From mboxrd@z Thu Jan 1 00:00:00 1970 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 C19B3C9832A for ; Tue, 29 Sep 2026 14:29:44 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 6883810EF1A; Tue, 29 Sep 2026 14:29:44 +0000 (UTC) Authentication-Results: gabe.freedesktop.org; dkim=pass (2048-bit key; unprotected) header.d=intel.com header.i=@intel.com header.b="lJ8UDy/K"; dkim-atps=neutral Received: from mgamail.intel.com (mgamail.intel.com [192.198.163.18]) by gabe.freedesktop.org (Postfix) with ESMTPS id CA93710EF0F for ; Tue, 29 Sep 2026 14:29:40 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1790692181; x=1822228181; h=from:to:cc:subject:date:message-id:in-reply-to: references:mime-version:content-transfer-encoding; bh=9oKxjENdRrL+KM1wwizQ6Y1KX9XurQ217VJdczcld0A=; b=lJ8UDy/KxSQneN7GSCugbhlXS3P0hK0wJc+NgUTWQrOvs3aNlMWG/Lau eQTDNeXMQ2mMSNtspzzKuf0BItojkLkrJ4OkHdaFZd1pRx5p1Ltl6qO0r dm/f91nS6fHHgkKOSe6/qdAaslDOqyFU2vQA5Ld2rH98Nlv6nLWiL0Cxm idGPZ1ujY7iCJZU9vlEASLeSSznYCrfJXxAJet1jBizhg3uBVR2VJOPYU zv/6N5jZw/byhhAwxvJjG4jaa5PipupjK3bb73wbhKey+rND+6df/c6VV K+LIwGkZpW3aOKMUsSG2AQXqGc+p7ajwImOens1ewYGaG7mptJLbF34qw A==; X-CSE-ConnectionGUID: gfa5KM7GQ1Gzm4F3h2LHug== X-CSE-MsgGUID: k1WpccRITx6uQ0e/30zncQ== X-IronPort-AV: E=McAfee;i="6800,10657,11920"; a="90541146" X-IronPort-AV: E=Sophos;i="6.27,130,1787036400"; d="scan'208";a="90541146" Received: from orviesa004.jf.intel.com ([10.64.159.144]) by fmvoesa112.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 29 Sep 2026 07:29:41 -0700 X-CSE-ConnectionGUID: TA8OuFAuSNaL2eDK078UeA== X-CSE-MsgGUID: QwCUBlHIQG2AS/dqJ07UAg== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.27,130,1787036400"; d="scan'208";a="278732063" Received: from abityuts-desk1.ger.corp.intel.com (HELO fedora) ([10.245.244.222]) by orviesa004-auth.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 29 Sep 2026 07:29:39 -0700 From: =?UTF-8?q?Thomas=20Hellstr=C3=B6m?= To: intel-xe@lists.freedesktop.org Cc: =?UTF-8?q?Thomas=20Hellstr=C3=B6m?= , Matthew Brost , Rodrigo Vivi , Matthew Auld Subject: [PATCH v4 1/3] drm/xe: Don't unload the driver until all drm devices are freed Date: Tue, 29 Sep 2026 16:29:08 +0200 Message-ID: <20260929142910.47480-2-thomas.hellstrom@linux.intel.com> X-Mailer: git-send-email 2.55.0 In-Reply-To: <20260929142910.47480-1-thomas.hellstrom@linux.intel.com> References: <20260929142910.47480-1-thomas.hellstrom@linux.intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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" xe, and shared helpers it uses such as drm_gpuvm, already take bare drm_device references (drm_dev_get()) from several contexts, for example GuC submission fence workers, EU stall, OA and PMU sampling code, and drm_gpuvm's own vm object lifetime, without pairing them with a module reference. Since xe_exit() is only invoked after the module's own refcount has dropped to zero, none of these references currently prevent `rmmod xe` from proceeding while they, or the underlying xe_device release path they can trigger, are still outstanding, i.e. driver code belonging to a module whose text is being freed could still end up executing. Close this gap by keeping a device-count and waiting for it to reach zero at module unload. The wait for the device-count to reach zero at module unload is unbounded and non-interruptible. Rather than blocking silently forever if a reference is ever leaked, use wait_var_event_timeout() with a 20s timeout, well above the typical maximum dma_fence signalling time, and warn once if devices still remain by then, before falling back to an unbounded wait_var_event() so a stuck rmmod is at least observable instead of an indefinite, silent hang. Note that if a reference genuinely leaks, this still ends up as an indefinite uninterruptible sleep, which may eventually trip the kernel's hung-task watchdog. The alternative would be for these bare drm_device references to also take a module reference, which would instead make the module unable to be unloaded unless all its devices are manually unbound first. The wait-based approach is chosen here since it keeps rmmod usable in the common case. xe_device_exit() is added as a new module exit hook. Its entry in the init_funcs[] table is placed between xe_destroy_wq_module_init and xe_register_pci_driver, so that (exit functions run in reverse array order) it executes after xe_unregister_pci_driver() has forced all devices to unbind, but before xe_destroy_wq_module_exit() tears