From 013e484dbd687a9174acf8f4450217bdb86ad788 Mon Sep 17 00:00:00 2001 From: Shuicheng Lin Date: Tue, 19 Aug 2025 15:39:51 +0000 Subject: [PATCH 1/8] drm/xe/tile: Release kobject for the failure path Call kobject_put() for the failure path to release the kobject v2: remove extra newline. (Matt) Fixes: e3d0839aa501 ("drm/xe/tile: Abort driver load for sysfs creation failure") Cc: Himal Prasad Ghimiray Reviewed-by: Matthew Brost Signed-off-by: Shuicheng Lin Link: https://lore.kernel.org/r/20250819153950.2973344-2-shuicheng.lin@intel.com Signed-off-by: Lucas De Marchi (cherry picked from commit b98775bca99511cc22ab459a2de646cd2fa7241f) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_tile_sysfs.c | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_tile_sysfs.c b/drivers/gpu/drm/xe/xe_tile_sysfs.c index b804234a6551..9e1236a9ec67 100644 --- a/drivers/gpu/drm/xe/xe_tile_sysfs.c +++ b/drivers/gpu/drm/xe/xe_tile_sysfs.c @@ -44,16 +44,18 @@ int xe_tile_sysfs_init(struct xe_tile *tile) kt->tile = tile; err = kobject_add(&kt->base, &dev->kobj, "tile%d", tile->id); - if (err) { - kobject_put(&kt->base); - return err; - } + if (err) + goto err_object; tile->sysfs = &kt->base; err = xe_vram_freq_sysfs_init(tile); if (err) - return err; + goto err_object; return devm_add_action_or_reset(xe->drm.dev, tile_sysfs_fini, tile); + +err_object: + kobject_put(&kt->base); + return err; } From fef8b64e48e836344574b85132a1c317f4260022 Mon Sep 17 00:00:00 2001 From: Michal Wajdeczko Date: Thu, 11 Sep 2025 00:24:39 +0200 Subject: [PATCH 2/8] drm/xe/pf: Drop rounddown_pow_of_two fair LMEM limitation MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This effectively reverts commit 4c3fe5eae46b ("drm/xe/pf: Limit fair VF LMEM provisioning") since we don't need it any more after non-contig VRAM allocations were fixed. This allows larger LMEM auto-provisioning for VFs, so instead: [ ] GT0: PF: LMEM available(14096M) fair(1 x 8192M) [ ] GT0: PF: VF1 provisioned with 8589934592 (8.00 GiB) LMEM or [ ] GT0: PF: LMEM available(14096M) fair(2 x 4096M) [ ] GT0: PF: VF1..VF2 provisioned with 4294967296 (4.00 GiB) LMEM we may get: [ ] GT0: PF: LMEM available(14096M) fair(1 x 14096M) [ ] GT0: PF: VF1 provisioned with 14780727296 (13.8 GiB) LMEM and [ ] GT0: PF: LMEM available(14096M) fair(2 x 7048M) [ ] GT0: PF: VF1..VF2 provisioned with 7390363648 (6.88 GiB) LMEM Fixes: 1e32ffbc9dc8 ("drm/xe/sriov: support non-contig VRAM provisioning") Signed-off-by: Michal Wajdeczko Reviewed-by: Piotr Piórkowski Link: https://lore.kernel.org/r/20250910222439.32869-1-michal.wajdeczko@intel.com (cherry picked from commit 95c1cfa306087142989bff34ea0e05dcd95ddc58) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 1 - 1 file changed, 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c index 494909f74eb2..d84831a03610 100644 --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c @@ -1632,7 +1632,6 @@ static u64 pf_estimate_fair_lmem(struct xe_gt *gt, unsigned int num_vfs) u64 fair; fair = div_u64(available, num_vfs); - fair = rounddown_pow_of_two(fair); /* XXX: ttm_vram_mgr & drm_buddy limitation */ fair = ALIGN_DOWN(fair, alignment); #ifdef MAX_FAIR_LMEM fair = min_t(u64, MAX_FAIR_LMEM, fair); From dcfd151d3277cc3bc73c94114e360556d47b992d Mon Sep 17 00:00:00 2001 From: Mallesh Koujalagi Date: Fri, 12 Sep 2025 17:04:58 +0530 Subject: [PATCH 3/8] drm/xe/hwmon: Remove type casting Refactor: eliminate type casts by using proper u32 declarations. v2: - Address review comments. (Karthik) v3: - Use the proper u32 type and drop cast. (Lucas De Marchi) - Modify variable when actually using u64 value. - Change r value to reg_value with u32 type. v4: - Remove newline between trailer and Signed-off-by. (Lucas De Marchi) - Change reg_val to val for more user-friendly logging. - Use mul_u32_u32 function since both values are u32. v5: - mul_u32_u32 function with shift. (Lucas De Marchi) Fixes: 7596d839f6228 ("drm/xe/hwmon: Add support to manage power limits though mailbox") Signed-off-by: Mallesh Koujalagi Reviewed-by: Lucas De Marchi Link: https://lore.kernel.org/r/20250912113458.2815172-1-mallesh.koujalagi@intel.com Signed-off-by: