt), e); > release_guc_id(guc, e); > @@ -889,7 +954,7 @@ static void guc_engine_fini_async(struct xe_engine *e) > bool kernel = e->flags & ENGINE_FLAG_KERNEL; > > INIT_WORK(&e->guc->fini_async, __guc_engine_fini_async); > - queue_work(system_unbound_wq, &e->guc->fini_async); > + queue_work(system_wq, &e->guc->fini_async); > > /* We must block on kernel engines so slabs are empty on driver unload */ > if (kernel) { > @@ -1078,6 +1143,9 @@ static int guc_engine_init(struct xe_engine *e) > if (err) > goto err_sched; > > + if (xe_engine_is_lr(e)) > + INIT_WORK(&e->guc->lr_tdr, xe_guc_engine_lr_cleanup); > + > mutex_lock(&guc->submission_state.lock); > > err = alloc_guc_id(guc, e); > @@ -1129,7 +1197,7 @@ static void guc_engine_kill(struct xe_engine *e) > { > trace_xe_engine_kill(e); > set_engine_killed(e); > - drm_sched_set_timeout(&e->guc->sched, MIN_SCHED_TIMEOUT); > + xe_guc_engine_trigger_cleanup(e); > } > > static void guc_engine_add_msg(struct xe_engine *e, struct drm_sched_msg *msg, > @@ -1280,10 +1348,11 @@ static void guc_engine_stop(struct xe_guc *guc, struct xe_engine *e) > drm_sched_run_wq_stop(sched); > > /* Clean up lost G2H + reset engine state */ > - if (engine_destroyed(e) && engine_registered(e)) { > - if (engine_banned(e)) > + if (engine_registered(e)) { > + if ((engine_banned(e) && engine_destroyed(e)) || > + xe_engine_is_lr(e)) > xe_engine_put(e); > - else > + else if (engine_destroyed(e)) > __guc_engine_fini(guc, e); > } > if (e->guc->suspend_pending) { > @@ -1503,7 +1572,8 @@ int xe_guc_deregister_done_handler(struct xe_guc *guc, u32 *msg, u32 len) > trace_xe_engine_deregister_done(e); > > clear_engine_registered(e); > - if (engine_banned(e)) > + > + if (engine_banned(e) || xe_engine_is_lr(e)) > xe_engine_put(e); > else > __guc_engine_fini(guc, e); > @@ -1540,7 +1610,7 @@ int xe_guc_engine_reset_handler(struct xe_guc *guc, u32 *msg, u32 len) > */ > set_engine_reset(e); > if (!engine_banned(e)) > - drm_sched_set_timeout(&e->guc->sched, MIN_SCHED_TIMEOUT); > + xe_guc_engine_trigger_cleanup(e); > > return 0; > } > @@ -1567,7 +1637,7 @@ int xe_guc_engine_memory_cat_error_handler(struct xe_guc *guc, u32 *msg, > /* Treat the same as engine reset */ > set_engine_reset(e); > if (!engine_banned(e)) > - drm_sched_set_timeout(&e->guc->sched, MIN_SCHED_TIMEOUT); > + xe_guc_engine_trigger_cleanup(e); > > return 0; > } > diff --git a/drivers/gpu/drm/xe/xe_trace.h b/drivers/gpu/drm/xe/xe_trace.h > index 2f8eb7ebe9a7..02861c26e145 100644 > --- a/drivers/gpu/drm/xe/xe_trace.h > +++ b/drivers/gpu/drm/xe/xe_trace.h > @@ -219,6 +219,11 @@ DEFINE_EVENT(xe_engine, xe_engine_resubmit, > TP_ARGS(e) > ); > > +DEFINE_EVENT(xe_engine, xe_engine_lr_cleanup, > + TP_PROTO(struct xe_engine *e), > + TP_ARGS(e) > +); > + > DECLARE_EVENT_CLASS(xe_sched_job, > TP_PROTO(struct xe_sched_job *job), > TP_ARGS(job),