down the module-lifetime xe_destroy_wq. This preserves xe_destroy_wq's existing teardown ordering relative to xe_sched_job_module_exit() and xe_hw_fence_module_exit(), which destroy kmem_caches that work drained from xe_destroy_wq relies on, while ensuring xe_destroy_wq itself is only torn down once xe_device_exit() has confirmed no more work can be queued onto it. v2: - Updated commit message to describe the actual implementation (a single 20s wait_var_event_timeout() followed by one pr_warn() and an unbounded wait_var_event(), rather than a loop retrying with a diagnostic every 10s) and its uninterruptible-sleep tradeoff (sashiko) v4: - No longer sync on drmm release callbacks finished execution Signed-off-by: Thomas Hellström Assisted-by: LLM Reviewed-by: Matthew Brost # v3 --- drivers/gpu/drm/xe/xe_device.c | 29 +++++++++++++++++++++++++++++ drivers/gpu/drm/xe/xe_device.h | 2 ++ drivers/gpu/drm/xe/xe_module.c | 18 +++++++++++++++++- 3 files changed, 48 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c index 205cb4e7f9e8..10436b5bd03f 100644 --- a/drivers/gpu/drm/xe/xe_device.c +++ b/drivers/gpu/drm/xe/xe_device.c @@ -372,6 +372,9 @@ bool xe_device_is_admin_only(const struct xe_device *xe) } #endif +/* Number of allocated struct xe_device */ +static atomic_t xe_device_count; + static void xe_device_destroy(struct drm_device *dev, void *dummy) { struct xe_device *xe = to_xe_device(dev); @@ -391,6 +394,9 @@ static void xe_device_destroy(struct drm_device *dev, void *dummy) destroy_workqueue(xe->destroy_wq); ttm_device_fini(&xe->ttm); + + if (atomic_dec_and_test(&xe_device_count)) + wake_up_var(&xe_device_count); } /** @@ -461,6 +467,7 @@ int xe_device_init_early(struct xe_device *xe) return err; xe_bo_dev_init(&xe->bo_device); + atomic_inc(&xe_device_count); err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL); if (err) return err; @@ -1501,3 +1508,25 @@ struct xe_vm *xe_device_asid_to_vm(struct xe_device *xe, u32 asid) return vm; } + +/** + * xe_device_exit() - Device subsystem exit function. + * + * Exit function to be called at module unload time. + */ +void xe_device_exit(void) +{ + /* + * Wait for all devices to be freed. 20s is well above the typical + * maximum dma_fence signalling time, so warn and keep waiting if + * we're still not done by then, since it may indicate a leaked + * xe_device reference is stalling module unload. + */ + if (!wait_var_event_timeout(&xe_device_count, + !atomic_read(&xe_device_count), + HZ * 20)) { + pr_warn("%s: Waiting for %d xe device(s) to be freed before unloading.\n", + DRIVER_NAME, atomic_read(&xe_device_count)); + wait_var_event(&xe_device_count, !atomic_read(&xe_device_count)); + } +} diff --git a/drivers/gpu/drm/xe/xe_device.h b/drivers/gpu/drm/xe/xe_device.h index 6d3d6d5eba29..83d6dafab53c 100644 --- a/drivers/gpu/drm/xe/xe_device.h +++ b/drivers/gpu/drm/xe/xe_device.h @@ -283,6 +283,8 @@ static inline bool xe_device_is_admin_only(const struct xe_device *xe) } #endif +void xe_device_exit(void); + /* * Occasionally it is seen that the G2H worker starts running after a delay of more than * a second even after being queued and activated by the Linux workqueue subsystem. This diff --git a/drivers/gpu/drm/xe/xe_module.c b/drivers/gpu/drm/xe/xe_module.c index 4bc28dfc1992..c61bd33546f2 100644 --- a/drivers/gpu/drm/xe/xe_module.c +++ b/drivers/gpu/drm/xe/xe_module.c @@ -12,7 +12,7 @@ #include #include "xe_defaults.h" -#include "xe_device_types.h" +#include "xe_device.h" #include "xe_drv.h" #include "xe_configfs.h" #include "xe_hw_fence.h" @@ -162,6 +162,22 @@ static const struct init_funcs init_funcs[] = { .init = xe_destroy_wq_module_init, .exit = xe_destroy_wq_module_exit, }, + /* + * xe_destroy_wq_module_exit() must run after xe_device_exit() + * (below), since freeing a device can still queue work on + * xe_destroy_wq that must be drained before the module can safely + * unload. At the same time, xe_device_exit() must run after + * xe_unregister_pci_driver() (below), and xe_destroy_wq_module_exit() + * must run before xe_sched_job_module_exit() and + * xe_hw_fence_module_exit() (above), whose kmem_caches are still used + * by work drained from xe_destroy_wq. Exit functions run in reverse + * array order, so this entry must sit between the + * xe_destroy_wq_module_init entry above and the xe_register_pci_driver + * entry below. + */ + { + .exit = xe_device_exit, + }, { .init = xe_register_pci_driver, .exit = xe_unregister_pci_driver, -- 2.55.0