Lucas De Marchi (cherry picked from commit 4e1d3b5e6423dc841acd8691d75626b3d3b2b6a8) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_hwmon.c | 35 +++++++++++++++++++---------------- 1 file changed, 19 insertions(+), 16 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_hwmon.c b/drivers/gpu/drm/xe/xe_hwmon.c index c17ed1ae8649..c5b63e10bb91 100644 --- a/drivers/gpu/drm/xe/xe_hwmon.c +++ b/drivers/gpu/drm/xe/xe_hwmon.c @@ -286,7 +286,7 @@ static struct xe_reg xe_hwmon_get_reg(struct xe_hwmon *hwmon, enum xe_hwmon_reg */ static void xe_hwmon_power_max_read(struct xe_hwmon *hwmon, u32 attr, int channel, long *value) { - u64 reg_val = 0, min, max; + u32 reg_val = 0; struct xe_device *xe = hwmon->xe; struct xe_reg rapl_limit, pkg_power_sku; struct xe_mmio *mmio = xe_root_tile_mmio(xe); @@ -294,7 +294,7 @@ static void xe_hwmon_power_max_read(struct xe_hwmon *hwmon, u32 attr, int channe mutex_lock(&hwmon->hwmon_lock); if (hwmon->xe->info.has_mbx_power_limits) { - xe_hwmon_pcode_read_power_limit(hwmon, attr, channel, (u32 *)®_val); + xe_hwmon_pcode_read_power_limit(hwmon, attr, channel, ®_val); } else { rapl_limit = xe_hwmon_get_reg(hwmon, REG_PKG_RAPL_LIMIT, channel); pkg_power_sku = xe_hwmon_get_reg(hwmon, REG_PKG_POWER_SKU, channel); @@ -304,19 +304,21 @@ static void xe_hwmon_power_max_read(struct xe_hwmon *hwmon, u32 attr, int channe /* Check if PL limits are disabled. */ if (!(reg_val & PWR_LIM_EN)) { *value = PL_DISABLE; - drm_info(&hwmon->xe->drm, "%s disabled for channel %d, val 0x%016llx\n", + drm_info(&hwmon->xe->drm, "%s disabled for channel %d, val 0x%08x\n", PWR_ATTR_TO_STR(attr), channel, reg_val); goto unlock; } reg_val = REG_FIELD_GET(PWR_LIM_VAL, reg_val); - *value = mul_u64_u32_shr(reg_val, SF_POWER, hwmon->scl_shift_power); + *value = mul_u32_u32(reg_val, SF_POWER) >> hwmon->scl_shift_power; /* For platforms with mailbox power limit support clamping would be done by pcode. */ if (!hwmon->xe->info.has_mbx_power_limits) { - reg_val = xe_mmio_read64_2x32(mmio, pkg_power_sku); - min = REG_FIELD_GET(PKG_MIN_PWR, reg_val); - max = REG_FIELD_GET(PKG_MAX_PWR, reg_val); + u64 pkg_pwr, min, max; + + pkg_pwr = xe_mmio_read64_2x32(mmio, pkg_power_sku); + min = REG_FIELD_GET(PKG_MIN_PWR, pkg_pwr); + max = REG_FIELD_GET(PKG_MAX_PWR, pkg_pwr); min = mul_u64_u32_shr(min, SF_POWER, hwmon->scl_shift_power); max = mul_u64_u32_shr(max, SF_POWER, hwmon->scl_shift_power); if (min && max) @@ -493,8 +495,8 @@ xe_hwmon_power_max_interval_show(struct device *dev, struct device_attribute *at { struct xe_hwmon *hwmon = dev_get_drvdata(dev); struct xe_mmio *mmio = xe_root_tile_mmio(hwmon->xe); - u32 x, y, x_w = 2; /* 2 bits */ - u64 r, tau4, out; + u32 reg_val, x, y, x_w = 2; /* 2 bits */ + u64 tau4, out; int channel = (to_sensor_dev_attr(attr)->index % 2) ? CHANNEL_PKG : CHANNEL_CARD; u32 power_attr = (to_sensor_dev_attr(attr)->index > 1) ? PL2_HWMON_ATTR : PL1_HWMON_ATTR; @@ -505,23 +507,24 @@ xe_hwmon_power_max_interval_show(struct device *dev, struct device_attribute *at mutex_lock(&hwmon->hwmon_lock); if (hwmon->xe->info.has_mbx_power_limits) { - ret = xe_hwmon_pcode_read_power_limit(hwmon, power_attr, channel, (u32 *)&r); + ret = xe_hwmon_pcode_read_power_limit(hwmon, power_attr, channel, ®_val); if (ret) { drm_err(&hwmon->xe->drm, - "power interval read fail, ch %d, attr %d, r 0%llx, ret %d\n", - channel, power_attr, r, ret); - r = 0; + "power interval read fail, ch %d, attr %d, val 0x%08x, ret %d\n", + channel, power_attr, reg_val, ret); + reg_val = 0; } } else { - r = xe_mmio_read32(mmio, xe_hwmon_get_reg(hwmon, REG_PKG_RAPL_LIMIT, channel)); + reg_val = xe_mmio_read32(mmio, xe_hwmon_get_reg(hwmon, REG_PKG_RAPL_LIMIT, + channel)); } mutex_unlock(&hwmon->hwmon_lock); xe_pm_runtime_put(hwmon->xe); - x = REG_FIELD_GET(PWR_LIM_TIME_X, r); - y = REG_FIELD_GET(PWR_LIM_TIME_Y, r); + x = REG_FIELD_GET(PWR_LIM_TIME_X, reg_val); + y = REG_FIELD_GET(PWR_LIM_TIME_Y, reg_val); /* * tau = (1 + (x / 4)) * power(2,y), x = bits(23:22), y = bits(21:17) From 7926ba2143d8fef40bb940232818c7363e33598c Mon Sep 17 00:00:00 2001 From: Nitin Gote Date: Thu, 11 Sep 2025 10:58:23 +0530 Subject: [PATCH 4/8] drm/xe: defer free of NVM auxiliary container to device release callback Do not kfree the intel_dg_nvm_dev in xe_nvm_fini() right after auxiliary_device_delete/uninit. The auxiliary_device embeds the device/kobject (and its name); freeing it too early can race with asynchronous device_del/udev processing and cause a use-after-free. Signed-off-by: Nitin Gote Fixes: c28bfb107dac ("drm/xe/nvm: add on-die non-volatile memory device") Reviewed-by: Lucas De Marchi Link: https://lore.kernel.org/r/20250911052823.226696-1-nitin.r.gote@intel.com Signed-off-by: Lucas De Marchi (cherry picked from commit d4c3ed963e41d488695cf91068eabb8eb9f538ec) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_nvm.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_nvm.c b/drivers/gpu/drm/xe/xe_nvm.c index 61b0a1531a53..2cfe9eb67391 100644 --- a/drivers/gpu/drm/xe/xe_nvm.c +++ b/drivers/gpu/drm/xe/xe_nvm.c @@ -35,6 +35,10 @@ static const struct intel_dg_nvm_region regions[INTEL_DG_NVM_REGIONS] = { static void xe_nvm_release_dev(struct device *dev) { + struct auxiliary_device *aux = container_of(dev, struct auxiliary_device, dev); + struct intel_dg_nvm_dev *nvm = container_of(aux, struct intel_dg_nvm_dev, aux_dev); + + kfree(nvm); } static bool xe_nvm_non_posted_erase(struct xe_device *xe) @@ -162,6 +166,5 @@ void xe_nvm_fini(struct xe_device *xe) auxiliary_device_delete(&nvm->aux_dev); auxiliary_device_uninit(&nvm->aux_dev); - kfree(nvm); xe->nvm = NULL; } From cbc7f3b4f6ca19320e2eacf8fc1403d6f331ce14 Mon Sep 17 00:00:00 2001 From: Dan Carpenter Date: Thu, 7 Aug 2025 18:53:41 +0300 Subject: [PATCH 5/8] drm/xe: Fix a NULL vs IS_ERR() in xe_vm_add_compute_exec_queue() The xe_preempt_fence_create() function returns error pointers. It never returns NULL. Update the error checking to match. Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Signed-off-by: Dan Carpenter Reviewed-by: Matthew Brost Link: https://lore.kernel.org/r/aJTMBdX97cof_009@stanley.mountain Signed-off-by: Rodrigo Vivi (cherry picked from commit 75cc23ffe5b422bc3cbd5cf0956b8b86e4b0e162) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_vm.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c index dc4f61e56579..5146999d27fa 100644 --- a/drivers/gpu/drm/xe/xe_vm.c +++ b/drivers/gpu/drm/xe/xe_vm.c @@ -240,8 +240,8 @@ int xe_vm_add_compute_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q) pfence = xe_preempt_fence_create(q, q->lr.context, ++q->lr.seqno); - if (!pfence) { - err = -ENOMEM; + if (IS_ERR(pfence)) { + err = PTR_ERR(pfence); goto out_fini; } From ff89a4d285c82813faa0e2e386d07120ae1f9c85 Mon Sep 17 00:00:00 2001 From: Zongyao Bai Date: Tue, 16 Sep 2025 05:47:15 +0800 Subject: [PATCH 6/8] drm/xe/sysfs: Add cleanup action in xe_device_sysfs_init On partial failure, some sysfs files created before the failure might not be removed. Add common cleanup step to remove them all immediately, as is should be harmless to attempt to remove non-existing files. Fixes: 0e414bf7ad01 ("drm/xe: Expose PCIe link downgrade attributes") Cc: Lucas De Marchi Cc: Stuart Summers Cc: Shuicheng Lin Cc: Michal Wajdeczko Signed-off-by: Zongyao Bai Reviewed-by: Shuicheng Lin Reviewed-by: Lucas De Marchi Link: https://lore.kernel.org/r/20250915214716.1327379-2-zongyao.bai@intel.com Signed-off-by: Lucas De Marchi (cherry picked from commit 1a869168d91f1a1a2b0db22cea0295c67908e5d8) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_device_sysfs.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_device_sysfs.c b/drivers/gpu/drm/xe/xe_device_sysfs.c index bd9015761aa0..3e3b2d9033a7 100644 --- a/drivers/gpu/drm/xe/xe_device_sysfs.c +++ b/drivers/gpu/drm/xe/xe_device_sysfs.c @@ -311,12 +311,16 @@ int xe_device_sysfs_init(struct xe_device *xe) if (xe->info.platform == XE_BATTLEMAGE) { ret = sysfs_create_files(&dev->kobj, auto_link_downgrade_attrs); if (ret) - return ret; + goto cleanup; ret = late_bind_create_files(dev); if (ret) - return ret; + goto cleanup; } return devm_add_action_or_reset(dev, xe_device_sysfs_fini, xe); + +cleanup: + xe_device_sysfs_fini(xe); + return ret; } From ae5fbbda341f92e605a9508a0fb18456155517f0 Mon Sep 17 00:00:00 2001 From: Daniele Ceraolo Spurio Date: Tue, 9 Sep 2025 15:12:40 -0700 Subject: [PATCH 7/8] drm/xe: Fix error handling if PXP fails to start Since the PXP start comes after __xe_exec_queue_init() has completed, we need to cleanup what was done in that function in case of a PXP start error. __xe_exec_queue_init calls the submission backend init() function, so we need to introduce an opposite for that. Unfortunately, while we already have a fini() function pointer, it performs other operations in addition to cleaning up what was done by the init(). Therefore, for clarity, the existing fini() has been renamed to destroy(), while a new fini() has been added to only clean up what was done by the init(), with the latter being called by the former (via xe_exec_queue_fini). Fixes: 72d479601d67 ("drm/xe/pxp/uapi: Add userspace and LRC support for PXP-using queues") Signed-off-by: Daniele Ceraolo Spurio Cc: John Harrison Cc: Matthew Brost Reviewed-by: John Harrison Signed-off-by: John Harrison Link: https://lore.kernel.org/r/20250909221240.3711023-3-daniele.ceraolospurio@intel.com (cherry picked from commit 626667321deb4c7a294725406faa3dd71c3d445d) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_exec_queue.c | 22 +++++--- drivers/gpu/drm/xe/xe_exec_queue_types.h | 8 ++- drivers/gpu/drm/xe/xe_execlist.c | 25 +++++---- drivers/gpu/drm/xe/xe_execlist_types.h | 2 +- drivers/gpu/drm/xe/xe_guc_exec_queue_types.h | 4 +- drivers/gpu/drm/xe/xe_guc_submit.c | 54 ++++++++++++-------- 6 files changed, 73 insertions(+), 42 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c index 8991b4aed440..c07edcda99c5 100644 --- a/drivers/gpu/drm/xe/xe_exec_queue.c +++ b/drivers/gpu/drm/xe/xe_exec_queue.c @@ -151,6 +151,16 @@ static int __xe_exec_queue_init(struct xe_exec_queue *q) return err; } +static void __xe_exec_queue_fini(struct xe_exec_queue *q) +{ + int i; + + q->ops->fini(q); + + for (i = 0; i < q->width; ++i) + xe_lrc_put(q->lrc[i]); +} + struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *vm, u32 logical_mask, u16 width, struct xe_hw_engine *hwe, u32 flags, @@ -181,11 +191,13 @@ struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *v if (xe_exec_queue_uses_pxp(q)) { err = xe_pxp_exec_queue_add(xe->pxp, q); if (err) - goto err_post_alloc; + goto err_post_init; } return q; +err_post_init: + __xe_exec_queue_fini(q); err_post_alloc: __xe_exec_queue_free(q); return ERR_PTR(err); @@ -283,13 +295,11 @@ void xe_exec_queue_destroy(struct kref *ref) xe_exec_queue_put(eq); } - q->ops->fini(q); + q->ops->destroy(q); } void xe_exec_queue_fini(struct xe_exec_queue *q) { - int i; - /* * Before releasing our ref to lrc and xef, accumulate our run ticks * and wakeup any waiters. @@ -298,9 +308,7 @@ void xe_exec_queue_fini(struct xe_exec_queue *q) if (q->xef && atomic_dec_and_test(&q->xef->exec_queue.pending_removal)) wake_up_var(&q->xef->exec_queue.pending_removal); - for (i = 0; i < q->width; ++i) - xe_lrc_put(q->lrc[i]); - + __xe_exec_queue_fini(q); __xe_exec_queue_free(q); } diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h b/drivers/gpu/drm/xe/xe_exec_queue_types.h index cc1cffb5c87f..1c9d03f2a3e5 100644 --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h @@ -166,8 +166,14 @@ struct xe_exec_queue_ops { int (*init)(struct xe_exec_queue *q); /** @kill: Kill inflight submissions for backend */ void (*kill)(struct xe_exec_queue *q); - /** @fini: Fini exec queue for submission backend */ + /** @fini: Undoes the init() for submission backend */ void (*fini)(struct xe_exec_queue *q); + /** + * @destroy: Destroy exec queue for submission backend. The backend + * function must call xe_exec_queue_fini() (which will in turn call the + * fini() backend function) to ensure the queue is properly cleaned up. + */ + void (*destroy)(struct xe_exec_queue *q); /** @set_priority: Set priority for exec queue */ int (*set_priority)(struct xe_exec_queue *q, enum xe_exec_queue_priority priority); diff --git a/drivers/gpu/drm/xe/xe_execlist.c b/drivers/gpu/drm/xe/xe_execlist.c index 788f56b066b6..f83d421ac9d3 100644 --- a/drivers/gpu/drm/xe/xe_execlist.c +++ b/drivers/gpu/drm/xe/xe_execlist.c @@ -385,10 +385,20 @@ static int execlist_exec_queue_init(struct xe_exec_queue *q) return err; } -static void execlist_exec_queue_fini_async(struct work_struct *w) +static void execlist_exec_queue_fini(struct xe_exec_queue *q) +{ + struct xe_execlist_exec_queue *exl = q->execlist; + + drm_sched_entity_fini(&exl->entity); + drm_sched_fini(&exl->sched); + + kfree(exl); +} + +static void execlist_exec_queue_destroy_async(struct work_struct *w) { struct xe_execlist_exec_queue *ee = - container_of(w, struct xe_execlist_exec_queue, fini_async); + container_of(w, struct xe_execlist_exec_queue, destroy_async); struct xe_exec_queue *q = ee->q; struct xe_execlist_exec_queue *exl = q->execlist; struct xe_device *xe = gt_to_xe(q->gt); @@ -401,10 +411,6 @@ static void execlist_exec_queue_fini_async(struct work_struct *w) list_del(&exl->active_link); spin_unlock_irqrestore(&exl->port->lock, flags); - drm_sched_entity_fini(&exl->entity); - drm_sched_fini(&exl->sched); - kfree(exl); - xe_exec_queue_fini(q); } @@ -413,10 +419,10 @@ static void execlist_exec_queue_kill(struct xe_exec_queue *q) /* NIY */ } -static void execlist_exec_queue_fini(struct xe_exec_queue *q) +static void execlist_exec_queue_destroy(struct xe_exec_queue *q) { - INIT_WORK(&q->execlist->fini_async, execlist_exec_queue_fini_async); - queue_work(system_unbound_wq, &q->execlist->fini_async); + INIT_WORK(&q->execlist->destroy_async, execlist_exec_queue_destroy_async); + queue_work(system_unbound_wq, &q->execlist->destroy_async); } static int execlist_exec_queue_set_priority(struct xe_exec_queue *q, @@ -467,6 +473,7 @@ static const struct xe_exec_queue_ops execlist_exec_queue_ops = { .init = execlist_exec_queue_init, .kill = execlist_exec_queue_kill, .fini = execlist_exec_queue_fini, + .destroy = execlist_exec_queue_destroy, .set_priority = execlist_exec_queue_set_priority, .set_timeslice = execlist_exec_queue_set_timeslice, .set_preempt_timeout = execlist_exec_queue_set_preempt_timeout, diff --git a/drivers/gpu/drm/xe/xe_execlist_types.h b/drivers/gpu/drm/xe/xe_execlist_types.h index 415140936f11..92c4ba52db0c 100644 --- a/drivers/gpu/drm/xe/xe_execlist_types.h +++ b/drivers/gpu/drm/xe/xe_execlist_types.h @@ -42,7 +42,7 @@ struct xe_execlist_exec_queue { bool has_run; - struct work_struct fini_async; + struct work_struct destroy_async; enum xe_exec_queue_priority active_priority; struct list_head active_link; diff --git a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h index a3f421e2adc0..c30c0e3ccbbb 100644 --- a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h +++ b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h @@ -35,8 +35,8 @@ struct xe_guc_exec_queue { struct xe_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; + /** @destroy_async: do final destroy async from this worker */ + struct work_struct destroy_async; /** @resume_time: time of last resume */ u64 resume_time; /** @state: GuC specific state for this xe_exec_queue */ diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c index cafb47711e9b..93fc7b290b65 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.c +++ b/drivers/gpu/drm/xe/xe_guc_submit.c @@ -1277,21 +1277,12 @@ guc_exec_queue_timedout_job(struct drm_sched_job *drm_job) return DRM_GPU_SCHED_STAT_NO_HANG; } -static void __guc_exec_queue_fini_async(struct work_struct *w) +static void guc_exec_queue_fini(struct xe_exec_queue *q) { - struct xe_guc_exec_queue *ge = - container_of(w, struct xe_guc_exec_queue, fini_async); - struct xe_exec_queue *q = ge->q; + struct xe_guc_exec_queue *ge = q->guc; struct xe_guc *guc = exec_queue_to_guc(q); - xe_pm_runtime_get(guc_to_xe(guc)); - trace_xe_exec_queue_destroy(q); - release_guc_id(guc, q); - if (xe_exec_queue_is_lr(q)) - cancel_work_sync(&ge->lr_tdr); - /* Confirm no work left behind accessing device structures */ - cancel_delayed_work_sync(&ge->sched.base.work_tdr); xe_sched_entity_fini(&ge->entity); xe_sched_fini(&ge->sched); @@ -1300,25 +1291,43 @@ static void __guc_exec_queue_fini_async(struct work_struct *w) * (timeline name). */ kfree_rcu(ge, rcu); +} + +static void __guc_exec_queue_destroy_async(struct work_struct *w) +{ + struct xe_guc_exec_queue *ge = + container_of(w, struct xe_guc_exec_queue, destroy_async); + struct xe_exec_queue *q = ge->q; + struct xe_guc *guc = exec_queue_to_guc(q); + + xe_pm_runtime_get(guc_to_xe(guc)); + trace_xe_exec_queue_destroy(q); + + if (xe_exec_queue_is_lr(q)) + cancel_work_sync(&ge->lr_tdr); + /* Confirm no work left behind accessing device structures */ + cancel_delayed_work_sync(&ge->sched.base.work_tdr); + xe_exec_queue_fini(q); + xe_pm_runtime_put(guc_to_xe(guc)); } -static void guc_exec_queue_fini_async(struct xe_exec_queue *q) +static void guc_exec_queue_destroy_async(struct xe_exec_queue *q) { struct xe_guc *guc = exec_queue_to_guc(q); struct xe_device *xe = guc_to_xe(guc); - INIT_WORK(&q->guc->fini_async, __guc_exec_queue_fini_async); + INIT_WORK(&q->guc->destroy_async, __guc_exec_queue_destroy_async); /* We must block on kernel engines so slabs are empty on driver unload */ if (q->flags & EXEC_QUEUE_FLAG_PERMANENT || exec_queue_wedged(q)) - __guc_exec_queue_fini_async(&q->guc->fini_async); + __guc_exec_queue_destroy_async(&q->guc->destroy_async); else - queue_work(xe->destroy_wq, &q->guc->fini_async); + queue_work(xe->destroy_wq, &q->guc->destroy_async); } -static void __guc_exec_queue_fini(struct xe_guc *guc, struct xe_exec_queue *q) +static void __guc_exec_queue_destroy(struct xe_guc *guc, struct xe_exec_queue *q) { /* * Might be done from within the GPU scheduler, need to do async as we @@ -1327,7 +1336,7 @@ static void __guc_exec_queue_fini(struct xe_guc *guc, struct xe_exec_queue *q) * this we and don't really care when everything is fini'd, just that it * is. */ - guc_exec_queue_fini_async(q); + guc_exec_queue_destroy_async(q); } static void __guc_exec_queue_process_msg_cleanup(struct xe_sched_msg *msg) @@ -1341,7 +1350,7 @@ static void __guc_exec_queue_process_msg_cleanup(struct xe_sched_msg *msg) if (exec_queue_registered(q)) disable_scheduling_deregister(guc, q); else - __guc_exec_queue_fini(guc, q); + __guc_exec_queue_destroy(guc, q); } static bool guc_exec_queue_allowed_to_change_state(struct xe_exec_queue *q) @@ -1574,14 +1583,14 @@ static bool guc_exec_queue_try_add_msg(struct xe_exec_queue *q, #define STATIC_MSG_CLEANUP 0 #define STATIC_MSG_SUSPEND 1 #define STATIC_MSG_RESUME 2 -static void guc_exec_queue_fini(struct xe_exec_queue *q) +static void guc_exec_queue_destroy(struct xe_exec_queue *q) { struct xe_sched_msg *msg = q->guc->static_msgs + STATIC_MSG_CLEANUP; if (!(q->flags & EXEC_QUEUE_FLAG_PERMANENT) && !exec_queue_wedged(q)) guc_exec_queue_add_msg(q, msg, CLEANUP); else - __guc_exec_queue_fini(exec_queue_to_guc(q), q); + __guc_exec_queue_destroy(exec_queue_to_guc(q), q); } static int guc_exec_queue_set_priority(struct xe_exec_queue *q, @@ -1711,6 +1720,7 @@ static const struct xe_exec_queue_ops guc_exec_queue_ops = { .init = guc_exec_queue_init, .kill = guc_exec_queue_kill, .fini = guc_exec_queue_fini, + .destroy = guc_exec_queue_destroy, .set_priority = guc_exec_queue_set_priority, .set_timeslice = guc_exec_queue_set_timeslice, .set_preempt_timeout = guc_exec_queue_set_preempt_timeout, @@ -1732,7 +1742,7 @@ static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q) if (exec_queue_extra_ref(q) || xe_exec_queue_is_lr(q)) xe_exec_queue_put(q); else if (exec_queue_destroyed(q)) - __guc_exec_queue_fini(guc, q); + __guc_exec_queue_destroy(guc, q); } if (q->guc->suspend_pending) { set_exec_queue_suspended(q); @@ -1989,7 +1999,7 @@ static void handle_deregister_done(struct xe_guc *guc, struct xe_exec_queue *q) if (exec_queue_extra_ref(q) || xe_exec_queue_is_lr(q)) xe_exec_queue_put(q); else - __guc_exec_queue_fini(guc, q); + __guc_exec_queue_destroy(guc, q); } int xe_guc_deregister_done_handler(struct xe_guc *guc, u32 *msg, u32 len) From 26caeae9fb482ec443753b4e3307e5122b60b850 Mon Sep 17 00:00:00 2001 From: Daniele Ceraolo Spurio Date: Fri, 5 Sep 2025 16:56:33 -0700 Subject: [PATCH 8/8] drm/xe/guc: Set RCS/CCS yield policy All recent platforms (including all the ones officially supported by the Xe driver) do not allow concurrent execution of RCS and CCS workloads from different address spaces, with the HW blocking the context switch when it detects such a scenario. The DUAL_QUEUE flag helps with this, by causing the GuC to not submit a context it knows will not be able to execute. This, however, causes a new problem: if RCS and CCS queues have pending workloads from different address spaces, the GuC needs to choose from which of the 2 queues to pick the next workload to execute. By default, the GuC prioritizes RCS submissions over CCS ones, which can lead to CCS workloads being significantly (or completely) starved of execution time. The driver can tune this by setting a dedicated scheduling policy KLV; this KLV allows the driver to specify a quantum (in ms) and a ratio (percentage value between 0 and 100), and the GuC will prioritize the CCS for that percentage of each quantum. Given that we want to guarantee enough RCS throughput to avoid missing frames, we set the yield policy to 20% of each 80ms interval. v2: updated quantum and ratio, improved comment, use xe_guc_submit_disable in gt_sanitize Fixes: d9a1ae0d17bd ("drm/xe/guc: Enable WA_DUAL_QUEUE for newer platforms") Signed-off-by: Daniele Ceraolo Spurio Cc: Matthew Brost Cc: John Harrison Cc: Vinay Belgaumkar Reviewed-by: John Harrison Tested-by: Vinay Belgaumkar Link: https://lore.kernel.org/r/20250905235632.3333247-2-daniele.ceraolospurio@intel.com (cherry picked from commit 88434448438e4302e272b2a2b810b42e05ea024b) Signed-off-by: Rodrigo Vivi [Rodrigo added #include "xe_guc_submit.h" while backporting] --- drivers/gpu/drm/xe/abi/guc_actions_abi.h | 1 + drivers/gpu/drm/xe/abi/guc_klvs_abi.h | 25 +++++++++ drivers/gpu/drm/xe/xe_gt.c | 3 +- drivers/gpu/drm/xe/xe_guc.c | 6 +-- drivers/gpu/drm/xe/xe_guc_submit.c | 66 ++++++++++++++++++++++++ drivers/gpu/drm/xe/xe_guc_submit.h | 2 + 6 files changed, 98 insertions(+), 5 deletions(-) diff --git a/drivers/gpu/drm/xe/abi/guc_actions_abi.h b/drivers/gpu/drm/xe/abi/guc_actions_abi.h index 81eb046aeebf..b9f67d7a00d8 100644 --- a/drivers/gpu/drm/xe/abi/guc_actions_abi.h +++ b/drivers/gpu/drm/xe/abi/guc_actions_abi.h @@ -117,6 +117,7 @@ enum xe_guc_action { XE_GUC_ACTION_ENTER_S_STATE = 0x501, XE_GUC_ACTION_EXIT_S_STATE = 0x502, XE_GUC_ACTION_GLOBAL_SCHED_POLICY_CHANGE = 0x506, + XE_GUC_ACTION_UPDATE_SCHEDULING_POLICIES_KLV = 0x509, XE_GUC_ACTION_SCHED_CONTEXT = 0x1000, XE_GUC_ACTION_SCHED_CONTEXT_MODE_SET = 0x1001, XE_GUC_ACTION_SCHED_CONTEXT_MODE_DONE = 0x1002, diff --git a/drivers/gpu/drm/xe/abi/guc_klvs_abi.h b/drivers/gpu/drm/xe/abi/guc_klvs_abi.h index 0366a9da5977..d7719d0e36ca 100644 --- a/drivers/gpu/drm/xe/abi/guc_klvs_abi.h +++ b/drivers/gpu/drm/xe/abi/guc_klvs_abi.h @@ -17,6 +17,7 @@ * | 0 | 31:16 | **KEY** - KLV key identifier | * | | | - `GuC Self Config KLVs`_ | * | | | - `GuC Opt In Feature KLVs`_ | + * | | | - `GuC Scheduling Policies KLVs`_ | * | | | - `GuC VGT Policy KLVs`_ | * | | | - `GuC VF Configuration KLVs`_ | * | | | | @@ -152,6 +153,30 @@ enum { #define GUC_KLV_OPT_IN_FEATURE_DYNAMIC_INHIBIT_CONTEXT_SWITCH_KEY 0x4003 #define GUC_KLV_OPT_IN_FEATURE_DYNAMIC_INHIBIT_CONTEXT_SWITCH_LEN 0u +/** + * DOC: GuC Scheduling Policies KLVs + * + * `GuC KLV`_ keys available for use with UPDATE_SCHEDULING_POLICIES_KLV. + * + * _`GUC_KLV_SCHEDULING_POLICIES_RENDER_COMPUTE_YIELD` : 0x1001 + * Some platforms do not allow concurrent execution of RCS and CCS + * workloads from different address spaces. By default, the GuC prioritizes + * RCS submissions over CCS ones, which can lead to CCS workloads being + * significantly (or completely) starved of execution time. This KLV allows + * the driver to specify a quantum (in ms) and a ratio (percentage value + * between 0 and 100), and the GuC will prioritize the CCS for that + * percentage of each quantum. For example, specifying 100ms and 30% will + * make the GuC prioritize the CCS for 30ms of every 100ms. + * Note that this does not necessarly mean that RCS and CCS engines will + * only be active for their percentage of the quantum, as the restriction + * only kicks in if both classes are fully busy with non-compatible address + * spaces; i.e., if one engine is idle or running the same address space, + * a pending job on the other engine will still be submitted to the HW no + * matter what the ratio is + */ +#define GUC_KLV_SCHEDULING_POLICIES_RENDER_COMPUTE_YIELD_KEY 0x1001 +#define GUC_KLV_SCHEDULING_POLICIES_RENDER_COMPUTE_YIELD_LEN 2u + /** * DOC: GuC VGT Policy KLVs * diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c index c8eda36546d3..17634195cdc2 100644 --- a/drivers/gpu/drm/xe/xe_gt.c +++ b/drivers/gpu/drm/xe/xe_gt.c @@ -41,6 +41,7 @@ #include "xe_gt_topology.h" #include "xe_guc_exec_queue_types.h" #include "xe_guc_pc.h" +#include "xe_guc_submit.h" #include "xe_hw_fence.h" #include "xe_hw_engine_class_sysfs.h" #include "xe_irq.h" @@ -97,7 +98,7 @@ void xe_gt_sanitize(struct xe_gt *gt) * FIXME: if xe_uc_sanitize is called here, on TGL driver will not * reload */ - gt->uc.guc.submission_state.enabled = false; + xe_guc_submit_disable(>->uc.guc); } static void xe_gt_enable_host_l2_vram(struct xe_gt *gt) diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c index b1d1d6da3758..270fc3792493 100644 --- a/drivers/gpu/drm/xe/xe_guc.c +++ b/drivers/gpu/drm/xe/xe_guc.c @@ -880,9 +880,7 @@ int xe_guc_post_load_init(struct xe_guc *guc) return ret; } - guc->submission_state.enabled = true; - - return 0; + return xe_guc_submit_enable(guc); } int xe_guc_reset(struct xe_guc *guc) @@ -1579,7 +1577,7 @@ void xe_guc_sanitize(struct xe_guc *guc) { xe_uc_fw_sanitize(&guc->fw); xe_guc_ct_disable(&guc->ct); - guc->submission_state.enabled = false; + xe_guc_submit_disable(guc); } int xe_guc_reset_prepare(struct xe_guc *guc) diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c index 93fc7b290b65..0104afbc941c 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.c +++ b/drivers/gpu/drm/xe/xe_guc_submit.c @@ -32,6 +32,7 @@ #include "xe_guc_ct.h" #include "xe_guc_exec_queue_types.h" #include "xe_guc_id_mgr.h" +#include "xe_guc_klv_helpers.h" #include "xe_guc_submit_types.h" #include "xe_hw_engine.h" #include "xe_hw_fence.h" @@ -316,6 +317,71 @@ int xe_guc_submit_init(struct xe_guc *guc, unsigned int num_ids) return drmm_add_action_or_reset(&xe->drm, guc_submit_fini, guc); } +/* + * Given that we want to guarantee enough RCS throughput to avoid missing + * frames, we set the yield policy to 20% of each 80ms interval. + */ +#define RC_YIELD_DURATION 80 /* in ms */ +#define RC_YIELD_RATIO 20 /* in percent */ +static u32 *emit_render_compute_yield_klv(u32 *emit) +{ + *emit++ = PREP_GUC_KLV_TAG(SCHEDULING_POLICIES_RENDER_COMPUTE_YIELD); + *emit++ = RC_YIELD_DURATION; + *emit++ = RC_YIELD_RATIO; + + return emit; +} + +#define SCHEDULING_POLICY_MAX_DWORDS 16 +static int guc_init_global_schedule_policy(struct xe_guc *guc) +{ + u32 data[SCHEDULING_POLICY_MAX_DWORDS]; + u32 *emit = data; + u32 count = 0; + int ret; + + if (GUC_SUBMIT_VER(guc) < MAKE_GUC_VER(1, 1, 0)) + return 0; + + *emit++ = XE_GUC_ACTION_UPDATE_SCHEDULING_POLICIES_KLV; + + if (CCS_MASK(guc_to_gt(guc))) + emit = emit_render_compute_yield_klv(emit); + + count = emit - data; + if (count > 1) { + xe_assert(guc_to_xe(guc), count <= SCHEDULING_POLICY_MAX_DWORDS); + + ret = xe_guc_ct_send_block(&guc->ct, data, count); + if (ret < 0) { + xe_gt_err(guc_to_gt(guc), + "failed to enable GuC sheduling policies: %pe\n", + ERR_PTR(ret)); + return ret; + } + } + + return 0; +} + +int xe_guc_submit_enable(struct xe_guc *guc) +{ + int ret; + + ret = guc_init_global_schedule_policy(guc); + if (ret) + return ret; + + guc->submission_state.enabled = true; + + return 0; +} + +void xe_guc_submit_disable(struct xe_guc *guc) +{ + guc->submission_state.enabled = false; +} + static void __release_guc_id(struct xe_guc *guc, struct xe_exec_queue *q, u32 xa_count) { int i; diff --git a/drivers/gpu/drm/xe/xe_guc_submit.h b/drivers/gpu/drm/xe/xe_guc_submit.h index 9b71a986c6ca..0d126b807c10 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.h +++ b/drivers/gpu/drm/xe/xe_guc_submit.h @@ -13,6 +13,8 @@ struct xe_exec_queue; struct xe_guc; int xe_guc_submit_init(struct xe_guc *guc, unsigned int num_ids); +int xe_guc_submit_enable(struct xe_guc *guc); +void xe_guc_submit_disable(struct xe_guc *guc); int xe_guc_submit_reset_prepare(struct xe_guc *guc); void xe_guc_submit_reset_wait(struct xe_guc *guc);